text
stringlengths 4
2.78M
| meta
dict |
|---|---|
---
abstract: 'We develop a theory of the low-temperature charge transfer between a superconductor and a hopping insulator. We show that the charge transfer is governed by the coherent two-electron – Cooper pair conversion process, *time reversal reflection*, where electrons tunnel into superconductor from the localized states in the hopping insulator located near the interface, and calculate the corresponding interface resistance. This process is an analog to conventional Andreev reflection process. We show that the time reversal interface resistance is accessible experimentally, and that in mesoscopic structures it can exceed the bulk hopping resistance.'
author:
- 'V. I. Kozub'
- 'A. A. Zyuzin'
- 'Y. M. Galperin'
- 'V. Vinokur'
title: Charge transfer between a superconductor and a hopping insulator
---
The transmission of the charge through the normal metal-superconductor interface occurs via the electron-hole conversion known as the Andreev reflection process: an electron incident from the metal side with an energy smaller than the energy gap in the superconductor is converted into a hole which moves backward with respect to the electron. The missing charge $2e$ (an electron has charge $-e$ and a hole $+e$) propagates as an electron pair into the superconductor and joins the Cooper pair condensate [@Andreev]. Correspondingly, a Cooper pair transfer from the superconductor is described as the Andreev reflection of a hole. This Andreev transport channel is characterized by the so-called Andreev interface contact resistance. Since transport current is introduced into a superconductor via normal leads, the Andreev reflection phenomenon is a foundation for most applications of superconductors (see Ref. [@Book] for a review).
There exists however an important experimental situation of the hopping insulator coupled to a measuring circuit via superconducting leads (see, for example, [@Rentch]), where the conventional Andreev reflection picture does not apply. The transport in hopping semiconductors occurs via localized (*non-propagating*) single particle states [@end1] with undefined momentum and therefore a Cooper pair on the superconductor side cannot form. A single particle transport through the interface is exponentially suppressed, $\propto e^ {-\Delta/T}$, where $\Delta$ is the superconductor gap, the temperature, $T$, being measured in energy units; therefore to explain the finite conductivity observed in experiments one needs Andreev-type processes capable to facilitate two particle transfer through the hopping insulator/superconductor interface allowing for Cooper pair formation. The possibility of such a transfer through the hopping-superconductor interface was discussed in Ref. [@Agrinskaya] but no quantitative theory of hopping transport - supercurrent conversion was presented.
In this Letter we develop a theory for the transport through the hopping insulator-superconductor interface and derive the corresponding contact resistance. We show that the low-temperature charge transfer occurs via the correlated processes mediated by the *pairs* of hopping centers located near the interface. We demonstrate that this process resembles the conventional Andreev electron-to-hole reflection into a normal metal, the exponential suppression of transport specific to a single-particle processes being lifted. Thus, despite the limitation in the number of coherent hopping centers that can carry Andreev transport, the resulting contact resistance can become low as compared to the resistance of the hopping insulator. However in mesoscopic structures the interface resistance can be comparable or even exceed the hopping resistance. The proposed mechanism resembles the so-called crossed Andreev charge transfer [@CA], discussed recently in connection with a superconductor-dot entangler [@ent; @ent1]. The difference is that in [CA,ent,ent1]{} the transport mediated by artificial quantum dots was considered. In our case, the transport occurs via randomly located sites in the hopping insulator (HI), and the main problem one has to solve is finding the optimal configuration of the sites responsible for the charge transfer. Hereafter we will refer to the proposed charge transfer mechanism as to the *time reversal reflection*.
Let a superconductor (S) and an HI to occupy the adjacent 3D semi-spaces separated by a tunneling barrier (B). The presence of the barrier simplifies calculations which will be made in the lowest non-vanishing approximation in the tunneling amplitude $T_{0}$. This models the Schottky barrier usually presenting at a semiconductor-metal interface. In the linear response theory the conductance is determined by the Kubo formula [@Kubo] for the susceptibility, $$\chi (\omega )=i\int\limits_{0}^{\infty }\left\langle \left[ \hat{I}^{+}(t),\hat{I}(0)\right] \right\rangle e^{i\omega t}\,dt \label{in}$$as $\mathcal{G}=\lim_{\omega \rightarrow 0}\omega ^{-1}{\mathop{\mathrm{Im}}}\chi (\omega )$. Here the current operator $\hat{I}(t)$ is defined as [@tunnel-current]: $$\hat{I}(t)=ied\,T_{0}\int d^{2}r\,[a^{+}(\mathbf{r},t)b(\mathbf{r},t)-\text{h.c.}]\,,$$where $\mathbf{r}$ is the coordinate in the interface plane, $a^{+}(\mathbf{r},t)$ and $b(\mathbf{r},t)$ are creation and annihilation operators in the semiconductor and superconductor, respectively, $d$ is the electron localization length under barrier. The susceptibility, $\chi (\omega )$, is calculated by analytical continuation of the Matsubara susceptibility [AGD]{}, $$\chi _{M}(\Omega )=\int_{0}^{\beta }\left\langle T_{\tau }I(\tau
)I(0)\right\rangle e^{i\Omega \tau }\,d\tau \,.$$Here $T_{\tau }$ means ordering in the imaginary time, $\beta \equiv 1/T$. In the expression for $\left\langle T_{\tau }I(\tau )I(0)\right\rangle $ one should expand to the second order with respect to the tunneling Hamiltonian, $$H_{T}(\tau )=dT_{0}\int d^{2}r\left[ a^{+}(\mathbf{r},\tau )b(\mathbf{r},\tau )+\text{h.c.}\right] \,. \label{TH}$$Keeping only those second order terms that contain $\left\langle T_{\tau }b(\mathbf{r},\tau )b(\mathbf{r}_{0},0)\right\rangle \left\langle T_{\tau
}b^{+}(\mathbf{r}_{1},\tau _{1})b^{+}(\mathbf{r}_{2},\tau _{2})\right\rangle
$ products and thus represent the time reversal scattering which we are interested in, one arrives at the expression $$\begin{aligned}
&&\left\langle T_{\tau }\hat{I}(\tau )\hat{I}(0)\right\rangle
=e^{2}|T_{0}|^{4}\int d\tau _{1}\,d\tau _{2}\prod_{i}d^{2}r_{i}(A+B)\,;
\notag \\
&&A(\{x_{i}\})=F(x-x_{0})F^{+}(x_{1}-x_{2})G(x_{1},x)G(x_{2},x_{0})\,,
\notag \\
&&B(\{x_{i}\})=F(x-x_{1})F^{+}(x_{0}-x_{2})[G(x_{0},x)G(x_{2},x_{1}) \notag
\\
&&\qquad \qquad \qquad -G(x_{0},x_{1})G(x_{2},x)]\,, \label{gf1}\end{aligned}$$where $x\equiv \{\mathbf{r},\tau \}$, $x_{0}\equiv \{\mathbf{r}_{0},0\}$, $x_{i}\equiv \{\mathbf{r}_{i},\tau _{i}\}$; $F(x-x^{\prime })=\langle T_{\tau
}b(\mathbf{r},\tau )b(\mathbf{r}^{\prime },\tau ^{\prime })\rangle $ is the anomalous Green function in the superconductor while $G(x,x^{\prime
})=-\langle T_{\tau }a(\mathbf{r},\tau )a^{+}(\mathbf{r}^{\prime },\tau
^{\prime })\rangle $ is the Green function in the hopping insulator. One can show that the Andreev-type process we are interested in is given by the first term of $B(\{x_{i}\})$ in Eq. (\[gf1\]). The relevant diagram is shown in Fig. \[fig:1\].
![The diagram describing the time reversal reflection. Lines with one arrow correspond to the Green’s functions in the hopping insulator. They are associated either with the center 1, or with the center 2. Lines with two arrows correspond to anomalous Green’s functions, see [[@AGD]]{}. Squares correspond to matrix elements of the tunneling Hamiltonian [(\[TH\])]{}. []{data-label="fig:1"}](fig1.eps){width="5cm"}
Keeping only this term and using the Matsubara frequency representation one obtains $$\begin{aligned}
&&\chi _{M}(\Omega )=2Te^{2}|T_{0}|^{4}d^{4}\int
\prod_{i}d^{2}r_{i}\sum_{\omega _{n}}F(\mathbf{r}-\mathbf{r}_{1},\omega _{n})
\notag \label{1} \\
&&\times F^{+}(\mathbf{r}_{0}-\mathbf{r}_{2},\omega _{n})G(\mathbf{r}_{0},\mathbf{r},\omega _{n}-\Omega _{m})G(\mathbf{r}_{2},\mathbf{r}_{1},-\omega
_{n}),\end{aligned}$$where $\Omega _{m}=2\pi mT$ and $\omega _{n}=(2n+1)\pi T$. The normal Green’s functions can be expressed through the wave functions of the localized states, $\varphi _{s}(\mathbf{r})=(\pi a^{3})^{-1/2}\exp (-|\mathbf{r}-\mathbf{r}_{s}|/a)$, as $$G(\mathbf{r},\mathbf{r}^{\prime },\omega _{n})=\sum_{s}\frac{\varphi
_{s}^{\ast }(\mathbf{r})\varphi _{s}(\mathbf{r}^{\prime })}{i\omega
_{n}-\varepsilon _{s}}\,. \\$$We have assumed that for all of the sites under consideration the voltage drops between the site and the superconductor are the same. This is true when the partial interface resistance due to a time reversal pair is much larger than the typical resistance of the bond forming the percolation cluster. This situation resembles that considered by Larkin and Shklovskii for the tunnel resistance between the hopping conductors [@Larkin].
The anomalous Green function, $F(R,\omega_n)$, is $$\begin{aligned}
F(R,\omega_n)&=&\int \frac{d^{3}p}{(2\pi \hbar)^3}\frac {\Delta}{\Delta^{2}+\xi^{2}_{\mathbf{p}}+\omega_n^{2}}\, e^{i\mathbf{p} \cdot \mathbf{R}/\hbar} \notag \\
&=& \frac{\pi g_{m} \Delta}{2\sqrt{\Delta^{2}+\omega_n^{2}}} \frac{\sin(Rk_{F})}{Rk_{F}} e^{-\frac{R}{\pi\xi}\frac{\sqrt {\Delta^{2}+\omega_n^{2}}}{\Delta}}\, .\end{aligned}$$ Here $\xi_{\mathbf{p}}=(p^{2}-p_F^2)/2m$, $g_{m}=mp_{F}/\pi^{2}\hbar^3$ is the density of states in a metal, $k_F=p_F/\hbar$, while $\xi$ is the coherence length in a superconductor. Since $F(R)$ oscillates with the period $2\pi/k_F
$ integration over spatial coordinates along the interface yields the factor $a^4/k_F^6|\bm{\rho}_{ls}|^2$. Here $\bm{\rho}_{ls}$ is projection of the vector $\mathbf{R}_{ls}$ connecting the centers on the interface plane. Note that the dependencies on $R$ and $\xi$ are similar to that given by Eq. (21) of the paper [@ent] for the pair of the quantum dots near the superconducting interface. However the latter equation does not specify the dependences of the transmission coefficients on the real physical parameters of the interface and the localized centers.
The summation over the Matsubara frequencies, $\omega_n$, is standard, $$T\sum_{\omega_n}f(\omega_n) =\oint \frac{d \varepsilon}{4\pi i}\,
f(\varepsilon)\tanh \frac{\varepsilon}{2T}\, .$$ The contour of integration closes the cuts $|\varepsilon | >\Delta$ along the real axis. Upon analytical continuation one arrives at the following expression for the conductance: $$\begin{aligned}
\label{3}
&&\mathcal{G} =\frac{\pi e^2 g_{m}^{2} |T_{0}|^{4}d^{4}}{2\hbar Tk_F^6 a^2}
\sum_{s \ne l} \frac{n(\varepsilon_s)n(\varepsilon_l)\Delta^2}{|\bm{\rho}_{ls}|^2(\Delta^2- \varepsilon_{s}^{2})} e^{-2(z_{s}+z_{l})/a} \notag \\
&&\times e^{-2|\bm{\rho}_{ls} |\sqrt{\Delta^{2}-\varepsilon_{s}^{2}}/\pi\xi
\Delta} \delta \left(\varepsilon_{s}+\varepsilon_{l}+U_c\right)\, .\end{aligned}$$ Here $n(\varepsilon) \equiv \left(e^{\varepsilon/T}+1\right)^{-1}$ is the Fermi distribution, $U_c$ is the energy of the inter-site Coulomb repulsion, and $z$-axis is perpendicular to the interface.
In what follows we will replace $\sum_{l,s}$ by $g^2 \int d^3 r_l\, d^3 r_s
\, d\varepsilon_l \, d \varepsilon_s$ where $g$ is the the effective density of states in the hopping insulator. This is the density of states in the layer adjacent to the interface. Due to screening by the superconductor it is not affected by the Coulomb gap and can be considered as constant. Since we are dealing with the pairs close to the interface, the Coulomb repulsion is suppressed by screening. This screening can be conveniently regarded as an interaction of the charged particle with its image having the opposite charge. Thus the Coulomb correlation manifest themselves as the dipole-dipole interaction and for $\rho_{sl} \gg a$ one arrives at $U_c=e^2a^2/\kappa \rho_{sl}^3$. Requiring it to me smaller than $T$ one obtains a cut-off $\rho_{sl} \ge \rho_T \equiv a (e^2/\kappa a T)^{1/3}$.
As a crude estimate, we take $d^4 \sim k_F^{-4}$, while $T_0 \approx T_p
e^{-\Lambda}$ with $T_p \sim \varepsilon_F$. Bearing this in mind one finds $g_m^2 T_p^2 /k_F^6 \sim g_m^2 \varepsilon_F^2 /k_F^6 \sim 1$. Since the ratio $T_p/(ak_F)^2$ is of the order of the typical energy of the localized state, $\varepsilon_d \sim \hbar^2/ma^2$, one arrives at the estimate $$\label{es}
\frac{\mathcal{G}}{\mathcal{G}_n}\sim ga \rho_T^2\varepsilon_d e^{-2
\Lambda}, \quad \mathcal{G}_n \sim \frac{e^2}{\hbar} \, g aS \varepsilon_d\,
e^{-2 \Lambda}.$$ Here $\mathcal{G}_n$ is the conductance of a boundary between a normal metal and a hopping insulator, while $S$ is the contact area. The product $gaS\varepsilon_d$ is nothing but the number of localized centers within the layer of a thickness $a$ near the interface and the factor $ga \rho_T^2
\varepsilon_d$ expresses the probability of finding a critical pair, i. e., a pair of nearly located hopping centers that dominates the time reversal reflection processes discussed above.
The above approach holds, as we have already mentioned, only if the resistance of the typical time reversal resistor (TRR) is much larger then that of the critical hopping resistor, $R_h =(h/e^2 \gamma)\,e^{\zeta}$, where $\gamma$ is a dimensionless factor depending on the mechanism of electron-phonon interaction and $\zeta$ is the hopping exponent [Shklovskii-Efros]{}, i. e., with the exponential accuracy, as long as $4\Lambda > \zeta$.
There are many realistic situations where the barrier strength, $\Lambda$, is not too large; the Schottky barrier at the natural interface [Agrinskaya]{} is certainly the case like that. Consequently, if $\zeta \gg 1$, i. e., if the system is far from the metal-to-insulator transition point, the procedure of summation over the localized states should be modified. Namely, the choice of the pairs facilitating the charge transfer depends on the structure of the bonds connecting critical pairs to the rest of percolation cluster.
According to the above considerations the voltage drops mainly on the bond connecting percolation cluster in the HI the critical TRR for which the distance between its pair components is less than the correlation length, $\mathcal{L}$, of the backbone cluster. The incoherent electron transport can be ensured by a single *one* bond connecting the cluster to any of the TRR sites. Thus, the ratio $\mathcal{G}/\mathcal{G}_n$ is the probability to find a TRR contacting the percolation cluster.
To estimate this probability let us consider the layer with the thickness of the typical hopping distance, $r_h$, near the interface where all the bonds of the backbone cluster have necessarily have a site within this layer. The total number of states in this layer is $gSr_h \varepsilon_h$ where $\varepsilon_h=T\zeta$ is the width of the hopping band. This product can be estimated as $(\beta/8)(S/r_h)^2$, where $\beta$ is a numerical constant [@Shklovskii-Efros]. For the case of Mott variable range hopping (VRH), $\beta \approx 20$.
The number of TRRs in this layer can be estimated as follows. Let us note first that the conserving energy, $\varepsilon_s+ \varepsilon_l+U_c$, from $\delta$-function in Eq. (\[3\]) is associated with the band given by the broadening $\nu =\nu_0 \exp(-2 r_d/a)$. Here $r_d$ is distance to the nearest neighbor in HI. Indeed, the most natural source for the broadening of the resonance is coupling of the localized states. Secondly, since both electrons escape from the TRR through a single bond, the in-plane distance, $\rho_{sl}$ should not exceed the typical distance between the hopping sites, $r_h =a\zeta/2 \ll \xi$. Keeping the exponential accuracy we arrive at the following criterion that the resistance of TRR is less than the resistance of a typical hopping resistor: $$\label{bind}
4 \Lambda +\ln \frac{T}{\nu_0}+\frac{\max(|\varepsilon_s|,|\varepsilon_l|)}{T} + \frac{2r_d}{a} + \frac{ 2(z_s + z_l)}{a}< \zeta\, .$$ One may consider this equation as a generalization of the “connectivity criterion" to include the TRR. Here we deal with the independent variables $\varepsilon_s, \varepsilon_l, r_d, z_s$ and $z_l$ over which the averaging procedure should be done with an account of the restriction of Eq. ([bind]{}), Thus the number of the relevant TRRs is $$\begin{aligned}
&&8\pi^2 g^3S \int \mathrm{d}\varepsilon_l\int \! \mathrm{d} \varepsilon_s
\int^{\varepsilon_s}_0 \! \! \mathrm{d}\varepsilon_d \int \! r_d^2\, \mathrm{d}r_d \int\! \mathrm{d} z_s \mathrm{d} z_l \int_0^{r_h} \! \! \rho\, \mathrm{d} \rho \notag \\
&& \qquad \times \Theta \left[\frac{\max(\varepsilon_s,\varepsilon_l)}{T} +
\frac{2r_d}{a} + \frac{ 2(z_s + z_l)}{a}- \alpha \zeta\right] \label{NAR}\end{aligned}$$ where $\alpha \equiv 1-[4\Lambda +\ln (T/\nu_0)]/\zeta < 1$. Let us now measure the energies in units of $\alpha \varepsilon_h$ and lengths in units of $\alpha r_h $, where $\varepsilon_h \equiv T\zeta$. Again, the product $gr_h^3 \varepsilon_h$ can be estimated as $\beta/8$, and we obtain the number of effective TRRs as $N_A \sim \mathcal{A}\alpha^7 S/r_h^2$, where $$\begin{aligned}
&&\mathcal{A}=4\pi^2 (\beta/8)^2 \int \mathrm{d}\epsilon_l\int \! \mathrm{d}
\epsilon_s \int^{\epsilon_s}_0 \! \! \mathrm{d}\epsilon_d \int \! \eta_d^2\,
\mathrm{d}r_d \int\! \mathrm{d} \eta_s \int\! \mathrm{d}\eta_l \notag \\
&& \times \Theta \left[\max(\epsilon_s,\epsilon_l) + \eta_d + \eta_s +
\eta_l) - 1\right]\approx 0.1 \, . \label{NAR1}\end{aligned}$$ Consequently, $\mathcal{G}/\mathcal{G}_n \sim \mathcal{A} \alpha^7 \ll 1$. One concludes that the difference between the “contact" resistances in normal and superconducting states is dominated by the contribution of TRRs. Since $\mathcal{G}_N \approx R_h^{-1}S/\mathcal{L}^2$, $$\label{eq:mr1}
\delta R \equiv \mathcal{G}_n^{-1}-\mathcal{G}^{-1}\approx -\mathcal{G}^{-1}=-R_h\, (\mathcal{L}^2/S \mathcal{A} \alpha^7)\, .$$ Note that the interface resistance is of the order of the resistance of HI layer with the thickness $\sim \mathcal{L}/(\mathcal{A}\alpha^7) >> \mathcal{L}$. Since $\mathcal{L} \sim a \zeta^2$, one concludes that for $\zeta > 10$ the interface resistance can be comparable or even exceed the hopping resistance if the thickness of the sample (or of the contact) is $\lesssim 10
$ $\mu$m.
The resistance estimated above can be experimentally measured as a magnetoresistance in magnetic fields higher than the critical field for superconductivity (similar effect for quasiparticle channel was studied in [@Agrinskaya]).
We were implicitly assuming so far that the variable range hopping occurs according to the Mott’s law. This assumption certainly holds near the interface where the Coulomb gap is screened by a superconductor. However in the bulk of HI the Efros-Shklovskii (ES) law [@CG; @Shklovskii-Efros] can become the dominant hopping mechanism. Then the value of $\zeta$ in the connectivity criterion (\[bind\]) will be controlled by the Coulomb gap, $\zeta \to \zeta_{\text{ES}}=(\beta_1 e^2/\kappa a T)^{1/2}$, where $\beta_1$ is a numerical constant [@CG; @Shklovskii-Efros]. In this case, it turns out that each bond of the ES backbone cluster finds some TRR ensuring charge transfer. Thus in the limiting case of a weak tunneling interface barrier the contact resistance will be the same for both normal metal and superconductor leads. This fact can be used to discriminate between Mott and Efros-Shklovskii laws in the situation when it is difficult to do so from temperature dependence.
Note that, in principle, the charge transfer involving double occupied localized states is possible. However, such a process would require an additional activation exponential factor, $\propto e^{-U/T}$, where $U$ is the on-site correlation energy. One can also consider processes where a double occupied center (so-called $D^-$-center) serves as an intermediate state for the phonon-assisted two-electron tunneling. This channel is unfavorable (at least in the case of a large interface barrier) because of the above-mentioned exponential factor and a small pre-exponential factor due to phonon-assisted tunneling. For the weak tunnel barriers the conductance is controlled by “typical" hopping sites. In this case $D^{-}$ channel is suppressed either by the additional tunneling exponential, $\propto e^{-4 r_h/a}$, or by a small probability to form a close triad of hopping sites. Therefore the $D^{-}$ channel can be also neglected.
To conclude, we have developed a theory of the low-temperature charge transfer between a superconductor and a hopping insulator and calculated the interface resistance. This resistance is dominated by time reversal reflection processes involving localized states in the insulator. It is the time reversal reflection process that allows the low-temperature measurements of hopping transport utilizing superconducting electrodes in the experimental setups. In the Efros-Shklovskii VRH regime, the corresponding interface resistance is small as compared to the bulk hopping resistance and is nearly equal to the resistance at the interface between the HI and normal metal. On the contrary, in the Mott hopping regime (relevant, in particular, for 2D gated structures), the interface resistance grows much larger and becomes commensurate (or even exceeds) to the bulk hopping resistance. This effect is especially pronounced in the mesoscopic samples. The contribution from the interface resistance can be detected by application of the external magnetic field: the relatively weak magnetic field will drive the superconductor into the normal state, but will not affect the hopping transport eliminating thus time reversal reflection process. This effect holds even in the case where the interface contribution is less than the typical resistance of the hopping system itself.
We are delighted to thank A. S. Ioselevich, F. W. J. Hekking, F. Pistolesi, and J. Bergli for discussion and critical reading the manuscript. We are also grateful to L. I. Glazman and B. I. Shklovskii for posing important questions. This work was supported by the U. S. Department of Energy Office of Science through contract No. W-31-109-ENG-38. A.A.Z. is grateful to International Center for fundamental physics in Moscow and to the Fund of noncommercial programs DYNASTY.
[99]{}
A. F. Andreev, Zh. Eksp. Teor. Fiz. **46** 1823 (1964) \[Sov. Phys.-JETP **19**, 1228 (1964)\].
V. P. Galaiko and A. F. Volkov, in *Nonequilibrium Superconductivity*, ed. by D. Langenberg and A. I. Larkin (North-Holland, 1986).
N. V. Agrinskaya, V. I. Kozub, and R. Rentzsch, JETP **84**, 814 (1997).
We assume that the on-site correlation (Hubburd) energy is sufficiently large to prevent double occupation of the sites.
N. V. Agrinskaya, V. I. Kozub, A. V. Chernyaev, D. V. Shamshur, A. A. Zuzin, cond-mat/0410697 (unpublished).
G. Deutsher and D. Feinberg, Appl. Phys. Lett. **76**, 487 (2000); G. Falci, D. Feinberg, and F. W. J. Hekking, Europhys. Lett. **54**, 255 (2001); D. Feinberg, Eur. Phys. J. B **36**, 419 (2003).
P. Recher, E. V. Sukhorukov, and D. Loss, **63**, 165314 (2001).
O. Sauret, T. Martin, D. Feinberg, , 245313 (2004); cond-mat/0410325 (unpublished).
R. Kubo, Journ. Phys. Soc. Japan, **12**, 570 (1957).
J. Bardeen, , 57 (1961).
A. A. Abrikosov, L. P. Gorkov, and I. E. Dzyaloshinskii, *Quantum field theoretical methods in statistical physics* (Pergamon Press, 1965).
A. I. Larkin and B. I. Shklovskii, Phys. Stat. Sol B **230**, 189 (2002).
B. I. Shklovskii and A. L. Efros, *Electronic properties of doped semiconductors* (Springer- Verlag, 1984).
A. L. Efros and B. I. Shklovskii, J. Phys. C **8**, L49 (1975).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In this paper, we construct efficient schemes based on the scalar auxiliary variable (SAV) block-centered finite difference method for the modified phase field crystal (MPFC) equation, which is a sixth-order nonlinear damped wave equation. The schemes are linear, conserve mass and unconditionally dissipate a pseudo energy. We prove rigorously second-order error estimates in both time and space for the phase field variable in discrete norms. We also present some numerical experiments to verify our theoretical results and demonstrate the robustness and accuracy.'
author:
- 'Xiaoli Li [^1].'
- 'Jie Shen [^2].'
bibliography:
- 'SAV\_MPFC.bib'
title: 'Efficient Linear and Unconditionally Energy Stable Schemes for the Modified Phase Field Crystal Equation[^3] '
---
Modified phase field crystal, scalar auxiliary variable (SAV), energy stability, error estimate, numerical experiments
35G25, 65M06, 65M12, 65M15
Introduction
============
The phase field crystal (PFC) model was developed in [@elder2004modeling; @elder2002modeling] to model the crystallization process in the purification of solid compounds. It has been used to model the evolution of the atomic-scale crystal growth on diffusive time scales. In the PFC model, the phase field variable is introduced to describe the phase transition from the liquid phase to the crystal phase. The model is versatile and able to simulate various phenomena, such as grain growth, epitaxial growth, reconstructive phase transitions, material hardness, and crack propagations. Numerical methods and simulations for the PFC model have been studied extensively, including finite element method [@gomez2012unconditionally], finite difference methods [@li2017efficient; @wise2009energy; @zhang2013adaptive], local discontinuous Galerkin method [@guo2016local] and Fourier-spectral method [@yang2017linearly].
The modified phase field crystal (MPFC) equation was introduced in [@stefanovic2006phase] to model phase-field crystals with elastic interactions. The MPFC equation can be viewed as a perturbed gradient flow with respect to a free energy, and is a sixth order nonlinear damped wave equation. However, as pointed out in [@wang2010global], the original free energy of the MPFC equation may increase in time on some time intervals. Thus A pseudo energy is introduced in [@wang2010global] and shown to be dissipative. There exist a number of work on the numerical approximations of the MPFC model. First and second order accurate nonlinear convex splitting schemes have been proposed in [@wang2010global; @baskaran2013convergence], and are proved to be unconditional energy stable and convergent. A nonlinear multigrid method is used to solve the nonlinear system at each time step [@baskaran2013energy]. Guo and Xu [@guo2018high] developed a first-order and a second-order nonlinear convex splitting, and a first-order linear energy stable fully discrete with local discontinuous Galerkin (LDG) methods. Very recently, Li and his coauthors [@li2019efficient] proposed unconditional energy stable schemes based on the “Invariant Energy Quadratization” (IEQ) approach for the MPFC model but without convergence proof. The convergence analysis is challenging due to the nonlinear hyperbolic nature of the MPFC equation. To our knowledge, there is no second-order convergence analysis on any linear scheme for the MPFC equation.
The main goals of this paper are to construct linear and unconditionally energy stable schemes based on the recently proposed scalar auxiliary variable (SAV) approach [@shen2018scalar; @shen2019new], and to carry out a rigorous error analysis. More specifically, we construct two SAV block-centered finite difference schemes for the MPFC equation based on the Euler backward and Crank-Nicolson schemes respectively, and show that they are unconditionally energy stable with a suitably defined pseudo energy, and we establish second-order convergence in both time and space in a discrete $L^{\infty}(0,T;H^3(\Omega))$ norm. The rest of the paper is organized as follows. In Section 2 we describe the MPFC model and reformulate it using the SAV approach. In Section 3 we construct fully discrete schemes for the reformulated MPFC equation by block-centered finite difference method, and show that the scheme conserves mass and is unconditionally energy stable. In Section 4 we derive the error estimate for the MPFC model. In Section 5 some numerical experiments are presented to verify the accuracy of the proposed numerical schemes.
The MPFC model and its semi-discretization in time {#sec:Preliminary}
==================================================
We describe in this section the MPFC model, its reformulation using the SAV approach, construct a second-order SAV semi-discretization scheme and show that it preserves mass and dissipates a pseudo energy.
The MPFC model and its SAV reformulation
----------------------------------------
Consider the free energy (cf. [@baskaran2013energy; @baskaran2013convergence; @guo2018high]) $$\label{e_definition of free energy}
\aligned
E(\phi)=\int_{\Omega}\{\frac{1}{2}(\Delta \phi)^2-|\nabla \phi|^2+\frac \alpha 2 \phi^2+F(\phi)\}d\textbf{x},
\endaligned$$ where $\Omega\subset \mathbb{R}^d \,(d=1,2,3)$. The phase field variable $\phi$ is introduced to represent the concentration field of a coarse-grained temporal average of the density of atoms. $F(\phi)=\frac 1 4 \phi^4$. Here $\alpha=1-\epsilon$ with $\epsilon\ll1$. Then the MPFC model, which is designed to describe the elastic interactions: $$\label{e_continuous_model}
\left\{
\begin{array}{l}
\displaystyle \frac{\partial^2 \phi}{\partial t^2}+\beta\frac{\partial \phi}{\partial t}=M\Delta\mu, \ \ \textbf{x}\in \Omega, t>0,\\
\displaystyle \mu=\Delta^2 \phi+2\Delta \phi+\alpha\phi+F^{\prime}(\phi), \ \ \textbf{x}\in \Omega, t>0,\\
\displaystyle \phi(\textbf{x},0)=\phi_0(\textbf{x}),
\end{array}
\right.$$ where $\beta> 0$. But we should note that the energy may actually increase on some time intervals. The PFC and MPFC equations have close relationship. However, we should keep in mind that the original energy of the MPFC equation may increase in time on some time intervals. Thus it is desirable to introduce a pseudo energy. Besides, we can observe that does not satisfy the mass conservation due to the term $\frac{\partial^2 \phi}{\partial t^2}$. However, it is possible to verify that $\int_{\Omega}\frac{\partial \phi}{\partial t}d\textbf{x}=0$ with a suitable initial condition for $\frac{\partial \phi}{\partial t}$.
To fix the idea, we consider the homogeneous Neumann boundary conditions: $$\label{e_boundary and initial condition}
\partial_\textbf{n}\phi|_{\partial \Omega}=0,\ \partial_\textbf{n}\Delta \phi|_{\partial \Omega}=0, \
\partial_\textbf{n}\mu|_{\partial \Omega}=0,$$ where $\textbf{n}$ is the unit outward normal vector of the domain $\Omega$.
The homogeneous Neumann boundary conditions are assumed to simplify the presentation. The algorithm and its analysis also hold for the periodic boundary conditions with very little modification. One can refer to [@wang2011energy Lemma 3.6] for more detail about the periodic boundary conditions. While we only present the algorithm and analysis for homogeneous Neumann boundary conditions, we do present some numerical results with periodic boundary conditions in Section 5.
To introduce an appropriate pseudo energy for the MPFC equation, we need to define the $H^{-1}$ inner-product [@baskaran2013energy]. Let $u_i \,(i=1,2) \in \{f\in L^2(\Omega)|\ \int_{\Omega}fd\textbf{x}=0\}:=L_0^2(\Omega)$, we define $\eta_{u_i}\in H^2(\Omega)\cap L_0^2(\Omega)$ to be the unique solution to the following problem: $$\label{e_pseudo_energy1}
\aligned
-\Delta \eta_{u_i}=u_i \ \ in \ \Omega, \ \ \partial_\textbf{n} \eta_{u_i} |_{\partial \Omega}=0.
\endaligned$$ Then we have $\eta_{u_i}=-\Delta^{-1}u_i$. Define $$\label{e_pseudo_energy2}
\aligned
(u_1,u_2)_{H^{-1}}:=(\nabla\eta_{u_1},\nabla\eta_{u_2})_{L^2}.
\endaligned$$ Using integration by parts, we can obtain $$\label{e_pseudo_energy3}
\aligned
(u_1,u_2)_{H^{-1}}=-(\Delta^{-1}u_1,u_2)_{L^2}=-(\Delta^{-1}u_2,u_1)_{L^2}=(u_2,u_1)_{H^{-1}}.
\endaligned$$ Then we define $\|u\|_{H^{-1}}=\sqrt{(u,u)_{H^{-1}}}$ for every $u\in L^2_0(\Omega)$.
In order to construct an efficient scheme for the MPFC equation , we first reformulate it using the so called SAV approach [@shen2018scalar]. Introducing two auxiliary functions as follows: $$\label{e_auxiliary1}
\aligned
\psi=\frac{\partial \phi}{\partial t},\ r=\sqrt{E_1(\phi)}:=\sqrt{\int_{\Omega}F(\phi)d\textbf{x}}.
\endaligned$$ Then the MPFC equation can be recast as the following system:
\[e\_model\_recast\] $$\begin{aligned}
&\frac{\partial \psi}{\partial t}+\beta\psi=M\Delta \mu, \label{e_model_recastA}\\
& \mu=\Delta^2 \phi+2\Delta \phi+\alpha\phi+\frac{r(t)}{\sqrt{E_1(\phi)}}F^{\prime}(\phi), \label{e_model_recastB}\\
&r_t=\frac{1}{2\sqrt{E_1(\phi)}}\int_{\Omega}F^{\prime}(\phi)\phi_t d\textbf{x}. \label{e_model_recastC}
\end{aligned}$$
Define the pseudo energy $$\label{e_pseudo_energy}
\aligned
\mathcal{E}(\phi,r,\psi)=\int_{\Omega}(\frac{1}{2}(\Delta \phi)^2-|\nabla \phi|^2+\frac \alpha 2 \phi^2)d\textbf{x}+r^2+\frac{1}{2M}\|\psi\|_{H^{-1}}^2,
\endaligned$$ which requires that $\int_{\Omega}\psi=0$ for well posedness. As long as $\psi=\frac{\partial \phi}{\partial t}$ is of mean zero, we can obtain the following dissipation law: $$\label{e_evolution of pseudo energy}
\aligned
\frac{d }{dt}\mathcal{E}(\phi,\xi,r,\psi)=&\int_{\Omega}\mu\frac{\partial \phi}{\partial t}d\textbf{x}-\frac 1 M\int_{\Omega}\Delta \eta_{\psi}\frac{\partial \eta_{\psi}}{\partial t}d\textbf{x}\\
=&(\mu,\frac{\partial \phi}{\partial t})-\frac 1 M(\psi,\Delta^{-1}\frac{\partial \psi}{\partial t})\\
=&\frac{\beta}{M}(\psi,\Delta^{-1}\psi)=-\frac{\beta}{M}\|\psi\|_{H^{-1}}^2\leq 0,
\endaligned$$ where $\eta_{\psi}=(-\Delta )^{-1}\psi$.
A second-order semi-discrete scheme
-----------------------------------
Let $N>0$ be a positive integer and $J=(0,T]$. Set $\Delta t=T/N,\ t^n=n\Delta t,\ \ \rm{for} \ n\leq N,$ where $T$ is the final time. The second-order semi-discrete scheme based on the Crank-Nicolsion method for is as follows:
Assuming $\phi^n$, $\psi^n$ and $r^n$ are known, then we update $\phi^{n+1}$, $\psi^{n+1}$ and $r^{n+1}$ by solving
\^[n+1]{}-\^[n]{}+t\^[n+1/2]{}=Mt\^[n+1/2]{}, \[e\_semi\_second-order1\]\
t\^[n+1/2]{}=\^[n+1]{}-\^n, \[e\_semi\_second-order2\]\
\^[n+1/2]{}=\^2 \^[n+1/2]{}+2\^[n+1/2]{}+\^[n+1/2]{}\
+F\^(\^[n+1/2]{}), \[e\_semi\_second-order3\]\
r\^[n+1]{}-r\^n=(F\^(\^[n+1/2]{}), \^[n+1]{}-\^n), \[e\_semi\_second-order4\]
where $f^{n+1/2}=(f^{n+1}+f^{n})/2$ and $\tilde{f }^{n+1/2}=(3 f ^n-f^{n-1})/2$ for any function $f$. For the case of $n=0$, we can computer $\tilde{\phi }^{1/2}$ by the first-order scheme.
\[thm: semi-discrete mass conservation\] The scheme (\[e\_semi\_second-order1\])-(\[e\_semi\_second-order4\]) is mass conserving, i.e., $ \int_{\Omega} \phi^{n+1} d\textbf{x}= \int_{\Omega} \phi^{n} d\textbf{x} $ for all $n$, and unconditionally stable in the sense that $$\label{semi-discrete energy stability1}
\aligned
\tilde{\mathcal{E}}(\phi^{n+1},r^{n+1},\psi^{n+1})-\tilde{\mathcal{E}}(\phi^n,r^n,\psi^n)\leq -\frac{\beta}{M}\Delta t\|\psi^{n+1/2}\|_{H^{-1}}^2,
\endaligned$$ where $\tilde{\mathcal{E}}(\phi^n,r^n,\psi^n)=\mathcal{E}(\phi^n,r^n,\psi^n)+\frac 1 2\|\nabla \phi^{n}-\nabla \phi^{n-1}\|^2$.
Taking the inner products of with $1$ leads to $$\label{semi-discrete mass conservation1}
\aligned
(\psi^{n+1}-\psi^{n},1)+\beta\Delta t(\psi^{n+1/2},1)=M\Delta t(\Delta \mu^{n+1/2},1).
\endaligned$$ Similarly, by taking the inner products of with $1$, we can obtain $$\label{semi-discrete mass conservation2}
\aligned
(\phi^{n+1}-\phi^n,1)=\Delta t(\psi^{n+1/2},1).
\endaligned$$ Using the integration by parts, the term on the right hand side of can be transformed into $$\label{semi-discrete mass conservation3}
\aligned
M\Delta t(\Delta \mu^{n+1/2},1)=-M\Delta t\left((\nabla \mu^{n+1/2},\nabla 1)+(\nabla \mu^{n+1/2}, \nabla 1) \right)=0.
\endaligned$$ Then can be recast as follows: $$\label{semi-discrete mass conservation4}
\aligned
(1+\frac \beta 2\Delta t)(\psi^{n+1},1)=(1-\frac \beta 2\Delta t)(\psi^{n},1).
\endaligned$$ Combining with the condition on the initial condition $(\psi^0,1)=0$ leads to $(\psi^{n+1},1)=0$ for all $n\leq 0$. Recalling , we have $(\phi^{n+1},1)=(\phi^{n},1)$.
Next, we prove . Taking the inner products of with $\mu^{n+1/2}$ gives $$\label{semi-discrete energy stability2}
\aligned
\Delta t(\psi^{n+1/2},\mu^{n+1/2})=(\phi^{n+1}-\phi^n,\mu^{n+1/2}).
\endaligned$$ Taking the inner products of with $\phi^{n+1}-\phi^n$, we have $$\label{semi-discrete energy stability3}
\aligned
&(\mu^{n+1/2},\phi^{n+1}-\phi^n)=(\Delta^2 \phi^{n+1/2}, \phi ^{n+1}-\phi^{n}) +2(\Delta \tilde{ \phi }^{n+1/2}, \phi^{n+1}-\phi^n)\\
&\ \ \ \ \ +\alpha( \phi ^{n+1/2}, \phi^{n+1}-\phi^{n}) +(\frac{r^{n+1/2}}{\sqrt{E_1(\tilde{ \phi }^{n+1/2})}}F^{\prime}(\tilde{ \phi }^{n+1/2}), \phi ^{n+1}-\phi ^n).
\endaligned$$ The first three terms on the right-hand side of can be estimated with the help of the integration by parts: $$\label{semi-discrete energy stability4}
\aligned
(\Delta^2 \phi^{n+1/2}, \phi^{n+1}-\phi^n) =\frac1 2(\|\Delta \phi^{n+1}\|^2-\|\Delta \phi^{n}\|^2),
\endaligned$$
$$\label{semi-discrete energy stability5}
\aligned
&2(\Delta \tilde{ \phi }^{n+1/2}, \phi^{n+1}- \phi^n) \\
=&-\|\nabla \phi ^{n+1}\|^2+\|\nabla \phi^{n}\|^2+\frac1 2(\|\nabla \phi^{n+1}-\nabla \phi^{n}\|^2-\|\nabla \phi^{n}-\nabla \phi^{n-1}\|^2)\\
&+\frac1 2\|\nabla \phi^{n+1}-2\nabla \phi^{n}+\nabla \phi^{n-1}\|^2,
\endaligned$$
and $$\label{semi-discrete energy stability5_add}
\aligned
\alpha(\phi^{n+1/2}, \phi^{n+1}-\phi^n) =\frac\alpha 2(\| \phi^{n+1}\| ^2-\| \phi^{n}\| ^2).
\endaligned$$ Multiplying by $(r^{n+1}+r^n)$ leads to $$\label{semi-discrete energy stability6}
\aligned
(r^{n+1})^2-(r^n)^2=(\frac{r^{n+1/2}}{\sqrt{E_1 (\tilde{ \phi }^{n+1/2})}}F^{\prime}(\tilde{ \phi }^{n+1/2} ), \phi^{n+1}-\phi^n).
\endaligned$$ Combining with and -, we have $$\label{semi-discrete energy stability7}
\aligned
&\frac1 2(\|\Delta \phi^{n+1}\|^2-\|\Delta \phi^{n}\| ^2)
-\|\nabla \phi^{n+1}\|^2+\|\nabla \phi^{n}\|^2+(r^{n+1})^2-(r^n)^2\\
+&\frac1 2(\|\nabla \phi^{n+1}-\nabla \phi^{n}\|^2-\|\nabla \phi^{n}-\nabla \phi^{n-1}\|^2)\\
+&\frac1 2\|\nabla \phi^{n+1}-2\nabla \phi^{n}+\nabla \phi^{n-1}\|^2+\frac\alpha 2(\|\phi^{n+1}\|^2-\| \phi^{n}\| ^2)\\
=&\Delta t(\psi^{n+1/2}, \mu^{n+1/2}).
\endaligned$$ Since recalling , we can derive $$\label{semi-discrete energy stability8}
\aligned
\frac{1}{2M}(\|\psi^{n+1}\|_{H^{-1}}^2-&\|\psi^n\|_{H^{-1}}^2)=\frac 1 M(\psi^{n+1}-\psi^n,\psi^{n+1/2})_{-1}\\
=&-\frac 1 M(\psi^{n+1}-\psi^n,\Delta^{-1}\psi^{n+1/2})\\
=&-\frac{\beta}{M}\Delta t\|\psi^{n+1/2}\|_{H^{-1}}^2-\Delta t( \mu^{n+1/2},\psi^{n+1/2}).
\endaligned$$ Finally, combining with gives the desired result.
Since the scheme (\[e\_semi\_second-order1\])-(\[e\_semi\_second-order4\]) is linear, one can also show that it admits a unique solution, and can be efficiently implemented. For the sake of brevity, we shall provide detail only for the fully discretized scheme presented in the next section.
Fully discrete schemes and their properties
===========================================
In this section, we construct two linear SAV block-centered finite difference schemes for the SAV reformulated MPFC equation .
Full discrete schemes based on block-centered finite difference method
----------------------------------------------------------------------
First we describe briefly the block-centered finite difference framework that we will employ to define and analyze our schemes. To fix the idea, we set $\Omega=(0,L_x)\times
(0,L_y)$, although the algorithm and analysis presented below apply also to the one- and three-dimensional rectangular domains.
We begin with the definitions of grid points and difference operators. Let $L_x=N_xh_x$ and $L_y=N_yh_y$, where $h_x$ and $h_y$ are grid spacings in $x$ and $y$ directions, and $N_x$ and $N_y$ are the number of grids along the $x$ and $y$ coordinates, respectively. The grid points are denoted by $$(x_{i+1/2},y_{j+1/2}),\ \ i=0,...,N_x,\ \ j=0,...,N_y,$$ and $$\aligned
&x_{i}=(x_{i-\frac{1}{2}}+x_{i+\frac{1}{2}})/2, \ \ i=1,...,N_x,\\
&y_{j}=(y_{j-\frac{1}{2}}+y_{j+\frac{1}{2}})/2, \ \ j=1,...,N_y.
\endaligned$$ Define $$\aligned
&[d_{x}g]_{i+\frac{1}{2},j}=(g_{i+1,j}-g_{i,j})/h_x,\\
&[d_{y}g]_{i,j+\frac{1}{2}}=(g_{i,j+1}-g_{i,j})/h_y,\\
&[D_{x}g]_{i,j}=(g_{i+\frac{1}{2},j}-g_{i-\frac{1}{2},j})/h_x,\\
&[D_{y}g]_{i,j}=(g_{i,j+\frac{1}{2}}-g_{i,j-\frac{1}{2}})/h_y,\\
&[\Delta_hg]_{i,j}=D_x(d_xg)_{i,j}+D_y(d_yg)_{i,j}.
\endaligned$$ Define the discrete inner products and norms as follows, $$\aligned
&(f,g)_{m}=\sum\limits_{i=1}^{N_{x}}\sum\limits_{j=1}^{N_{y}}h_xh_yf_{i,j}g_{i,j},\\
&(f,g)_{x}=\sum\limits_{i=1}^{N_{x}-1}\sum\limits_{j=1}^{N_{y}}h_xh_yf_{i+\frac{1}{2},j}g_{i+\frac{1}{2},j},\\
&(f,g)_{y}=\sum\limits_{i=1}^{N_{x}}\sum\limits_{j=1}^{N_{y}-1}h_xh_yf_{i,j+\frac{1}{2}}g_{i,j+\frac{1}{2}}.
\endaligned$$
\[le\_discrete-integration-by-part\] Let $q_{i,j},w_{1,i+1/2,j}\ and \ w_{2,i,j+1/2} $ be any values such that $w_{1,1/2,j}=w_{1,N_x+1/2,j}=w_{2,i,1/2}=w_{2,i,N_y+1/2}=0$, then $$(q,D_xw_1)_m=-(d_xq,w_1)_x,$$ $$(q,D_yw_2)_m=-(d_yq,w_2)_y.$$
Next we define the discrete $H^{-1}$ inner-product. Suppose $\eta_{\phi_i}\in \{f|(f,1)_m=0\}:=\mathcal{H}$ to be the unique solution to the following problem: $$\label{e_discrete_inner-product1}
\aligned
-\Delta_h\eta_{\phi_i}=\phi_i,
\endaligned$$ where $\eta_{\phi_i}$ satisfies the discrete homogenous Neumann boundary condition $$\label{e_first-order10}
\left\{
\begin{array}{l}
\displaystyle (\eta_{\phi_i})_{0,j}=(\eta_{\phi_i})_{1,j}, \ (\eta_{\phi_i})_{N_x+1,j}=(\eta_{\phi_i})_{N_x,j}, \ \ j=1,2,\ldots,N_y,\\
\displaystyle (\eta_{\phi_i})_{k,0}=(\eta_{\phi_i})_{k,1}, \ \ (\eta_{\phi_i})_{k,N_y+1}=(\eta_{\phi_i})_{k,N_y}, \ \ k=1,2,\ldots,N_x.
\end{array}
\right.$$ We define the bilinear form $$(\phi_1,\phi_2)_{-1}=(d_x\eta_{\phi_1},d_x\eta_{\phi_2})_x+(d_y\eta_{\phi_1},d_y\eta_{\phi_2})_y,$$ for any $\phi_1,\phi_2\in \mathcal{H}$. Then we can obtain that $(\phi_1,\phi_2)_{-1}$ is an inner product on the space $\mathcal{H}$. Moreover, we have $$(\phi_1,\phi_2)_{-1}=-(\phi_1,\Delta^{-1}_h\phi_2)_m=-(\Delta^{-1}_h\phi_1,\phi_2)_m.$$ Then we can define the discrete $H^{-1}$ norm $\|\phi\|_{-1}=\sqrt{(\phi,\phi)_{-1}}$.
Hereafter, we use $C$, with or without subscript, to denote a positive constant, which could have different values at different appearances.
Let us denote by $\{Z^n, W^n, R^n,\Psi^n\}_{n=0}^{N}$ the block-centered finite difference approximations to $\{\phi^n,\mu^n, r^n,\psi^n\}_{n=0}^{N}$. The second-order scheme defined by the Crank-Nicolsion method for is as follows:
Set the boundary condition as $$\label{e_first-order1}
\left\{
\begin{array}{l}
\displaystyle Z_{0,j}=Z_{1,j}, \ Z_{N_x+1,j}=Z_{N_x,j}, \ \ j=1,2,\ldots,N_y,\\
\displaystyle Z_{i,0}=Z_{i,1}, \ \ Z_{i,N_y+1}=Z_{i,N_y}, \ \ i=1,2,\ldots,N_x, \\
\displaystyle W_{0,j}=W_{1,j}, \ W_{N_x+1,j}=W_{N_x,j}, \ \ j=1,2,\ldots,N_y,\\
\displaystyle W_{i,0}=W_{i,1}, \ \ W_{i,N_y+1}=W_{i,N_y}, \ \ i=1,2,\ldots,N_x, \\
\displaystyle \Delta_h Z_{0,j}=\Delta_h Z_{1,j}, \ \Delta_h Z_{N_x+1,j}=\Delta_h Z_{N_x,j}, \ \ j=1,2,\ldots,N_y,\\
\displaystyle \Delta_h Z_{i,0}=\Delta_h Z_{i,1}, \ \ \Delta_h Z_{i,N_y+1}=\Delta_h Z_{i,N_y}, \ \ i=1,2,\ldots,N_x.
\end{array}
\right.$$ We find $\{Z^{n+1}, W^{n+1}, R^{n+1},\Psi^{n+1}\}_{n=0}^{N-1}$ such that
\^[n+1]{}-\^[n]{}+t\^[n+1/2]{}=Mt\_hW\^[n+1/2]{}, \[e\_second-order1\]\
t\^[n+1/2]{}=Z\^[n+1]{}-Z\^n, \[e\_second-order2\]\
W\^[n+1/2]{}=\_h\^2Z\^[n+1/2]{}+2\_h\^[n+1/2]{}+Z\^[n+1/2]{}\
+F\^(\^[n+1/2]{}), \[e\_second-order3\]\
R\^[n+1]{}-R\^n=(F\^(\^[n+1/2]{}), Z\^[n+1]{}-Z\^n)\_m, \[e\_second-order4\]
where $f^{n+1/2}=(f^{n+1}+f^{n})/2,\ f=W,\Psi, R$ and $\tilde{Z}^{n+1/2}=(3Z^n-Z^{n-1})/2$. For the case of $n=0$, we can computer $\tilde{Z}^{1/2}$ by the first-order scheme.
Efficient implementation {#eff_impl}
------------------------
A remarkable property about the above schemes is that it can be solved very efficiently. We demonstrate the detail procedure to solve the second-order SAV scheme (\[e\_second-order1\])-(\[e\_second-order4\]). Indeed, we can eliminate $\Psi^{n+1}$, $W^{n+1}$, $R^{n+1}$ from (\[e\_second-order1\])-(\[e\_second-order4\]) to obtain $$\label{e_implementation1}
\aligned
&(\frac{2}{\Delta t}+\beta)\frac{Z^{n+1}-Z^n}{\Delta t}-\frac{2 }{\Delta t} \Psi^{n}=M( \frac{1}{2} \Delta_h^3Z^{n+1}+ \frac{1}{2} \Delta_h^3Z^{n} + 2\Delta_h^2 \tilde{Z}^{n+1/2} \\
&\ \ \ \ \ \ + \frac{\alpha}{2} \Delta_hZ^{n+1} + \frac{\alpha}{2} \Delta_hZ^{n} )
+M \frac{\Delta_hF^{\prime}(\tilde{Z}^{n+1/2} )}{ \sqrt{E_1^h(\tilde{Z}^{n+1/2} )}}\left(R^n+ (\frac{F^{\prime}(\tilde{Z}^{n+1/2} )}{4\sqrt{E_1^h(\tilde{Z}^{n+1/2} )}},Z^{n+1}-Z^n)_m\right).
\endaligned$$ Let $b^n=\frac{F^{\prime}(\tilde{Z}^{n+1/2} )}{\sqrt{E_1^h(\tilde{Z}^{n+1/2} )}}$, then the above equation can be transformed into the following: $$\label{e_implementation2}
\aligned
&\mathcal{A}Z^{n+1}-\frac{M}{4}(b^n,Z^{n+1})_m\Delta_h b^n=f^n,
\endaligned$$ where $\mathcal{A}=(\frac{2 }{\Delta t^2}+\frac{\beta}{\Delta t})I-\frac{M}{2} \Delta_h^3-\frac{M}{2} \alpha\Delta_h$ and the right term $$\aligned
f^n=& \frac{2}{\Delta t}\Psi^n+\left( (\frac{2}{\Delta t^2}+\frac{\beta}{\Delta t})I+ \frac{M}{2} \Delta_h^3 + \frac{M}{2} \alpha\Delta_h \right)Z^n \\
& +2M\Delta_h^2 \tilde{Z}^{n+1/2} +M\left(R^n-\frac 1 4(b^n,Z^n)_m\right)\Delta_hb^n.
\endaligned$$ In order to solve the above equation, we should determine $(b^n,Z^{n+1})_m$ first. To this end, multiplying by $\mathcal{A}^{-1}$ leads to $$\label{e_implementation3}
\aligned
&Z^{n+1}-\frac{M}{4}(b^n,Z^{n+1})_m\mathcal{A}^{-1}\Delta_h b^n=\mathcal{A}^{-1}f^n.
\endaligned$$ Multiplying by $b^n_{i,j}h_xh_y$, and making summation on $i,j$ for $1\leq i\leq N_x,\ 1\leq j\leq N_y$, we have $$\label{e_implementation5}
\aligned
(b^n,Z^{n+1})_m=\frac{(b^n,\mathcal{A}^{-1}f^n)_m}{1-\frac{M}{4 }(\mathcal{A}^{-1}\Delta_h b^n,b^n)_m}.
\endaligned$$ Since $M>0$ and for $\alpha,\,\beta\ge 0$, $\mathcal{A}^{-1}\Delta_h$ is negative definite. So $(b^n,Z^{n+1})_m$ can be uniquely determined from above. Finally, we can get $Z^{n+1}$ by . Since the scheme is linear, the above procedure shows that it admits a unique solution.
In conclusion, the second-order SAV scheme (\[e\_second-order1\])-(\[e\_second-order4\]) can be effectively implemented in the following algorithm:
**Given:** $\Psi^n$, $Z^n$, $R^n$ and $b^n$.
Step 1. Computer $(\mathcal{A}^{-1}\Delta_h b^n,b^n)_m$. This can be accomplished by solving a sixth-order equation with constant coefficients.\
Step 2. Calculate $(b^n,Z^{n+1})_m$ using , which requires solving another sixth-order equation $\mathcal{A}^{-1}f^n$ with constant coefficients.\
Step 3. Update $Z^{n+1}$ by $Z^{n+1}=\frac{M}{4 }(b^n,Z^{n+1})_m\mathcal{A}^{-1}\Delta_h b^n+\mathcal{A}^{-1}f^n.
$
While the second-order scheme above is suitable in most situations, there are cases, e.g., when only steady state solutions are desired, where a first-order scheme is preferred. For the readers’ convenience, we list the first-order SAV scheme below:
We find $\{Z^{n+1}, W^{n+1}, R^{n+1},\Psi^{n+1}\}_{n=0}^{N-1}$ such that
\^[n+1]{}-\^[n]{}+t\^[n+1]{}=Mt\_hW\^[n+1]{}, \[e\_first-order2\]\
t\^[n+1]{}=Z\^[n+1]{}-Z\^n, \[e\_first-order3\]\
W\^[n+1]{}=\_h\^2Z\^[n+1]{}+2\_hZ\^[n]{}+Z\^[n+1]{}+F\^(Z\^n), \[e\_first-order4\]\
R\^[n+1]{}-R\^n=(F\^(Z\^[n]{}), Z\^[n+1]{}-Z\^n)\_m, \[e\_first-order5\]
where the discrete form of $E_1(Z^{n})$ is defined as follows: $$E_1^h(Z^{n})=\sum\limits_{i=1}^{N_{x}}\sum\limits_{j=1}^{N_{y}}h_xh_yF(Z^{n}_{i,j}).$$
Mass conservation and unconditional energy stability
----------------------------------------------------
Define the discrete pseudo energy $$\label{e_discrete pseudo energy}
\aligned
\mathcal{E}_d(Z^n,R^n,\Psi^n)=\frac 1 2 \|\Delta_hZ^n\|_m^2-\|\nabla_hZ^n\|^2+\frac \alpha 2 \|Z^{n+1}\|_m^2+R^2+\frac{1}{2M}\|\Psi^n\|_{H^{-1}}^2,
\endaligned$$ where $\|\nabla_hZ\|=\sqrt{(d_xZ,d_xZ)_x+(d_yZ,d_yZ)_y}$.
\[thm: second-order mass conservation\] The scheme (\[e\_second-order1\])-(\[e\_second-order4\]) admits a unique solution, is mass conserving, i.e., $(Z^{n+1},1)_m=(Z^{n},1)_m$ for all $n$, and unconditionally stable in the sense that $$\label{second-order energy stability1}
\aligned
\tilde{\mathcal{E}}_d(Z^{n+1},R^{n+1},\Psi^{n+1})-\tilde{\mathcal{E}}_d(Z^n,R^n,\Psi^n)\leq -\frac{\beta}{M}\Delta t\|\Psi^{n+1/2}\|_{H^{-1}}^2,
\endaligned$$ where $\tilde{\mathcal{E}}_d(Z^n,R^n,\Psi^n)=\mathcal{E}_d(Z^n,R^n,\Psi^n)+\frac 1 2\|\nabla_hZ^{n}-\nabla_hZ^{n-1}\|^2$.
Since the scheme (\[e\_second-order1\])-(\[e\_second-order4\]) is linear, the algorithm describes in Subsection \[eff\_impl\] indicates that it admits a unique solution. The proof for mass conservation and energy dissipation is essentially the same as that for the semi-discrete case. For the readers’ convenience, we still provide details below.
Summing on $i,j$ for $1\leq i\leq N_x,~1\leq j\leq N_y$ leads to $$\label{second-order mass conservation1}
\aligned
(\Psi^{n+1}-\Psi^{n},1)_m+\beta\Delta t(\Psi^{n+1/2},1)_m=M\Delta t(\Delta_hW^{n+1/2},1)_m.
\endaligned$$ Similarly, by summing , we can obtain $$\label{second-order mass conservation2}
\aligned
(Z^{n+1}-Z^n,1)_m=\Delta t(\Psi^{n+1/2},1)_m.
\endaligned$$ Taking notice of Lemma \[le\_discrete-integration-by-part\] and the boundary condition , the term on the right hand side of can be transformed into $$\label{second-order mass conservation3}
\aligned
M\Delta t(\Delta_hW^{n+1/2},1)_m=-M\Delta t\left((d_xW^{n+1/2},d_x1)_x+(d_yW^{n+1/2},d_y1)_y\right)=0.
\endaligned$$ Then can be estimated as follows: $$\label{second-order mass conservation4}
\aligned
(1+\frac \beta 2\Delta t)(\Psi^{n+1},1)_m=(1-\frac \beta 2\Delta t)(\Psi^{n},1)_m.
\endaligned$$ Combining with the condition on the initial condition $(\Psi^0,1)=0$ leads to $(\Psi^{n+1},1)_m=0$ for all $n\leq 0$. Recalling , we have $(Z^{n+1},1)_m=(Z^{n},1)_m$.
Next, we prove . Multiplying by $W_{i,j}^{n+1/2}h_xh_y$, and making summation on $i,j$ for $1\leq i\leq N_x,\ 1\leq j\leq N_y$, we have $$\label{second-order energy stability2}
\aligned
\Delta t(\Psi^{n+1/2},W^{n+1/2})_m=(Z^{n+1}-Z^n,W^{n+1/2})_m.
\endaligned$$ Multiplying by $(Z^{n+1}_{i,j}-Z^n_{i,j})h_xh_y$, and making summation on $i,j$ for $1\leq i\leq N_x,\ 1\leq j\leq N_y$, we have $$\label{second-order energy stability3}
\aligned
&(W^{n+1/2},Z^{n+1}-Z^n)_m=(\Delta_h^2Z^{n+1/2},Z^{n+1}-Z^{n})_m+2(\Delta_h\tilde{Z}^{n+1/2},Z^{n+1}-Z^n)_m\\
&\ \ \ \ \ +\alpha(Z^{n+1/2},Z^{n+1}-Z^{n})_m+(\frac{R^{n+1/2}}{\sqrt{E_1^h(\tilde{Z}^{n+1/2})}}F^{\prime}(\tilde{Z}^{n+1/2}),Z^{n+1}-Z^n)_m.
\endaligned$$ The first three terms on the right-hand side of can be dealt with the help of Lemma \[le\_discrete-integration-by-part\] and the boundary condition : $$\label{second-order energy stability4}
\aligned
(\Delta_h^2Z^{n+1/2},Z^{n+1}-Z^n)_m=\frac1 2(\|\Delta_hZ^{n+1}\|_m^2-\|\Delta_hZ^{n}\|_m^2).
\endaligned$$
$$\label{second-order energy stability5}
\aligned
&2(\Delta_h\tilde{Z}^{n+1/2},Z^{n+1}-Z^n)_m\\
=&-\|\nabla_hZ^{n+1}\|^2+\|\nabla_hZ^{n}\|^2+\frac1 2(\|\nabla_hZ^{n+1}-\nabla_hZ^{n}\|^2-\|\nabla_hZ^{n}-\nabla_hZ^{n-1}\|^2)\\
&+\frac1 2\|\nabla_hZ^{n+1}-2\nabla_hZ^{n}+\nabla_hZ^{n-1}\|^2.
\endaligned$$
$$\label{second-order energy stability5_add}
\aligned
\alpha(Z^{n+1/2},Z^{n+1}-Z^n)_m=\frac\alpha 2(\|Z^{n+1}\|_m^2-\|Z^{n}\|_m^2).
\endaligned$$
Multiplying by $(R^{n+1}+R^n)$ leads to $$\label{second-order energy stability6}
\aligned
(R^{n+1})^2-(R^n)^2=(\frac{R^{n+1/2}}{\sqrt{E_1^h(\tilde{Z}^{n+1/2} )}}F^{\prime}( \tilde{Z}^{n+1/2} ),Z^{n+1}-Z^n)_m.
\endaligned$$ Combining with and -, we have $$\label{second-order energy stability7}
\aligned
&\frac1 2(\|\Delta_hZ^{n+1}\|_m^2-\|\Delta_hZ^{n}\|_m^2)
-\|\nabla_hZ^{n+1}\|^2+\|\nabla_hZ^{n}\|^2+(R^{n+1})^2-(R^n)^2\\
+&\frac1 2(\|\nabla_hZ^{n+1}-\nabla_hZ^{n}\|^2-\|\nabla_hZ^{n}-\nabla_hZ^{n-1}\|^2)\\
+&\frac1 2\|\nabla_hZ^{n+1}-2\nabla_hZ^{n}+\nabla_hZ^{n-1}\|^2+\frac\alpha 2(\|Z^{n+1}\|_m^2-\|Z^{n}\|_m^2)\\
=&\Delta t(\Psi^{n+1/2},W^{n+1/2})_m.
\endaligned$$ Since recalling , we can derive $$\label{second-order energy stability8}
\aligned
\frac{1}{2M}(\|\Psi^{n+1}\|_{H^{-1}}^2-&\|\Psi^n\|_{H^{-1}}^2)=\frac 1 M(\Psi^{n+1}-\Psi^n,\Psi^{n+1/2})_{-1}\\
=&-\frac 1 M(\Psi^{n+1}-\Psi^n,\Delta_h^{-1}\Psi^{n+1/2})_m\\
=&-\frac{\beta}{M}\Delta t\|\Psi^{n+1/2}\|_{H^{-1}}^2-\Delta t(W^{n+1/2},\Psi^{n+1/2})_m.
\endaligned$$ Finally, combining with gives the desired result.
Error analysis
==============
In this section, we carry out a rigorous error analysis for the second-order scheme -. Set $$\aligned
& e_{\phi}^n=Z^n-\phi^n,\ e_{\psi}^n=\Psi^n-\psi^n,\\
& e_{\mu}^{n}=W^{n}-\mu^n,\ e_{r}^n=R^n-r^n.
\endaligned$$
We start by proving the following lemma which will be used to control the backward diffusion term in the error analysis.
\[le: control of backward diffusion term\] Suppose that $\phi$ and $\Delta_h\phi$ are satisfy the homogeneous Neumann boundary conditions, then we have $$\label{second-order error estimate1}
\aligned
\|\Delta_h\phi\|_m^2\leq \frac{1}{3\epsilon^2}\|\phi\|_m^2+\frac{2\epsilon}{3}\|\nabla_h(\Delta_h\phi)\|^2.
\endaligned$$
The proof for the homogeneous Neumann boundary condition is essentially the same as for the periodic boundary condition. One can refer to [@wise2009energy Lemma 3.10] for more detail.
\[thm: second-order error estimate\] We assume that $\phi\in W^{4,\infty}(J;L^{\infty}(\Omega)) \cap L^{\infty}(J;W^{6,\infty}(\Omega))\cap W^{2,\infty}(J;W^{4,\infty}(\Omega))$. Let $\Delta t\leq C(h_x+h_y)$, then for the discrete scheme -, there exists a positive constant $C$ independent of $h_x$, $h_y$ and $\Delta t$ such that $$\label{second-order error estimate2}
\aligned
&\|Z^{k+1}-\phi^{k+1}\|_m+\|\nabla_h(\Delta_hZ^{k+1})-\nabla_h(\Delta_h\phi^{k+1})\|\\
&+\|\Delta_hZ^{k+1}-\Delta_h\phi^{k+1}\|_m+
|R^{k+1}-r^{k+1}|\\
\leq&C(\|\phi\|_{W^{4,\infty}(J;L^{\infty}(\Omega))}+\|\phi\|_{W^{2,\infty}(J;W^{4,\infty}(\Omega))})\Delta t^2\\
&+C\|\phi\|_{L^{\infty}(J;W^{8,\infty}(\Omega))}(h_x^2+h_y^2),\quad \forall 0\le k\le N-1.
\endaligned$$
Subtracting equation from equation , we obtain $$\label{second-order error estimate3}
\aligned
\frac{e_{\psi}^{n+1}-e_{\psi}^{n}}{\Delta t}+\beta e_{\psi}^{n+1/2}=M\Delta_he_{\mu}^{n+1/2}+T_1^{n+1/2},
\endaligned$$ where $$\label{second-order error estimate4}
\aligned
T_{1}^{n+1/2}=\frac{\partial \psi}{\partial t}\big|_{t=n+1/2}-\frac{\psi^{n+1}-\psi^{n}}{\Delta t}
\leq
C\|\psi\|_{W^{3,\infty}(J;L^{\infty}(\Omega))}\Delta t^2.
\endaligned$$ Recalling and , we have $$\label{second-order error estimate5}
\aligned
e_{\psi}^{n+1/2}=\frac{e_{\phi}^{n+1}-e_{\phi}^n}{\Delta t}+T_2^{n+1/2},
\endaligned$$ where $$\label{second-order error estimate6}
\aligned
T_{2}^{n+1/2}=\frac{\phi^{n+1}-\phi^{n}}{\Delta t}-\frac{\partial \phi}{\partial t}\big|_{t=n+1/2}
\leq C\|\phi\|_{W^{3,\infty}(J;L^{\infty}(\Omega))}\Delta t^2.
\endaligned$$ Subtracting from leads to $$\label{second-order error estimate7}
\aligned
e_{\mu}^{n+1/2}=&\Delta_h^2e_{\phi}^{n+1/2}+2\Delta_h\tilde{e}_{\phi}^{n+1/2}+\alpha
e_{\phi}^{n+1/2}+\frac{R^{n+1/2}}{\sqrt{E_1^h(\tilde{Z}^{n+1/2})}}F^{\prime}(\tilde{Z}^{n+1/2})\\
&-\frac{r^{n+1/2}}{\sqrt{E_1(\phi^{n+1/2})}}F^{\prime}(\phi^{n+1/2})+T_3^{n+1/2},
\endaligned$$ where $$\label{second-order error estimate8}
\aligned
T_{3}^{n+1/2}=&\Delta_h^2\phi^{n+1/2}-\Delta^2\phi^{n+1/2}+2\Delta_h\tilde{\phi}^{n+1/2}-2\Delta\phi^{n+1/2}\\
\leq &C(\|\phi\|_{L^{\infty}(J;W^{6,\infty}(\Omega))}+\|\phi\|_{L^{\infty}(J;W^{4,\infty}(\Omega))})(h_x^2+h_y^2)\\
&+C\|\phi\|_{W^{2,\infty}(J;W^{2,\infty}(\Omega))}\Delta t^2.
\endaligned$$ Subtracting from gives that $$\label{second-order error estimate9}
\aligned
\frac{e_r^{n+1}-e_r^{n}}{\Delta t}=&\frac{1}{2\sqrt{E_1^h(\tilde{Z}^{n+1/2})}}(F^{\prime}(\tilde{Z}^{n+1/2}),
\frac{Z^{n+1}-Z^{n}}{\Delta t})_m\\
&-\frac{1}{2\sqrt{E_1(\phi^{n+1/2})}}\int_{\Omega}F^{\prime}(\phi^{n+1/2})\phi^{n+1/2}_t d\textbf{x}+T_{4}^{n+1/2},
\endaligned$$ where $$\label{second-order error estimate10}
\aligned
T_4^{n+1/2}=r_t^{n+1/2}-\frac{r^{n+1}-r^n}{\Delta t}\leq
C\|r\|_{W^{3,\infty}(J)}\Delta t^2.
\endaligned$$ Multiplying by $e_{\psi,i,j}^{n+1/2}h_xh_y$, and making summation on $i,j$ for $1\leq i\leq N_x,~1\leq j\leq N_y$, we have $$\label{second-order error estimate11}
\aligned
(\frac{e_{\psi}^{n+1}-e_{\psi}^{n}}{\Delta t},e_{\psi}^{n+1/2})_m+\beta\|e_{\psi}^{n+1/2}\|_m^2
=&M(\Delta_he_{\mu}^{n+1/2},e_{\psi}^{n+1/2})_m+(T_1^{n+1/2}, e_{\psi}^{n+1/2})_m.
\endaligned$$ The first term on the left-hand side of can be transformed into the following $$\label{second-order error estimate12}
\aligned
(\frac{e_{\psi}^{n+1}-e_{\psi}^{n}}{\Delta t},e_{\psi}^{n+1/2})_m=\frac{\|e_{\psi}^{n+1}\|_m^2-\|e_{\psi}^{n}\|_m^2}{2\Delta t}.
\endaligned$$ Taking notice of , we can write the first term on the right-hand side of as $$\label{second-order error estimate13}
\aligned
&M(\Delta_he_{\mu}^{n+1/2},e_{\psi}^{n+1/2})_m=M(\Delta_h^3e_{\phi}^{n+1/2},e_{\psi}^{n+1/2})_m+2M(\Delta_h^2\tilde{e}_{\phi}^{n+1/2},e_{\psi}^{n+1/2})_m\\
&\ \ \ \ \ +M(\frac{R^{n+1/2}}{\sqrt{E_1^h(\tilde{Z}^{n+1/2})}}\Delta_hF^{\prime}(\tilde{Z}^{n+1/2})-\frac{r^{n+1/2}}{\sqrt{E_1(\phi^{n+1/2})}}\Delta_hF^{\prime}(\phi^{n+1/2}),e_{\psi}^{n+1/2})_m\\
& \ \ \ \ \ +\alpha(\Delta_he_{\phi}^{n+1/2},e_{\psi}^{n+1/2})_m+M(\Delta_hT_3^{n+1/2},e_{\psi}^{n+1/2})_m.
\endaligned$$ Using Lemma \[le\_discrete-integration-by-part\] and the boundary condition , we can write the first and second terms on the right-hand side of as $$\label{second-order error estimate14}
\aligned
M(\Delta_h^3e_{\phi}^{n+1/2},e_{\psi}^{n+1/2})_m=&-M(\nabla_h(\Delta_he_{\phi}^{n+1/2}),\nabla_h(\Delta_he_{\psi}^{n+1/2}))\\
=&-M\frac{\|\nabla_h(\Delta_he_{\phi}^{n+1})\|^2-\|\nabla_h(\Delta_he_{\phi}^{n})\|^2
}{2\Delta t}.
\endaligned$$
$$\label{second-order error estimate15}
\aligned
&2M(\Delta_h^2\tilde{e}_{\phi}^{n+1/2},e_{\psi}^{n+1/2})_m=M(\Delta_h(3e_{\phi}^n-e_{\phi}^{n-1}), \Delta_he_{\psi}^{n+1/2})_m\\
=&\frac{M}{\Delta t}\left(\|\Delta_he_{\phi}^{n+1}\|_m^2-\|\Delta_he_{\phi}^{n}\|_m^2-\frac{1}{2}(\|\Delta_he_{\phi}^{n+1}-\Delta_he_{\phi}^{n}\|_m^2-\|\Delta_he_{\phi}^{n}-\Delta_he_{\phi}^{n-1}\|_m^2)\right)\\
&-\frac{M}{2\Delta t}\|\Delta_he_{\phi}^{n+1}-2\Delta_he_{\phi}^{n}+\Delta_he_{\phi}^{n-1}\|_m^2.
\endaligned$$
The third term on the right-hand side of can be estimated by $$\label{second-order error estimate16}
\aligned
&M(\frac{R^{n+1/2}}{\sqrt{E_1^h(\tilde{Z}^{n+1/2})}}\Delta_hF^{\prime}(\tilde{Z}^{n+1/2})-\frac{r^{n+1/2}}{\sqrt{E_1(\phi^{n+1/2})}}\Delta_hF^{\prime}(\phi^{n+1/2}),e_{\psi}^{n+1/2})_m\\
=&Mr^{n+1/2}(\frac{\Delta_hF^{\prime}(\tilde{Z}^{n+1/2})}{\sqrt{E_1^h(\tilde{Z}^{n+1/2})}}-\frac{\Delta_hF^{\prime}(\tilde{\phi}^{n+1/2})}{\sqrt{E_1^h(\tilde{\phi}^{n+1/2})}}, e_{\psi}^{n+1/2})_m\\
&+Mr^{n+1/2}(\frac{\Delta_hF^{\prime}(\tilde{\phi}^{n+1/2})}{\sqrt{E_1^h(\tilde{\phi}^{n+1/2})}}-\frac{\Delta_hF^{\prime}(\phi^{n+1/2})}{\sqrt{E_1(\phi^{n+1/2})}}, e_{\psi}^{n+1/2})_m\\
&+Me_{r}^{n+1/2}(\frac{\Delta_hF^{\prime}(\tilde{Z}^{n+1/2})}{\sqrt{E_1^h(\tilde{Z}^{n+1/2})}},
e_{\psi}^{n+1/2})_m.
\endaligned$$
Below we shall first assume that there exist three positive constants $C_1$, $C_2$ and $C_3$ such that $$\label{second-order error estimate17_hypotheses}
\aligned
\|Z^n\|_{{L^{\infty}(\Omega)}}\leq C_1, \ \ \|\nabla_hZ^n\|_{{L^{\infty}(\Omega)}}\leq C_2, \ \|\Delta_hZ^n\|_{{L^{\infty}(\Omega)}}\leq C_3,\quad\forall 0\le n\le N,
\endaligned$$ which will be verified late in the proof.
Applying Lemma \[le: control of backward diffusion term\], the first term on the right-hand side of can be controlled similar to the estimates in [@wang2011energy] by $$\label{second-order error estimate18}
\aligned
&Mr^{n+1/2}(\frac{\Delta_hF^{\prime}(\tilde{Z}^{n+1/2})}{\sqrt{E_1^h(\tilde{Z}^{n+1/2})}}-\frac{\Delta_hF^{\prime}(\tilde{\phi}^{n+1/2})}{\sqrt{E_1^h(\tilde{\phi}^{n+1/2})}}, e_{\psi}^{n+1/2})_m\\
\leq &C(\|e_{\phi}^n\|_m^2+\|\Delta_he_{\phi}^n\|_m^2)+C(\|e_{\phi}^{n-1}\|_m^2+\|\Delta_he_{\phi}^{n-1}\|_m^2)+C\|e_{\psi}^{n+1/2}\|_m^2\\
\leq&C(\|e_{\phi}^n\|_m^2+\|\nabla_h(\Delta_he_{\phi}^n)\|^2)+C(\|e_{\phi}^{n-1}\|_m^2+\|\nabla_h(\Delta_he_{\phi}^{n-1})\|^2)+C\|e_{\psi}^{n+1/2}\|_m^2,
\endaligned$$ where $C$ is dependent on $\|r\|_{L^{\infty}(J)},\ \|Z^n\|_{L^{\infty}(\Omega)}, \|\nabla_hZ^n\|_{L^{\infty}(\Omega)}$.
The second term on the right-hand side of can be handled by: $$\label{second-order error estimate19}
\aligned
&Mr^{n+1/2}(\frac{\Delta_hF^{\prime}(\tilde{\phi}^{n+1/2})}{\sqrt{E_1^h(\tilde{\phi}^{n+1/2})}}-\frac{\Delta_hF^{\prime}(\phi^{n+1/2})}{\sqrt{E_1(\phi^{n+1/2})}}, e_{\psi}^{n+1/2})_m\\
=&Mr^{n+1/2}(\frac{\Delta_hF^{\prime}(\tilde{\phi}^{n+1/2})}{\sqrt{E_1^h(\tilde{\phi}^{n+1/2})}}-\frac{\Delta_hF^{\prime}(\phi^{n+1/2})}{\sqrt{E_1^h(\tilde{\phi}^{n+1/2})}}, e_{\psi}^{n+1/2})_m\\
&+Mr^{n+1/2}(\frac{\Delta_hF^{\prime}(\phi^{n+1/2})}{\sqrt{E_1^h(\tilde{\phi}^{n+1/2})}}-\frac{\Delta_hF^{\prime}(\phi^{n+1/2})}{\sqrt{E_1(\phi^{n+1/2})}}, e_{\psi}^{n+1/2})_m\\
\leq &C\|e_{\psi}^{n+1/2}\|_m^2+C\|\phi\|_{W^{2,\infty}(J;W^{2,\infty}(\Omega))}\Delta t^4\\
&+C\|\phi\|_{L^{\infty}(J;W^{3,\infty}(\Omega))}(h_x^4+h_y^4).
\endaligned$$ The last term on the right-hand side of can be directly controlled by Cauchy-Schwarz inequality: $$\label{second-order error estimate20}
\aligned
Me_{r}^{n+1/2}(\frac{\Delta_hF^{\prime}(\tilde{Z}^{n+1/2})}{\sqrt{E_1^h(\tilde{Z}^{n+1/2})}},
e_{\psi}^{n+1/2})_m\leq C\|e_{\psi}^{n+1/2}\|_m^2+C(e_r^{n+1/2})^2,
\endaligned$$ where $C$ is dependent on $\|Z^n\|_{L^{\infty}(\Omega)},\ \|\nabla_hZ^n\|_{L^{\infty}(\Omega)}, \ \|\Delta_hZ^n\|_{L^{\infty}(\Omega)}$. Applying estimates - yields $$\label{second-order error estimate21}
\aligned
&M(\frac{R^{n+1/2}}{\sqrt{E_1^h(\tilde{Z}^{n+1/2})}}\Delta_hF^{\prime}(\tilde{Z}^{n+1/2})-\frac{r^{n+1/2}}{\sqrt{E_1(\phi^{n+1/2})}}\Delta_hF^{\prime}(\phi^{n+1/2}),e_{\psi}^{n+1/2})_m\\
\leq&C(\|e_{\phi}^n\|_m^2+\|\nabla_h(\Delta_he_{\phi}^n)\|^2)+C(\|e_{\phi}^{n-1}\|_m^2+\|\nabla_h(\Delta_he_{\phi}^{n-1})\|^2)\\
&+C\|e_{\psi}^{n+1/2}\|_m^2+C(e_r^{n+1/2})^2+C\|\phi\|^2_{W^{2,\infty}(J;W^{2,\infty}(\Omega))}\Delta t^4\\
&+C\|\phi\|^2_{L^{\infty}(J;W^{3,\infty}(\Omega))}(h_x^4+h_y^4).
\endaligned$$ The last term on the right-hand side of can be estimated by $$\label{second-order error estimate22}
\aligned
&M(\Delta_hT_3^{n+1/2},e_{\psi}^{n+1/2})_m\leq C\|e_{\psi}^{n+1/2}\|_m^2+C\|\phi\|^2_{W^{2,\infty}(J;W^{4,\infty}(\Omega))}\Delta t^4\\
&\ \ \ \ \ +C(\|\phi\|^2_{L^{\infty}(J;W^{8,\infty}(\Omega))}+\|\phi\|^2_{L^{\infty}(J;W^{6,\infty}(\Omega))})(h_x^4+h_y^4).
\endaligned$$ Combining with the above equations leads to $$\label{second-order error estimate23}
\aligned
&\frac{\|e_{\psi}^{n+1}\|_m^2-\|e_{\psi}^{n}\|_m^2}{2\Delta t}+\beta\|e_{\psi}^{n+1/2}\|_m^2+M\alpha\frac{\|e_{\phi}^{n+1}\|_m^2-\|e_{\phi}^{n}\|_m^2
}{2\Delta t}\\
&+M\frac{\|\nabla_h(\Delta_he_{\phi}^{n+1})\|^2-\|\nabla_h(\Delta_he_{\phi}^{n})\|^2
}{2\Delta t}\\
&+\frac{M}{2\Delta t}(\|\Delta_he_{\phi}^{n+1}-\Delta_he_{\phi}^{n}\|_m^2-\|\Delta_he_{\phi}^{n}-\Delta_he_{\phi}^{n-1}\|_m^2)\\
\leq &C(\|e_{\phi}^n\|_m^2+\|\nabla_h(\Delta_he_{\phi}^n)\|^2)+C(\|e_{\phi}^{n-1}\|_m^2+\|\nabla_h(\Delta_he_{\phi}^{n-1})\|^2)\\
&+C\|e_{\psi}^{n+1/2}\|_m^2+C(e_r^{n+1/2})^2+\frac{M}{\Delta t}(\|\Delta_he_{\phi}^{n+1}\|_m^2-\|\Delta_he_{\phi}^{n}\|_m^2)\\
&+C(\|\phi\|^2_{L^{\infty}(J;W^{8,\infty}(\Omega))}+\|\phi\|^2_{L^{\infty}(J;W^{6,\infty}(\Omega))})(h_x^4+h_y^4)\\
&+C(\|\phi\|^2_{W^{4,\infty}(J;L^{\infty}(\Omega))}+\|\phi\|^2_{W^{2,\infty}(J;W^{4,\infty}(\Omega))})\Delta t^4.
\endaligned$$ Next we give the error estimate of auxiliary function $r$. Multiplying by $e_{r}^{n+1}+e_{r}^{n}$ leads to $$\label{second-order error estimate24}
\aligned
\frac{(e_r^{n+1})^2-(e_r^{n})^2}{\Delta t}=&
\frac{e_r^{n+1/2}}{\sqrt{E_1^h(\tilde{Z}^{n+1/2})}}(F^{\prime}(\tilde{Z}^{n+1/2}),
d_tZ^{n+1})_m\\
&-\frac{e_r^{n+1/2}}{\sqrt{E_1(\phi^{n+1/2})}}\int_{\Omega}F^{\prime}(\phi^{n+1/2})\phi^{n+1/2}_t d\textbf{x}\\
&+T_{4}^{n+1/2}\cdot (e_{r}^{n+1}+e_{r}^{n}).
\endaligned$$ The first two terms on the right-hand side of can be transformed into: $$\label{second-order error estimate25}
\aligned
&\frac{e_r^{n+1/2}}{\sqrt{E_1^h(\tilde{Z}^{n+1/2})}}(F^{\prime}(\tilde{Z}^{n+1/2}),
d_tZ^{n+1})_m-\frac{e_r^{n+1/2}}{\sqrt{E_1(\phi^{n+1/2})}}\int_{\Omega}F^{\prime}(\phi^{n+1/2})\phi^{n+1/2}_t d\textbf{x}\\
=&\frac{e_{r}^{n+1/2}}{\sqrt{E_1(\phi^{n+1/2})}}\left((F^{\prime}(\phi^{n+1/2}),d_t\phi^{n+1})_m-\int_{\Omega}F^{\prime}(\phi^{n+1/2})\phi^{n+1/2}_t d\textbf{x}\right)\\
&+e_{r}^{n+1/2}(\frac{F^{\prime}(\tilde{Z}^{n+1/2})}{\sqrt{E_1^h(\tilde{Z}^{n+1/2})}}-
\frac{F^{\prime}(\phi^{n+1/2})}{\sqrt{E_1(\phi^{n+1/2})}},d_t\phi^{n+1})_m\\
&+\frac{e_{r}^{n+1/2}}{\sqrt{E_1^h(\tilde{Z}^{n+1/2})}}(F^{\prime}(\tilde{Z}^{n+1/2}),
d_te_{\phi}^{n+1})_m,
\endaligned$$ which can be handled in a similar way as in [@li2019energy]. Thus we have $$\label{second-order error estimate26}
\aligned
&\frac{e_r^{n+1/2}}{\sqrt{E_1^h(\tilde{Z}^{n+1/2})}}(F^{\prime}(\tilde{Z}^{n+1/2}),
d_tZ^{n+1})_m-\frac{e_r^{n+1/2}}{\sqrt{E_1(\phi^{n+1/2})}}\int_{\Omega}F^{\prime}(\phi^{n+1/2})\phi^{n+1/2}_t d\textbf{x}\\
\leq &C(e_r^{n+1/2})^2+C\|\phi\|^2_{W^{1,\infty}(J;L^{\infty}(\Omega))}(\|e_{\phi}^{n}\|_m^2+\|e_{\phi}^{n-1}\|_m^2)\\
&+C\|e_{\psi}^{n+1/2}\|_m^2+
C\|\phi\|^2_{W^{1,\infty}(J;W^{2,\infty}(\Omega))}(h_x^4+h_y^4).
\endaligned$$ Substituting into and applying Cauchy-Schwartz inequality, we can obtain $$\label{second-order error estimate27}
\aligned
\frac{(e_r^{n+1})^2-(e_r^{n})^2}{\Delta t}\leq&C(e_r^{n+1/2})^2+C\|\phi\|^2_{W^{1,\infty}(J;L^{\infty}(\Omega))}(\|e_{\phi}^{n}\|_m^2+\|e_{\phi}^{n-1}\|_m^2)\\
&+C\|e_{\psi}^{n+1/2}\|_m^2+
C\|\phi\|^2_{W^{1,\infty}(J;W^{2,\infty}(\Omega))}(h_x^4+h_y^4)\\
&+C\|r\|^2_{W^{3,\infty}(J)}\Delta t^4.
\endaligned$$ Combining with and multiplying by $2\Delta t$, summing over $n,~n=0,1,\ldots,k$, we have $$\label{second-order error estimate28}
\aligned
&\|e_{\psi}^{k+1}\|_m^2+\beta\sum_{n=0}^{k}\Delta t\|e_{\psi}^{n+1/2}\|^2+M\alpha\|e_{\phi}^{k+1}\|^2+M\|\nabla_h(\Delta_he_{\phi}^{k+1})\|^2+2(e_r^{k+1})^2\\
\leq &2M\|\Delta_he_{\phi}^{k+1}\|_m^2+C\sum_{n=0}^{k}\Delta t\|e_{\phi}^n\|_m^2+C\sum_{n=0}^{k}\Delta t\|\nabla_h(\Delta_he_{\phi}^n)\|_m^2+C\sum_{n=0}^{k}\Delta t\|e_{\psi}^{n}\|^2\\
&+C\sum_{n=0}^{k}\Delta t(e_r^n)^2+C(\|\phi\|^2_{L^{\infty}(J;W^{8,\infty}(\Omega))}+\|\phi\|^2_{L^{\infty}(J;W^{6,\infty}(\Omega))})(h_x^4+h_y^4)\\
&+C(\|\phi\|^2_{W^{4,\infty}(J;L^{\infty}(\Omega))}+\|\phi\|^2_{W^{2,\infty}(J;W^{4,\infty}(\Omega))})\Delta t^4.
\endaligned$$ To carry out further analysis, we should give the following inequality first. Applying $e_{\phi}^k=e_{\phi}^0+\sum\limits_{l=1}^{k}\Delta te_{\psi}^l$ and using Cauchy-Schwarz inequality, we obtain that $$\label{second-order error estimate29}
\aligned
\|e_{\phi}^k\|_m^2\leq &2\|e_{\phi}^0\|_m^2+2\|\sum\limits_{l=1}^{k}\Delta te_{\psi}^l \|_m^2\\
\leq &2\|e_{\phi}^0\|_m^2+2T\sum\limits_{l=1}^{k}\Delta t\|e_{\psi}^l\|_m^2.
\endaligned$$ Applying Lemma \[le: control of backward diffusion term\] and , the first term on the right-hand side of can be transformed into $$\label{second-order error estimate30}
\aligned
2M\|\Delta_he_{\phi}^{k+1}\|_m^2\leq &C\|e_{\phi}\|_m^2+\frac{M}{2}\|\nabla_h(\Delta_he_{\phi}^{k+1})\|^2\\
\leq&C\sum\limits_{l=1}^{k}\Delta t\|e_{\psi}^l\|_m^2+\frac{M}{2}\|\nabla_h(\Delta_he_{\phi}^{k+1})\|^2.
\endaligned$$ Then using the discrete Gronwall inequality and Lemma \[le: control of backward diffusion term\], can be estimated as follows: $$\label{second-order error estimate31}
\aligned
&\|e_{\psi}^{k+1}\|_m^2+\|e_{\phi}^{k+1}\|_m^2+\|\Delta_he_{\phi}^{k+1}\|_m^2+
\|\nabla_h(\Delta_he_{\phi}^{k+1})\|^2+(e_r^{k+1})^2\\
\leq &C(\|\phi\|^2_{W^{4,\infty}(J;L^{\infty}(\Omega))}+\|\phi\|^2_{W^{2,\infty}(J;W^{4,\infty}(\Omega))})\Delta t^4\\
&+C\|\phi\|^2_{L^{\infty}(J;W^{8,\infty}(\Omega))}(h_x^4+h_y^4),\quad\forall 0\le k\le N-1.
\endaligned$$ It remains to verify the hypothesis . Actually this part of the proof follows a similar procedure as in our previous works [@li2018block; @li2019energy]. For the readers’ convenience, we still provide a detail proof for $\|Z^n\|_{{L^{\infty}(\Omega)}} \leq C_1$ in the following two steps by using the mathematical induction.
***Step 1*** (Definition of $C_1$): Using the scheme - for $n=0$ and applying the inverse assumption, we can get the approximation $Z^1$ with the following property: $$\aligned
& \|Z^1\|_{{L^{\infty}(\Omega)}} \leq \|Z^1-\phi^1\|_{{L^{\infty}(\Omega)}} +\|\phi^1\|_{{L^{\infty}(\Omega)}} \\
\leq& \|Z^1-\Pi_h\phi^1\|_{{L^{\infty}(\Omega)}} +\|\Pi_h\phi^1-\phi^1\|_{{L^{\infty}(\Omega)}} +\|\phi^1\|_{{L^{\infty}(\Omega)}} \\
\leq&Ch^{-1}(\|Z^1-\phi^1 \|_m+\|\phi^1-\Pi_h\phi^1\|_m)+\|\Pi_h\phi^1-\phi^1\|_{{L^{\infty}(\Omega)}} +\|\phi^1\|_{{L^{\infty}(\Omega)}} \\
\leq&C(h+h^{-1}\Delta t^2)+\|\phi^1\|_{{L^{\infty}(\Omega)}} \leq C.
\endaligned$$ where $h=\max\{h_x,h_y\}$ and $\Pi_h$ is an bilinear interpolant operator with the following estimate: $$\label{e_boundedness_added2}
\aligned
\|\Pi_h\phi^1-\phi^1\|_{{L^{\infty}(\Omega)}} \leq Ch^2.
\endaligned$$ Thus we can choose the positive constant $C_1$ independent of $h$ and $\Delta t$ such that $$\begin{aligned}
C_1&\geq \max\{\|Z^{1}\|_{{L^{\infty}(\Omega)}} , 2\|\phi(t^n)\|_{{L^{\infty}(\Omega)}} \}.\end{aligned}$$
***Step 2*** (Induction): By the definition of $C_1$, it is trivial that hypothesis $\|Z^l\|_{{L^{\infty}(\Omega)}}\leq C_1$ holds true for $l=1$. Supposing that $\|Z^{l-1}\|_{{L^{\infty}(\Omega)}} \leq C_1$ holds true for an integer $l=1,\cdots,k+1$, with the aid of the estimate , we have that $$\|Z^{l}-\phi^l\|_m\leq C(\Delta t^2+h^2).$$ Next we prove that $\|Z^{l}\|_{{L^{\infty}(\Omega)}} \leq C_1$ holds true. Since $$\label{e_boundedness_added3}
\aligned
& \|Z^l\|_{{L^{\infty}(\Omega)}} \leq \|Z^l-\phi^l\|_{{L^{\infty}(\Omega)}} +\|\phi^l\|_{{L^{\infty}(\Omega)}} \\
\leq & \|Z^l-\Pi_h\phi^l\|_{{L^{\infty}(\Omega)}} +\|\Pi_h\phi^l-\phi^l\|_{{L^{\infty}(\Omega)}} +\|\phi^l\|_{{L^{\infty}(\Omega)}} \\
\leq&Ch^{-1}(\|Z^l-\phi^l \|_m+\|\phi^l-\Pi_h\phi^l\|_m)+\|\Pi_h\phi^l-\phi^l\|_{{L^{\infty}(\Omega)}} +\|\phi^l\|_{{L^{\infty}(\Omega)}} \\
\leq&C_4(h+h^{-1}\Delta t^2)+\|\phi^1\|_{{L^{\infty}(\Omega)}} .
\endaligned$$ Let $\Delta t\leq C_5h$ and a positive constant $h_1$ be small enough to satisfy $$C_4(1+C_5^2)h_1\leq\frac{C_1}{2}.$$ Then for $h\in (0,h_1],$ we derive from that $$\aligned
\|Z^l\|_{{L^{\infty}(\Omega)}}
\leq&C_4(h+h^{-1}\Delta t^2)+\|\phi^l\|_{{L^{\infty}(\Omega)}} \\
\leq &C_4(h_1+C_5^2h_1)+\frac{C_1}{2}
\leq C_1.
\endaligned$$ This indicates that $\|Z^n\|_{{L^{\infty}(\Omega)}} \leq C_1$ for all $n$. The proof for the other two inequalities in is essentially identical with the above procedure so we skip it for the sake of brevity.
Numerical results and discussions
=================================
In this section, we carry out some numerical experiments with the proposed scheme for the MPFC equation. We first verify the order of convergence. Then we plot evolutions of the original energy as well as the pseudo energy to show that the pseudo energy is indeed dissipative while the original energy is not.
Accuracy tests
--------------
We take $\Omega=(0,1)\times(0,1)$, $T=0.5$, $\epsilon=0.25$, $\beta=0.9$, $M=0.001$ and the initial solution $\phi_0=\cos(2\pi x)\cos(2\pi y)$ with the homogenous Neumann boundary conditions. We use the second-order scheme - and measure the Cauchy error since we do not know the exact solution. Specifically, the error between two different grid spacings $h$ and $\frac{h}{2}$ is calculated by $\|e_{\zeta}\|=\|\zeta_h-\zeta_{h/2}\|$. We take the time step to be $\Delta t=\frac{T}{N}$ with $N=N_x=N_y$, and list the results in Table \[table2\_example1\]. For simplicity, we define $\|e_{f}\|_{\infty}=\max\limits_{0\leq l\leq k}\|e_{f}^{l}\|$. We observe a solid second order convergence rate, which are consistent with the error estimates in Theorem \[thm: second-order error estimate\].
$N_x\times N_y$ $\|e_{\phi}\|_{\infty,m}$ Rate $\|\nabla_h(\Delta_he_{\phi})\|_{\infty}$ Rate $\|e_{r}\|_{\infty}$ Rate
----------------- --------------------------- ------ ------------------------------------------- ------ ---------------------- ------
$20\times 20$ 1.15E-1 — 79.6E-0 — 2.15E-2 —
$40\times 40$ 3.15E-2 1.87 22.0E-0 1.85 6.62E-3 1.70
$80\times 80$ 8.02E-3 1.97 5.62E-0 1.97 1.28E-3 2.38
$160\times 160$ 2.11E-3 1.93 1.48E-0 1.93 2.32E-4 2.46
: Errors and convergence rates for the scheme -.[]{data-label="table2_example1"}
Energy stability test
---------------------
In this example, we set $\Omega=(0,128)\times(0,128)$, $M=1$, $\epsilon=0.025$, $\beta=0.1$, and consider the MPFC model with the periodic boundary conditions. The initial condition is taken as follows [@baskaran2013energy; @guo2018high]: $$\label{numerical_initial}
\aligned
\phi_0(x,y)=&0.07-0.02\cos(\frac{2\pi(x-12)}{32})\sin(\frac{2\pi(y-1)}{32})+0.02\cos^2(\frac{\pi(x+10)}{32})\\
&\cos^2(\frac{\pi(y+3)}{32})-0.01\sin^2(\frac{4\pi x}{32})\sin^2(\frac{4\pi(y-6)}{32}).
\endaligned$$ We take $\Delta t=0.05$ and evolve the system to the final time $T=10$. The evolutions of discrete original energy and pseudo energy using the second-order scheme are plotted in Figure \[fig: pseudo energy\]. We observe that the discrete original energy may increase on some time intervals, while the pseudo energy are non-increasing at all times, which is consistent with our analysis.
![The discrete original energy and pseudo energy plotted as functions of time[]{data-label="fig: pseudo energy"}](fig_energy.eps)
Summary
-------
We constructed in this paper two efficient schemes for the MPFC model based on the SAV approach and block finite-difference method. Since the original energy of the MPFC equation may increase in time on some time intervals, we introduced a pseudo energy that is dissipative for all times. It is shown that our schemes conserve mass and are unconditionally energy stable with respect to the pseudo energy. We also established rigorously second-order error estimates in both time and space for our second-order SAV block-centered finite difference method. Finally some numerical experiments are presented to validate our theoretical results.
[^1]: School of Mathematical Sciences and Fujian Provincial Key Laboratory on Mathematical Modeling and High Performance Scientific Computing, Xiamen University, Xiamen, Fujian, 361005, China. Email: [email protected]
[^2]: Department of Mathematics, Purdue University, West Lafayette, IN 47907, USA. Email: [email protected]
[^3]: The work of X. Li is supported by the National Natural Science Foundation of China under grant number 11901489, 11971407 and Postdoctoral Science Foundation of China under grant numbers BX20190187 and 2019M650152. The work of J. Shen is supported in part by NSF grant DMS-1720442 and AFOSR grant FA9550-16-1-0102.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We have analytically solved the spin wave excitations for the intercalated ternary iron-selenide AFe$_{1.5}$Se$_2$ (A=K, Tl) in the $4\times 2$ collinear antiferromagnetic order. It is found that there are one acoustic branch (gapless Goldstone mode) and two gapful optical branches of spin wave excitations with each in double degeneracy. By examining the non-imaginary excitation frequency condition, we can determine the corresponding phase boundary. The exchange couplings between Fe moments in AFe$_{1.5}$Se$_2$ are derived based on the first-principles total energy calculations. The Fe spin is found to be $S=\frac{3}{2}$ through computing the antiferromagnetic quantum fluctuation. And it is further found that a very small spin-orientation anisotropy can remarkably suppress the antiferromagnetic quantum fluctuation. The spin dynamical structure factors are calculated and discussed in associated with neutron inelastic scattering experiment.'
author:
- 'Miao Gao$^{1}$'
- 'Xun-Wang Yan$^{1,2}$'
- 'Zhong-Yi Lu$^{1}$'
title: 'Spin Wave Excitations in AFe$_{1.5}$Se$_2$ (A=K, Tl): Analytical Study'
---
Introduction
============
The discovery of iron-based superconductors [@kamihara] has stimulated great interest on the investigation of unconventional superconducting mechanism, in which magnetism is considered to play a substantial role. It is well-known that the parent compound of a cuprate superconductor is an antiferromagnetic (AFM) Mott insulator with a checkerboard AFM order on the copper square lattice. In contrast, the parent compound of an iron-based superconductor was found to be an AFM semi-metal [@lu0] with either collinear [@cruz; @ma1] or bi-collinear [@ma; @bao; @shi] AFM order on the iron square lattice. Regarding the nature of the magnetism, there are basically two contradictive views. The one [@mazin] is based on itinerant electron picture, in which the Fermi surface nesting is responsible for the collinear AFM order. On the contrary, the other one is based on local moment interactions which can be described by the $J_1$-$J_2$ frustrated Heisenberg model. [@yildirim; @si; @ma1] And it was further shown [@ma1] that the underlying driving force herein is the As-bridged AFM superexchange interaction between a pair of next-nearest-neighboring fluctuating Fe local moments embedded in itinerant electrons. There are now more and more evidences in favor of the fluctuating Fe local moment picture. Especially, the neutron inelastic scattering experiments have shown that the low-energy magnetic excitations can be well described by the spin waves based on the $J_1$-$J_2$ Heisenberg model.[@dai-0; @dai-1; @dai-2; @dai-3]
The newly discovered intercalated ternary iron-selenide superconductors A$_y$Fe$_x$Se$_2$ (A=K, Tl)[@chen; @Cs; @fang] show rich phase diagrams and many unusual physical properties that have not been found in other iron-based superconductors, for example, the superconductivity was found to coexist with a strong AFM order with a giant magnetic moment of 3.31 $\mu_B$/Fe formed below a Neel temperature of 559$K$ [@muSR; @bao1] and to be proximity to an AFM insulating phase.[@fang] These reveal the close relationship between unconventional superconductivity and antiferromagnetism once more, and have triggered another surge of interest for the investigation of unconventional superconducting mechanism.
The compounds A$_y$Fe$_x$Se$_2$ have the ThCr$_{2}$Si$_{2}$ type crystal structure, isostructural with 122-type iron pnictides.[@rotter] However, the stable structures of A$_y$Fe$_x$Se$_2$ contain Fe vacancies ordered in either $\sqrt{5}\times \sqrt{5}$ or $4\times 2$ superstructure due to the balance required in chemical valences, which respectively correspond to A$_{0.8}$Fe$_{1.6}$Se$_2$ with one-fifth Fe vacancies or AFe$_{1.5}$Se$_2$ with one-quarter Fe vacancies. For A$_{0.8}$Fe$_{1.6}$Se$_2$, the neutron observation has found that it has a $\sqrt{5}\times \sqrt{5}$ blocked checkerboard AFM order to match the Fe vacancy superstructure.[@bao1] For AFe$_{1.5}$Se$_2$, the first-principles electronic structure calculations predicted[@yan] that its ground state is in a $4\times 2$ collinear AFM order, as shown in Fig. \[fig:Mag\](d). Moreover, the calculations further showed [@yan; @yan11] that both compounds AFe$_{1.5}$Se$_2$ and A$_{0.8}$Fe$_{1.6}$Se$_2$ are antiferromagnetic semiconductors with band gaps of dozens and hundreds meV, respectively. Such band gaps have been confirmed by the recent ARPES and transport measurements.[@arpes] It was further proposed that the parent compound of an A$_y$Fe$_x$Se$_2$ superconductor is an AFM semiconductor either A$_{0.8}$Fe$_{1.6}$Se$_2$ or AFe$_{1.5}$Se$_2$.[@yan; @yan11] The latest neutron diffraction experiment shows that the parent compound is likely AFe$_{1.5}$Se$_2$ with a $4\times 2$ collinear AFM order.[@zhao] In order to well understand the magnetism in the compound AFe$_{1.5}$Se$_2$, we have studied the magnetic excitations and spin dynamical structure factors, which can be directly detected by neutron experiments. Our approach is based on the linearized spin wave theory upon the effective spin Heisenberg model.
Effective model
===============
As schematically shown in Fig. \[fig:Mag\](d), in the ground state of the compound AFe$_{1.5}$Se$_2$,[@yan] the Fe vacancies are ordered in a rhombus structure, in which unit cell there are two inequivalent Fe atoms according to the number of neighboring Fe atoms, namely 2-Fe-neighbored and 3-Fe-neighbored Fe atoms respectively. The corresponding magnetic order is a $4\times 2$ collinear AFM order (also called A-col AFM order here), as shown in Fig. \[fig:Mag\](d), in which the Fe moments are antiferromagnetically ordered along the lines without Fe vacancies and ferromagnetically ordered along the lines perpendicular.
To describe the magnetism in the compound AFe$_{1.5}$Se$_2$, considering the Se-bridging effect,[@yan; @ma1] we adopt the spin Heisenberg model with the nearest and next-nearest neighbor exchange couplings between the Fe moments with quantum spin $\vec{S}$ as follows, $$\label{eq:Ham}
\hat{H}=J_{1a}\sum_{\langle i,\delta_a \rangle}\vec{S}_i\cdot\vec{S}_{\delta_a} +J_{1b}\sum_{\langle i,\delta_b \rangle}\vec{S}_i\cdot\vec{S}_{\delta_b} + J_2\sum_{ \ll ij \gg}\vec{S}_i\cdot\vec{S}_j,$$ whereas $\langle i,\delta_a \rangle$, $\langle i,\delta_b \rangle$, and $\ll ij \gg$ denote the summation over the nearest neighbors along the lines without Fe vacancies, the nearest neighbors along the lines with Fe vacancies, and the next-nearest neighbors, respectively. To be more specific, the exchange coupling $J_{1a}$ links a pair of 3-Fe-neighbored Fe atoms and the exchange coupling $J_{1b}$ links a 2-Fe-neighbored Fe atom and a 3-Fe-neighbored Fe atom, while the exchange coupling $J_2$ connects a 2-Fe-neighbored Fe atom with a 3-Fe-neighbored Fe atom (see Fig. \[fig:Mag\](d)).
![(Color online) Different magnetic orders on an Fe-Fe square layer with one-quarter Fe vacancies ordered in rhombus: (a) ferromagnetic order; (b) Néel antiferromagnetic (AFM) order, in which all pairs of the nearest-neighboring Fe moments are in antiparallel order; (c) P-collinear AFM order, in which the Fe moments are in antiparallel order along the lines with the Fe vacancies; (d) A-collinear AFM order (ground state), in which the Fe moments are in antiparallel order along the lines without the Fe vacancies. The filled circles represent the Fe atoms while the empty circles represent the Fe vacancies. The red (blue) arrows represent the up-spins (down-spins). The A-collinear AFM order can be divided into six ferromagnetic sublattices, labeled successively as $a_1$ to $a_6$. $J_{1a}$ and $J_{1b}$ are the exchange couplings along the nearest neighbor Fe-Fe directions without and with the Fe vacancies, respectively. $J_2$ is the next-nearest neighbor exchange coupling. The rectangle enclosed by the thick blue solid lines denote a $4\times 2$ magnetic unit cell. The $x$ and $y$ axes are also shown.[]{data-label="fig:Mag"}](fig1){width="8.0cm"}
The magnetic phase diagram of the Hamiltonian has been studied by classical Monte Carlo simulations in Ref. . In the shadowed part as shown in Fig. \[fig:diagram\], the ground state of the Hamiltonian is in the A-col AFM phase with a $4\times 2$ collinear AFM order, whose spin dynamics will be studied below.
Spin dynamics
=============
For a spin Heisenberg model with a long-range magnetic order in its ground state, the linearized spin wave theory with the Holstein-Primakoff (HP) transformation[@HP] is a standard approach to obtain the spin wave excitations and other dynamical properties. In the case of a simple magnetic unit cell, namely no more than two spins per cell, the approach can easily and directly give an analytical solution, for example, the well-known quadratic and linear dispersion behaviors in low energy for ferromagnetic and antiferromagnetic orders, respectively. However, for a complex magnetic structure with a magnetic unit cell containing more than two spins, it is still a severely challenging task to analytically solve the spin wave excitations even though lots of efforts have been devoted in the past.[@Meyer; @Wallace; @Zhang; @Zhi-dong; @Milica] One thus has to be satisfied with numerical solutions. The underlying difficulty is as follows. After the linearized HP transformation, the spin Heisenberg model is transformed into a quadratic Bosonic Hamiltonian. Conventionally one attempts to construct a Bogoliubov transformation to diagonalize the Bosonic Hamiltonian.[@Bogoliubov] In general, there is no practical procedure to analytically construct such a Bogoliubov transformation for a Bosonic Hamiltonian with more than two component bosons, corresponding to a case of more than two spins per unit cell.
Nevertheless an analytical solution is usually in desire and very helpful to understand physics, especially more helpful in the case of a spin Heisenberg model in comparison with neutron measurement, for example, to determine the exchange couplings. Here by studying the Hamiltonian , rather than constructing a Bogoliubov transformation as conventionally, we are going to illustrate an alternative method, which is able to analytically solve spin wave excitations for the cases of multi-spin unit cells. This method employs the equation of motion to construct a secular equation, i.e. an algebraic equation. According to the algebraic basic theorem, an algebraic equation can be analytically solved up to the fourth power, which corresponds to an antiferromagnetic order with eight spins per unit cell. To our knowledge, the advantage of this method applied on a complex magnetic structure has not been well recognized.
. Black solid and red dashed lines represent the classical Monte Carlo calculation (Ref. ) and our result determined by examining the non-imaginary excitation frequency condition, respectively. Blue and green solid dots stand for the positions of KFe$_{1.5}$Se$_2$ and TlFe$_{1.5}$Se$_2$ respectively.[]{data-label="fig:diagram"}](fig2){width="8.0cm"}
Spin wave excitations
---------------------
We now consider the Hamiltonian in the A-col AFM phase with a $4\times 2$ collinear AFM order. As shown in Fig. \[fig:Mag\](d), the magnetic unit cell contains two vacancies and six spins, i.e. three up spins and three down spins, which are labeled by $a_\xi$ with $\xi$ ($\xi=1,\ldots,6$) being odd number or even number, corresponding to up or down spins respectively. In other words, the original antiferromagnetic lattice can be divided into two vacancy-sublattices and six ferromagnetic sublattices, in each of which we can perform the linearized Holstein-Primakoff transformation for the quantum spin $\vec{S}$ at each site $i$ as follows, $$\nonumber
\begin{array}{cc}
\left\{
\begin{array}{l}
S_i^+=\hat{a}_{\xi i}\sqrt{2S} \\
S_i^-=\hat{a}_{\xi i}^{\dagger}\sqrt{2S} \\
S_i^z=S-\hat{a}_{\xi i}^{\dagger}\hat{a}_{\xi i} \\
\xi\in \text{odd number}
\end{array}
\right.
&
~~~~
\left\{
\begin{array}{l}
S_i^+=\hat{a}_{\xi i}^{\dagger}\sqrt{2S} \\
S_i^-=\hat{a}_{\xi i}\sqrt{2S} \\
S_i^z=\hat{a}_{\xi i}^{\dagger}\hat{a}_{\xi i}-S \\
\xi\in \text{even number}
\end{array}
\right.
\end{array}$$ where $\hat{a}_{\xi i}^{\dagger}$ ($\hat{a}_{\xi i}$) is the $\xi$-th component boson creation (annihilation) operator and $i$ is the site index in the $\xi$-th sublattice. We then perform the following Fourier transformation, $$\nonumber
\begin{array}{cc}
\left\{
\begin{array}{l}
\hat{a}_{\xi i}=\frac{1}{\sqrt{N}}\sum_{\textbf{k}} e^{i\textbf{k}\cdot\textbf{R}_i}\hat{a}_{\xi \textbf{k}} \\
\hat{a}_{\xi i}^{\dagger}=\frac{1}{\sqrt{N}}\sum_{\textbf{k}} e^{-i\textbf{k}\cdot\textbf{R}_i}\hat{a}_{\xi\textbf{k}}^{\dagger} \\
\xi\in \text{odd number}
\end{array}
\right.
&
\left\{
\begin{array}{l}
\hat{a}_{\xi i}=\frac{1}{\sqrt{N}}\sum_\textbf{k} e^{-i\textbf{k}\cdot \textbf{R}_i}\hat{a}_{\xi \textbf{k}} \\
\hat{a}_{\xi i}^{\dagger}=\frac{1}{\sqrt{N}}\sum_\textbf{k} e^{i\textbf{k}\cdot \textbf{R}_i}\hat{a}_{\xi \textbf{k}}^{\dagger} \\
\xi\in \text{even number}
\end{array}
\right.
\end{array}$$ where $\textbf{R}_i$ denotes the position vector of site $i$ in the $\xi$-th sublattice, $\textbf{k}=(k_x,k_y)$ is a wave vector in the magnetic Brillouin zone, and $N$ is the total site number of each sublattice.
It turns out that the spin Hamiltonian is now transformed into a quadratic Bosonic Hamiltonian in the momentum space as follows,[@explain] $$\label{eq:Fe1.5-H}
\begin{array}{l}
\hat{H}=E_0+\hat{H}_1 ,\\
E_0=(4J_{1b}-4J_{1a}-8J_2)NS(S+1) ,\\
\hat{H}_1=S\sum_\textbf{k}\hat{\psi}_\textbf{k}^\dag H_1^\textbf{k}\hat{\psi}_\textbf{k} , \\
\end{array}$$ whereas $$\begin{array}{l}
H_1^\textbf{k}=
\left(
\begin{array}{cccccc}
D_1 & 0 & A_\textbf{k} & B_\textbf{k} & 0 & C_\textbf{k} \\
0 & D_1 & B_\textbf{k} & A_\textbf{k} & C_\textbf{k} & 0 \\
A_\textbf{k}^* & B_\textbf{k}^* & D_2 & 0 & A_\textbf{k} & B_\textbf{k} \\
B_\textbf{k}^* & A_\textbf{k}^* & 0 & D_2 & B_\textbf{k} & A_\textbf{k} \\
0 & C_\textbf{k} & A_\textbf{k}^* & B_\textbf{k}^* & D_1 & 0 \\
C_\textbf{k} & 0 & B_\textbf{k}^* & A_\textbf{k}^* & 0 & D_1 \\
\end{array}
\right), \\
\begin{array}{l}
D_1=-J_{1b}+2J_{1a}+2J_2 ,\\
D_2=-2J_{1b}+4J_2 ,\\
A_\textbf{k}=J_{1b}e^{i\theta_y}, \\
B_\textbf{k}=2J_2\cos(\theta_x)e^{-i\theta_y} , \\
C_\textbf{k}=2J_{1a}\cos(\theta_x) ,\\
\theta_x=k_xl , ~~ \theta_y=k_yl ,\\
\hat{\psi}_\textbf{k}^\dag=(\hat{a}_{1\textbf{k}}^{\dag}~\hat{a}_{2\textbf{k}}~
\hat{a}_{3\textbf{k}}^{\dag}
~\hat{a}_{4\textbf{k}}~\hat{a}_{5\textbf{k}}^{\dag}~\hat{a}_{6\textbf{k}}) , \\
\end{array}
\end{array}$$ where $l$ is the length of two nearest neighbor sites, namely the Fe-Fe bond length. In the following we take $l$ as the length unit for convenience.
Next we construct a secular equation to solve the eigenvalues of the Hamiltonian by using the equation of motion method. We first consider the following linear combination, $$\hat{\alpha}_{\textbf{k}}=-\sum_{\xi\in \text{odd}} u_{\textbf{k}\xi}\hat{a}_{\xi \textbf{k}}^\dag+
\sum_{\xi\in \text{even}} u_{\textbf{k}\xi}\hat{a}_{\xi \textbf{k}} ,$$ or $$\hat{\beta}_{\textbf{k}}=\sum_{\xi\in \text{odd}} v_{\textbf{k}\xi}\hat{a}_{\xi \textbf{k}}-
\sum_{\xi\in \text{even}} v_{\textbf{k}\xi}\hat{a}_{\xi \textbf{k}}^\dag .\\$$ Without loss of generality, the coefficients $u_{\textbf{k}\xi}$ and $v_{\textbf{k}\xi}$ are assumed to be complex numbers. Here $\hat{\alpha}_{\textbf{k}}$ and $\hat{\beta}_{\textbf{k}}$ are the boson operators, which satisfy the relationship of the boson commutation, namely $[\hat{\alpha}_{\textbf{k}}, \hat{\alpha}_{\textbf{k}'}^\dag]=\delta_{\textbf{k}\textbf{k}'}$ and $[\hat{\beta}_{\textbf{k}}, \hat{\beta}_{\textbf{k}'}^\dag]=\delta_{\textbf{k}\textbf{k}'}$. This imposes the following constraints upon the coefficients $u_{\textbf{k}\xi}$ and $v_{\textbf{k}\xi}$, $$\begin{array}{r}
-\sum_{\xi\in \text{odd}} u_{\textbf{k}\xi}u_{\textbf{k}\xi}^*+
\sum_{\xi\in \text{even}} u_{\textbf{k}\xi}u_{\textbf{k}\xi}^*=1 , \\
\sum_{\xi\in \text{odd}} v_{\textbf{k}\xi}v_{\textbf{k}\xi}^*-
\sum_{\xi\in \text{even}} v_{\textbf{k}\xi}v_{k\xi}^*=1 ,\\
\end{array}$$ whereas $*$ means the complex conjugate and remains the same meaning as follows.
Now we assume that $\hat{\alpha}_{\textbf{k}}$ or $\hat{\beta}_{\textbf{k}}$ describes an eigen-mode of the Hamiltonian , it is then required that $\hat{\alpha}_{\textbf{k}}$ or $\hat{\beta}_{\textbf{k}}$ fulfils the following equations of motion respectively, $$i\hbar\dot{\hat{\alpha}}_{\textbf{k}}=[\hat{\alpha}_{\textbf{k}}, \hat{H}]=\lambda_\textbf{k} \alpha_{\textbf{k}} \quad
\text{or} ~~~
i\hbar\dot{\hat{\beta}}_{\textbf{k}}=[\hat{\beta}_{\textbf{k}}, \hat{H}]=\lambda_\textbf{k}' \hat{\beta}_{\textbf{k}} ,$$ where $\lambda_\textbf{k}$ and $\lambda_\textbf{k}'$ are the respective eigenvalues. This gives rise to a generalized eigenvalue problem at each wave vector $\textbf{k}$ in the magnetic Brillouin zone, $$\label{eq:alpha-eq}
H_1^\textbf{k} U_\textbf{k}=-\lambda_\textbf{k}\sigma_3U_\textbf{k}, \quad U_\textbf{k}=(u_{\textbf{k}1}, u_{\textbf{k}2}, u_{\textbf{k}3}, u_{\textbf{k}4}, u_{\textbf{k}5}, u_{\textbf{k}6})^T ,\\$$ or $$\label{eq:beta-eq}
H_1^{\textbf{k}*} V_\textbf{k}=\lambda_\textbf{k}'\sigma_3V_\textbf{k}, \quad V_\textbf{k}=(v_{\textbf{k}1}, v_{\textbf{k}2}, v_{\textbf{k}3}, v_{\textbf{k}4}, v_{\textbf{k}5}, v_{\textbf{k}6})^T ,\\$$ where the $6\times 6$ matrix $H_1^{\textbf{k}}$ is given in Eq. , and $\sigma_3$ is a $6\times 6$ diagonal matrix with diagonal elements being (1, -1, 1, -1, 1, -1) respectively.
Eqs. and will show that each eigenvalue of the Hamiltonian is in double degeneracy, which reflects the symmetry or equivalence between the spin-up and spin-down in the ground state. For convenience, we only deal with Eq. in detail. We will not diagonalize Eq. directly since this yields a set of quadratic equations with at least 36 unknowns, which is unlikely to give an analytical expression. Instead, we come to solve the secular equation of the Hamiltonian , namely zeroing the determinant of the matrix $H_1^{\bf{k}}+\lambda_\textbf{k}\sigma_3$, which is the condition for there being a nonzero vector $U_\textbf{k}$. Actually, the Hamiltonian has the same secular equation. The secular equation turns out to be an algebraic equation with one unknown $\lambda_\textbf{k}$, namely $$\label{eq:cubic}
\lambda_\textbf{k}^6+b\lambda_\textbf{k}^4+c\lambda_\textbf{k}^2+d=0,$$ whereas the coefficients $b$, $c$, and $d$ all are real functions of the wave vector $\textbf{k}$, and composed of the exchange couplings. For the detailed expressions on $b$, $c$, and $d$, please refer to Appendix I.
Setting $\gamma=\lambda_\textbf{k}^2$, we then have $\gamma^3+b\gamma^2+c\gamma+d=0$ from Eq. . Further setting $\gamma=\alpha-\frac{b}{3}$, the equation reduces to $\alpha^3+p\alpha+q=0$, whereas $p=-\frac{b^2}{3}+c$, and $q=\frac{2b^3}{27}-\frac{bc}{3}+d$. When the discriminant $\Delta=[(\frac{q}{2})^2+(\frac{p}{3})^3] \leq 0$, there are three real roots for this cubic equation, namely $
\alpha_1=2\sqrt[3]{r}\cos\phi ,~
\alpha_2=2\sqrt[3]{r}\cos(\phi+\frac{2\pi}{3}) ,~
\alpha_3=2\sqrt[3]{r}\cos(\phi+\frac{4\pi}{3}) ,
$ whereas $r=\sqrt{-(\frac{p}{3})^3}$ and $\phi=\frac{1}{3}\arccos(-\frac{q}{2r})$. Now we arrive at the condition, which makes all six roots of Eq. being real, is that $\alpha_1 -\frac{b}{3} \ge 0$ and $\alpha_2-\frac{b}{3}\ge 0$ and $\alpha_3-\frac{b}{3}\ge 0$. In the end, we analytically obtain six branches of eigenvalues for the Hamiltonian , in three pairs of positive and negative, similar to the case of calculating lattice phonon excitations. And we keep the three positive branches to describe the spin wave excitations as follows, $$\label{spinwave}
\left\{
\begin{array}{l}
\lambda_{1\textbf{k}}=\hbar\omega_1(\textbf{k})=S\sqrt{\alpha_1-b/3} ,\\
\lambda_{2\textbf{k}}=\hbar\omega_2(\textbf{k})=S\sqrt{\alpha_2-b/3} ,\\
\lambda_{3\textbf{k}}=\hbar\omega_3(\textbf{k})=S\sqrt{\alpha_3-b/3} .\\
\end{array}
\right.$$
Likewise we can analytically solve Eq. to obtain the same set of eigenvalues, but with the different eigenvectors. Lets take the complex conjugate of Eq. , namely $H_1^{\textbf{k}*}U_\textbf{k}^*=-\lambda_\textbf{k}\sigma_3U_\textbf{k}^*$. Comparing with Eq. , we can see that a positive or negative eigenvalue ($\pm\lambda_\textbf{k}$) with eigenvector $U_\textbf{k}$ of Eq. is a negative or positive eigenvalue ($\mp\lambda_\textbf{k}$) with eigenvector $V_{\textbf{k}}=U_\textbf{k}^*$ of Eq. . This directly shows that the spin wave excitations are in double degeneracy.
The Hamiltonian is derived from the spin Hamiltonian on the assumption of the ground state being in the $4\times 2$ collinear AFM order (see Fig. \[fig:Mag\](d)). As shown above, the eigenvalues of the Hamiltonian may be imaginary. Such a case, if happening, indicates that the $4\times 2$ collinear AFM order is unstable for the spin Hamiltonian . This is similar to the case of imaginary phonon vibration modes, which indicates that the corresponding crystal structure is unstable. Thus we can determine the boundary of A-col phase in the phase diagram of the spin Hamiltonian by checking the condition whether the eigenvalues of the Hamiltonian being real or not. Figure \[fig:diagram\] shows our calculated phase boundary by the red dashed line, which is quite close to the one given by the classical Monte Carlo simulations. The difference between the two is that the AFM quantum fluctuations are included in our spin wave calculations but not in the classical Monte Carlo simulations.
In order to calculate physical quantities, we need to know the eigenstates corresponding to the spin wave excitations, which but are not directly $U_\textbf{k}$ and $V_\textbf{k}$ of Eqs. and because of the boson’s characteristic. Let us denote the eigenvectors corresponding to the positive and negative eigenvalues of Eq. by $U_{\eta,\textbf{k}}^{(+)}$ and $U_{\eta,\textbf{k}}^{(-)}$ respectively, with $\eta=1,2,3$ being the branch index of the spin wave excitations. Then the Hamiltonian will be diagonalized into $$\nonumber
H=E_0+\sum_{\eta \bf{k}}\hbar\omega_\eta(\hat{\alpha}_{\eta \bf{k}}^\dag\hat{\alpha}_{\eta \bf{k}}
+\hat{\beta}_{\eta \bf{k}}^\dag\hat{\beta}_{\eta \bf{k}}+1)$$ after the following linear transformation, $$\nonumber
\left\{
\begin{array}{l}
\hat{\alpha}_{\eta \bf{k}}=-\sum_{\xi\in \text{odd}} u_{\eta \xi,\textbf{k}}^{(+)} \hat{a}_{\xi \bf{k}}^\dag+
\sum_{\xi\in \text{even}} u_{\eta \xi,\textbf{k}}^{(+)} \hat{a}_{\xi \bf{k}}, \\
\hat{\beta}_{\eta \bf{k}}^\dag=\sum_{\xi\in \text{odd}} u_{\eta \xi,\textbf{k}}^{(-)} \hat{a}_{\xi \bf{k}}^\dag-
\sum_{\xi\in \text{even}} u_{\eta \xi,\textbf{k}}^{(-)} \hat{a}_{\xi \bf{k}}. \\
\end{array}
\right.$$ The above set of equations can be rewritten in a matrix form as $$\label{eq:eigenstates}
(\hat{\alpha}_{1\bf{k}}~\hat{\beta}_{1\bf{k}}^{\dag}~ \hat{\alpha}_{2\bf{k}}~\hat{\beta}_{2\bf{k}}^{\dag} ~ \hat{\alpha}_{3\bf{k}}~\hat{\beta}_{3\bf{k}}^{\dag})
=(\hat{a}_{1\bf{k}}^{\dag}~\hat{a}_{2\bf{k}}~ \hat{a}_{3\bf{k}}^{\dag}~\hat{a}_{4\bf{k}} ~ \hat{a}_{5\bf{k}}^{\dag}~\hat{a}_{6\bf{k}})Q_{\textbf{k}}.$$ Here the six columns of the matrix $Q_{\textbf{k}}$ are nothing but the eigenstates corresponding to the spin wave excitations.
Spin dynamical structure factor
-------------------------------
As shown above, $S^z$ just contributes to elastic scattering rather than inelastic scattering since it does not change the number of magnons. In contrast, $S^x$ and $S^y$ contribute to inelastic scattering by changing the number of magnons. In zero temperature the spin dynamical structure factor (SDSF) in inelastic scattering process through single magnon excitations is defined as $$\label{eq:dynamic factor}
\begin{array}{lll}
S(\textbf{k},\omega)& = & \sum_f\sum_{i=x,y}|\langle f|S^i(\textbf{k})|0\rangle|^2\delta(\omega-\omega_f) \\
& = & S\sum_{\gamma}|Q^{-1}_{\gamma,\textbf{k}}|^2\delta(\omega-\omega_f),
\end{array}$$ where $|0\rangle$ is the vacuum state, $|f\rangle$ denotes the final states of a spin system with excitation energy $\omega_f$, and $Q^{-1}_{\gamma,\textbf{k}}$ means the sum of the elements in the $\gamma$-th row of the matrix $Q^{-1}_{\textbf{k}}$ given in Eq. .
Sublattice magnetization
------------------------
In the $\xi$-th sublattice the AFM quantum fluctuation reduces the staggered magnetization from its classical value $S$ by the following quantity $\Delta S_\xi$, $$\label{eq:submag}
\begin{array}{ll}
\Delta S_\xi=S-\frac{1}{N}\langle\sum_{i=1}^N(S-\hat{a}_{\xi i}^\dag \hat{a}_{\xi i})\rangle=
\frac{1}{N}\langle\sum_{\bf{k}} \hat{a}_{\xi\bf{k}}^\dag \hat{a}_{\xi\bf{k}}\rangle \\
=\left\{
\begin{array}{ll}
\frac{1}{N}\sum_{\bf{k}}\sum_{\nu=1,3,5}|Q^{-1}_{\nu\xi,\textbf{k}}|^2, \quad\text{for} ~ \xi \in \text{odd number}, \\
\frac{1}{N}\sum_{\bf{k}}\sum_{\nu=2,4,6} |Q^{-1}_{\nu\xi,\textbf{k}}|^2, \quad\text{for} ~ \xi \in \text{even number}. \\
\end{array}
\right.
\end{array}$$ where $Q^{-1}_{\nu\xi,\textbf{k}}$ denotes the element of the matrix $Q^{-1}_{\textbf{k}}$ at the crossing of the $\nu$-th row and the $\xi$-th column. The summation of $\bf{k}$ can be replaced by a two dimensional integral over the whole Brillouin Zone. It is well-known that the AFM quantum fluctuation is strongly influenced by the spatial dimension and the number of nearest neighbors of a site, namely the coordination number. There are two kinds of two Fe atoms (sites) in a magnetic unit cell in AFe$_{1.5}$Se$_2$, one is 2-Fe-neighbored and the other is 3-Fe-neighbored. Thus we need to calculate the AFM quantum fluctuation on these two kinds of sites separately. Moreover, there usually exists small spin-orientation anisotropy in realistic materials. To account for such anisotropy, one standard approach is to add the term of $-\Lambda_{z}\sum S_i^{z2}$ into the Hamiltonian . This turns out to add $2\Lambda_z$ to all diagonal term of $H_1^{\bf{k}}$ and $-3\Lambda_zNS(S+1)$ to $E_0$ in Eq. respectively.
Results and discussion
======================
The above analytical study will be very helpful both experimentally and theoretically. Experimentally in comparison with neutron inelastic scattering on AFe$_{1.5}$Se$_2$, we can determine the exchange couplings $J_{1a}$, $J_{1b}$, and $J_2$ through the spin wave excitations (Eqs. and ). This will help us better understand the magnetism in AFe$_{1.5}$Se$_2$. On the other hand, these exchange couplings can be theoretically derived from the relative energies of different magnetic states with respect to the non-magnetic state. For the detailed derivation, please refer to the appendix in Ref. . To be specific, we need to first obtain total energies of the four different magnetic states, i.e. the ferromagnetic state ($E_{F}$), Néel AFM state ($E_N$), P-collinear AFM state ($E_P$), and A-collinear AFM state ($E_A$), as shown respectively in Fig. \[fig:Mag\]. The energy differences among these different states result in three linearly independent equations as follows, $$\nonumber
\left\{
\begin{array}{ll}
E_{N}-E_{F}=-4(J_{1a}+J_{1b})/3 ,\\
E_{A}-E_{N}=4(J_{1b}-2J_2)/3 ,\\
E_{A}-E_{P}=-4(J_{1a}-J_{1b})/3 , \\
\end{array}
\right.$$ from which the exchange couplings $J_{1a}$, $J_{1b}$, and $J_2$ can be uniquely determined.
To obtain the total energies, we have carried out the first-principles electronic structure calculations on the compounds AFe$_{1.5}$Se$_2$, which were reported in Ref. . The calculated total energies and subsequently derived exchange couplings for the compounds AFe$_{1.5}$Se$_2$ are listed in Table \[table:energy\]. Accordingly, in Fig. \[fig:diagram\] we mark the positions of AFe$_{1.5}$Se$_2$ in the magnetic phase diagram, both of which are in the A-col phase, but close to the phase boundary. Thus the ground states of AFe$_{1.5}$Se$_2$ both are in the $4\times 2$ collinear AFM order, as observed in the neutron elastic scattering experiment, but they may be readily destructed by doping or applying pressure.
-------------------- -------- -------- -------- -------- ---------- ---------- -------
Compounds $E_F$ $E_N$ $E_P$ $E_A$ $J_{1a}$ $J_{1b}$ $J_2$
KFe$_{1.5}$Se$_2$ -157.1 -253.0 -325.3 -370.0 52.7 19.2 53.5
TlFe$_{1.5}$Se$_2$ -79.4 -176.0 -255.7 -283.8 46.8 25.7 53.3
-------------------- -------- -------- -------- -------- ---------- ---------- -------
: Calculated energies of the four different magnetic states, i.e. the ferromagnetic state ($E_{F}$), Néel antiferromagnetic state ($E_N$), P-collinear antiferromagnetic state ($E_P$), and A-collinear antiferromagnetic state ($E_A$), as shown in Fig. \[fig:Mag\], for AFe$_{1.5}$Se$_2$ (A=K, Tl) (unit: meV/Fe). The energy of the nonmagnetic state is set to zero. The calculated exchange couplings $J_{1a}$, $J_{1b}$, and $J_2$ are also listed (unit: meV/S$^2$).[]{data-label="table:energy"}
From Table \[table:energy\], we also see that the exchange couplings $J_2$ are almost the same for both KFe$_{1.5}$Se$_2$ and TlFe$_{1.5}$Se$_2$, but the former is more anisotropic than the latter on the exchange couplings $J_1$, namely the difference between $J_{1a}$ and $J_{1b}$. Such a magnetic anisotropy will help lessen the AFM quantum fluctuations. With the help of Eq. , we are now able to calculate the AFM quantum fluctuations in AFe$_{1.5}$Se$_2$. We summarize the calculated results in Table \[table:quantum\]. As we find, the amount of the AFM quantum fluctuation in KFe$_{1.5}$Se$_2$ is less than the one in TlFe$_{1.5}$Se$_2$ by about 20%. From Table II, we further find that the AFM quantum fluctuation on a 2-Fe-neighbored Fe atom is, as expected, rather larger than the one on a 3-Fe-neighbored atom, by about 10% and 20% for KFe$_{1.5}$Se$_2$ and TlFe$_{1.5}$Se$_2$ respectively. Moreover a very small spin-orientation anisotropy can remarkably reduce the AFM quantum fluctuations, as shown in the case of $\Lambda_z=0.02J_2$, in which the AFM quantum fluctuation is already reduced from 0.2521 (0.3162) to 0.1943 (0.2337) for KFe$_{1.5}$Se$_2$ (TlFe$_{1.5}$Se$_2$) by approximately 20%. Experimentally the neutron measurements have shown that there is a small spin-orientation anisotropy in KFe$_{1.5}$Se$_2$ and the other iron pnictides.[@cruz; @bao; @shi; @bao1; @zhao] In addition, the effective ordering moment in KFe$_{1.5}$Se$_2$ was found to be 2.8$\mu_B$ per Fe atom.[@zhao] Here setting $S=\frac{3}{2}$, our calculations show that the ordering moment is 2.61$\mu_B$ with the spin anisotropy of $\Lambda_z=0.02J_2$ for KFe$_{1.5}$Se$_2$, in good agreement with the neutron observation. We thus arrive with $S=\frac{3}{2}$ for each Fe atom in KFe$_{1.5}$Se$_2$. In the following discussion, we set $S=\frac{3}{2}$ for each lattice site.
-------------------- ----------------- ----------------- ------------ ----------------- ----------------- ------------
Compounds
$\Delta S_{2n}$ $\Delta S_{3n}$ $\Delta S$ $\Delta S_{2n}$ $\Delta S_{3n}$ $\Delta S$
KFe$_{1.5}$Se$_2$ 0.2685 0.2439 0.2521 0.2093 0.1868 0.1943
TlFe$_{1.5}$Se$_2$ 0.3665 0.2911 0.3162 0.2780 0.2115 0.2337
-------------------- ----------------- ----------------- ------------ ----------------- ----------------- ------------
: Antiferromagnetic quantum fluctuations for AFe$_{1.5}$Se$_2$ (A=K, Tl) with and without spin anisotropy respectively. $\Delta S_{2n}$ and $\Delta S_{3n}$ denote the antiferromagnetic quantum fluctuation on a 2-Fe-neighbored and 3-Fe-neighbored Fe atom respectively. There are two 2-Fe-neighbored and four 3-Fe-neighbored Fe atoms in a $4\times 2$ magnetic unit cell (see Fig. \[fig:Mag\](d)). $\Delta S$ denotes the average antiferromagnetic quantum fluctuation on an Fe atom.[]{data-label="table:quantum"}
). (a) KFe$_{1.5}$Se$_2$, in which $J_{1a}=23.4$, $J_{1b}=8.5$, and $J_2=23.8$ meV ($S=\frac{3}{2}$). (b) TlFe$_{1.5}$Se$_2$, in which $J_{1a}=20.8$, $J_{1b}=11.4$, and $J_2=23.7$ meV ($S=\frac{3}{2}$). Note $k_x$ is of spin-antiparallel direction without the Fe vacancies aligned and $k_y$ is of spin-parallel direction with the Fe vacancies aligned. Here the Fe-Fe bond length $l$ is taken as the length unit for convenience.[]{data-label="fig:spectrum"}](fig3){width="8.0cm"}
For an AFM structure with a magnetic unit cell containing six spins, physically there are six branches of spin wave excitations with every two in degeneracy, namely three 2-degenerated branches, among which one is the gapless Goldstone mode and the other two are gapful optical modes. With Table \[table:energy\] and Eqs. , we can plot the spin wave spectra in the extended Brillouin zone for the compound AFe$_{1.5}$Se$_2$ in Fig. \[fig:spectrum\]. As we see, there are exactly one gapless Goldstone mode and two gapful optical modes, and each branch of spin wave excitation is much more dispersive along the spin-antiparallel direction than along the spin-parallel direction. This is attributed to the Fe vacancies hindering the spin wave propagations, which are located along the spin-parallel direction rather than the spin-antiparallel direction (see Fig. \[fig:Mag\](d)). Furthermore, such hindering becomes more severe, also the spin wave diffraction becomes weaker, for the higher spin wave excitations because of the corresponding wavelengthes being shorter. Especially, the highest optical mode is almost dispersionless along the spin-parallel direction.
![(Color online) Constant-energy cuts of the spin dynamical structure factor (SDSF) for KFe$_{1.5}$Se$_2$, in which $J_{1a}=23.4$, $J_{1b}=8.5$, and $J_2=23.8$ meV ($S=\frac{3}{2}$). (a)-(f) are the SDSF at the excitation energies of 20, 40, 60, 80, 100, and 120 meV, respectively. $k_x$ and $k_y$ both vary from -$\pi$ to $\pi$. Note $k_x$ is of spin-antiparallel direction and $k_y$ is of spin-parallel direction.[]{data-label="fig:DSF-K"}](fig4){width="8.0cm"}
![(Color online) Constant-energy cuts of the spin dynamical structure factor (SDSF) for TlFe$_{1.5}$Se$_2$, in which $J_{1a}=20.8$, $J_{1b}=11.4$, and $J_2=23.7$ meV ($S=\frac{3}{2}$). (a)-(f) are the SDSF at the excitation energies of 20, 40, 60, 80, 100, and 120 meV, respectively. $k_x$ and $k_y$ both vary from -$\pi$ to $\pi$. Note $k_x$ is of spin-antiparallel direction and $k_y$ is of spin-parallel direction.[]{data-label="fig:DSF-Tl"}](fig5){width="8.0cm"}
To compute the spin dynamical structure factor by using Eq. , the function $\delta(\omega-\omega_f)$ is smoothed by a Gaussian distribution function, which is centered at the excitation energy $\omega_f$ with the half-width at the half-maximum (HWHM) assigned to $0.2J_2$. We plot the spin dynamical structure factor at a constant cut energy, ranging from 20 to 120 meV for both compounds, as shown in Figs. 4 and 5 respectively.
For KFe$_{1.5}$Se$_2$, referring to Fig. 3, one can see that the diffraction peaks come from the acoustic (Goldstone mode) spin waves for the excitation energy $\omega_f<$ 50 meV, the acoustic and the first optical spin waves for 50 meV $< \omega_f <$ 100 meV, and the first and second optical spin waves for $\omega_f >$ 100 meV, respectively. When $\omega_f=20$ meV, Figure \[fig:DSF-K\] shows that the strongest diffraction peaks are located at ($\pm\pi$, 0). In addition there are two small half elliptical diffraction peaks near $\Gamma$ point and four almost invisible diffraction peaks at the four corners of the extended Brillouin zone. When $\omega_f$ increases from 20 to 80 meV, these ellipses extend larger. At 100 meV, the diffraction peaks become line-shaped. When $\omega_f=120$ meV, the diffraction peaks become helical lines along the $k_y$ directions, namely spin-parallel direction. These features are consistent with the strong anisotropy of the spin wave excitations along the spin-parallel and spin-antiparallel directions due to the Fe vacancies.
For TlFe$_{1.5}$Se$_2$, we also find the quite similar features on the spin dynamical structure factor (see Fig. \[fig:DSF-Tl\]). Since the spin wave dispersions become further weaker along the spin-parallel direction in TlFe$_{1.5}$Se$_2$ than in KFe$_{1.5}$Se$_2$, as shown in Fig. 3, the diffraction peaks thus become line-shaped along $k_y$ at 60 meV rather than 100 meV for TlFe$_{1.5}$Se$_2$. Moreover, when $\omega_f$ increases from 80 to 120 meV, the diffraction peaks become pear-shaped patterns. The large green areas near the right and left boundaries of the extended Brillouin zone at 100 meV come from the flatness of the highest optical branch.
Conclusion
==========
We have shown that we can analytically solve the spin wave excitations for a complex magnetic structure by using the equation of motion method in the framework of the linearized spin wave theory, which is illustrated by studying the magnetism of the newly discovered intercalated ternary iron-selenide AFe$_{1.5}$Se$_2$ (A=K, Tl). We find that there are one acoustic branch (gapless Goldstone mode) and two gapful optical branches of the spin wave excitations with each in double degeneracy in AFe$_{1.5}$Se$_2$. The phase boundary of AFe$_{1.5}$Se$_2$ in the $4\times 2$ collinear antiferromagnetic order is determined by examining the non-imaginary excitation frequency condition, which incorporates the antiferromagnetic quantum fluctuations. We also derive the exchange couplings between the Fe moments based on the first-principles total energy calculations, so that we can calculate and discuss the spin dynamical structure factors in connection with neutron inelastic scattering experiment. By computing the antiferromagnetic quantum fluctuations, we find that the Fe spin is $S=\frac{3}{2}$ in AFe$_{1.5}$Se$_2$ and a very small spin-orientation anisotropy can remarkably suppress the antiferomagnetic quantum fluctuations.
This work is supported by National Program for Basic Research of MOST of China (Grant No. 2011CBA00112) and National Natural Science Foundation of China (Grant Nos. 11190024 and 91121008).
Appendix I: coefficients in Eq. (5)
===================================
$$%\begin{equation}
\left\{
\begin{array}{l}
b=-4J_{1b}^2+16J_2^2\cos(\theta_x)^2+2C_{\bf{k}}^2-2D_1^2-D_2^2 ,\\
\\
c=4J_{1b}^4-32J_{1b}^2J_2^2\cos(\theta_x)^2+64J_2^4\cos(\theta_x)^4-4J_{1b}^2C_{\bf{k}}^2
+16J_2^2\cos(\theta_x)^2C_{\bf{k}}^2+C_{\bf{k}}^4+4J_{1b}^2D_1^2-16J_2^2\cos(\theta_x)^2D_1^2 \\
\qquad -2C_{\bf{k}}^2D_1^2+D_1^4+16J_{1b}J_2\cos(\theta_x)C_{\bf{k}}D_2-4J_{1b}^2D_1D_2
-16J_2^2\cos(\theta_x)^2D_1D_2-2C_{\bf{k}}^2D_2^2+2D_1^2D_2^2 ,\\
\\
d=2J_{1b}^4\cos(4\theta_y)C_{\bf{k}}^2+2J_{1b}^4C_{\bf{k}}^2+
16J_{1b}^2J_2^2\cos(\theta_x)^2\cos(4\theta_y)C_{\bf{k}}^2
-16J_{1b}^2J_2^2\cos(\theta_x)^2C_{\bf{k}}^2+32J_2^4\cos(\theta_x)^4\cos(4\theta_y)C_{\bf{k}}^2 \\ \qquad
-32J_{1b}^2J_2^2\cos(\theta_x)^2C_{\bf{k}}^2+32J_2^4\cos(\theta_x)^4C_{\bf{k}}^2
-16J_{1b}^3J_2\cos(\theta_x)\cos(4\theta_y)C_{\bf{k}}D_1+16J_{1b}^3J_2\cos(\theta_x)C_{\bf{k}}D_1 \\ \qquad
-64J_{1b}J_2^3\cos(\theta_x)^3\cos(4\theta_y)C_{\bf{k}}D_1
+64J_{1b}J_2^3\cos(\theta_x)^3C_{\bf{k}}D_1-4J_{1b}^4D_1^2
+32J_{1b}^2J_2^2\cos(\theta_x)^2\cos(4\theta_y)D_1^2 \\ \qquad
-64J_2^4\cos(\theta_x)^4D_1^2+16J_{1b}J_2\cos(\theta_x)C_{\bf{k}}^3D_2
-4J_{1b}^2C_{\bf{k}}^2D_1D_2-16J_2^2\cos(\theta_x)^2C_{\bf{k}}^2D_1D_2-
16J_{1b}J_2\cos(\theta_x)C_{\bf{k}}D_1^2D_2 \\ \qquad
+4J_{1b}^2D_1^3D_2+16J_2^2\cos(\theta_x)^2D_1^3D_2-C_{\bf{k}}^4D_2^2
+2C_{\bf{k}}^2D_1^2D_2^2-D_1^4D_2^2 ,
\end{array}
\right.$$
where $C_{\bf{k}}$, $D_1$, $D_2$, $\theta_x$, and $\theta_y$ are defined in Eq. in the text.
Y. Kamihara, T. Watanabe, M. Hirano, and H. Hosono, J. Am. Chem. Soc. [**130**]{}, 3296 (2008).
F. Ma and Z. Y. Lu, Phys. Rev. B [**78**]{}, 033111 (2008).
C. de la Cruz, Q. Huang, J. W. Lynn, J. Li, W. Ratcliff II, J. L. Zarestky, H. A. Mook, G.F. Chen, J. L. Luo, N. L. Wang, and P. Dai, Nature [**453**]{}, 899 (2008).
F. Ma, Z. Y. Lu, and T. Xiang, Phys. Rev. B [**78**]{}, 224517 (2008).
F. Ma, W. Ji, J. P. Hu, Z. Y. Lu, and T. Xiang, Phys. Rev. Lett. [**102**]{}, 177003 (2009). W. Bao [*et al.*]{}, Phys. Rev. Lett. [**102**]{}, 247001 (2009).
S. L. Li [*et al.*]{}, Phys. Rev. B [**79**]{}, 054503 (2009).
I. I. Mazin, D. J. Singh, M. D. Johannes, and M. H. Du, Phys. Rev. Lett. [**101**]{}, 057003 (2008).
T. Yildirim, Phys. Rev. Lett. [**101**]{}, 057010 (2008).
Q. Si and E. Abrahams, Phys. Rev. Lett. [**101**]{}, 076401 (2008).
J. Zhao, D. X. Yao, S. Li, T. Hong, Y. Chen, S. Chang, W. Ratcliff, II, J. W. Lynn, H. A. Mook, G. F. Chen, J. L. Luo, N. L. Wang, E. W. Carlson, J. Hu, and P. Dai, Phys. Rev. Lett. [**101**]{}, 167203 (2008).
J. Zhao, D. T. Adroja, D. X. Yao, R. Bewley, S. Li, X. F. Wang, G. Wu, X. H. Chen, J. Hu, and P. Dai, Nature Physics [**5**]{}, 555 (2009).
O. J. Lipscombe, G. F. Chen, C. Fang, T. G. Perring, D. L. Abernathy, A. D. Christianson, T. Egami, N. Wang, J. Hu, and P. Dai, Phys. Rev. Lett. [**106**]{}, 057004 (2011).
M. Wang, C. Fang, D. X. Yao, G. T. Tan, L. W. Harriger, Y. Song, T. Netherton, C. Zhang, M. Wang, M. B. Stone, W. Tian, J. Hu, and P. Dai, Nature Communications [**2**]{}, 580 (2011).
J. Guo, S. Jin, G. Wang, S. Wang, K. Zhu, T. Zhou, M. He, and X. Chen, Phys. Rev. B [**82**]{}, 180520(R) (2010).
A. Krzton-Maziopa, [*et al.*]{}, J. Phys.: Condens. Matter [**23**]{}, 052203 (2011).
M. H. Fang, H. D. Wang, C. H. Dong, Z. J. Li, C. M. Feng, J. Chen, and H. Q. Yuan, Europhysics Letters [**94**]{}, 27009 (2011). Z. Shermadini, [*et al.*]{}, Phys. Rev. Lett. [**106**]{}, 117602 (2011).
W. Bao, Q. Huang, G. F. Chen, M. A. Green, D. M. Wang, J. B. He, X. Q. Wang, and Y. Qiu, Chin. Phys. Lett. [**28**]{}, 086104 (2011). M. Rotter, M. Tegel, and D. Johrendt, Phys. Rev. Lett. [**101**]{}, 107006 (2008).
X. W. Yan, M. Gao, Z. Y. Lu, and T. Xiang, Phys. Rev. Lett. [**106**]{}, 087005 (2011).
X. W. Yan, M. Gao, Z. Y. Lu, and T. Xiang, Phys. Rev. B [**83**]{}, 233205 (2011).
Y. J. Yan, [*et al.*]{}, Phys. Rev. X [**1**]{}, 021020 (2011).
J. Zhao, H. Cao, E. Bourret-Courchesne, D. -H. Lee, and R. J. Birgeneau, arXiv:1205.5992.
T. Holstein and H. Primakoff, Phys. Rev. [**58**]{}, 1098 (1940).
H. Meyer and A. B. Harris, J. Appl. Phys. 31, 49S (1960). D. C. Wallace, Phys. Rev. [**128**]{}, 1614 (1962). Z. D. Zhang, Phys. Rev. B [**53**]{}, 2569 (1996). M. Pavkov, M. Škrinjar, D. Kapor, and S. Stojanović, Phys. Rev. B [**62**]{}, 6385 (2000). N. N. Bogoliubov, J. Phys. (USSR), [**11**]{}, 23 (1947). R. Yu, P. Goswami, and Q. Si, Phys. Rev. B [**84**]{}, 094451 (2011). There are numerous arrangements available for the Hamiltonian matrix elements. The current arrangement $H_1^{\textbf{k}}$ is found to well fit the derivation of Eqs. and .
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We study the interaction energy between two graphene nanoribbons by first principles calculations, including van der Waals interactions and spin polarization. For ultranarrow zigzag nanoribbons, the direct stacking is even more stable than the Bernal stacking, competing in energy for wider ribbons. This behavior is due to the magnetic interaction between edge states. We relate the reduction of the magnetization in zigzag nanoribbons with increasing ribbon width to the structural changes produced by the magnetic interaction, and we show that when deposited on a substrate, zigzag bilayer ribbons remain magnetic for larger widths.'
author:
- 'H. Santos$^{1,2,3}$'
- 'A. Ayuela$^{3}$'
- 'L. Chico$^1$'
- 'Emilio Artacho$^{2,4,5,6}$'
title: Van der Waals interaction in magnetic bilayer graphene nanoribbons
---
Introduction
============
Charge carriers in graphene follow a linear dispersion relation close to the Fermi energy. For this reason, they are considered as massless fermions obeying Dirac’s equation. [@transport] When several layers of graphene are piled up together, their electronic and transport properties can be dramatically modified, depending on the stacking arrangement and the number of layers.[@Heinz] There are several possible stacking arrangements in bilayer graphene, the most symmetric cases being direct (AA) and Bernal (AB) stackings. Most theoretical studies have focused on the AB stacking, because it is that of graphite, being the lowest energy configuration for the three-dimensional crystal. [@charlier1] However, the AA stacking has been observed in experiments on few-layer graphene, and it should also be considered in bilayer stackings.[@Norimatsu; @jae_kap; @Liu; @Ohto] For example, the AA and AB stackings have been observed indistinctly at the graphene edges in samples grown on SiC. [@Norimatsu] In fact, bilayer graphene with AA stacking has also been synthesized, and observed by transmission electron microscopy (TEM). [@jae_kap; @Liu] Due to the differences in the electronic properties of bilayer AA and AB, the change between stackings by a relative displacement of the layers has even been proposed as the key mechanism for a switch device.[@Ohto; @Jhon1; @Jhon2]
Graphene nanoribbons are graphene strips of nanometric width and arbitrary length, with electronic properties depending on their edges and widths.[@son_nature; @Brey_2006] They are considered as potential materials for future nanoelectronics because they can behave as metals or semiconductors, making possible the design of electronic elements based solely on them. The simplest nanoribbon geometries are those with zigzag and armchair edges, which have been studied extensively.[@Brey_2006] Other edges terminations are possible, but they can be mapped onto three basic types, the armchair being the only one without edge states.[@Akhmerov; @Jaskolski2011] These localized states close to the Fermi energy are responsible for the magnetic and transport properties of zigzag graphene ribbons, and they are the origin of defect-related interface bands in graphene junctions.[@Santos2009] Within a simple tight-binding model, armchair graphene nanoribbons (AGNRs) can be either metallic or semiconducting depending on their width, [@nakada] whereas zigzag graphene nanoribbons (ZGNRs) are metallic with edge states. [@son_2006] More realistic calculations yield all semiconducting armchair ribbons. [@Ezawa; @Barone] With regard to ZGNRs, the inclusion of electronic interactions reveals a ferromagnetic order of the magnetic moments at each edge, with an edge-edge antiferromagnetic coupling that opens a small gap. In fact, this magnetic characteristic makes ZGNRs interesting for spintronic devices. [@alicante; @santos]
In bilayer graphene nanoribbons ([*b*]{}-GNR) both, edges and stacking order, determine the electronic and magnetic properties. Even though in few-layer samples there are multiple possibilities for the stacking arrangements, the majority of previous theoretical studies have focused on the AB stacking.[@McCann; @McCann2; @Nilsson; @banerjee; @fazzio] The interaction between the edges of zigzag bilayer graphene ribbons determines the survival of magnetism. A combined first-principles and tight-binding approach was used to study the electronic properties in armchair and zigzag GNR.[@banerjee] Because these authors find an important dependence on the functional employed, they fix their distance to graphite for zigzag GNR. Their magnetism is thus masked, as its survival depends on the layer-layer distance. An attempt to relax the edges was considered using a local spin density (LSD) approach within density functional theory.[@Kim] They found that, for wide bilayer zigzag nanoribbons, the total magnetic moment is zero. Previous works do not consider van der Waals (vdW) forces. In order to relax bilayer ribbons, an explicit description of the vdW interaction must be included beyond LSD. When these long-range interactions are included, the electronic densities between the layers are rearranged, and this yields variations on the interlayer distances. Such vdW interaction is included at a simple level in Ref. and the edge magnetism disappears for small ribbon widths. However, we should note that vdW interaction is included in an effective way, modifying the atomic potentials. Other implementations using a fully nonlocal van der Waals density functional must thus be checked.
In this work we study the properties of bilayer zigzag ribbons where all the edge carbon atoms are passivated by hydrogen. We include van der Waals dispersion forces with the fully non-local density functional recently proposed from first-principles,[@LMKLL] within the family of functionals based on Ref. , as recently factorized for efficiency. In Section II we describe the computational details. We investigate the stability of Bernal and direct stackings in [*b*]{}-GNR, focusing on the magnetic interaction between edges and on the interplay between magnetism and structural changes in narrow zigzag ribbons. Section III describes the systems studied, and shows our results, presenting the binding energies, and magnetic and structural changes in zigzag bilayer nanoribbons. Our main conclusion is that direct stacking competes with Bernal stacking below a critical ribbon width, and we show that the magnetic coupling between edge states in the different ribbons plays a key role in such competition. Indeed, for ultranarrow ribbons, the direct stacking has the lowest total energy and largest binding energy. Furthermore, the structural distorsion at the edges due to this interaction makes the magnetization negligible in bilayer ribbons, causing metallization. However, when deposited on a substrate, the structural deformation is reduced, thus maintaining the edge magnetism for larger ribbon widths. We finish with a brief summary in Section IV.
Computational details {#sec:comp}
=====================
First principles calculations are performed using the SIESTA code with spin polarization.[@siesta1] We use the van der Waals functional parametrized by Lee [*et al.*]{} (vdW-DF2),[@LMKLL] which is a second version of the original vdW-DF functional by Dion and coworkers.[@DRSLL] The factorization proposed in Ref. represents a very substantial efficiency improvement in the evaluation of the exchange-correlation potential and energy, thus enabling first-principles van der Waals calculations for any system accessible to usual generalized gradient approximations GGAs. We check that the interlayer space in graphite is in agreement with previous calculations.[@LMKLL] The results presented below have been performed using the functional vdW-DF2, but we have checked that other functionals implemented in the SIESTA code preserve the main features found employing vdW-DF2. Our choice of functional is motivated by the fact that vdW-DF2 gives more realistic binding energies for bilayer graphene when compared to experimental works. The electron-ion interactions use norm-conserving nonlocal Troullier-Martins pseudopotentials[@Troullier] generated with the atomic configuration \[He\]$2s^2$$2p^2$ taken as reference with a radius cutoff of 1.25 Å for $s$, $p$, $d$ and $f$ orbitals. Spin polarized calculations normally require a fine sampling of the Brillouin zone, which we performed with a Monkhorst-Pack scheme of $30\times1\times1$ $k$-points.[@Monkhorst] The real-space grid for matrix-element computations [@siesta1] uses an energy cutoff of 350 Ry. The structure was relaxed by conjugate gradients optimization until forces were smaller than 0.01 eV/Å. Periodic boundary conditions were applied, so we use large enough supercell parameters (15 Å) in the directions perpendicular to the ribbon’s long axis to avoid spurious interactions between adjacent ribbons. All the carbon atoms at the edges are passivated by hydrogen.
Results {#sec:rec}
=======
Ingredients: monolayer zigzag nanoribbons
-----------------------------------------
Before undertaking the calculation of bilayer nanoribbons, we have first verified that our approach gives reasonable results for monolayer zigzag nanoribbons. For these, the key parameters to determine the electronic behavior at the Fermi energy are both the edge shape and the ribbon width. We have used two initial magnetic configurations for the edges of ZGNR: either ferromagnetic ([*fm*]{}), with aligned spin polarizations, or antiferromagnetic ([*afm*]{}), with antiparallel spin polarizations. Notice that all atoms in the same edge are ferromagnetic coupled. [@son_2006] We have performed calculations of ZGNRs with several widths and both [*afm*]{} or [*fm*]{} orderings. We found that the [*afm*]{} order is always more stable than the [*fm*]{}, and that their energy difference decreases with increasing ribbon width, as in previous calculations.[@son_2006]
![(Color online) (a) Schematic packing of zigzag bilayer graphene nanoribbons: AA (above), AB$_\beta$ (middle), and AB$_\alpha$(below). The dark gray layer corresponds to the bottom layer and the light gray (blue) layer correspond to the upper layer. Hydrogen atoms are denoted by white balls. The ribbon width $N$ is given by the number of dimers (zigzag chains) from edge to edge. (b) Magnetic configurations of the edge states in [*b*]{}-ZGNR. Interlayer (intralayer) coupling can be either ferromagnetic \[FM([*fm*]{})\] or antiferromagnetic \[AFM([*afm*]{})\]. Notice that one single edge is always ferromagnetic, i.e., all the spins along the same edge are parallel.[]{data-label="fig1"}](fig_1.eps){width="8.4cm"}
Bilayer zigzag nanoribbons
--------------------------
### Binding energies and stable configurations
As in infinite bilayer graphene, we have to look at different stacking orders for bilayer graphene nanoribbons ([*b*]{}-ZGNRs). Fig. \[fig1\] (a) shows the three stackings considered in this work. The top panel of the figure depicts an example of direct (AA) stacking. Two types of AB stacking have to be considered, according to the relative position of their edges. The medium and bottom panels of Fig.\[fig1\] (a) show the so-called AB$_\beta$ and AB$_\alpha$ stackings for zigzag ribbons. We identify the ribbon width by $N$, being the number of zigzag chains from edge to edge. [@nakada]
As the edges of [*b*]{}-ZGNRs have magnetization, we have to study both the intralayer and interlayer (i.e., layer-to-layer) magnetic couplings in these ribbons. We study four possible magnetic configurations for all the stackings considered, as depicted in Fig. \[fig1\] (b). In order to distinguish in our notation between intralayer and interlayer coupling, we use capital letters for the layer-to-layer coupling, and lower-case letters to label the intralayer coupling. The AFM-[*afm*]{} configuration (upper left diagram of Fig. \[fig1\] (b)) thus has both antiferromagnetic intralayer and interlayer coupling. The bottom-left diagram, the FM-[*afm*]{}, shows two [*afm*]{}-coupled GNR layers with FM interlayer coupling. The third and fourth configurations, AFM-[*fm*]{} and FM-[*fm*]{}, shown in the top and bottom right diagrams of Fig. \[fig1\] (b), have both [*fm*]{} intralayer coupling with AFM or FM interlayer coupling respectively.
We start from one of the four initial spin configurations for [*b*]{}-ZGNRs described above in a range of distances, and then we relax the minimum geometry so that the system evolves in principle towards similar magnetic configurations with lower energy. The converged solutions for different initial guesses are very close in energy. Each converged magnetic configuration can be viewed as a possible metastable solution.[@Marom] It is likely that external conditions, such as magnetic and electric fields, can stabilize the system into in a configuration different from the energy minimum. However, for the AA stacking, the FM-([*afm*]{}, [*fm*]{}) magnetic initial guesses do not yield a stable solution: as the layers become closer, the electron density flips during self-consistency to the AFM ground state. The same happens for the AB stacking, where the FM-[*afm*]{} cases flip to AFM-[*afm*]{} solutions. Since for the AB stackings the atoms of an edge are not exactly on top of the atoms of the other, we obtained a large number of metastable magnetic alignments.
From the total energies, we calculate the binding energy (BE) as the difference between the energy of the coupled bilayer and the two isolated monolayers in the most stable configuration, i.e., the antiferromagnetic ([*afm*]{}) alignment.[@BE] The binding energy is related to the strength of the layer-layer interaction; it is given in meV per atom in a layer.[@Grimme; @Podeszwa] Figure \[fig2\] shows the binding energy of [*b*]{}-ZGNRs with widths $N$ ranging from 2 to 10, for several magnetic configurations between edges. The binding energies that we obtain for bilayer graphene are 43.3 meV/atom and 50.6 meV/atom for AA and AB stacking, respectively. These values are in good agreement with the ones obtained from experimental works, namely, the exfoliation of graphite determines a binding energy between graphene layers of about 43 meV/atom,[@Girifalco] and that estimated for the separation of polyaromatic hydrocarbons is about 52 meV/atom. [@Zacharia] As expected, we see in Fig. \[fig2\] that the binding energy increases in absolute value with the nanoribbon width. Interestingly, the increase is not monotonous and for certain widths it shows an enhanced stability, see for instance the $N=4$ case.
We find that the largest binding energy and most stable configuration for $N> 4$ is the AB$_\alpha$ in the AFM-[*afm*]{} configuration. Remarkably, for ultranarrow ribbons, $N=2,4$, the AA stacking with AFM-[*afm*]{} coupling is more favorable, albeit with a very close value of the BE to that of the AB$_\alpha$ in the AFM-[*afm*]{} magnetic ordering.
We now analyze in more detail the role of magnetic configurations on the stacking of ribbons. The cases AB$_\alpha$ and AB$_\beta$ which have [*fm*]{} intralayer coupling are in the same energetic range, with the binding energy of AB$_\beta$ larger than AB$_\alpha$ for all widths. With the exception of the ultranarrow widths, all the AB$_\alpha$ cases, as well as the AB$_\beta$ with either FM or [*fm*]{} couplings, are rather close as to the binding energies. This shows a relationship between stackings and magnetic configurations of the edges.
To elucidate the role of the magnetic interactions between edges, we have calculated the binding energies of ZGNRs on graphene. In such a case, because we are suppressing a ribbon, we are focusing on the edge-graphene interaction instead of the edge-edge interlayer interaction, and we only distinguish between AB and AA stackings. We find that the binding energy is lower in Bernal stacking for all the studied widths, as it has been previously found in bulk graphite, where the Bernal stacking is more stable than the AA. This indicates that our finding on the greater stability and stronger binding energies for ultranarrow 4-ZGNR with AA stacking is related to the edge-edge interlayer coupling.
### Structural and magnetic changes: quenching of the edge magnetic moments
The differences in binding energies between stacking orderings can be related to structural and magnetic changes in the bilayer ribbons. In our fully relaxed simulations for [*b*]{}-ZGNR, all the AB$_\beta$ cases, and most of the AB$_\alpha$ cases, the layers remain flat. Only the AA and AB$_\alpha$ stackings with AFM-[*afm*]{} coupling change their geometrical structure. Fig. \[fig3\] (a) shows two examples for $N=10$. Notice that the edges are bent inwards; in the AA case, the layers bend symmetrically, becoming convex at their center, and in the AB$_\alpha$ case, for which the ribbons are laterally displaced, the edges approach maintaining a flat central region. The converged geometries of the ribbons are distorted, but still they have relevant symmetries, which are preserved within tolerance ($\approx 0.02 $ Å): the AA stacking with AFM-[*afm*]{} configuration shows a mirror symmetry and $C_2$ rotation with an axis parallel to the ribbons, while AB$_\alpha$ with AFM-[*afm*]{} shows only $C_2$ symmetry.
![(Color online) (a) Structural distortions of the zigzag ribbon with $N=10$ for the most stable stackings AA and AB$_\alpha$ in the AFM-[*afm*]{} configurations after full relaxation. The systems with higher binding energy (in absolute value) consist of non-planar, distorted, graphene ribbons. (b) Interlayer distances at the center $h_c$ (empty diamonds) and edges h$_e$ (full diamonds) for [*b*]{}-ZGNRs with AA (black) and AB$_\alpha$ \[blue, gray\] stackings in the AFM-[*afm*]{} configuration as a function of the ribbon width. []{data-label="fig3"}](fig_3.eps){width="8.5cm"}
To quantify these distortions, in Fig. \[fig3\] (b) we plot the distances between the two ribbons at their center, $h_c$, and at their edges, $h_e$, as a function of the bilayer width. The distances at the central part $h_c$ remain constant for the AB$_\alpha$ series, whereas they show larger changes for the AA stackings. For very small widths, up to $N=4$, the central distances $h_c$ for AA stackings are lower than for the AB cases. This is not what happens in bulk bilayer graphene, where the layer-to-layer distance is smaller for Bernal stacking than for AA. This is an indication of the strong interaction in these ultranarrow ribbons with AA stacking. For bilayer ribbons with $ N \geq 6$, the behavior is as expected, i.e., with central interlayer distances $h_c$ smaller in the AB$_\beta$ ribbons than for the AA cases. On the contrary, the distance between edges, $h_e$, is notably different from $h_c$ when N $>$ 6. This deviation can be as large as 0.6 Å, indicating a strong edge-edge interaction. Note that for the $AB_\alpha$ cases, these structural distortions are accompanied by a lateral sliding, but the values we find, e,g., 0.1 Å for $N=10$, are much smaller than those reported previously. [@fazzio] These interlayer distances $h_c$ and $h_e$ show that for ultranarrow ribbons, up to $N=6$, the bilayer ribbons behave rigidly, becoming more flexible for larger widths.
We now focus on the changes of magnetization at the edges $M$ due to the edge-edge interaction, as also shown in Fig. \[figmag\]. It is defined as $M=N_{up}-N_{down}$, where $N_{up}$ ($N_{down}$) is the number of electrons with spin $up$ ($down$) per edge atom. The total magnetization by $95 \%$ corresponds to $p_z$ orbitals. We find that the magnetic moments are mainly located at the edges and decay exponentially when moving into the central part of the ribbon, in agreement with previous results.[@son_nature; @mananes] Notice that the interlayer interaction between edges suppresses the site magnetization: for the planar cases, the magnetization value is about 0.25 $\mu _b$. This is the case for all the magnetic configurations with AB$_\beta$ stacking, and the AB$_\alpha$ ones with FM intralayer coupling, see Fig. \[figmag\]. However, for the AA and AB$_\alpha$ stackings with AFM-[*afm*]{} couplings, when the interlayer edge distance $h_e$ is reduced, a strong interaction appears between the $p_z$ orbitals at opposite edges, and the spin cloud evolves towards nonmagnetic configurations. In fact, the spin polarization for $ N\geq10$ is almost quenched.
### Implications of the magnetic quenching for calculations and experiments
[*Narrowing of gaps at large widths*]{}. These magnetic and structural changes are associated with other variations of the electronic properties of ribbons. The band structures of the [*b*]{}-ZGNRs of width $N=8$ for AA and AB$_\alpha$ stackings in the AFM-[*afm*]{} configuration are shown in Fig. \[fig4\] (a). The gaps of the ribbons with AA stacking are smaller than those of the AB$_\alpha$ ones. When increasing the ribbon width $N$ the gap narrows as $\sim 1/N$. [@banerjee]
![(Color online) (a) Band structure of a [*b*]{}-ZGNR with $N=8$ in the most stable stackings, AA and AB$_\alpha$, and AFM-[*afm*]{} magnetic configuration. The Fermi energy is set to zero. Note the narrowing of the gap. (b) Gaps versus ribbon widths for the two most stable stackings and magnetic configurations, namely, AA and AB$_\alpha$ all with AFM-[*afm*]{}.[]{data-label="fig4"}](fig_5.eps){height="8.5cm"}
Note the sharp drop in the energy gap after $N=6$, related to the sudden decrease of the edge magnetization and the subsequent metallization of the bilayer nanoribbon, with $E_g < 0.05$ eV for $N=8$ in the AA stacking, barely visible in Fig. \[fig4\] (a).[@son_nature; @mananes] It should be noted that the nanoribbon gaps may be underestimated in a GGA calculation.[@Heyd] The use of other functionals, such as the hybrid Heyd-Scuseria-Ernzerhof[@HSE] would correct this effect; in any case, the gap decrease with increasing width and the abrupt jumps associated with the magnetic changes will certainly hold in calculations employing other functionals, but with the quenching of magnetic moments taking place at larger widths.
[*Magnetoelastic switching*]{}. Our results show a strong relationship between structure deformation and magnetic configuration. However, a question that remains is on the reversibility of structural changes with respect to the magnetic configuration, which can be relevant for experiments. To address it, as well as to corroborate the interplay between magnetism and structural changes, we have chosen the case of $N=10$ in the ground-state, i.e., stacking AB$_\alpha$ and AFM-[*afm*]{} configuration. As it is shown in the previous Section, the ribbons in this structure are strongly curved. The application of a magnetic field perpendicular to the layer flips the magnetic moments at the edges from an AFM to an FM interlayer configuration; when we flip the magnetic moments from AFM to FM interlayer coupling in the curved structure and we relax it, it converges to a planar geometry in the FM configuration. We consider this finding to be an indication of a change from curved to planar geometry driven by magnetic fields. In principle, in a magneto-mechanical device based on these ribbons, one could control the edge deformation with magnetic fields, which in turn can produce other electronic changes such as gap narrowing.
[*Effect of a substrate*]{}. We have studied the magnetic interaction of the edges in bilayer ribbons when deposited on a graphene substrate. Figure \[such\] shows the geometric and magnetic structure of the [*b*]{}-ZGNR with $N=10$ deposited on graphene in the AFM-[*afm*]{} magnetic configuration with AB$_\alpha$ stacking. The interlayer distance in the center of the structure is overestimated, as it also occurs in bulk graphite[@LMKLL]; this is a consequence of the vdW-DF2 functional used. The top nanoribbon is deformed most, while the lower nanoribbon is nearly planar, due to the competing interaction between the graphene layer and the top ribbon. Due to this flat geometry, the associated magnetization at the edges has higher values than the those obtained for bilayer nanoribbons, close to the ones of a single strip. As the structural deformation of the sandwiched layer is impeded by the substrate interaction, the magnetic quenching is also precluded. Therefore, bilayer nanoribbons on substrates will remain magnetic for larger values of $N$ than when suspended. Bilayer nanoribbons with widths about $20$ nm can therefore act as a spintronic device, maybe not in the free standing geometry, but certainly on substrates.
Summary
=======
Bilayer zigzag graphene nanoribbons have been studied by first principles DFT calculations including a vdW-DF2 van der Waals functional. Four possible magnetic configurations have been explored for the three more symmetric stackings, the AA and two Bernal (AB$_\alpha$ and AB$_\beta$).
Our results show that the AA stacking is more favorable for ultranarrow ribbons in the AFM-[*afm*]{} configuration, competing in energy with the Bernal AB$_\alpha$ for larger [*b*]{}-ZGNRs. The edge interaction bends their structure inwards for the AA and AB$_\alpha$ stackings with an intralayer and interlayer antiferromagnetic configuration, but this bending is reduced for the smallest widths. With increasing ribbon width, the structural deformation at the edge is larger, leading to a reduction of the edge magnetic moments and the metallization of the [*b*]{}-ZGNRs. A magnetic external field can modify the structural changes, flattening the ribbons. We have also studied the effect of a graphene substrate. In this case Bernal stacking is more favorable, and the bilayer ribbons maintain their magnetization for larger widths. This is due to the reduction of the structural deformation because of the graphene substrate.
Acknowledgments
===============
This work has been partially supported by the Spanish DGES under Grants No. FIS2009-08744 and FIS2010-19609-C02-02, the Basque Departamento de Educación and the UPV/EHU (Grant No. IT-366-07), and the Nanoiker project (Grant No. IE11-304) under the ETORTEK program funded by the Basque Research Departament of Industry. L. C. and H. S. gratefully acknowledge helpful discussions with Juan Manuel Rodríguez Puerta and the hospitality of the Donostia International Physics Center. Calculations were partly done using the Camgrid high-throughput facility of the University of Cambridge. H. S. acknowledges the hospitality of the Department of Earth Sciences of the University of Cambridge.
[23]{} natexlab\#1[\#1]{}bibnamefont \#1[\#1]{}bibfnamefont \#1[\#1]{}citenamefont \#1[\#1]{}url \#1[`#1`]{}urlprefix\[2\][\#2]{} \[2\]\[\][[\#2](#2)]{} A. H. Castro Neto, F. Guinea, N. M. R. Peres, K. S. Novoselov, and A. K. Geim, Rev. Mod. Phys. [**81**]{}, 109 (2009). K. F. Mak, J. Shan, and Tony F. Heinz, Phys. Rev. Lett. [**104**]{}, 176404 (2010).
J.-C. Charlier, X. Gonze, and J.-P. Michenaud, Europhys. Lett., [**28**]{}, 403 (1994).
W. Norimatsu and M. Kusunoki, Phys. Rev. B [**81**]{}, 161410(R) (2010).
J.-K. Lee, S.-C Lee, J.-P. Ahn, S.-C. Kim, J. I. B. Wilson, and P. John, J. Chem. Phys. [**129**]{}, 234709 (2008).
Z. Liu, K. Suenaga, P. J. F. Harris, and S. Iijima, Phys. Rev. Lett. [**102**]{}, 015501 (2009).
T. Ohta, A. Bostwick, T. Seyller, K. Horn, and R. Rotenberg, Science [**313**]{}, 951 (2006).
J. W. González, H. Santos, M. Pacheco, L. Chico, and L. Brey, Phys. Rev. B [**81**]{}, 195406 (2010). J. W. González, H. Santos, E. Prada, L. Brey, and L. Chico, Phys. Rev. B [**83**]{}, 205402 (2011).
Y. W. Son, M. L. Cohen, and S. G. Louie, Nature (London) [**444**]{}, 347 (2006). L. Brey, and H. A. Fertig, Phys. Rev. B [**73**]{}, 235411 (2006).
A. R. Akhmerov and C. W. J. Beenakker, Phys. Rev. B [**77**]{}, 085423 (2008).
W. Jaskólski, A. Ayuela, M. Pelc, H. Santos, and L. Chico, Phys. Rev. B [**83**]{}, 235424 (2011).
H. Santos, A. Ayuela, W. Jaskólski, M. Pelc, and L. Chico, Phys. Rev. B [**80**]{}, 035436 (2009).
K. Nakada, M. Fujita, G. Dresselhaus, and M.S. Dresselhaus, Phys. Rev. B [**54**]{}, 17954 (1996). Y.-W. Son, M. L. Cohen, and S. G. Louie, Phys. Rev. Lett. [**97**]{}, 216803 (2006). M. Ezawa, Phys. Rev. B [**73**]{}, 045432 (2006).
V. Barone, O. Hod and G. E. Scuseria, Nano Lett. [**6**]{}, 2748 (2006).
F. Muñoz-Rojas, J. Fernández-Rossier, and J. J. Palacios, Phys. Rev. Lett. [**102**]{}, 136810 (2009). H. Santos, L. Chico, and L. Brey, Phys. Rev. Lett. [**103**]{}, 086801 (2009).
E. McCann and V. I. Fal’ko, Phys Rev. Lett. [**96**]{}, 086805 (2006). E. McCann, Phys. Rev. B [**74**]{}, 161403(R) (2006). J. Nilsson, A. H. Castro Neto, N. M. R. Peres, and F. Guinea, Phys. Rev. B [**73**]{}, 214418 (2006).
B. Sahu, H. Min, A. H. MacDonald, and S. K. Banerjee, Phys. Rev. B, [**78**]{}, 045404 (2008).
M. P. Lima, A. Fazzio, and A. J. R. da Silva, Phys. Rev. B, [**79**]{}, 153401 (2009).
G. Kim, and S.-H. Jhi, Appl. Phys. Lett. [**97**]{}, 263114 (2010).
K. Lee, E. D. Murray, L. Kong, B. I. Lundqvist, and D. C. Langreth, Phys. Rev. B. [**82**]{}, 081101(R) (2010).
M. Dion, H. Rydberg, E. Schroder, D. C. Langreth, and B. I. Lundqvist, Phys. Rev. Lett. [**92**]{}, 246401 (2004).
G. Román-Pérez and J. M. Soler, Phys. Rev. Lett. [**103**]{}, 096102 (2009).
J. M. Soler, E. Artacho, J. D. Gale, A. García, J. Junquera, P. Ordejón, and D. Sánchez-Portal, J. Phys.: Cond. Matt. [**14**]{}, 2745 (2002).
N. Troullier and J. L. Martins, Phys. Rev. B [**43**]{}, 1993 (1991).
H. J. Monkhorst and J. D. Park, Phys. Rev. B [**13**]{}, 5188 (1976).
N. Marom, J. Bernstein, J. Garel, A. Tkatchenko, E. Joselevich, L. Kronik, and O. Hod, Phys. Rev. Lett. [**105**]{}, 046801 (2010).
The binding energy is defined as $E_{\rm BE}= E_{\rm total} - 2 E_{\rm ML}$, where $E_{\rm total}$ is the total energy of the [*b*]{}-GNR and $E_{\rm ML}$ is the total energy of an individual layer in the most stable magnetic configuration, i. e., [*afm*]{}.
S. Grimme, C. M[ü]{}ck-Lichtenfeld, and J. Antony, J. Phys. Chem. C [**111**]{}, 11199 (2007). R. Podeszwa, J. Chem. Phys. [**132**]{}, 044704 (2010).
L. A. Girifalco and R. A. Lad, J. Chem. Phys. [**25**]{}, 693 (1956). R. Zacharia, H. Ulbricht, and T. Hertel, Phys. Rev. B [**69**]{}, 155406 (2004).
A. Mañanes, F. Duque, A. Ayuela, M. J. López, and J. A. Alonso, Phys. Rev. B [**78**]{}, 035432 (2008).
J. Heyd and G. E. Scuseria, J. Chem. Phys. [**121**]{}, 1187 (2004).
J. Heyd, G. E. Scuseria and M. Ernzerhof. J. Chem. Phys. [**118**]{}, 8207 (2003).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We study the family of $\alpha$-connections of Amari-Chentsov on the homogeneous space $\mathcal{D}(M)/\mathcal{D}_\mu(M)$ of diffeomorphisms modulo volume-preserving diffeomorphims of a compact manifold $M$. We show that in some cases their geodesic equations yield completely integrable Hamiltonian systems.'
address:
- 'J.L.: Department of Mathematics, Baylor University, Waco, TX 76798, USA'
- 'G.M.: Department of Mathematics, University of Notre Dame, IN 46556, USA'
author:
- Jonatan Lenells
- Gerard Misiołek
bibliography:
- 'is.bib'
title: 'Amari-Chentsov connections and their geodesics on homogeneous spaces of diffeomorphism groups'
---
[: 53C21, 58D05, 62B10.]{}
[: Diffeomorphism groups, geometric statistics, geodesics, integrable systems.]{}
Introduction
============
In a recent paper [@klmp] several key notions of geometric statistics were developed on diffeomorphism groups and their quotient spaces equipped with right-invariant metrics. For example, the metric defined by (the homogeneous part of) the Sobolev $H^1$ inner product of vector fields on the underlying compact manifold was shown to induce on the quotient of the diffeomorphism group by its subgroup of volume-preserving diffeomorphisms an infinite-dimensional analogue of the Fisher-Rao metric while geodesics of its Levi-Civita connection - when the underlying manifold is the circle - were shown to be related to solutions of a well-known one-dimensional completely integrable equation [@km; @le]. Furthermore, the authors also described analogues of the so-called $\alpha$-connections introduced in geometric statistics by Chentsov [@ch] and Amari [@AN] and pointed out integrability of another geodesic equation corresponding to $\alpha=-1$.
Our goals in this paper are to provide a proof of the construction in [@klmp] of Amari-Chentsov connections for circle diffeomorphisms (Theorem \[thm1\]), to generalize this construction to diffeomorphism groups of higher-dimensional manifolds (Theorem \[thm3\]) and finally, as a by-product, to show integrability of the geodesic equations corresponding to $\alpha=1$ (Theorem \[affineremark\], Corollary \[affineremarkndim\]).
Diffeomorphism groups and the Fisher-Rao metric
-----------------------------------------------
Let $M$ be a compact Riemannian manifold without boundary. Let $\mathcal{D}(M)$ denote the group of smooth diffeomorphisms of $M$ and let $\mathcal{D}_\mu(M)$ be its subgroup of diffeomorphisms preserving the volume form $\mu$ on $M$. It is well-known that the completion of $\mathcal{D}(M)$ (resp. $\mathcal{D}_\mu(M)$) in the $H^s$ Sobolev norm with $s>n/2 +1$ can be equipped with the structure of a smooth Hilbert manifold whose tangent space at the identity diffeomorphism $e$ consists of all $H^s$ vector fields (resp. all divergence-free $H^s$ vector fields) on $M$. However, for what follows it will be sufficient to work with the smooth category. It will also be convenient to normalize the Riemannian volume $\mu(M)=1$. For any $\eta \in \mathcal{D}(M)$ and any $V, W \in T_\eta\mathcal{D}(M)$ we set $$\begin{aligned}
\label{H1met}
\langle V, W \rangle_{\dot{H}^1}
=
\frac{1}{4} \int_M \mathrm{div}\, v \, \mathrm{div}\, w \, d\mu \end{aligned}$$ where $V=v\circ\eta$ and $W=w\circ\eta$ with $v, w \in T_e\mathcal{D}(M)$ to obtain a right-invariant (degenerate) $\dot{H}^1$ metric on $\mathcal{D}(M)$.
The geometry of this metric turns out to be particularly remarkable. The homogeneous space $\mathcal{D}(M)/\mathcal{D}_\mu(M)$ can be naturally identified with the set of smooth probability densities, i.e., smooth functions $\rho >0$ on $M$ satisfying the condition $\int_M \rho \, d\mu = 1$. If $N$ is an $n$-dimensional submanifold of such densities $\rho = \rho_{t_1\dots t_n}$ parameterized by $(t_1, \dots , t_n) \in \mathbb{R}^n$ then recall that the Fisher-Rao metric on $N$ is given by the formula $$\begin{aligned}
\label{FRmet}
g_{ij} = \int_M \frac{\partial\log\rho}{\partial t_i} \frac{\partial\log\rho}{\partial t_j} \rho\, d\mu
\quad\qquad
1 \leq i,j \leq n. \end{aligned}$$ It turns out that the right-invariant $\dot{H}^1$ metric defined by on $\mathcal{D}(M)$ descends to a (nondegenerate) metric on $\mathcal{D}(M)/\mathcal{D}_\mu(M)$ and the map $\eta \mapsto \sqrt{\mathrm{Jac}_\mu\eta}$ defines an isometry between $\mathcal{D}(M)/\mathcal{D}_\mu(M)$ and a subset of the unit sphere in $L^2(M,d\mu)$ with the canonical round metric. Furthermore, the restriction of the $\dot{H}^1$ metric to any finite-dimensional submanifold $N$ of $\mathcal{D}(M)/\mathcal{D}_\mu(M)$ coincides with the Fisher-Rao metric on $N$ while its Riemannian distance is the spherical Hellinger distance between probability densities on $M$. Proofs of these statements can be found in Section 3 of [@klmp].
Thus, in the framework of diffeomorphism groups, information geometry associated with the Fisher-Rao metric and its spherical Hellinger distance can be viewed as an $\dot{H}^1$ analogue of standard optimal transport associated with the metric[^1] on $\mathcal{D}(M)/\mathcal{D}_\mu(M)$ induced by the (non-invariant) $L^2$ metric on $\mathcal{D}(M)$ and whose Riemannian distance is the celebrated Kantorovich (or Wasserstein) distance, cf. [@v].
Divergence functions and dual connections
-----------------------------------------
Recall that a divergence on an $n$-dimensional manifold $N$ is a smooth function $
D: N \times N \to \mathbb{R}
$ satisfying $
D(p \| q) \geq 0
$ with equality if and only if $p=q$ and such that the matrix $g^D_{ij}$ defined in a chart at $p \in N$ by $$\label{eq:Dmetric}
g^D_{ij}(p)
=
-\tfrac{\partial}{\partial p_i}\tfrac{\partial}{\partial q_j} D(p \| q)\vert_{p=q}
\qquad\qquad
1 \leq i,j \leq n$$ is strictly positive definite for every $p \in N$. Equation (\[eq:Dmetric\]) defines a Riemannian metric on $N$ with covariant derivative determined by $$\label{eq:Dconn}
\Gamma^D_{ij,k}
=
- \tfrac{\partial}{\partial p_i} \tfrac{\partial}{\partial p_j} \tfrac{\partial}{\partial q_k} D(p \| q)\vert_{p=q}
\qquad\qquad
1 \leq i,j,k \leq n.$$ In what follows we shall consider on $\mathcal{D}(M)\times\mathcal{D}(M)$ the functions $$\begin{aligned}
\label{Dalpha}
&D^{(\alpha)}(\xi \| \eta)
=
\frac{1}{1 - \alpha^2} \left( 1 - \int_{M}
( \mathrm{Jac}_\mu \xi )^{\frac{1-\alpha}{2}} ( \mathrm{Jac}_\mu \eta )^{\frac{1+\alpha}{2}}
d\mu\right),
\qquad
-1 < \alpha < 1
\\ \label{Dalpha1}
&D^{(-1)}(\xi \| \eta) = D^{(1)}(\eta\|\xi)
=
\frac{1}{4} \int_M \big( \log\mathrm{Jac}_\mu\xi - \log\mathrm{Jac}_\mu\eta \big) \mathrm{Jac}_\mu\xi \, d\mu. \end{aligned}$$ These functions are well-defined on the homogeneous space $\mathcal{D}(M)/\mathcal{D}_\mu(M)$ and satisfy $D^{(\alpha)}(\xi\|\eta) \geq 0$ with equality if and only if $\xi$ and $\eta$ project onto the same probability density on $M$.
Furthermore, recall that two affine connections $\nabla$ and $\nabla^*$ on a Riemannian manifold $N$ are called dual (or conjugate) relative to the Riemannian metric if for any vector fields $U$, $V$ and $W$ they satisfy $
W \langle U, V \rangle
=
\langle \nabla_W U, V \rangle + \langle U, \nabla^*_W V \rangle.
$ Basic facts about dual connections and divergences can be found in [@AN] or [@kv].
The one-dimensional case: $\mathcal{D}(S^1)/{\text{\upshape Rot}}(S^1)$ {#section2}
=======================================================================
We first consider the case when the underlying manifold is the circle $S^1 = \mathbb{R}/\mathbb{Z}$. In this case $\mathcal{D}_\mu(S^1)$ is the space of rigid rotations ${\text{\upshape Rot}}(S^1) \simeq S^1$. It will be convenient to identify the homogeneous space $\mathcal{D}(S^1)/{\text{\upshape Rot}}(S^1)$ with the subgroup of circle diffeomorphisms which fix a prescribed point, for example with $\{ \eta \in \mathcal{D}(S^1): \eta(0)=0 \}$. Its tangent space at the identity can be identified with the space of smooth periodic functions vanishing at $0$. Furthermore, given any such function $u(x)$ we define the operator $$\label{A}
A^{-1}u(x) = - \int_0^x \int_0^y u(z) \, dz dy + x \int_0^1 \int_0^y u(z) \, dz dy$$ i.e., the inverse of $A = -\partial_x^2$.
The following is a reformulation of Theorem 6.2 stated (without proof) in [@klmp]. Our first objective is to provide a proof of this result.
\[thm1\]
1. Each divergence $D^{(\alpha)}$ induces on $\mathcal{D}(S^1)/{\text{\upshape Rot}}(S^1)$ the $\dot{H}^1$ metric and an affine connection $\nabla^{(\alpha)}$ whose Christoffel symbols are given by $$\label{Gammaformula}
\Gamma^{(\alpha)}_\eta (W, V)
=
-\frac{1+\alpha}{2} \Big\{ A^{-1}\partial_x \big[ (V\circ\eta^{-1})_x (W\circ\eta^{-1})_x \big]\Bigr\} \circ\eta, \quad -1 \leq \alpha \leq 1.$$
2. For any $\alpha$ the connections $\nabla^{(\alpha)}$ and $\nabla^{(-\alpha)}$ are dual with respect to the $\dot{H}^1$ metric and $\nabla^{(0)}$ is the self-dual Levi-Civita connection.
3. The geodesic equation of $\nabla^{(\alpha)}$ on $\mathcal{D}(S^1)/{\text{\upshape Rot}}(S^1)$ is the generalized Proudman-Johnson equation $$\label{PJ}
u_{txx} + (2-\alpha)u_x u_{xx} + u u_{xxx} = 0.$$ In particular, $\alpha =0$ yields the completely integrable Hunter-Saxton equation $$\label{HS}
u_{txx} + 2u_x u_{xx} + u u_{xxx} =0$$ and $\alpha =-1$ yields the completely integrable $\mu$-Burgers equation $$\label{muB}
u_{txx} + 3u_x u_{xx} + u u_{xxx} = 0.$$
The equation corresponding to $\alpha = 1$ also turns out to be integrable and its solutions can be given explicitly, see Theorem \[affineremark\] below.
The metrics induced by $D^{(\alpha)}$ and their connections can be calculated essentially as in finite dimensions using formulas and . We first assume that $\alpha \neq \pm 1$. Given any tangent vectors $V$ and $W$ at $\eta \in \mathcal{D}(S^1)$ let $\eta(s, t)$ be a two-parameter family of diffeomorphisms in $\mathcal{D}(S^1)$ such that $\eta(0,0) = \eta$ with $\frac{\partial}{\partial s} \eta(0,0) = V$ and $\frac{\partial}{\partial t} \eta (0,0) = W$. From (\[eq:Dmetric\]) and we have $$\begin{aligned}
\nonumber
\langle V, W \rangle_{\alpha}
&=
- \tfrac{\partial}{\partial s}\big|_{s = 0} \tfrac{\partial}{\partial t}\big|_{t= 0}
D^{(\alpha)}(\eta(s,0) \| \eta(0,t) )
\\ \label{alphametric}
&=
\frac{1}{1 - \alpha^2} \,
\tfrac{\partial}{\partial s}\big|_{s = 0} \tfrac{\partial}{\partial t}\big|_{t= 0}
\int_{S^1} \eta_x(s,0)^{\frac{1-\alpha}{2}} \eta_x(0,t)^{\frac{1+\alpha}{2}} dx
\\ \nonumber
&=
\frac{1}{4} \int_{S^1} V_x W_x \, \eta_x^{\frac{-1-\alpha}{2}} \eta_x^{\frac{-1+\alpha}{2}} dx
=
\frac{1}{4} \int_{S^1} \frac{V_x W_x}{\eta_x} \, dx
\\ \nonumber
&=
\langle V, W \rangle_{\dot{H}^1}. \end{aligned}$$ Now suppose that $W$ is a vector field on $\mathcal{D}(S^1)$ defined in a neighborhood of $\eta$. Let $\eta(s,t,r)$ be a three-parameter family of diffeomorphisms such that $\eta(0,0,0)=\eta$ with $\frac{\partial}{\partial s}\eta(0,0,0)=V$, $\frac{\partial}{\partial t}\eta(s,0,0)= W_{\eta(s,0,0)}$ for all sufficiently small $s$, and $\frac{\partial}{\partial r}\eta(0,0,0)=Z$. It is clear that such a map $\eta(s,t,r)$ exists. From the formulas , and we have $$\begin{aligned}
\nonumber
\langle \nabla_V^{(\alpha)} &W, Z \rangle_{\alpha}
=
\frac{1}{4}\int_{S^1} \frac{(\nabla_V^{(\alpha)} W)_x Z_x}{\eta_x} \, dx
\\ \nonumber
& =
-\tfrac{\partial}{\partial s}\big|_{s = 0}
\tfrac{\partial}{\partial t}\big|_{t = 0}
\tfrac{\partial}{\partial r}\big|_{r = 0}
D^{(\alpha)} \big( \eta(s, t, 0) \| \eta(0, 0, r) \big)
\\ \nonumber
&=
\frac{1}{1 - \alpha^2}
\tfrac{\partial}{\partial s}\big|_{s= 0}
\tfrac{\partial}{\partial t}\big|_{t = 0}
\tfrac{\partial}{\partial r}\big|_{r = 0}
\int_{S^1} \eta_x(s,t,0)^{\frac{1-\alpha}{2}} \eta_x(0,0,r)^{\frac{1+\alpha}{2}} dx
\\ \label{alphaconn}
&=
\frac{1}{4} \,
\tfrac{\partial}{\partial s}\big|_{s = 0}
\int_{S^1} W_x(\eta(s, 0,0)) \eta_x(s, 0,0)^{\frac{-1-\alpha}{2}} Z_x \eta_x^{\frac{-1+\alpha}{2}} dx
\\ \nonumber
&=
\frac{1}{4}
\int_{S^1} (DW \cdot V)_x Z_x \eta_x^{-1} dx
-
\frac{1+\alpha}{8} \int_{S^1} W_x V_x \eta_x^{\frac{-3-\alpha}{2}} Z_x \eta_x^{\frac{-1+\alpha}{2}} dx
\\ \nonumber
&=
\frac{1}{4} \int_{S^1}
\Big\{ ((DW \cdot V \big) \circ \eta^{-1} )_x
-
\frac{1+\alpha}{2} (V\circ\eta^{-1})_x (W\circ\eta^{-1})_x \Big\}
(Z\circ\eta^{-1})_x \, dx \end{aligned}$$ and integrating by parts and using the fact that $Z$ is arbitrary we find that $$\begin{aligned}
(\nabla^{(\alpha)}_V W)_\eta
=
(DW \cdot V)(\eta) - \Gamma^{\alpha}_\eta (W, V) \end{aligned}$$ where the Christoffel map is given by .
We will use the same notation for calculations in the remaining cases $\alpha= \pm 1$. From and we have $$\begin{aligned}
\nonumber
\langle V, W \rangle_{-1}
&=
\langle V, W \rangle_{1}
=
-\tfrac{\partial}{\partial s}\big|_{s = 0}
\tfrac{\partial}{\partial t}\big|_{t = 0}
D^{(-1)}( \eta(s, 0) \| \eta(0, t) )
\\ \label{alphaPMmetric}
&=
-\frac{1}{4}
\tfrac{\partial}{\partial s}\big|_{s = 0}
\tfrac{\partial}{\partial t}\big|_{t = 0}
\int_{S^1} \eta_x(s, 0) \big( \log \eta_x(s, 0) - \log \eta_x(0, t) \big) dx
\\ \nonumber
&=
\frac{1}{4}
\tfrac{\partial}{\partial s}\big|_{s = 0}
\int_{S^1} \eta_x(s, 0) \frac{W_x}{\eta_x} \, dx
=
\frac{1}{4}
\int_{S^1} \frac{V_xW_x}{\eta_x} dx
\\ \nonumber
&=
\langle V, W \rangle_{\dot{H}^1}. \end{aligned}$$ The corresponding affine connections can be obtained as in using , and . When $\alpha = -1$ we have $$\begin{aligned}
\int_{S^1} \frac{ (\nabla^{(-1)}_V W)_x Z_x }{ \eta_x } \, dx
&=
-\tfrac{\partial}{\partial s}\big|_{s = 0}
\tfrac{\partial}{\partial t}\big|_{t = 0}
\tfrac{\partial}{\partial r}\big|_{r = 0}
\int_{S^1} \eta_x (s, t, 0) \log\frac{\eta_x(s,t,0)}{\eta_x(0,0,r)} \, dx
\\
&=
\tfrac{\partial}{\partial s}\big|_{s = 0}
\tfrac{\partial}{\partial t}\big|_{t = 0}
\int_{S^1} \eta_x(s, t, 0) \frac{Z_x}{\eta_x} \, dx
\\
& =
\tfrac{\partial}{\partial s}\big|_{s = 0}
\int_{S^1} W_x(s, 0, 0) \frac{Z_x}{\eta_x} \,dx
\\
&=
\int_{S^1} ( DW \cdot V)_x \frac{Z_x}{\eta_x} \, dx \end{aligned}$$ from which we deduce that $$\label{flat-1}
\Gamma^{(-1)}_\eta (W, V) = 0.$$ When $\alpha =1$ an analogous calculation gives $$\label{flat1}
\Gamma^{(1)}_\eta (W, V)
=
- A^{-1}\partial_x \left\{ (V \circ \eta^{-1})_x (W \circ \eta^{-1})_x ) \right\} \circ \eta.$$ This establishes the first part of the theorem.
For the second part we need to verify that for any vector fields $X$, $Y$ and $Z$ on $\mathcal{D}(S^1)/{\text{\upshape Rot}}(S^1)$ we have $$\begin{aligned}
\label{eq:duality}
X \langle Y, Z \rangle_{\dot{H}^1}
=
\langle \nabla^{(\alpha)}_X Y, Z \rangle_{\dot{H}^1} + \langle Y, \nabla^{(-\alpha)}_X Z \rangle_{\dot{H}^1}. \end{aligned}$$ This can be done either by a direct calculation as above or else it can be deduced from general properties of divergences of the type and which are discussed in Chapter 3 of [@AN]. The fact that $\nabla^{(0)}$ is the Levi-Civita connection of the $\dot{H}^1$ metric follows at once from .
The equation for geodesics of $\nabla^{(\alpha)}$ on $\mathcal{D}(S^1)/{\text{\upshape Rot}}(S^1)$ has the form $$\begin{aligned}
\frac{d^2\gamma}{dt^2} = \Gamma^{(\alpha)}_\gamma \Big( \frac{d\gamma}{dt}, \frac{d\gamma}{dt} \Big). \end{aligned}$$ Setting $d\gamma/dt = u \circ \gamma$ defines a time-dependent vector field $u$ on the circle $S^1$ (i.e., a periodic function vanishing at $x=0$). Differentiating this relation with respect to $t$ and eliminating the first and second derivatives of $\gamma$ from the geodesic equation gives $$(u_t + uu_x)\circ\gamma
=
\Gamma^{(\alpha)}_\gamma(u \circ \gamma, u \circ \gamma).$$ Using and composing both sides with $\gamma^{-1}$ we obtain a nonlinear pseudodifferential equation $$u_t + uu_x
=
- \frac{1 + \alpha}{2}A^{-1} \partial_x (u_x^2)$$ which we can rewrite as a nonlinear PDE $$-u_{txx} - 3u_xu_{xx} - uu_{xxx} = - (1 + \alpha)u_xu_{xx}$$ yielding .
The Hunter-Saxton equation can be alternatively derived by observing that it is is the Euler-Arnold equation of $\nabla^{(0)}$ on $T_e\mathcal{D}(S^1)/{\text{\upshape Rot}}(S^1)$ and as such it is obtained from the geodesic equation of the right-invariant $\dot{H}^1$ metric by a standard reduction procedure, see [@km].
Another form of the Proudman-Johnson equation can be obtained by integrating in the $x$ variable $$u_{tx} + uu_{xx} + \frac{1 - \alpha}{2} \, u_x^2 = C(t)$$ where $C(t) = -\frac{1 + \alpha}{2} \int_{S^1} u_x^2 \, dx$. Observe that $C(t)$ is a conserved integral of the equation when $\alpha = 0$.
Using the Christoffel symbols it is possible to calculate the curvature of the $\alpha$-connections. It turns out to be proportional to the curvature of the $\dot{H}^1$ metric, i.e. for any vector fields $X$, $Y$ and $Z$ on $\mathcal{D}(S^1)/{\text{\upshape Rot}}(S^1)$ we have $$\label{Ralpha}
R^{(\alpha)}(X,Y)Z
=
(1-\alpha^2) \Big( X \langle Y, Z \rangle_{\dot{H}^1} + Y\langle X,Z\rangle_{\dot{H}^1} \Big).$$ This formula can be computed as in finite dimensions; see [@mc] where a different choice of parameters is made.
It turns out that the geodesic equation corresponding to with $\alpha = 1$ can be integrated as well, albeit indirectly, by constructing affine coordinates for $\nabla^{(1)}$. Observe that from we already know that the connections $\nabla^{(-1)}$ and $\nabla^{(1)}$ are flat. In the former case this is also evident from .
\[affineremark\] The geodesic equation $$\label{(1)}
u_{txx} + u_xu_{xx} + uu_{xxx} = 0$$ of $\nabla^{(1)}$ is integrable. Its general solution is given by $$\label{formulas}
u = \frac{d\eta}{dt} \circ\eta^{-1}
\quad
\mbox{where}
\quad
\eta(t,x) = \frac{\int_0^x e^{a(y)t + b(y)} dy}{\int_{S^1} e^{a(x)t + b(x)} dx}$$ and where $a, b$ are smooth mean-zero functions on $S^1$.
We will construct a chart on $\mathcal{D}(S^1)/{\text{\upshape Rot}}(S^1)$ in which the Christoffel symbols of $\nabla^{(1)}$ vanish identically. Consider the map $$\label{fi}
\eta \mapsto \phi(\eta) = \log \eta_x - \int_{S^1} \log \eta_x \, dx$$ from $\mathcal{D}(S^1)/{\text{\upshape Rot}}(S^1)$ to the space of smooth periodic mean-zero functions. To see how the Christoffel symbols transform under $\eta \mapsto \tilde\eta = \phi(\eta)$ we first compute the derivatives $$D_\eta\phi (W)
=
\frac{W_x}{\eta_x} - \int_{S^1} \frac{W_x}{\eta_x} \, dx,
\qquad
D^2_\eta \phi (W, V) = -\frac{V_x W_x}{\eta_x^2} + \int_{S^1} \frac{V_x W_x}{\eta_x^2} \, dx$$ for any $V, W \in T_\eta \mathcal{D}(S^1)/{\text{\upshape Rot}}(S^1)$. Next, from and we obtain $$\tilde{\Gamma}^{(1)}_{\phi(\eta)} \big( D_\eta\phi (W), D_\eta\phi (V) \big)
=
D^2_\eta\phi (W, V) + D_\eta\phi \big( \Gamma^{(1)}_\eta (W, V) \big)$$ $$\begin{aligned}
&=
-\frac{ V_x W_x }{\eta_x^2}
+
\int_{S^1} \frac{ V_x W_x }{\eta_x^2} \, dx
-
\frac{ ( A^{-1} (v_x w_x)_x \circ \eta )_x }{\eta_x}
+
\int_{S^1} \frac{ (A^{-1} (v_x w_x)_x \circ \eta)_x }{\eta_x} \, dx
\\
&=
- (v_x w_x)\circ\eta
+
\int_{S^1} (v_x w_x)\circ \eta \, dx
-
(A^{-1} (v_x w_x)_x)_x \circ \eta
+
\int_{S^1} (A^{-1} (v_x w_x)_x)_x \circ \eta \, dx
\\
&=
- (v_x w_x)\circ\eta + \int_{S^1} (v_x w_x)\circ \eta \, dx
-
\Big( - v_x w_x + \int_{S^1} v_x w_x \, dx \Big) \circ \eta
\\
&\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad
+
\int_{S^1} \Big( - v_x w_x + \int_{S^1} v_x w_x dx \Big) \circ \eta \, dx
= 0,\end{aligned}$$ where $v = V\circ\eta^{-1}$ and $w = W\circ\eta^{-1}$.
We can now solve as follows. Since $\tilde\Gamma^{(1)} \equiv 0$ all geodesics of $\nabla^{(1)}$ in the affine coordinates are the straight lines which can be written as $$t \to \tilde{\eta}(t,x) = a(x) t + b(x)
\qquad\qquad\qquad
x \in S^1$$ for some smooth mean-zero periodic functions $a$ and $b$. Thus, given any such functions to construct a general solution $u$ it is sufficient to (i) invert the map $\phi$ in to obtain the flow $t \to \eta(t) = \phi^{-1}\tilde{\eta}(t)$ and (ii) right-translate the velocity vector of $\eta(t)$ to the tangent space at the identity in $\mathcal{D}(S^1)/{\text{\upshape Rot}}(S^1)$.[^2] The required formulas are those in .
It is worth pointing out that the above proof manifests integrability of (\[(1)\]) in that it provides an explicit change of coordinates that linearizes the flow in the same spirit as the inverse scattering transform formalism.
The $n$-dimensional case: $\mathcal{D}(M)/\mathcal{D}_\mu(M)$
=============================================================
We now turn to the general case when $M$ is an $n$-dimensional compact Riemannian manifold without boundary and work with the coset space $\mathcal{D}(M)/\mathcal{D}_\mu(M)$. Our next result is stated in analogy with Theorem \[thm1\].
\[thm3\]
1. Each divergence $D^{(\alpha)}$ induces on the quotient $\mathcal{D}(M)/\mathcal{D}_\mu(M)$ the $\dot{H}^1$ metric and an affine connection $\nabla^{(\alpha)}$ given for a right-invariant vector field $W_\eta = w \circ \eta$ and a tangent vector $V = v \circ \eta$ by $$\begin{aligned}
\label{ndimnabla}
( \nabla^{(\alpha)}_V W)_\eta
=
-\left\{ \Delta^{-1}d \Bigl(
d\operatorname{div}w \cdot v + \frac{1-\alpha}{2} \operatorname{div}{w} \operatorname{div}{v}
\Bigr) \right\}^\sharp \circ \eta \end{aligned}$$ where $\alpha \in [-1, 1]$ and $\Delta = d\delta + \delta d$ denotes the Laplace-de Rham operator.
2. For any $\alpha$ the connections $\nabla^{(\alpha)}$ and $\nabla^{(-\alpha)}$ are dual with respect to the $\dot{H}^1$ metric and $\nabla^{(0)}$ is the Levi-Civita connection.
3. The geodesic equation of $\nabla^{(\alpha)}$ on $\mathcal{D}(M)/\mathcal{D}_\mu(M)$ is equivalent to the following nonlinear PDE $$\begin{aligned}
\label{PJn}
d\varphi_t + d\iota_{\displaystyle u} d\varphi + (1-\alpha) \varphi \, d\varphi = 0,
\qquad
\varphi = \mathrm{div}\, u. \end{aligned}$$
Despite the fact that we are now working with cosets the computations involved in the proof of statements (i) and (ii) are similar to the one-dimensional case in Section \[section2\]. We will prove (i) for $\alpha \in (-1,1)$; the proofs for $\alpha = \pm 1$ are analogous and will be omitted.
Let $W$ be a right-invariant vector field on $\mathcal{D}(M)$ defined in a neighborhood of $\eta \in \mathcal{D}(M)$. Given $V, Z \in T_\eta \mathcal{D}(M)$ let $\eta(s,t,r)$ be a three-parameter family of diffeomorphisms such that $\eta(0,0,0)=\eta$ with $\frac{\partial}{\partial s}\eta(0,0,0)=V$, $\frac{\partial}{\partial t}\eta(s,0,0)= W_{\eta(s,0,0)}$ for all sufficiently small $s$ and $\frac{\partial}{\partial r}\eta(0,0,0)=Z$. Using the identity $$\begin{aligned}
\tfrac{\partial}{\partial t}\big|_{t = 0} {\text{\upshape Jac}}_\mu\eta(s,t,0)
& = \operatorname{div}(W_{\eta(s,0,0)} \circ \eta^{-1}(s,0,0)) \circ \eta(s,0,0) \, {\text{\upshape Jac}}_\mu\eta(s,0,0)
\\
& = \operatorname{div}w \circ \eta(s,0,0) \, {\text{\upshape Jac}}_\mu\eta(s,0,0),\end{aligned}$$ where $W_\eta = w \circ \eta$ and similar computations for the partial derivatives in $s$ and $r$ we have $$\begin{aligned}
\nonumber
\langle & \nabla_V^{(\alpha)} W, Z \rangle_{\alpha}
=
\frac{1}{4}\int_{M} \operatorname{div}\bigl((\nabla_V^{(\alpha)} W) \circ \eta^{-1}\bigr) \operatorname{div}(Z \circ \eta^{-1}) \, d\mu
\\ \nonumber
&=
-\tfrac{\partial}{\partial s}\big|_{s = 0}
\tfrac{\partial}{\partial t}\big|_{t = 0}
\tfrac{\partial}{\partial r}\big|_{r = 0}
D^{(\alpha)} \big( \eta(s, t, 0) \| \eta(0, 0, r) \big)
\\ \nonumber
&=
\frac{1}{1 - \alpha^2}
\tfrac{\partial}{\partial s}\big|_{s= 0}
\tfrac{\partial}{\partial t}\big|_{t = 0}
\tfrac{\partial}{\partial r}\big|_{r = 0}
\int_{M} ({\text{\upshape Jac}}_\mu\eta(s,t,0))^{\frac{1-\alpha}{2}} ({\text{\upshape Jac}}_\mu \eta(0,0,r))^{\frac{1+\alpha}{2}} d\mu
\\ \nonumber
&=
\frac{1}{4} \,
\tfrac{\partial}{\partial s}\big|_{s = 0}
\int_{M} \operatorname{div}w \circ \eta(s,0,0) \, ({\text{\upshape Jac}}_\mu \eta(s,0,0))^{\frac{1-\alpha}{2}}
\operatorname{div}(Z \circ \eta^{-1}) \circ\eta \, ({\text{\upshape Jac}}_\mu \eta)^{\frac{1+\alpha}{2}} d\mu
\\ \nonumber
&=
\frac{1}{4}
\int_M \Bigl\{ d\operatorname{div}w \circ \eta \cdot V
\\ \nonumber
&\hspace{2cm}+ \frac{1-\alpha}{2} \operatorname{div}w \circ\eta
\operatorname{div}(V\circ\eta^{-1})\circ\eta\Bigr\} \operatorname{div}(Z\circ\eta^{-1})\circ\eta \, {\text{\upshape Jac}}_\mu\eta \, d\mu
\\ \nonumber
&=
\frac{1}{4}
\int_M \Bigl\{ d\operatorname{div}w \cdot v
+
\frac{1-\alpha}{2} \operatorname{div}{w} \operatorname{div}{v}\Bigr\} \operatorname{div}(Z\circ\eta^{-1}) \, d\mu,\end{aligned}$$ where $V = v \circ \eta$. Using integration by parts $$\int_M f \operatorname{div}X \, d\mu = - \int_M f \delta X^\flat d\mu = - \int_M df (X) \, d\mu$$ and the fact that $Z$ is arbitrary, we find $$\begin{aligned}
d\operatorname{div}\bigl\{(\nabla_V^{(\alpha)} W) \circ \eta^{-1}\bigr\}
=
df \quad \text{i.e.} \quad \operatorname{div}\bigl\{(\nabla_V^{(\alpha)} W) \circ \eta^{-1}\bigr\}
=
f - \int_M f d\mu,\end{aligned}$$ where $f = d\operatorname{div}w \cdot v + \frac{1-\alpha}{2} \operatorname{div}{w} \operatorname{div}{v}$. The expression in (\[ndimnabla\]) now follows from the next lemma.
Let $\Delta$ be the Laplace-de Rham operator. Then $$\operatorname{div}(-(\Delta^{-1} df)^\sharp) = f - \int_M f d\mu, \qquad f \in C^\infty(M).$$
If $g = \operatorname{div}\bigl(-(\Delta^{-1} df)^\sharp\bigr) = \delta \Delta^{-1} df$ then $\Delta g = \delta df = \Delta f$ and since the kernel of $\Delta$ acting on functions consists of the constants, we deduce that $g = f + c$ for some constant $c$. Integrating over $M$ yields $0 = \int_M f d\mu + c$, which determines $c$.
Calculation of $\nabla_V^{(\alpha)} W$ for an arbitrary (not necessarily right-invariant) vector field $W$ can be reduced to the right-invariant case as follows. Let $\eta \in \mathcal{D}(M)$ and let $W^R$ be the right-invariant vector field with $W^R_\eta = W_\eta$. Then $$\begin{aligned}
\label{generalnabladef}
( \nabla_V^{(\alpha)} W )_\eta
=
(\nabla_V^{(\alpha)} (W - W^R))_\eta + ( \nabla_V^{(\alpha)} W^R )_\eta
=
(\mathcal{L}_V(W - W^R))_\eta + ( \nabla_V^{(\alpha)} W^R )_\eta,\end{aligned}$$ where $\mathcal{L}_V$ denotes the Lie derivative in the direction of any vector field $\tilde{V}$ such that $\tilde{V}_\eta = V$. Indeed, if $Z$ is a vector fields satisfying $Z_\eta = 0$, then locally $$(\nabla_V Z)_\eta = DZ(\eta) \cdot V - \Gamma_\eta(Z_\eta, V) = DZ(\eta) \cdot V = [V, Z]_\eta = (\mathcal{L}_V Z)_\eta.$$ Using (\[ndimnabla\]) and (\[generalnabladef\]) we can compute the affine connection $\nabla^{(\alpha)}$ for arbitrary fields on $\mathcal{D}(M)$ and this completes the proof of (i).
As in the proof of Theorem \[thm1\], (ii) can be established by a direct calculation or, alternatively, it can be deduced from the general properties of divergences of the type and .
Regarding (iii) let $\gamma(t)$ be a curve in $\mathcal{D}(M)$ with $\gamma(0) = e$ and $\dot{\gamma}= d\gamma/dt = u \circ \gamma$. In the appendix we prove that implies $$\begin{aligned}
\label{nablarelation}
\nabla_{\dot{\gamma}}^{(\alpha)}\dot{\gamma}
=
u_t \circ \gamma + \nabla_{\dot{\gamma}}^{(\alpha)} \dot{\gamma}^R.\end{aligned}$$ Using (\[ndimnabla\]) and (\[nablarelation\]) the geodesic equation $\nabla_{\dot{\gamma}}^{(\alpha)} \dot{\gamma} = 0$ can be written as $$u_t
=
(\Delta^{-1}df)^\sharp
\quad
\mbox{where}
\quad
f = d\operatorname{div}u \cdot u + \frac{1-\alpha}{2} (\operatorname{div}{u})^2$$ so that setting $\varphi = \mathrm{div}\, u$ and taking the exterior derivative we get $$-d\varphi_t + \delta d u_t^\flat
=
d\iota_ud\varphi + (1-\alpha)\varphi d\varphi.$$ The relation $u_t^\flat = \Delta^{-1} df$ implies $\Delta d u_t^\flat = d \Delta u_t^\flat = 0$ hence, in particular, $\delta du_t^\flat = 0$. This proves (\[PJn\]).
Finally, we turn to integrability of the geodesic equations in which to the best of our knowledge has not been studied in the literature before, with the exception of the case $\alpha =0$ in [@klmp].
\[affineremarkndim\] The geodesic equations of $\nabla^{(\alpha)}$ on $\mathcal{D}(M)/\mathcal{D}_\mu(M)$ corresponding to with $\alpha =0, \pm 1$ are integrable equations in any dimension $n$.
Equation (\[PJn\]) with $\alpha =0$ was derived and shown to be integrable in [@klmp]. Since $\nabla^{(1)}$ and $\nabla^{(-1)}$ are flat, integrability of the other two equations can be established similarly to their one-dimensional analogues in Section \[section2\]. We will consider the more complicated case $\alpha = 1$ with the corresponding equation $$\label{(1ndim)}
d\varphi_t + d\iota_u d\varphi = 0,
\qquad
\varphi = \operatorname{div}{u}.$$ Proceeding as in the proof of Theorem \[affineremark\] we first show that the map $$\eta \mapsto \phi(\eta) = \log {\text{\upshape Jac}}_\mu\eta - \int_{M} \log {\text{\upshape Jac}}_\mu \eta \, d\mu$$ from $\mathcal{D}(M)/\mathcal{D}_\mu(M)$ to the space of smooth mean-zero functions on $M$ defines an affine chart for $\nabla^{(1)}$. Thus, as before, the geodesics of $\nabla^{(1)}$ in the affine coordinates defined by $\phi$ are the straight lines $$t \to \tilde{\eta}(t,x) = a(x) t + b(x)
\qquad\qquad\quad
x \in M$$ where $a$ and $b$ are smooth mean-zero functions on $M$. From this we find $${\text{\upshape Jac}}_\mu \eta(t,x) = \frac{e^{a(x) t + b(x)}}{\int_M e^{a(x) t + b(x)} d\mu}
\qquad\qquad
x \in M$$ and combining this expression with the identity $
\frac{d}{dt}{\text{\upshape Jac}}_\mu(\eta) = (\varphi \circ \eta) {\text{\upshape Jac}}_\mu\eta,
$ we obtain $$\label{eq:sol}
\varphi(t, \eta(t,x))
=
a(x) - \frac{\int_M a(x) e^{a(x) t + b(x)} d\mu}{\int_M e^{a(x) t + b(x)} d\mu}.$$ Note that the time derivative of $\varphi \circ \eta$ is independent of $x$, so that $$d\varphi_t + d\iota_u d\varphi
=
d\bigl((\varphi \circ \eta)_t \circ \eta^{-1}\bigr)
= 0$$ which shows that solves the equations in .
Proof of (\[nablarelation\]) {#app}
============================
Let $\gamma(t)$ be a curve in $\mathcal{D}(M)$ with $\dot{\gamma}(t) = u(t) \circ \gamma(t)$. For any $t_0 >0$ we compute $$\begin{aligned}
\nonumber
u_t(t_0) \circ \gamma(t_0) & = \biggl\{\frac{d}{dt}\bigg|_{t = t_0} \left(\dot{\gamma}(t) \circ \gamma(t)^{-1}\right)\biggr\} \circ \gamma(t_0)
\\\label{A1}
& = \ddot{\gamma}(t_0) + D(\dot{\gamma}(t_0)) \cdot \biggl\{\frac{d}{dt}\bigg|_{t=t_0} \gamma(t)^{-1}\biggr\}\circ \gamma(t_0).\end{aligned}$$ But differentiating the relation $\gamma(t)^{-1} \circ \gamma(t) = id$ with respect to $t$ we find $$\biggl\{\frac{d}{dt}\bigg|_{t=t_0} \gamma(t)^{-1}\biggr\}\circ \gamma(t_0)
+ D(\gamma(t_0)^{-1})\circ\gamma(t_0) \cdot \dot{\gamma}(t_0) = 0.$$ Thus, (\[A1\]) yields $$\begin{aligned}
u_t(t_0) \circ \gamma(t_0)
& = \ddot{\gamma}(t_0) - D(\dot{\gamma}(t_0)) \cdot D(\gamma(t_0)^{-1})\circ\gamma(t_0) \cdot \dot{\gamma}(t_0)
\\
& = \frac{d}{dt}\bigg|_{t = t_0} \left\{\dot{\gamma}(t) - \dot{\gamma}(t_0) \circ \gamma(t_0)^{-1} \circ \gamma(t)\right\}
= D(\dot{\gamma} - \dot{\gamma}^R) \cdot \dot{\gamma}(t_0).\end{aligned}$$ Equation (\[nablarelation\]) now follows from (\[generalnabladef\]).
[9999]{} S.-I. Amari and H. Nagaoka, *Methods of Information Geometry*, American Mathematical Society, Providence, RI 2000.
V. I. Arnold, Sur la géometrie différentielle des groupes de Lie de dimension infinie et ses application à l’hydrodynamique des fluides parfaits, [*Ann. Inst. Fourier (Grenoble)*]{} [**16**]{} (1966), 319–361.
N. N. Chentsov, *Statistical Decision Rules and Optimal Inference*, American Mathematical Society, Providence, RI 1982.
R. Kass and P. Vos, *Geometrical Foundations of Asymptotic Inference*, Wiley-Interscience, New York 1997.
B. Khesin, J. Lenells, G. Misiołek and S. Preston, *Geometry of diffeomorphism groups, complete integrability and geometric statistics*, to appear in Geom. funct. anal. (2012).
B. Khesin and G. Misiołek, *Euler equations on homogeneous spaces and Virasoro orbits*, Adv. Math. **176** (2003), 116–144.
J. Lenells, *The Hunter-Saxton equation describes the geodesic flow on a sphere*, J. Geom. Phys. **57** (2007), 2049–2064.
E. A. Morozova and N. N. Chentsov, *Natural geometry of families of probability laws* (in Russian), Itogi Nauki i Tekhniki, Akad. Nauk SSSR, **83** (1991), 133–265.
C. Villani, *Optimal Transport: Old and New*, Springer, New York 2009.
[^1]: This metric is sometimes called Otto’s metric.
[^2]: In the language of fluid dynamics the second step corresponds to going from Lagrangian to Eulerian coordinates.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Building on a model recently proposed by F. Calogero, we postulate the existence of a universal Keplerian tremor for any stable classical system. Deriving the characteristic unit of action $\alpha$ for each classical interaction, we obtain in all cases $\alpha \cong h$, the Planck action constant, suggesting that quantum corrections to classical dynamics can be simulated through a fluctuative hypothesis of purely classical origin.'
---
Semiclassical Aspects of Quantum Mechanics\
by Classical Fluctuations
\
Salvatore De Martino$^{*}$ [^1], Silvio De Siena$^{*}$ [^2], and Fabrizio Illuminati$^{*}$ $^{a)}$ [^3]\
\* [ *Dipartimento di Fisica, Università di Salerno;\
INFM, Unità di Salerno; and INFN, Sezione\
di Napoli – Gruppo Collegato di Salerno,\
I–84081 Baronissi (Salerno), Italy*]{}\
$a)$ [*Fakultät für Physik, Universität Konstanz,\
D–78457 Konstanz, Germany*]{}
21 January 1998
PACS Numbers: 03.65.Bz; 05.45.+b; 05.40.+j
In a recent paper F. Calogero has put forward an intriguing conjecture on the possible gravitational origin of quantization [@calogero].
The scheme followed by Calogero is simple but appealing. He suggests that the origin of quantization be attributed to the universal interaction of every particle with the background gravitational force due to all other particles in the Universe. Such background interaction generates a chaotic component in the motion of each single particle, with a characteristic constant $\tau$ measuring the time scale of stochasticity (Zitterbewegung).
Assuming a basic granularity of the Universe, made up of nucleons (or hydrogen atoms) of mass $m$, Calogero derives an expression for Planck’s constant, $h \cong G^{1/2}m^{3/2}R^{1/2}$, with $G$ the Newtonian gravitational constant and $R$ the observed Radius of the Universe. This formula, which connects the fundamental constant of quantum theory with the fundamental gravitational constants, was already known since some time [@weinberg] but its meaning and implications were so far unexplained; Calogero provided the first derivation of it from a mechanical model, so that one could very well name it the Calogero–Weinberg formula.
Now, the crucial point in the procedure carried out by Calogero is that the characteristic time $\tau$ of the stochastic motion per particle, being associated to a collective chaotic effect, should be inversely proportional to the square root of $N$, the total number of particles in the Universe [@calogero]: $$\tau \cong N^{-1/2}{\cal{T}} \, ,$$
with ${\cal{T}}$ the characteristic global time unit associated with a Universe of total mass $M$. Defining the energy per particle $\epsilon \cong E/N$, with $E$ total energy of the Universe, and a global unit of action for the Universe $A=E{\cal{T}}$, Calogero defines the unit of action per particle $$\alpha = \epsilon \tau \cong N^{-3/2}A \, .$$
Replacing $N$ with the ratio of the global and the granular amount of sources $M/m$, imposing that $\alpha$ be independent of extensive quantities and performing an elementary dimensional analysis for the combination of the nucleon mass $m$, the Radius of the Universe $R$ and the Newtonian gravitational constant $G$, Calogero finally arrives at the expression $$\alpha \cong G^{1/2}m^{3/2}R^{1/2} \, ;$$
inserting the numerical values $m \cong 10^{-27} kg$, $G \cong 10^{-11}
kg^{-1} \cdot m^{3} \cdot s^{-2}$ and the most updated cosmological estimate for the observed Radius of the Universe $R \cong 10^{30}m$ [@peebles], eq.(3) yields $\alpha \cong h$, the Planck action constant (we warn the reader that in the presente work we are neglecting in the numerical computations all those factors that do not substantially affect the order of magnitude of the estimated quantities).
In conclusion, Calogero suggests that quantization might be explained via classical gravitation by assuming the existence of a chaotic component of the individual particles’ motion due to a “universal coherent tremor” associated to the extremely large number $N$ of the elementary components making up the Universe. This conjecture is mathematically implemented introducing Maxwell–Boltzmann fluctuations proportional to $1/\sqrt{N}$.
In the present letter we take a closer look at the model provided by Calogero. We first show that his procedure can be equivalently reformulated by replacing the fluctuative law eq.(1) with a fractal space–time relation $l \sim \tau^{2/3}$ which is the mathematical expression of a “universal Keplerian tremor”.
We then move on to show that the scheme of Calogero can be applied to all the other known interactions (electromagnetic, strong, etc.) and that it leads in all cases to a variety of formulas again linking Planck’s constant with the proper fundamental constants associated to each considered interaction.
In this way we achieve two purposes: on the one hand we clarify that the fluctuative hypothesis of Calogero actually holds also for systems with few degrees of freedom, since in our Keplerian reformulation it does not involve the number $N$ of elementary constituents. On the other hand, we show that the mechanism is universal, in the sense that it allows to derive an expression for Planck’s constant for any physical system confined on the typical space–time region associated to any of the fundamental interactions known in Nature.
In particular, one can apply the Keplerian formulation of the Calogero model to “Gedanken Universes” made of an arbitrary number $N$ of gravitationally interacting particles, even with $N=2$. In any instance, one always obtains $\hbar$ as the unit of action and the observed Radius of our actual Universe as the typical length scale of the system, irrespective of the assumed number $N$ of elementary constituents. Therefore the Keplerian fluctuation and its relationship with the elementary quantum of action $\hbar$ seem to have little or nothing at all to do with a chaotic dynamics induced by the enormous number $N$ of elementary constituents that interact gravitationally in the actually observed Universe.
We can then draw the following conclusions:
1\) The mechanism envisaged by Calogero seems to capture some essential aspects of the interplay between classical and quantum mechanics.
2\) The universality of the mechanism as shown in the present letter and its insensitivity to $N$ being large or small strongly suggest that the simple and appealing interpretation of Calogero is untenable: quantum mechanics is in fact the most fundamental theory known up to now, and the derivation of Planck’s constant for each known interaction by this procedure rules out the possibility of a privileged role of classical gravitation as the “origin” of quantum mechanics.
3\) Still it is remarkable that a simple qualitative argument (the universal Keplerian tremor) allows to capture some relevant quantum features by purely classical considerations.
4\) From the point of view of quantum mechanics we can then interpret Calogero’s and our results as a first embryonal, still qualitative, step towards the possibility of recovering semiclassical aspects of quantum mechanics starting directly from classical mechanics and implementing (simulating) quantum corrections in terms of suitable, purely classical, stochastic fluctuations. This is at variance with the usual approximation schemes of quantum mechanics, such as the WKBJ procedure, which recover the semiclassical and classical domains starting from the “deep” quantum domain. It is rather a first operative definition of a semi–quantal approximation scheme, hopefully to be further developed.
We begin by showing that the tremor hypothesis of Calogero, eq. (1) is equivalent to a generalization on the microscopic scale of Kepler third law in the form $l \sim \tau^{2/3}$, where, by introducing the total volume of the Universe $V$ and the mean allowed volume per particle (specific volume) $v_{s}\cong V/N$, we have defined the mean free path of the individual constituents $l \cong v_{s}^{1/3}$. In fact, we can immediately rewrite Calogero’s fundamental relation eq. (1) as: $$\frac{{\cal{T}}^{2}}{V} \equiv
const. \cong \frac{\tau^{2}}{v_{s}}
\sim \frac{\tau^{2}}{l^{3}} \, ,$$
or, equivalently, $$l \sim \tau^{2/3} \, .$$
Note that, from $V \sim R^{3}$, the first member of eq. (4) is Kepler third law on the scale of the Universe.
One should note that the most updated cosmological scenarios [@peebles] lead to a recessing away law of galaxies in the expanding Universe of the form $L \sim t^{2/3}$ (with $L$ the distance between galaxies). The congruence of this phenomenon on large cosmological scales with our Keplerian version eq. (5) of Calogero’s tremor hypothesis eq. (1) implies the extension of validity of a Kepler–like third law for gravitational interactions ranging from small to large scales. The resulting picture clearly ignores the structure of the system in its finest details, being based on a sort of mean field description.
At this point we might draw the first provisional conclusion: if one assumes, as usual, that quantum mechanics is the fundamental theory, then the above analysis implies that the stability of the Universe on the scale of its observed Radius $R$ is ruled, via the Calogero mechanism as formulated in eq. (5), by the Planck quantum of action, in complete analogy with the stable confined systems associated to the other known interactions.
This interpretation is of course totally at variance with the one originally given by Calogero, and to see whether it is tenable, we should move on to apply his model to the other stable systems on different scales that are associated to the other relevant known interactions beyond gravity.
We expect in this way to obtain again formulas linking Planck’s action constant to the typical radius $R$ of stability and to the fundamental masses and interaction constants associated to the systems being considered.
As we will show below, it turns out that this is indeed the case. This fact seems to imply that a simple qualitative picture based on purely classical tools allows to recover some fundamental aspects of quantum mechanics in its semiclassical domain in connection with the stability of matter.
We thus proceed to apply the scheme already adopted for gravitation to aggregates of charged particles interacting electromagnetically and to systems of confined quarks in the nucleons. We do this by assuming the tremor hypothesis in the form of eq.(5) to hold for any stable aggregate of particles, relying on the fact that for these confined systems one can certainly introduce well defined characteristic global units of time ${\cal{T}}$ and volume $V$.
[*Electromagnetic Interactions*]{}. Let us consider first the case of a stable aggregate of charged particles interacting via electromagnetic forces. The fundamental constants involved are the electrostatic constant $K \equiv 1/4\pi\epsilon_{0} \cong 10^{10}
N\cdot m^{2}\cdot C^{-2}$, the elementary unit of charge $e \cong 10^{-19}C$ and the velocity of light $c \cong 10^{8} m \cdot
s^{-1}$.
Since such aggregates are in general made of collections of electrons and protons, we can take as the natural unit of elementary mass the reduced mass $\mu$, which substantially coincides with the mass of the electron $m \cong 10^{-30} kg$. Let us consider also the characteristic linear dimension $R$ of the stable aggregate; this characteristic global scale of length can vary from $R \cong 10^{-2} m$ (macroscopic dimensions) to $R \cong 10^{-10} m$ (atomic dimensions).
Expressing $N$ as the ratio of the global and the granular amount of sources $Q/e$, with $Q$ total charge of the aggregate, imposing eq. (2) and requiring the independence of the unit of action $\alpha$ on extensive quantities, we obtain $A=Q^{3/2}{\tilde A}$; by dimensional considerations ${\tilde A} = f(K,m,c,R)$, and we finally arrive at $$\alpha \cong e^{3/2}K^{3/4}m^{1/4}c^{-1/2}
R^{1/4} \, .$$
Inserting numbers in eq. (6) we have then in all cases, up to at most one order of magnitude, $\alpha \cong 10^{-34} J \cdot s \cong h$, i.e., once more, Planck constant.
[*Quarks*]{}. We now move on to consider a hadron having as granular costituents a collection of bound quarks. The interaction we consider is the “string law" described by the typical confining potential $V = kr$ with the strength constant $k$ varying in the range $k \cong 0.1 GeV \cdot fm^{-1}
\div 10 GeV \cdot fm^{-1}$ (values compatible with the experimental bounds [@perkins]). Let us also introduce the quark masses $m \cong 0.01 GeV \cdot c^{-2}
\div 10 GeV \cdot
c^{-2}$ [@particledata], the velocity of light $c$ and the radius $R \cong 10^{-15} m$, which is the range of nuclear forces.
Expressing $N$ as $N = M/m$, $M$ total mass of the hadron, we obtain, following the usual procedure, $A=M^{3/2}{\tilde A}$ and, finally, $$\alpha \cong (mc^{2})^{3/2}c^{-1}k^{-1/2}
R^{1/2} \, .$$
Inserting numbers, we have again, up to at most one or two orders of magnitude, $\alpha \cong h$.
This numerical equivalence with Planck constant of the elementary unit of action per particle for any classical fundamental interaction on each scale seems very significant, and can hardly be thought of being casual. We again stress that one always obtains the same order of magnitude of $\alpha$ ($\cong h$) for any force law on its typical scale (universality of Planck constant).
Since the above procedure is of a grossly qualitative nature, it seems important to provide other consistency checks of the fundamental Keplerian tremor law eq. (5), to yield further support to its universal validity.
In fact, the natural question arises on how to determine the order of magnitude, for each particular system, of the characteristic time $\tau$. The latter must obviously depend both on the universal elementary unit of action $\alpha \cong h$ and on the details of the chosen aggregate.
It is well known that the typical time scale of quantum fluctuations is defined as the ratio between $h$ and a suitable energy describing the equilibrium state of the given system on its characteristic dimensions. This leads naturally to identify this energy with the thermal energy $k_{B}T$, with $k_{B}$ the Boltzmann constant and $T$ the absolute temperature.
On the other hand, in our model such time scale coincides with the fluctuative time $\tau$; we therefore write $$\tau \cong \frac{h}{k_{B}T} \, .$$
With the above definition we can rewrite the universal tremor hypothesis eq. (1) in the form: $$T \cong \frac{h}{k_{B}}
\cdot \frac{\sqrt{N}}{{\cal{T}}} \, .$$
We can adopt the point of view that the above equation be the definition of the absolute temperature for the system one is considering. This definition, as one can see, connects the temperature to the global and the granular length scales, and to the characteristic velocity associated to the given aggregate.
We now exploit this definition, applying it to some well established thermodynamic phenomena, as a further test of validity of our theoretical scheme.
[*a) Emittance associated to charged beams in particle accelerators*]{}. Among the stable aggregates of charged particles, a paradigmatic role is played by the charged beams in particle accelerators. Such systems are very interesting from our point of view because they are generally described in classical terms, since they exist on a length scale that ranges approximately from thirtytwo to thirtyfive orders of magnitude above the Planck scale.
However, our analysis shows that their stability, as in the gravitational case, is ruled by Planck’s constant, and it is thus ultimately of quantum rather than classical origin.
The bunch consists solely of charges of the same sign, and stability (confinement) can be achieved only through the action of an external focusing potential (magnetic field). Our analysis applies, for instance, by considering the reference frame comoving with the synchronous particle, yielding again $h$ as the unit of action per particle (note that replacing electrons with protons does not affect in a appreciable way the order of magnitude of $\alpha$ due to the $m^{1/4}$ dependence in eq. (6)).
The emittance ${\cal{E}}$ is a scale of length (or ,equivalently, of “temperature”) associated to charged beams, whose numerical value in units of the Compton length $\lambda_{c}=h/mc$ lies, for typical accelerators (for instance electron machines), in the range ${\cal{E}}
\cong 10^{6} \lambda_{c} \div 10^{9} \lambda_{c}$ [@fedele]. Following our scheme, we identify the characteristic unit of emittance as the characteristic action associated to the charged beam divided by $mc$. The characteristic action associated to charged beams is, in our framework, $(k_{B} T) {\cal{T}} $, and therefore, by eq. (9), the associated emittance is estimated as: $${\cal{E}} \cong
\lambda_{c} \sqrt{N} \, .$$
We note that the above expression connects, at least in the leading semiclassical order, the characteristic emittance with the number of particles in a nontrivial way. Since in a typical bunch $N \cong 10^{11} \div
10^{12}$ [@particledata], we finally obtain ${\cal{E}} \cong 10^{6} \lambda_{c}
\cong 10^{-6} m$ in good agreement with the phenomenological order of magnitude estimated by other theoretical methods [@fedele].
Therefore the Calogero model supplemented by eq.(8) naturally supports a quantum–like description of the dynamis of charged particle beams with the correct order of magnitude for the emittance.
[*b) Temperature of macroscopic systems*]{}. In this case we know that $N \cong 10^{23}
mol^{-1} \div 10^{24} mol^{-1}$ (Avogadro constant), and ${\cal{T}} \cong 10^{-2} s \div 10^{-3} s$, corresponding to a rms velocity $v_{T} \cong 10^{2} m \cdot s^{-1}
\div 10^{3} m \cdot s^{-1}$ for gases around room temperature. Inserting the numerical values of $h/k_{B}$ we obtain $T \cong 10^{2} K \div 10^{3} K$, as it should be.
[*c) Temperature of quarks inside nucleons*]{}. In this case $N \cong 1$. The typical energy scale $E$ of light quarks in a nucleon is of the order of $\Lambda_{QCD} \cong 0.1 GeV$ [@particledata], corresponding to a temperature $T = E/k_{B} \cong
10^{12} K$. With the characteristic velocity of the order of the velocity of light $c$, and the global scale of length $R \cong 10^{-15} m$, we have ${\cal{T}} \cong R/c \cong 10^{-23} s$. Inserting numbers into eq. (9) we obtain just $T \cong 10^{12} K$.
[*d) Bose–Einstein condensation*]{}. Recently, Bose–Einstein condensation has been experimentally observed in a gas of rubidium and sodium atoms [@condensation]. The condensate has linear dimension $R \cong
10^{-4} m$ at a temperature $T \cong 10^{-6} K$ and it contains $N \cong 10^{7}$ atoms. Letting ${\cal{T}} \cong R/v$, with $v$ the characteristic velocity, eq. (9) yields $$v \cong \frac{k_{B}}{h} \cdot \frac{T}{\sqrt{N}} \cdot
R \cong 10^{-3} m \cdot s^{-1} \div 10^{-2} m
\cdot s^{-1} \, .$$
The characteristic velocity thus is smaller by a factor of the order $10^{-6}$ compared to the rms velocity of the gas observed at room temperature, in agreement with the theoretical prediction of a macroscopic “zero” momentum.
Therefore, the definition of temperature derived from the universal Keplerian tremor hypothesis seems to be consistent. We then move on to apply it to other two significant cases.
[*e) Cosmic background radiation*]{}. In the framework of our hypothesis it seems quite reasonable to interpret the measured temperature associated to the cosmic background radiation, $T = 2.7 K$, as the characteristic “temperature of the Universe”. Consequently, we insert in eq. (9) a characteristic global time ${\cal{T}} \cong R/v $, with $R$ the Radius of the Universe and $v$ a characteristic velocity. This velocity cannot be defined unambigously, therefore we take it in the wide range that goes from $10^{5} m \cdot s^{-1}$ (the circular velocity of hydrogen clouds surrounding galaxies) up to the velocity of light.
We now exploit eq. (9) to determine the order of magnitude of $N$, the total number of particles in the Universe. Inserting numbers: $$N \cong 10^{66} \div 10^{72} \, ,$$
which, in our crudely qualitative framework, is compatible, within the error range, with the value $N \cong 10^{\nu}$, $\nu = 78 \pm 8$, estimated by cosmological arguments [@peebles].
Some conclusive remarks are now due. We first want to stress that, according to our point of view, it is impossible to discriminate the gravitational system, through the observed Radius of the Universe $R$, from the other systems (e.g. charged particles and quarks) by claiming that its characteristic dimensions are not [*a priori*]{} determined by quantum mechanics while the dimensions of the other ones are.
In fact, one should note that there are about $20$ orders of magnitude separating the smallest scale of length considered (that of the quarks confined in the nucleons) and the Planck scale representing the fundamental quantum mechanical scale of length. Furthermore, if one considers [**macroscopic**]{} aggregates like charged beams in particle accelerators, this difference reaches up to $35$ orders of magnitude.
It seems then obvious to us that if one accepts, as it is commonly asserted, that the influence of quantum mechanics should manifest itself all the way through this huge difference of length scales, it should as well manifest itself also on the scale of length of the Universe.
Therefore, the significant aspect that we single out in the scheme put forward by Calogero is that it allows for any dynamical system to obtain a quantum correction to classical dynamics starting from a fluctuative hypothesis of purely classical origin. This model seems then worth to be developed and improved, since it could be of great conceptual and computational relevance in the study of the interplay between classical and quantum domains, a fundamental issue in modern physics. We will report in a forthcoming paper how the application of these ideas can be made already now quantitative in the study of the quantum–like dynamics of charged particle beams in accelerators.
>From both the conceptual and the calculational point of view the next task would then be to develop the present qualitative model into a fully quantitative scheme of approximation (semi–quantal approximation scheme) that would allow, in principle, to determine all the higher–order quantum corrections, i.e. a systematic reconstruction of quantum fetures from classical dynamics beyond the leading semiclassical order.
[**Aknowledgements**]{}.
It is very hard for all of us to express in words our deep gratitude to Francesco Guerra for his invaluable teachings and for his profound and far reaching insights, during our many–years acquaintance, and in many enlightening discussions on all aspects of the present work. We also gratefully aknowledge very useful conversations with Francesco Calogero on his model and on an early draft of the present paper.
One of us (F.I.) aknowledges the Alexander von Humboldt–Stiftung for financial support and the Fakültät für Physik der Universität Konstanz for hospitality while on leave of absence from the Dipartimento di Fisica dell’Università di Salerno.
[99]{}
F. Calogero, Phys. Lett. [**A 228**]{}, 335 (1997).
S. Weinberg, [*Gravitation and Cosmology: Principles and Applications of the General Theory of Relativity*]{} (Wiley, New York, 1972).
P. J. Peebles, [*Principles of Physical Cosmology*]{} (Princeton University Press, Princeton N. J., 1993).
D. H. Perkins, [*Introduction to High Energy Physics*]{} (Addison Wesley, Menlo Park, 1987).
R. Fedele, G. Miele, and L. Palumbo, Phys. Lett. [**A 194**]{}, 113 (1994).
Aa. Vv., [*Review of Particle Physics*]{}, Phys. Rev. [**D 54**]{}, 1 (1996).
M. H. Anderson, J. R. Ensher, M. R. Matthews, C. E. Wieman and E. A. Cornell, Science [**269**]{}, 198 (1995);
W. Ketterle, M. R. Andrews, K. B. Davis, D. S. Durfee, D. M. Kurn, M.–O. Mewes and N. J. van Druten, Physica Scripta [**T 66**]{}, 31 (1996).
[^1]: E–Mail: [email protected]
[^2]: E–Mail: [email protected]
[^3]: E–Mail: [email protected], [email protected]
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
The main conclusion of long-standing discussions concerning the role of solutions with degenerate metric ($g\equiv
det(g_{\mu\nu})=0$ and even with $g_{\mu\nu}=0$) was that in the first order formalism they are physically acceptable and must be included in the path integral. In particular, they may describe topology changes and reduction of “metrical dimension” of space-time. The latter implies disappearance of the volume element $\sqrt{-g}d^4x$ of a 4-D space-time in a neighborhood of the point with $g=0$. We pay attention that besides $\sqrt{-g}$, the 4-D space-time differentiable manifold possesses also a “manifold volume measure” (MVM) described by a 4-form which is sign indefinite and generically independent of the metric. The first order formalism proceeds with originally independent connection and metric structures of the space-time manifold. In this paper we bring up the question whether the first order formalism should be supplemented with degrees of freedom of the space-time differentiable manifold itself, e.g. by means of the MVM. It turns out that adding the MVM degrees of freedom to the action principle in the first order formalism one can realize very interesting dynamics. Such Two Measures Field Theory enables radically new approaches to resolution of the cosmological constant problem. We show that fine tuning free solutions describing a transition to $\Lambda =0$ state involve oscillations of $g_{\mu\nu}$ and MVM around zero. The latter can be treated as a dynamics involving changes of orientation of the space-time manifold. As we have shown earlier, in realistic scale invariant models (SIM), solutions formulated in the Einstein frame satisfy all existing tests of General Relativity (GR). Here we reveal surprisingly that in SIM, all ground state solutions with $\Lambda\neq 0$ appear to be degenerate either in $g_{00}$ or in MVM. Sign indefiniteness of MVM in a natural way yields a dynamical realization of a phantom cosmology ($w<-1$). It is very important that for all solutions, the metric tensor rewritten in the Einstein frame has regularity properties exactly as in GR. We discuss new physical effects which arise from this theory and in particular strong gravity effect in high energy physics experiments.
address: 'Physics Department, Ben Gurion University of the Negev, Beer Sheva 84105, Israel'
author:
- 'E. I. Guendelman [^1] and A. B. Kaganovich [^2]'
title: |
Transition to Zero Cosmological Constant and Phantom Dark Energy\
as Solutions Involving Change of Orientation of Space-Time Manifold
---
Introduction: Degenerate Metric, Manifold Volume Measure and Orientation of the Space-Time Manifold
===================================================================================================
Solutions with degenerate metric were a subject of a long-standing discussions starting probably with the paper by Einstein and Rosen[@Einstein]. In spite of some difficulty interpreting solutions with degenerate metric in classical theory of gravitation, the prevailing view was that they have physical meaning and must be included in the path integral[@Hawking1979],[@D'Auria-Regge],[@Tseytlin1982]. As it was shown in Refs.[@Hawking1979],[@Horowitz], in the first order formulation of an appropriately extended general relativity, solutions with $g(x)\equiv \det(g_{\mu\nu})=0$ can describe changes of the space-time topology. Similar idea is realized also in the Ashtekar’s variables[@Ashtekar-3-in-Jacobson],[@Jacobson-2-in-Jacobson]. There are known also classical solutions[@Dray-0]-[@Senovilla] with change of the signature of the metric tensor.
The space-time regions with $g(x)=0$ can be treated as having [*’metrical dimension’*]{} $D<4$ (using terminology by Tseytlin[@Tseytlin1982]).
The simplest solution with $g(x)=0$ is $g_{\mu\nu}=0$ while the affine connection is arbitrary (or, in the Einstein-Cartan formulation, the vierbein $e_{\mu}^a =0$ and $\omega_{\mu}^{ab}$ is arbitrary). Such solutions have been studied by D’Auria and Regge[@D'Auria-Regge], Tseytlin[@Tseytlin1982]), Witten[@Witten-15-and-16-in-Horowitz], Horowitz[@Horowitz], Giddings[@Giddings], Bañados[@Banados]; it has been suggested that $g_{\mu\nu}=0$ should be interpreted as essentially non-classical phase in which diffeomorphism invariance is unbroken and it is realized at high temperature and curvature.
Now we would like to bring up a question: whether the equality $g(x)=0$ really with a necessity means that the dimension of the space-time manifold in a small neighborhood of the point $x$ may become $D<4$? At first sight it should be so because the volume element is $$dV_{(metrical)}=\sqrt{-g}d^{4}x.
\label{dVg}$$ Note that the latter is the “metrical” volume element, and the possibility to describe the volume of the space-time manifold in this way appears after the 4-dimensional differentiable manifold $M_4$ is equipped with the metric structure. For a solution with $g_{\mu\nu} =0$, the situation with description of the space-time becomes even worse . However, in spite of lack of the metric, the manifold $M_4$ may still possess a nonzero volume element and have the dimension $D=4$. The well known way to realize it consists in the construction of a differential 4-form build for example by means of four differential 1-forms $d\varphi_a$, ($a=1,2,3,4$): $d\varphi_1\wedge d\varphi_2\wedge d\varphi_3\wedge d\varphi_4$. Each of the 1-forms $d\varphi_a$ may be defined by a scalar field $\varphi_a(x)$. The appropriate volume element of the 4-dimensional differentiable manifold $M_4$ can be represented in the following way $$dV_{(manifold)} = {4!}d\varphi_{1}\wedge d\varphi_{2}\wedge
d\varphi_{3}\wedge d\varphi_{4}\equiv \Phi d^{4}x \label{dV}$$ where $$\Phi \equiv \varepsilon_{abcd} \varepsilon^{\mu\nu\lambda\sigma}
(\partial_{\mu}\varphi_{a}) (\partial_{\nu}\varphi_{b})
(\partial_{\lambda}\varphi_{c}) (\partial_{\sigma}\varphi_{d}).
\label{Phi}$$ is the volume measure independent of $g_{\mu\nu}$ as opposed to the case of the metrical volume measure $\sqrt{-g}$. In order to emphasize the fact that the volume element (\[dV\]) is metric independent we will call it a [*manifold volume element*]{} and the measure $\Phi$ - a [*manifold volume measure*]{}.
If $\Phi(x)\neq 0$ one can think of four scalar fields $\varphi_a(x)$ as describing a homeomorphism of an open neighborhood of the point $x$ on the 4-dimensional Euclidean space $R^4$. However if one allows a dynamical mechanism of metrical dimensional reduction of the space-time by means of degeneracy of the metrical volume measure $\sqrt{-g}$, there is no reason to ignore a possibility of a similar effect permitting degenerate manifold volume measure $\Phi$. The possibility of such (or even stronger, with a sign change of $\Phi$) dynamical effect seems to be here more natural since [*the manifold volume measure $\Phi$ is sign indefinite*]{} (in Measure Theory, sign indefinite measures are known as signed measures[@signed]) . Note that the metrical and manifold volume measures are not obliged generically to be simultaneously nonzero.
The original idea to use differential forms as describing dynamical degrees of freedom of the space-time differentiable manifold has been developed by Taylor in his attempt[@Taylor] to quantize the gravity. Taylor argued that quantum mechanics is not compatible with a Riemannian metric space-time; moreover, in the quantum regime space-time is not even an affine manifold. Only in the classical limit the metric and connection emerge, that one allows then to construct a traditional space-time description. Of course, the transition to the classical limit is described in Ref.[@Taylor] rather in the form of a general prescription. Thereupon we would like to pay attention to the additional possibility which was ignored in Ref.[@Taylor]. Namely, in the classical limit not only the metric and connection emerge but also some of the differential forms could keep (or restore) certain dynamical effect in the classical limit. In such a case, the traditional space-time description may occur to be incomplete. [**Our key idea**]{} is that one of such lost differential forms, the 4-form (\[dV\]) survives in the classical limit as describing dynamical degrees of freedom of the volume measure of the space-time manifold, and hence can affect the gravity theory on the classical level too[^3].
If we add four scalar fields $\varphi_a(x)$ as new variables to a set of usual variables (like metric, connection and matter degrees of freedom) which undergo variations in the action principle[^4] then one can expect an effect of gravity and matter on the manifold volume measure $\Phi$ and vice versa. We will see later in this paper that in fact such effects exist and in particular classical cosmology solutions of a significant interest exist where $\Phi$ vanishes and changes sign.
As is well known, the 4-dimensional differentiable manifold is orientable if it possesses a differential form of degree 4 which is nonzero at every point on the manifold. Therefore two possible signs of the manifold volume measure (\[Phi\]) are associated with two possible orientations of the space-time manifold. The latter means that besides a dimensional reduction and topology changes on the level of the differentiable manifold, the incorporation of the manifold volume measure $\Phi$ allows to realize solutions describing [*dynamical change of the orientation of the space-time manifold*]{}.
In the light of existence of two volume measures, the simplest way to take into account this fact in the action principle consists in the modification of the action which should now consist of two terms, one with the usual measure $\sqrt{-g}$ and another - with the measure $\Phi$, $$S_{mod} = \int \left(\Phi L_1 +\sqrt{-g}L_2\right) d^{4}x,
\label{sec-2-S-modif}$$ where two Lagrangians $L_1$ and $L_2$ coupled with manifold and metrical volume measures appear respectively. According to our previous experience[@GK1]-[@GK10] in Two Measures Field Theory (TMT) we will proceed with an additional basic assumption that, at least on the classical level, the Lagrangians $L_1$ and $L_2$ are independent of the scalar fields $\varphi_a(x)$, i.e. the manifold volume measure degrees of freedom enter into TMT only through the manifold volume measure $\Phi$. In such a case, the action (\[sec-2-S-modif\]) possesses an infinite dimensional symmetry $$\varphi_{a}\rightarrow\varphi_{a}+f_{a}(L_{1}), \label{IDS}$$ where $f_{a}(L_{1})$ are arbitrary functions of $L_{1}$ (see details in Ref.[@GK3]). One can hope that this symmetry should prevent emergence of the scalar fields $\varphi_a(x)$ dependence in $L_{1}$ and $L_{2}$ after quantum effects are taken into account.
Note that Eq.(\[sec-2-S-modif\]) is just a convenient way for presentation of the theory in a general form. In concrete models studied in the present paper, we will see that the action (\[sec-2-S-modif\]) can be always rewritten in an equivalent form where [*each term in the action has its own total volume measure and the latter is a linear combination of $\Phi$ and $\sqrt{-g}$*]{}.
In the next section it will be shown that the space-time geometry described in terms of the original metric and connection of the underlying action (\[sec-2-S-modif\]) is not generically Riemannian. However by making use a change of variables to the Einstein frame one can represent the resulting equations of motion in the Riemannian (or Einstein-Cartan) space-time.
In our previous investigations we have shown that TMT enables radically new approaches to resolution of the cosmological constant (CC) problem[@GK3],[@G1],[@GK9] (for an alternative approach see Ref.[@Comelli]). Intrinsic features of TMT allow to realize a scalar field dark energy model where all dependence of the scalar field appears as a result of spontaneous breakdown of the dilatation symmetry. Solutions of this model formulated in the Einstein frame satisfy all existing tests of General Relativity (GR)[@GK6],[@GK7],[@GK10]. A new sort of dynamical protection from the initial singularity of the curvature becomes possible[@GK9]. It also allows us to realize a phantom dark energy in the late time universe without explicit introducing phantom scalar field[@GK9].
In contrast to all our previous investigations of TMT, [*the purpose of the present paper*]{} is to study the dynamics of the metric $g_{\mu\nu}$ and the manifold volume measure $\Phi$ (used in the underlying action (\[sec-2-S-modif\])) in a number of TMT models. The main attention is concentrated on the analysis of the amazing features of “irregularity” of $g_{\mu\nu}$ and $\Phi$ (involving change of orientation of the space-time manifold) in the course of transitions to a ground state and in the phantom dark energy. It is very important to note immediately that in the Einstein frame the metric tensor has regularity properties exactly as in GR. The organization of the paper is the following. In Sec. II we discuss general features of classical dynamics in TMT. In Sec. III we consider the pure gravity model. In Secs. IV and V, in the framework of a simple scalar field model I, we analyze in detail the behavior of two volume measures in the course of transition to the ground state with zero CC. In Sec. VI we study a (generically broken) intrinsic TMT symmetry which is restored in the ground states; the relation of this symmetry restoration to the old CC problem[@Weinberg1] is also analyzed; a discussion of this effect is continued in Sec.VIII. In Sec.VII we shortly present the scalar field model II with spontaneously broken global scale invariance[@G1] studied in detail in Ref.[@GK9]. In the framework of such class of models, an interesting dynamics of the metric and the manifold volume measure in the course of transition to ground states is analyzed in Sec.VIII. In section IX we reveal that a possibility to realize a phantom dark energy without explicit introducing a phantom scalar field (demonstrated in [@GK9]) has the origin in a sign indefiniteness of the manifold volume measure (\[Phi\]). Finally, since one cannot check directly whether a tiny/zero cosmological constant is fine-tuned or not, in Sec.X we discuss the new physical effects which arise from this theory and in particular a strong gravity effect in high energy physics experiments.
Classical equations of motion
=============================
Varying the measure fields $\varphi_{a}$, we get $B^{\mu}_{a}\partial_{\mu}L_{1}=0$ where $B^{\mu}_{a}=\varepsilon^{\mu\nu\alpha\beta}\varepsilon_{abcd}
\partial_{\nu}\varphi_{b}\partial_{\alpha}\varphi_{c}
\partial_{\beta}\varphi_{d}$. Since $Det (B^{\mu}_{a}) = \frac{4^{-4}}{4!}\Phi^{3}$ it follows that if $\Phi\neq 0$ the constraint $$L_{1}=sM^{4} =const.
\label{varphi}$$ must be satisfied, where $s=\pm 1$ and $M$ is a constant of integration with the dimension of mass. Variation of the metric $g^{\mu\nu}$ gives $$\zeta\frac{\partial L_1}{\partial g^{\mu\nu}}+\frac{\partial
L_2}{\partial g^{\mu\nu}}-\frac{1}{2}g_{\mu\nu}L_2 =0,
\label{g-mu-nu-varying}$$ where $$\zeta\equiv \frac{\Phi}{\sqrt{-g}}
\label{zeta}$$ is the scalar field build of the scalar densities $\Phi$ and $\sqrt{-g}$.
We study models with the Lagrangians of the form $$L_1=-\frac{1}{b_g\kappa}R(\Gamma, g)+L_1^m, \quad
L_2=-\frac{1}{\kappa}R(\Gamma, g)+L_2^m \label{L1L2}$$ where $\Gamma$ stands for affine connection, $R(\Gamma,
g)=g^{\mu\nu}R_{\mu\nu}(\Gamma)$, $R_{\mu\nu}(\Gamma)=R^{\lambda}_{\mu\nu\lambda}(\Gamma)$ and $R^{\lambda}_{\mu\nu\sigma}(\Gamma)\equiv \Gamma^{\lambda}_{\mu\nu
,\sigma}+ \Gamma^{\lambda}_{\alpha\sigma}\Gamma^{\alpha}_{\mu\nu}-
(\nu\leftrightarrow\sigma)$. Dimensionless factor $b_g^{-1}$ in front of $R(\Gamma, g)$ in $L_1$ appears because there is no reason for couplings of the scalar curvature to the measures $\Phi$ and $\sqrt{-g}$ to be equal. We choose $b_g>0$ and $\kappa
=16\pi G$, $G$ is the Newton constant. $L_1^m$ and $L_2^m$ are the matter Lagrangians which can include all possible terms used in regular (with only volume measure $\sqrt{-g}$) field theory models.
Since the measure $\Phi$ is sign indefinite, the total volume measure $(\Phi/b_g +\sqrt{-g})$ in the gravitational term $-\kappa^{-1}\int R(\Gamma, g)(\Phi/b_g +\sqrt{-g})d^4x$ is generically also sign indefinite.
Variation of the connection yields the equations we have solved earlier[@GK3]. The result is $$\Gamma^{\lambda}_{\mu\nu}=\{
^{\lambda}_{\mu\nu}\}+\frac{1}{2}(\delta^{\alpha}_{\mu}\sigma,_{\nu}
+\delta^{\alpha}_{\nu}\sigma,_{\mu}-
\sigma,_{\beta}g_{\mu\nu}g^{\alpha\beta}) \label{GAM2}$$ where $\{ ^{\lambda}_{\mu\nu}\}$ are the Christoffel’s connection coefficients of the metric $g_{\mu\nu}$ and $$\sigma,_{\mu}\equiv \frac{1}{\zeta
+b_g}\zeta,_{\mu},\label{sigma-mu}$$
If $\zeta\neq const.$ the covariant derivative of $g_{\mu\nu}$ with this connection is nonzero (nonmetricity) and consequently geometry of the space-time with the metric $g_{\mu\nu}$ is generically non-Riemannian. The gravity and matter field equations obtained by means of the first order formalism contain both $\zeta$ and its gradient as well. It turns out that at least at the classical level, the measure fields $\varphi_a$ affect the theory only through the scalar field $\zeta$.
For the class of models (\[L1L2\]), the consistency of the constraint (\[varphi\]) and the gravitational equations (\[g-mu-nu-varying\]) has the form of the following constraint $$(\zeta
-b_g)(sM^4-L_1^m)+g^{\mu\nu}\left(\zeta \frac{\partial
L_{1m}}{\partial g^{\mu\nu}}+\frac{\partial L_2^m}{\partial
g^{\mu\nu}}\right)-2L_2^m=0, \label{Constr-original}$$ which determines $\zeta(x)$ (up to the chosen value of the integration constant $sM^4$) as a local function of matter fields and metric. Note that the geometrical object $\zeta(x)$ does not have its own dynamical equation of motion and its space-time behavior is totally determined by the metric and matter fields dynamics via the constraint (\[Constr-original\]). Together with this, since $\zeta$ enters into all equations of motion, it generically has straightforward effects on dynamics of the matter and gravity through the forms of potentials, variable fermion masses and selfinteractions[@GK1]-[@GK10].
For understanding the structure of TMT it is important to note that TMT (where, as we suppose, the scalar fields $\varphi_a$ enter only via the measure $\Phi$) is a constrained dynamical system. In fact, the volume measure $\Phi$ depends only upon the first derivatives of fields $\varphi_a$ and this dependence is linear. The fields $\varphi_a$ do not have their own dynamical equations: they are auxiliary fields. All their dynamical effect is displayed only in the following two ways: a) in generating the constraint (\[varphi\]) (or (\[Constr-original\])); b) in the appearance of the scalar field $\zeta$ and its gradient in all equations of motion.
Pure Gravity TMT model
=======================
Let us start from the simplest TMT model with action (\[sec-2-S-modif\]) where $$L_1=-\frac{1}{b_g\kappa}R(\Gamma, g)-\Lambda_1, \quad
L_2=-\frac{1}{\kappa}R(\Gamma, g)-\Lambda_2 \label{L1L2pure}$$ and $\Lambda_1$, $\Lambda_2$ are constants. Note that $\Lambda_1
=const.$ cannot have a physical contribution to the field equations (in this model - only gravitational) because $\Phi
\Lambda_1$ is a total derivative. Nevertheless we keep $\Lambda_1$ to see explicitly how $\Lambda_1$ appears in the result. $\Lambda_2/2$ would have a sense of the cosmological constant in the regular, non TMT, theory (i.e. with the only measure $\sqrt{-g}$).
Following the procedure described in Sec.II we obtain the gravitational equations (\[g-mu-nu-varying\]) and the constraint (\[Constr-original\]) in the following form: $$R_{\mu\nu}(\Gamma)=\frac{\kappa}{2}\frac{b_g\Lambda_2}{\zeta
-b_g}g_{\mu\nu}
\label{Grav.eq.Pure}$$ $$\zeta=b_g-\frac{2\Lambda_2}{sM^4+\Lambda_1}=const.,
\label{Constr.Pure}$$ where $sM^4$ is the constant of integration that appears in Eq.(\[varphi\]) and we have assumed that the total volume measure in the gravitational term of the action is nonzero, that is $\Phi/b_g+\sqrt{-g}\neq 0$.
Since $\zeta =const.$ the connection $\Gamma^{\lambda}_{\mu\nu}$, Eq.(\[GAM2\]), coincides with the Christoffel’s connection coefficients of the metric $g_{\mu\nu}$. Therefore in the model under consideration, the space-time with the metric $g_{\mu\nu}$ is (pseudo) Riemannian. It follows from Eqs.(\[Grav.eq.Pure\]) and (\[Constr.Pure\]) the resulting Einstein equations $$G_{\mu\nu}(g)=\frac{\kappa}{2}\Lambda g_{\mu\nu}; \qquad \Lambda
=\frac{b_g}{2}(sM^4+\Lambda_1) \label{Ein.eq.Pure}$$
Constancy of $\zeta(x)$ on the mass shell, Eq.(\[Constr.Pure\]), means that for the described solution the manifold and metrical volume measures coincide up to a normalization factor. However, this is true only on the mass shell; if we were try to start from this assumption in the underlying action the resulting solution would be different completely.
The model possesses a few interesting features in what it concerns the CC:
\(1) The effective CC $\Lambda$ appears as a constant of integration (as we noticed above, the parameter $\Lambda_1$ has no a physical meaning and it can be absorbed by the constant of integration). The effective regular, non TMT, gravity theory provides the same equations if the cosmological constant is added explicitly.
\(2) The effective cosmological constant $\Lambda$ does not depend at all on the CC-like parameter $\Lambda_2$ (which should describe a total vacuum energy density including vacuum fluctuations of all matter fields). The latter resembles the situation in the unimodular theory[@unimodular-1],[@unimodular-2].
\(3) Note that $\Lambda$ becomes very small if the integration constant is chosen such that $sM^4+\Lambda_1$ is very small. The latter is equivalent to a solution with $\Phi/b_g\gg \sqrt{-g}$. In the limit where the metrical volume measure $\sqrt{-g}\to 0$ while the manifold volume measure $\Phi$ remains nonzero, we get $\Lambda\to 0$. Thus a $\Lambda =0$ state is realized for a solution which involves a reduction of the metrical dimension to $D^{(g)}<4$ and at the same time the dimension of the space-time as a differentiable manifold remains $D^{(m)}=4$.
\(4) In the limit where the free parameter $b_g\to \infty$, the gravitational term in the underlying action (Eqs.(\[sec-2-S-modif\]) and (\[L1L2pure\])) with coupling to the manifold measure $\Phi$ approaches zero; then TMT takes the form of a regular (non TMT) field theory, but the effective cosmological constant $\Lambda$ becomes infinite. If we wish to reach a very small value of $\Lambda$ keeping the integration constant arbitrary, one should take the opposite limit where $b_g\ll 1$. Then in the underlying action, the weight of the gravitational term with coupling to the manifold volume measure $\Phi$ increases with respect to the regular one with coupling to the metrical measure $\sqrt{-g}$.
The above speculations can be regarded as a strong indication that TMT possesses a potential for resolution of the CC problem. In the next sections we will study this issue in more realistic models.
Scalar Field Model I
====================
Let us now study a model including gravity as in Eqs.(\[L1L2\]) and a scalar field $\phi$. The action has the same structure as in Eq.(\[sec-2-S-modif\]) but it is more convenient to write down it in the following form $$S_{mod}^{(1)}=\frac{1}{b_g}\int d^4x \left[-\frac{1}{\kappa}(\Phi
+b_g\sqrt{-g})R(\Gamma,g)+(\Phi
+b_{\phi}\sqrt{-g})\frac{1}{2}g^{\mu\nu}\phi_{,\mu}\phi_{,\nu}
-\Phi V_1(\phi)-\sqrt{-g}\,V_2(\phi)\right]
\label{S-model-scalar.f.}$$ The appearance of the dimensionless factor $b_{\phi}$ is explained by the fact that without fine tuning it is impossible in general to provide the same coupling of the $\phi$ kinetic term to the measures $\Phi$ and $\sqrt{-g}$. $V_1(\phi)$ and $V_2(\phi)$ are potential-like functions; we will see below that the physical potential of the scalar $\phi$ is a complicated function of $V_1(\phi)$ and $V_2(\phi)$.
The constraint (\[Constr-original\]) reads now $$(\zeta
-b_g)[sM^4+V_1(\phi)]+2V_2(\phi)+b_g\frac{\delta}{2}g^{\alpha\beta}\phi_{,\alpha}\phi_{,\beta}
=0, \label{Scalar-f-Constr-original}$$ where $$\delta =\frac{b_g-b_{\phi}}{b_g} \label{delta}$$ Since $\zeta\neq const.$ the connection (\[GAM2\]) differs from the connection of the metric $g_{\mu\nu}$. Therefore the space-time with the metric $g_{\mu\nu}$ is non-Riemannian. To see the physical meaning of the model we perform a transition to a new metric $$\tilde{g}_{\mu\nu}=(\zeta +b_{g})g_{\mu\nu}, \label{gmunuEin}$$ where the connection $\Gamma^{\lambda}_{\mu\nu}$ becomes equal to the Christoffel connection coefficients of the metric $\tilde{g}_{\mu\nu}$ and the space-time turns into (pseudo) Riemannian. This is why the set of dynamical variables using the metric $\tilde{g}_{\mu\nu}$ we call the Einstein frame. One should point out that [*the transformation*]{} (\[gmunuEin\]) [*is not a conformal*]{} one since $(\zeta +b_{g})$ is sign indefinite. But $\tilde{g}_{\mu\nu}$ is a regular pseudo-Riemannian metric. For the action (\[S-model-scalar.f.\]), gravitational equations (\[g-mu-nu-varying\]) in the Einstein frame take canonical GR form with the same $\kappa =16\pi G$ $$G_{\mu\nu}(\tilde{g}_{\alpha\beta})=\frac{\kappa}{2}T_{\mu\nu}^{eff}
\label{gef}$$ Here $G_{\mu\nu}(\tilde{g}_{\alpha\beta})$ is the Einstein tensor in the Riemannian space-time with the metric $\tilde{g}_{\mu\nu}$ and the energy-momentum tensor reads $$\begin{aligned}
T_{\mu\nu}^{eff}&=&\frac{\zeta +b_{\phi}}{\zeta +b_{g}}
\left(\phi_{,\mu}\phi_{,\nu}- \tilde{g}_{\mu\nu}X\right)
-\tilde{g}_{\mu\nu}\frac{b_{g}-b_{\phi}}{(\zeta +b_{g})} X
+\tilde{g}_{\mu\nu}V_{eff}(\phi;\zeta,M)
\label{Tmn-2}\end{aligned}$$ where $$X\equiv\frac{1}{2}\tilde{g}^{\alpha\beta}\phi_{,\alpha}\phi_{,\beta}
\label{X}$$ and the function $V_{eff}(\phi;\zeta,M)$ is defined as following: $$V_{eff}(\phi;\zeta ,M)= \frac{b_g\left[sM^{4}+V_{1}(\phi)\right]
-V_{2}(\phi)}{(\zeta +b_{g})^{2}}. \label{Veff1}$$
The scalar $\phi$ field equation following from Eq.(\[S-model-scalar.f.\]) and rewritten in the Einstein frame reads $$\frac{1}{\sqrt{-\tilde{g}}}\partial_{\mu}\left[\frac{\zeta
+b_{\phi}}{\zeta
+b_{g}}\sqrt{-\tilde{g}}\tilde{g}^{\mu\nu}\partial_{\nu}\phi\right]
+\frac{\zeta V_1^{\prime}+ V_2^{\prime}}{(\zeta +b_{g})^2}=0
\label{phief}$$
The scalar field $\zeta$ in Eqs.(\[Tmn-2\])-(\[phief\]) is determined by means of the consistency equation (\[Constr-original\]) which in the Einstein frame (\[gmunuEin\]) takes the form $$(\zeta -b_{g})[sM^{4}+V_1(\phi)]+ 2V_2(\phi)+\delta\cdot
b_{g}(\zeta +b_{g})X =0.\label{constraint2-1}$$
Manifold Measure and Old Cosmological Constant Problem: Cosmological Dynamics with $|\Phi|/\sqrt{-g}\to \infty$
===============================================================================================================
It is interesting to see the role of the manifold volume measure in the resolution of the CC problem. We accomplish this now in the framework of the scalar field model I of previous section. The $\zeta$-dependence of $V_{eff}(\phi;\zeta ,M)$, Eq.(\[Veff1\]), in the form of inverse square like $(\zeta +b_{g})^{-2}$ has a key role in the resolution of the old CC problem in TMT. One can show that if quantum corrections to the underlying action generate nonminimal coupling like $\propto R(\Gamma,g)\phi^2$ in both $L_1$ and $L_2$, the general form of the $\zeta$-dependence of $V_{eff}$ remains similar: $V_{eff}\propto (\zeta +f(\phi))^{-2}$, where $f(\phi)$ is a function. The fact that only such type of $\zeta$-dependence emerges in $V_{eff}(\phi;\zeta ,M)$, and a $\zeta$-dependence is absent for example in the numerator of $V_{eff}(\phi;\zeta ,M)$, is a direct result of our basic assumption that $L_1$ and $L_2$ in the action (\[sec-2-S-modif\]) are independent of the manifold measure fields $\varphi_a$.
Generically, in the action (\[S-model-scalar.f.\]), $b_{\phi}\neq b_g$ that yields a nonlinear kinetic term (i.e. the $k$-essence type dynamics) in the Einstein frame[^5]. But for purposes of this section it is enough to take a simplified model with $b_{\phi}= b_g$ (which is in fact a fine tuning) since the nonlinear kinetic term has no qualitative effect on the zero CC problem. In such a case solving the constraint (\[constraint2-1\]) for $\zeta$ and substituting into Eqs.(\[Tmn-2\])-(\[phief\]) we obtain equations for scalar-gravity system which can be described by the regular GR effective action with the scalar field potential $$V_{eff}(\phi)=
\frac{(sM^{4}+V_{1})^2}{4[b_g(sM^4+V_1(\phi))-V_{2}(\phi)]}
\label{Veff3}$$ For an arbitrary nonconstant function $V_1(\phi)$ there exist infinitely many values of the integration constant $sM^4$ such that $V_{eff}(\phi)$ has the [*absolute minimum*]{} at some $\phi=\phi_0$ with $V_{eff}(\phi_0)=0$ (provided $b_g[sM^4+V_1(\phi)]-V_2(\phi)>0$). This effect takes place as $sM^4+V_1(\phi_0)=0$ [**without fine tuning of the parameters and initial conditions**]{}. Note that the choice of the scalar field potential in the GR effective action in a form proportional to a perfect square like emerging in Eq.(\[Veff3\]) would mean a fine tuning.
For illustrative purpose let us consider the model with $$V_1(\phi)=\frac{1}{2}\mu_1^2\phi^2, \qquad
V_2(\phi)=V_2^{(0)}+\frac{1}{2}\mu_2^2\phi^2. \label{V12model}$$ Recall that adding a constant to $V_1$ does not effect equations of motion, while $V_2^{(0)}$ absorbs the bare CC and all possible vacuum contributions. We take negative integration constant, i.e. $s=-1$, and the only restriction on the values of the integration constant $M$ and the parameters is that denominator in (\[Veff3\]) is positive[^6].
Consider spatially flat FRW universe with the metric in the Einstein frame $$\tilde{g}_{\mu\nu}=diag(1,-a^2,-a^2,-a^2), \label{FRW}$$ where $a=a(t)$ is the scale factor. Each cosmological solution ends with the transition to a $\Lambda =0$ state via damping oscillations of the scalar field $\phi$ towards its absolute minimum $\phi_0$. The appropriate oscillatory regime in the phase plane is presented in Fig. 1.
![Typical phase curve (in the phase plane ($\phi$,$\dot{\phi})$) of the scalar field $\phi$ during the transition to $\Lambda =0$ state. For illustrative purposes the parameters are chosen such that $V_{eff}=(M^2/2b_g)(\phi^2-M^2)^2/(\phi^2+4M^2)$ and $\phi_0=\pm
M$ and $\delta =0$. In the case without fine tuning of the parameters $b_g\neq b_{\phi}$, i.e. $\delta\neq 0$, the phase portrait is qualitatively the same.[]{data-label="fig1"}](phase_plane.eps){width="10.0cm" height="6.0cm"}
It follows from the constraint (\[constraint2-1\]) (where we took $\delta =0$) that $|\zeta|\to\infty$ as $\phi\to\phi_0$. More exactly, oscillations of $sM^{4}+V_1$ around zero are accompanied with a singular behavior of $\zeta$ each time when $\phi$ crosses $\phi_0$ $$\frac{1}{\zeta}\sim sM^{4}+V_1(\phi)\to 0 \qquad as \qquad
\phi\to\phi_0 \label{modul-zeta-infty}$$ and $\zeta^{-1}$ oscillates around zero together with $sM^{4}+V_1(\phi)$. Taking into account that the metric in the Einstein frame $\tilde{g}_{\mu\nu}$, Eq.(\[FRW\]), is regular we deduce from Eq.(\[gmunuEin\]) that the metric $g_{\mu\nu}$ used in the underlying action (\[S-model-scalar.f.\]) becomes degenerate each time when $\phi$ crosses $\phi_0$ $$g_{00}=\frac{\tilde{g}_{00}}{\zeta +b_g}\sim \frac{1}{\zeta}\to 0;
\qquad g_{ii}=\frac{\tilde{g}_{ii}}{\zeta +b_g}\sim
-\frac{1}{\zeta}\to 0 \qquad as \qquad \phi\to\phi_0,
\label{g00-degen}$$ where we have taken into account that the energy density approaches zero and therefore for this cosmological solution the scale factor $a(t)$ remains finite in all times $t$. Therefore $$\sqrt{-g}\sim \frac{1}{\zeta^2}\to 0 \qquad and \qquad \Phi
=\zeta\sqrt{-g}\sim \frac{1}{\zeta}\to 0\qquad as \qquad
\phi\to\phi_0\label{sqrtg-degen}$$
The detailed behavior of $\zeta$, the manifold measure $\Phi$ and $g_{\mu\nu}$ - components[^7] are shown in Fig. 2.
![Oscillations of the measure $\Phi$, the original metric $g_{\mu\nu}$ and the r.h.s. of Eq.(\[covar-conserv-J\]) during the transition to $\Lambda =0$ state. []{data-label="fig2"}](Phi_g00_gii_Current.eps){width="10.0cm" height="7.0cm"}
Recall that the manifold volume measure $\Phi$ is a signed measure[@signed] and therefore it is not a surprise that it can change sign. But TMT shows that including the manifold degrees of freedom into the dynamics of the scalar-gravity system we discover an interesting dynamical effect: a transition to zero vacuum energy is accompanied by oscillations of $\Phi$ around zero. Similar oscillations[^8] simultaneously occur with all components of the metric $g_{\mu\nu}$ used in the underlying action (\[S-model-scalar.f.\]).
The measure $\Phi$ and the metric $g_{\mu\nu}$ pass zero only in a discrete set of moments in the course of transition to the $\Lambda =0$ state. Therefore there is no problem with the condition $\Phi\neq 0$ used for the solution (\[varphi\]). Also there is no problem with singularity of $g^{\mu\nu}$ in the underlying action since $$\lim_{\phi\to\phi_0}\Phi g^{\mu\nu}=finite \qquad and \qquad
\sqrt{-g} g^{\mu\nu}\sim\frac{1}{\zeta}\to 0 \qquad as \qquad
\phi\to\phi_0 \label{lim}$$ The metric in the Einstein frame $\tilde{g}_{\mu\nu}$ is always regular because degeneracy of $g_{\mu\nu}$ is compensated in Eq.(\[gmunuEin\]) by singularity of the ratio of two measures $\zeta\equiv\Phi/\sqrt{-g}$.
Restoration of Intrinsic TMT Symmetry in the Course of Transition to Zero Cosmological Constant State
=====================================================================================================
Let us now turn to intrinsic symmetry of TMT which can reveal itself in a model with only the manifold volume measure $\Phi$. Indeed, if in Eq.(\[L1L2\]) we take the limit[^9] $b_g\to 0$ and $L_2^m\to 0$ then Eq.(\[Constr-original\]) reads $$L_1^m-g^{\mu\nu}\frac{\partial L_1^m}{\partial g^{\mu\nu}}=sM^4,
\quad {\text if} \quad\zeta\neq 0. \label{Constr-L2=0}$$ If in addition $L_1^m$ is homogeneous of degree 1 in $g^{\mu\nu}$ then the integration constant $M$ must be zero. The simplest example of a model for $L_1^m$ satisfying this property is the massless scalar field. In such a case the theory is invariant under transformations in the space of the scalar fields $\varphi_{a}$ $$\varphi_{a}\rightarrow\varphi^{\prime}_{a}=
\varphi^{\prime}_{a}(\varphi_{b}) \label{LES1}$$ resulting in the transformation of the manifold volume measure $\Phi$ $$\Phi(x)\rightarrow\Phi^{\prime}(x)=J(x)\Phi(x), \qquad J(x)=
Det(\frac{\partial\varphi^{\prime}_{a}}{\partial\varphi_{b}})
\label{LES2}$$ simultaneously with the local transformation of the metric $$g_{\mu\nu}(x)\rightarrow g^{\prime}_{\mu\nu}(x)=J(x)g_{\mu\nu}(x).
\label{LES3}$$ This symmetry was studied in earlier pulications[@GK1] where we called it the local Einstein symmetry (LES).
Consider now linear transformations in the space of the scalar fields $\varphi_{a}$ $$\varphi_{a}\rightarrow\varphi^{\prime}_{a}= A_a^b\varphi_{b}+C_b,
\quad a,b=1,2,3,4 \label{linear-trans}$$ where $A_a^b=constants$, $C_b=constants$. Then LES (\[LES1\])-(\[LES3\]) is reduced to transformations of the global Einstein symmetry (GES) with $J=det(A_a^b)=const$. Notice that the Einstein symmetry contains a $\Bbb Z_2$ subgroup of the sign inversions when $J=-1$: $$\Phi\rightarrow -\Phi, \quad g_{\mu\nu}\rightarrow
-g_{\mu\nu}\label{reflection}$$
LES as well as GES appear to be explicitly broken if $L_1^m$ is not a homogeneous function of degree 1 in $g^{\mu\nu}$, for example as in the model where $L_1^m$ describes a scalar field with a nontrivial potential[^10]. The Lagrangian $L_2^m$ generically breaks the Einstein symmetry too. The transformation of GES originated by the infinitesimal linear transformations $\varphi_a(x)\rightarrow\varphi^{\prime}_a(x)=
(1+\epsilon/4)\varphi_a(x)$, $\epsilon =const.$, yields the following variation of the action $(\ref{sec-2-S-modif})$ written in the form $S=\int{\cal L}d^4x$ where ${\cal L}=\Phi
L_1+\sqrt{-g}L_2$: $$\delta S=\int\left[-\frac{\partial{\cal L}}{\partial
g^{\mu\nu}}g^{\mu\nu}+
L_1\frac{\partial\Phi}{\partial\varphi_{a,\mu}}\varphi_{a,\mu}\right]\epsilon
d^4x. \label{deltaS}$$ The first term in (\[deltaS\]) equals zero on the mass shell giving the gravitational equation (\[g-mu-nu-varying\]); recall that we proceed in the first order formalism. Integrating the second term by part, using Eq.(\[varphi\]) and the definition (\[Phi\]) of the measure $\Phi$, we reduce the variation (\[deltaS\]) to $\delta S=\epsilon\int\partial_{\mu}j^{\mu}d^4x$ where $\partial_{\mu}j^{\mu}=sM^4\Phi$ and $j^{\mu}=sM^4B^{\mu}_a\varphi_a$. In the presence of topological defects with $\Phi =0$, Eq.(\[varphi\]) does not hold anymore all over space-time, and one should keep $L_1$ in the definition of the current: $j^{\mu}=L_1B^{\mu}_a\varphi_a$. In Subs.VIII.D we will see how such a situation may be realized.
To present the current conservation in the generally coordinate invariant form one has to use the covariant divergence. However when doing this using the original metric $g_{\mu\nu}$ we encounter the non-metricity. It is much more transparent to use the Einstein frame (\[gmunuEin\]) where the space-time becomes pseudo-Riemannian and the covariant derivative of the metric $\tilde{g}_{\mu\nu}$ equals zero identically. Thus with the definition $j^{\mu}=\sqrt{-\tilde{g}}J^{\mu}$, using the definition of $\zeta$ in Eq.(\[g-mu-nu-varying\]) and the transformation to the Einstein frame (\[gmunuEin\]) we obtain $$\tilde{\nabla}_{\mu}J^{\mu}\equiv\frac{1}{\sqrt{-\tilde{g}}}\partial_{\mu}
\left(\sqrt{-\tilde{g}}J^{\mu}\right)=sM^4\frac{\zeta}{(\zeta
+b_g)^2} \label{covar-conserv-J}$$
As one should expect, when $L_2\equiv 0$ and $L_1^m$ is homogeneous of degree 1 in $g^{\mu\nu}$, i.e. in the case of unbroken GES, the current is conserved because in this case the integration constant $M=0$.
As we have seen in the framework of the scalar field model of Sec.V, the dynamical evolution pushes $|\zeta|\equiv|\Phi|/\sqrt{-g}\to\infty$ as the gravity$+$scalar field $\phi$ -system approaches (without fine tuning) the $\Lambda
=0$ ground state $\phi =\phi_0$. Therefore according to Eq.(\[covar-conserv-J\]), $$\tilde{\nabla}_{\mu}J^{\mu}\to 0 \qquad as \qquad \phi\to\phi_0.
\label{covar-conserv-J-vacuum}$$ For the model of Sec.V, the damping oscillations of the r.h.s. of Eq.(\[covar-conserv-J\]) around zero are shown in Fig. 2. Thus, the [*GES explicitly broken in the underlying action, emerges in the vacuum which, as it turns out, has zero energy density. And vice versa, emergence of GES due to $|\zeta|\to\infty$ implies, according to Eq.(\[Veff1\]), a transition to a $\Lambda =0$ ground state*]{}.
Other way to reach the same conclusion is to look at the underlying action (\[S-model-scalar.f.\]). In virtue of Eq.(\[lim\]), it is evident that in the course of transition to the ground state, the terms in (\[S-model-scalar.f.\]) coupled to the metric volume measure $\sqrt{-g}$ become negligible in comparison with the corresponding terms coupled to the manifold volume measure $\Phi$; besides the term $-\int V_1(\phi)\Phi
d^4x$ (which also breaks the GES) disappears as $\phi\to\phi_0$. The only terms surviving in the transition to the $\Lambda =0$ ground state are the following $$\frac{1}{b_g}\int\Phi d^4x \left[-\frac{1}{\kappa}
R(\Gamma,g)+\frac{1}{2}g^{\mu\nu}\phi_{,\mu}\phi_{,\nu} \right]
\label{S-model-scalar.f.GES}$$ and they respect the GES.
One should notice however that one can regard the GES as the symmetry responsible for a zero CC only if TMT is taken in the strict framework formulated in Sec.I. In fact, let us consider for example a modified TMT model where the manifold volume measure degrees of freedom enter in the Lagrangian $L_1$ in contrast to our additional basic assumption made Sec.I ( after Eq.(\[sec-2-S-modif\])). Namely let us assume that the Lagrangian $L_1$ in Eq.(\[sec-2-S-modif\]) involves a term proportional to $\Phi/\sqrt{-g}$ that explicitly breaks the infinite dimensional symmetry (\[IDS\]). To be more concrete we consider a model with the action $$S=S_{mod}^{(1)}-\lambda\int\frac{\Phi^2}{\sqrt{-g}}d^4x
\label{S+Delta}$$ where $S_{mod}^{(1)}$ is the action defined in Eq.(\[S-model-scalar.f.\]). Such an addition to the action (\[S-model-scalar.f.\]) respects the GES but it is easy to see that it affects the theory in such a way that without fine tuning it is impossible generically to reach a zero CC (see Appendix B).
Scalar Field Model II. Global Scale Invariance
==============================================
Let us now turn to the analyze of the results of the TMT model possessing a global scale invariance studied early in detail[@G1]-[@GKatz],[@GK5]-[@GK9]. The scalar field $\phi$ playing the role of a model of dark energy appears here as a dilaton, and a spontaneous breakdown of the scale symmetry results directly from the presence of the manifold volume measure $\Phi$. In other words, this SSB is an intrinsic feature of TMT.
In the context of the present paper this model is of significant interest because cosmological solutions of the FRW universe exhibit two unexpected results: (a) the ground state as well as the asymptotic of quintessence-like evolution (in co-moving frame) possess certain degeneracies in $\Phi$ or $g_{\mu\nu}$; (b) superaccelerating expansion of the universe (phantom cosmology) appears as the direct dynamical effect when $\Phi<0$, i.e. as the orientation of the space-time manifold is opposite to the regular one. In this section we present the model and some of its relevant results. Regimes (a) and (b) will be analyzed in the next two sections.
The action of the model reads $$\begin{aligned}
&S=&\frac{1}{b_g}\int d^{4}x e^{\alpha\phi /M_{p}}
\left[-\frac{1}{\kappa}R(\Gamma ,g)(\Phi +b_{g}\sqrt{-g})+(\Phi
+b_{\phi}\sqrt{-g})\frac{1}{2}g^{\mu\nu}\phi_{,\mu}\phi_{,\nu}-e^{\alpha\phi
/M_{p}}\left(\Phi V_{1} +\sqrt{-g}V_{2}\right)\right]
\label{totaction}\end{aligned}$$ and it is invariant under the global scale transformations ($\theta =const.$): $$g_{\mu\nu}\rightarrow e^{\theta }g_{\mu\nu}, \quad
\Gamma^{\mu}_{\alpha\beta}\rightarrow \Gamma^{\mu}_{\alpha\beta},
\quad \varphi_{a}\rightarrow \lambda_{ab}\varphi_{b}\quad
\text{where} \quad \det(\lambda_{ab})=e^{2\theta}, \quad
\phi\rightarrow \phi-\frac{M_{p}}{\alpha}\theta . \label{st}$$ The appearance of the dimensionless parameters $b_g$ and $b_{\phi}$ is explained by the same reasons we mentioned after Eqs.(\[L1L2\]) and (\[S-model-scalar.f.\]). In contrast to the model of Sec.IV, now we deal with exponential (pre-) potentials where $V_1$ and $V_2$ are constant dimensionfull parameters. The remarkable feature of this TMT model is that Eq.(\[varphi\]), being the solution of the equation of motion resulting from variation of the manifold volume measure degrees of freedom, breaks spontaneously the scale symmetry (\[st\]): this happens due to the appearance of a dimensionfull integration constant $sM^4$ in Eq.(\[varphi\]). One can show[@GK9] that in the case of the negative integration constant ($s=-1$) and $V_1>0$, the ground state appears to be again (as it was in the scalar field model I of Sec.IV) a zero CC state without fine tuning of the parameters and initial conditions. The behavior of $\Phi$ and $g_{\mu\nu}$ in the course of transition to the $\Lambda =0$ state is qualitatively the same as we observed in Sec.V for the scalar field model I. Therefore in the present paper studying the model (\[totaction\]) we restrict ourself with the choice $s=+1$ and $V_1>0$.
Similar to the model of Sec.IV, equations of motion resulting from the action (\[totaction\]) are noncanonical and the space-time is non Riemannian when using the original set of variables. This is because all the equations of motion and the solution for the connection coefficients include terms proportional to $\partial_{\mu}\zeta$. However, when working with the new metric ($\phi$ remains the same) $$\tilde{g}_{\mu\nu}=e^{\alpha\phi/M_{p}}(\zeta +b_{g})g_{\mu\nu},
\label{ct}$$ which we call the Einstein frame, the connection becomes Riemannian and general form of all equations becomes canonical. Since $\tilde{g}_{\mu\nu}$ is invariant under the scale transformations (\[st\]), spontaneous breaking of the scale symmetry is reduced in the Einstein frame to the [*spontaneous breakdown of the shift symmetry*]{} $$\phi\rightarrow\phi +const.
\label{phiconst}$$
After the change of variables to the Einstein frame (\[ct\]) the gravitational equation takes the standard GR form with the same Newton constant as in the action (\[totaction\]) $$G_{\mu\nu}(\tilde{g}_{\alpha\beta})=\frac{\kappa}{2}T_{\mu\nu}^{eff}
\label{gef1}$$ where $G_{\mu\nu}(\tilde{g}_{\alpha\beta})$ is the Einstein tensor in the Riemannian space-time with the metric $\tilde{g}_{\mu\nu}$. The energy-momentum tensor $T_{\mu\nu}^{eff}$ reads $$\begin{aligned}
T_{\mu\nu}^{eff}&=&\frac{\zeta +b_{\phi}}{\zeta +b_{g}}
\left(\phi_{,\mu}\phi_{,\nu}-\frac{1}{2}
\tilde{g}_{\mu\nu}\tilde{g}^{\alpha\beta}\phi_{,\alpha}\phi_{,\beta}\right)
-\tilde{g}_{\mu\nu}\frac{b_{g}-b_{\phi}}{2(\zeta +b_{g})}
\tilde{g}^{\alpha\beta}\phi_{,\alpha}\phi_{,\beta}
+\tilde{g}_{\mu\nu}V_{eff}(\phi,\zeta;M)
\label{Tmn}\end{aligned}$$ where the function $V_{eff}(\phi,\zeta;M)$ is defined as following: $$V_{eff}(\phi,\zeta;M)=
\frac{b_{g}\left[M^{4}e^{-2\alpha\phi/M_{p}}+V_{1}\right]
-V_{2}}{(\zeta +b_{g})^{2}}. \label{Veff2}$$ Note that the $\zeta$-dependence of $V_{eff}(\phi,\zeta;M)$ is the same as in Eq.(\[Veff1\]) of the model of Sec.IV.
The scalar field $\zeta$ is determined by means of the constraint similar to Eq.(\[constraint2-1\]) of Sec.IV $$\begin{aligned}
&&(b_{g}-\zeta)\left[M^{4}e^{-2\alpha\phi/M_{p}}+
V_{1}\right]-2V_{2}-\delta\cdot b_{g}(\zeta +b_{g})X
=0\label{constraint2}\end{aligned}$$ where $$X\equiv\frac{1}{2}\tilde{g}^{\alpha\beta}\phi_{,\alpha}\phi_{,\beta}
\qquad \text{and} \qquad \delta =\frac{b_{g}-b_{\phi}}{b_{g}}
\label{delta}$$
The dilaton $\phi$ field equation in the Einstein frame is reduced to the following $$\frac{1}{\sqrt{-\tilde{g}}}\partial_{\mu}\left[\frac{\zeta
+b_{\phi}}{\zeta
+b_{g}}\sqrt{-\tilde{g}}\tilde{g}^{\mu\nu}\partial_{\nu}\phi\right]-\frac{2\alpha\zeta}{(\zeta
+b_{g})^{2}M_{p}}M^{4}e^{-2\alpha\phi/M_{p}} =0.
\label{phi-after-con}$$ where again $\zeta$ is a solution of the constraint (\[constraint2\]). Note that the dilaton $\phi$ dependence in all equations of motion in the Einstein frame appears [*only*]{} in the form $M^{4}e^{-2\alpha\phi/M_{p}}$, i.e. it results only from the spontaneous breakdown of the scale symmetry (\[st\]).
The effective energy-momentum tensor (\[Tmn\]) can be represented in a form of that of a perfect fluid $$T_{\mu\nu}^{eff}=(\rho +p)u_{\mu}u_{\nu}-p\tilde{g}_{\mu\nu},
\qquad \text{where} \qquad
u_{\mu}=\frac{\phi_{,\mu}}{(2X)^{1/2}}\label{Tmnfluid}$$ with the following energy and pressure densities resulting from Eqs.(\[Tmn\]) and (\[Veff2\]) after inserting the solution $\zeta =\zeta(\phi,X;M)$ of Eq.(\[constraint2\]): $$\rho(\phi,X;M) =X+ \frac{(M^{4}e^{-2\alpha\phi/M_{p}}+V_{1})^{2}-
2\delta b_{g}(M^{4}e^{-2\alpha\phi/M_{p}}+V_{1})X -3\delta^{2}
b_{g}^{2}X^2}{4[b_{g}(M^{4}e^{-2\alpha\phi/M_{p}}+V_{1})-V_{2}]},
\label{rho1}$$ $$p(\phi,X;M) =X- \frac{\left(M^{4}e^{-2\alpha\phi/M_{p}}+V_{1}+
\delta b_{g}X\right)^2}
{4[b_{g}(M^{4}e^{-2\alpha\phi/M_{p}}+V_{1})-V_{2}]}. \label{p1}$$
In a spatially flat FRW universe with the metric $\tilde{g}_{\mu\nu}=diag(1,-a^2,-a^2,-a^2)$ filled with the homogeneous scalar field $\phi(t)$, the $\phi$ field equation of motion takes the form $$Q_{1}\ddot{\phi}+ 3HQ_{2}\dot{\phi}- \frac{\alpha}{M_{p}}Q_{3}
M^{4}e^{-2\alpha\phi/M_{p}}=0 \label{phi1}$$ where $H$ is the Hubble parameter and we have used the following notations $$\dot{\phi}\equiv \frac{d\phi}{dt} \label{phidot-vdot}$$ $$Q_1=2[b_{g}(M^{4}e^{-2\alpha\phi/M_{p}}+V_{1})-V_{2}]\rho_{,X}
=(b_{g}+b_{\phi})(M^{4}e^{-2\alpha\phi/M_{p}}+V_{1})-
2V_{2}-3\delta^{2}b_{g}^{2}X \label{Q1}$$ $$Q_2=2[b_{g}(M^{4}e^{-2\alpha\phi/M_{p}}+V_{1})-V_{2}]p_{,X}=
(b_{g}+b_{\phi})(M^{4}e^{-2\alpha\phi/M_{p}}+V_{1})-
2V_{2}-\delta^{2}b_{g}^{2}X\label{Q2}$$ $$Q_{3}=\frac{1}{[b_{g}(M^{4}e^{-2\alpha\phi/M_{p}}+V_{1})-V_{2}]}
\left[(M^{4}e^{-2\alpha\phi/M_{p}}+V_{1})
[b_{g}(M^{4}e^{-2\alpha\phi/M_{p}}+V_{1})-2V_{2}] +2\delta
b_{g}V_{2}X+3\delta^{2}b_{g}^{3}X^{2}\right] \label{Q3}$$
The non-linear $X$-dependence appears here in the framework of the fundamental theory without exotic terms in the Lagrangians $L_1$ and $L_2$. This effect follows just from the fact that there are no reasons to choose the parameters $b_{g}$ and $b_{\phi}$ in the action (\[totaction\]) to be equal in general; on the contrary, the choice $b_{g}=b_{\phi}$ would be a fine tuning. Thus the above equations represent an explicit example of $k$-essence[@k-essence] resulting from first principles. The system of equations (\[gef\]), (\[rho1\])-(\[phi1\]) accompanied with the functions (\[Q1\])-(\[Q3\]) and written in the metric $\tilde{g}_{\mu\nu}=diag(1,-a^2,-a^2,-a^2)$ can be obtained from the k-essence type effective action $$S_{eff}=\int\sqrt{-\tilde{g}}d^{4}x\left[-\frac{1}{\kappa}R(\tilde{g})
+p\left(\phi,X;M\right)\right] \label{k-eff},$$ where $p(\phi,X;M)$ is given by Eq.(\[p1\]). In contrast to the simplified models studied in literature[@k-essence], it is impossible here to represent $p\left(\phi,X;M\right)$ in a factorizable form like $\tilde{K}(\phi)\tilde{p}(X)$. The scalar field effective Lagrangian, Eq.(\[p1\]), can be represented in the form $$p\left(\phi,X;M\right)=K(\phi)X+ L(\phi)X^2-U(\phi)
\label{eff-L-ala-Mukhanov}$$ where the potential $$U(\phi)=\frac{[V_{1}+M^{4}e^{-2\alpha\phi/M_{p}}]^{2}}
{4[b_{g}\left(V_{1}+M^{4}e^{-2\alpha\phi/M_{p}}\right)-V_{2}]}
\label{eff-L-ala-Mukhanov-potential}$$ and $K(\phi)$ and $L(\phi)$ depend on $\phi$ only via $M^{4}e^{-2\alpha\phi/M_{p}}$. Notice that $U(\phi)>0$ for any $\phi$ provided $$b_g>0, \qquad V_1>0 \qquad and \qquad b_{g}V_{1}\geq V_{2}
\label{bV1>V2},$$ that we will assume in what follows. Note that besides the presence of the effective potential $U(\phi)$, the Lagrangian $p\left(\phi,X;M\right)$ differs from that of Ref.[@k-inflation-Mukhanov] by the sign of $L(\phi)$: in our case $L(\phi)<0$ provided the conditions (\[bV1>V2\]). This result cannot be removed by a choice of the parameters of the underlying action (\[totaction\]) while in Ref.[@k-inflation-Mukhanov] the positivity of $L(\phi)$ was an essential [*assumption*]{}. This difference plays a crucial role for a possibility of a dynamical protection from the initial singularity of the curvature studied in detail in Ref[@GK9].
The model allows a power law inflation (where the dilaton $\phi$ plays the role of the inflaton) with a graceful exit to a zero or tiny cosmological constant state. In what it concerns to primordial perturbations of $\phi$ and their evolution, there are no difference with the usual (i.e. one-measure) model with the action (\[k-eff\])-(\[eff-L-ala-Mukhanov-potential\]).
In the model under consideration, the conservation law corresponding to the GES (\[linear-trans\]) has the form $$\tilde{\nabla}_{\mu}J^{\mu}\equiv\frac{1}{\sqrt{-\tilde{g}}}\partial_{\mu}
\left(\sqrt{-\tilde{g}}J^{\mu}\right)=sM^4\frac{\zeta}{(\zeta
+b_g)^2}e^{-2\alpha\phi/M_{p}} \label{covar-conserv-J-model2}$$ with the same definition of the current $J^{\mu}$ as in Sec.VI.
Degeneracies of $g_{00}$ and $\Phi$ in $\Lambda\neq 0$ Ground States
=====================================================================
Fine-Tuned $\delta =0$ Models
-----------------------------
We are going now to analyze some of the cosmological solutions for the late universe in the framework of the scale invariant model of the previous section. These solutions surprisingly exhibit that asymptotically, as $t\to\infty$, either $g_{00}\to 0$ or $\Phi\to
0$.
In the late universe, the kinetic energy $X\to 0$. Therefore in many cases the role of the nonlinear $X$ dependence becomes qualitatively unessential. This is why, for simplicity, in this section we can restrict ourself with the fine tuned model with $\delta =0$. In such a case the constraint (\[constraint2\]) yields $$\zeta
=\frac{b_gV_1-2V_2+b_gM^{4}e^{-2\alpha\phi/M_{p}}}{V_1+M^{4}e^{-2\alpha\phi/M_{p}}},
\label{zeta-without-ferm-delta=0}$$ The energy density and pressure take then the form $$\rho^{(0)}=\rho|_{\delta =0}=\frac{1}{2}\dot{\phi}^{2}+U(\phi);
\qquad p^{(0)}=p|_{\delta =0}=\frac{1}{2}\dot{\phi}^{2}-U(\phi),
\label{rho-delta=0}$$ where $U(\phi)$ is determined by Eq.(\[eff-L-ala-Mukhanov-potential\]). The $\phi$-equation (\[phi1\]) is reduced to $$\ddot{\phi}+3H\dot{\phi}+\frac{dU(\phi)}{d\phi}=0.
\label{eq-phief-without-ferm-delta=0}$$
Applying this model to of the late time cosmology of the spatially flat universe and assuming that the scalar field $\phi\rightarrow\infty$ as $t\rightarrow\infty$, it is convenient to rewrite the potential $U(\phi)$ in the form $$U(\phi)=\Lambda +V(\phi), \label{rho-without-ferm}$$ where $$\Lambda
=\frac{V_{1}^{2}} {4(b_{g}V_{1}-V_{2})}. \label{lambda}$$ is the positive cosmological constant and $$V(\phi)
=\frac{(b_{g}V_{1}-2V_{2})V_{1}M^{4}e^{-2\alpha\phi/M_{p}}+
(b_{g}V_{1}-V_{2})M^{8}e^{-4\alpha\phi/M_{p}}}
{4(b_{g}V_{1}-V_{2})[b_{g}(V_{1}+
M^{4}e^{-2\alpha\phi/M_{p}})-V_{2}]}.
\label{V-quint-without-ferm-delta=0}$$ It is evident that if $b_{g}V_1>
2V_2$ or $b_{g}V_1= 2V_2$ then $V(\phi)$ is a sort of a quintessence-like potential and therefore quintessence-like scenarios can be realized. This means that the dynamics of the late time universe is governed by the dark energy which consists of both the cosmological constant and the potential slow decaying to zero as $\phi\to\infty$. In the opposite case, $b_{g}V_1<2V_2$, the potential $V(\phi)$, and also $U(\phi)$, has an absolute minimum at some finite value of $\phi$, and therefore the cosmological scenario is different from the quintessence-like. Details of the cosmological evolution starting from the early inflation and up to the late time universe governed by the potential $U(\phi)$ have been studied in Ref.[@GK9] for each of these three cases. Here we want to analyze what kind of degeneracy appears in ground state depending on the region in the parameter space.
The case $b_{g}V_1>2V_2$
------------------------
Let us consider the case when the relation between the parameters $V_1$ and $V_2$ satisfies the condition $b_{g}V_1>2V_2$. It follows from Eq.(\[zeta-without-ferm-delta=0\]) that $$\zeta \to\frac{b_gV_1-2V_2}{V_1}=const
> 0 \quad as \quad \phi\to\infty \label{zeta-asympt-delta=0}$$
By making use the $(00)$ component of Eq.(\[ct\]), we see that $$g_{00}=\frac{e^{-\alpha\phi/M_{p}}}{\zeta +b_g}\to 0
\label{g00-general-asympt-delta=0}$$ In order to get the asymptotic time dependence of $g_{00}$ and the spatial components of the metric $$g_{ii}=-\frac{e^{-\alpha\phi/M_{p}}}{\zeta +b_g}\,a(t)^2,
\qquad i=1,2,3
\label{g-phi-sqrt-general-asympt-delta=0}$$ as $t\to\infty$, we have to know a solution $a=a(t)$, $\phi
=\phi(t)$. We can find analytically the asymptotic (as $\phi\to\infty$) behavior of a cosmological solution for a particular value of the parameter $\alpha =\sqrt{3/8}$. In such a case, keeping only the leading contribution of the $\phi$-exponent in Eq.(\[V-quint-without-ferm-delta=0\]), we deal with the following system of equations $$\left(\frac{\dot{a}}{a}\right)^{2}=\frac{1}{3M_{p}^{2}}\rho^{(0)}
\label{FRW-eq1}$$ $$\ddot{\phi}+3\frac{\dot{a}}{a}\dot{\phi} -\frac{2\alpha
}{M_{p}}\frac{V_1(b_gV_1-2V_2)}{4(b_gV_1-V_2)^2}M^{4}e^{-2\alpha\phi/M_{p}}
=0, \label{phi-eq-cosm}$$ where $\rho^{(0)}$ is determined by Eq.(\[rho-delta=0\]). The exact analytic solution for these equations is as follows[@GK3]: $$\phi(t)=const. + \frac{M_p}{2\alpha}\ln(M_pt), \qquad a(t)\propto
t^{1/3}e^{\lambda t}, \qquad \lambda
=\frac{1}{M_p}\sqrt{\frac{\Lambda}{3}} \label{a-phi-exact}$$ where $\Lambda$ is determined by Eq.(\[lambda\]). Therefore we obtain for the asymptotic cosmic time behavior of the components of the metric $g_{\mu\nu}$ $$g_{00}\sim\frac{1}{t^{1/2}}\to 0; \qquad g_{ii}\sim
-t^{1/6}e^{2\lambda t} \qquad as \qquad t\to\infty
\label{g00-gii-exact-asympt-delta=0}$$ So in the course of the expansion of the very late universe, only $g_{00}$ asymptotically vanishes while the space components $g_{ii}$ behave qualitatively in the same manner as the space components of the metric in the Einstein frame $\tilde{g}_{ii}$. Respectively, the asymptotic behavior of the volume measures is as follows: $$\Phi\approx \frac{b_{g}V_1-2V_2}{V_{1}}\sqrt{-g}\sim e^{3\lambda
t} \qquad as \qquad t\to\infty \label{Phi-g-exact-asympt-delta=0}$$
The GES is asymptotically restored that can be seen from the asymptotic time behavior of the conservation law (\[covar-conserv-J-model2\]) $$\tilde{\nabla}_{\mu}J^{\mu}\sim const\cdot
e^{-2\alpha\phi/M_{p}}\sim \frac{1}{t}.
\label{covar-conserv-J-model2-1}$$
The case $b_{g}V_1=2V_2$
------------------------
In this case the asymptotic form of $V(\phi)$ is $$V(\phi) \approx\frac{M^8}{2b_gV_1}e^{-4\alpha\phi/M_{p}}
\label{V-bgV1equal2V2-delta=0}$$ and $\zeta$ asymptotically approaches zero according to $$\zeta
=\frac{b_gM^4e^{-2\alpha\phi/M_p}}{V_1+M^4e^{-2\alpha\phi/M_p}}
\to 0 \qquad as \qquad \phi\to\infty.
\label{zeta-asympt-delta=0-1}$$ Similar to the previous subsection, the analytic form of the asymptotic (as $\phi\gg M_p$) cosmological solution exists for a particular value of the parameter $\alpha
=\sqrt{3/32}$: $$\phi(t)=const. + \frac{M_p}{4\alpha}\ln(M_pt), \qquad a(t)\propto
t^{1/3}e^{\lambda t}, \qquad \lambda
=\frac{1}{M_p}\sqrt{\frac{\Lambda}{3}},
\label{a-phi-exact-bV1equal2V2}$$ where now $$\Lambda =\frac{V_1}{2b_g} \label{Lambda-bV1equal2V2}$$
For this solution we obtain the following asymptotic cosmic time behavior for the components of the metric $g_{\mu\nu}$ and volume measures: $$g_{00}\sim\frac{1}{t^{1/4}}\to 0; \qquad g_{ii}\sim
-t^{5/12}e^{2\lambda t} \qquad as \qquad t\to\infty
\label{bgV1equal2V2g00-gii}$$ $$\sqrt{-g}\sim \sqrt{t}e^{3\lambda t}, \qquad \Phi\sim e^{3\lambda
t} \qquad as \qquad t\to\infty \label{bgV1equal2V2Phi-g}$$
The asymptotic time behavior of the conservation law describing the asymptotic restoration of the GES is the same as in Eq.(\[covar-conserv-J-model2-1\]).
The case $0<b_{g}V_1<2V_2$
--------------------------
In this case the potential $U(\phi)$, Eq.(\[eff-L-ala-Mukhanov-potential\]), has an absolute minimum $$\Lambda =U(\phi_{min})=\frac{V_{2}}{b_{g}^{2}} \qquad \text{at}
\qquad \phi =\phi_{min}=
-\frac{M_{p}}{2\alpha}\ln\left(\frac{2V_{2}-b_{g}V_{1}}{b_{g}M^4}\right).
\label{minVeff}$$ The spatially flat universe described in the Einstein frame with the metric $\tilde{g}_{\mu\nu}=diag(1,-a^2,-a^2,-a^2)$, in a finite time[@GK2] reaches this ground state where it expands exponentially $$a\propto e^{\lambda t}, \qquad \lambda =M_p^{-1}\sqrt{\Lambda/3}
\label{a(t)minVeff}$$ and $\Lambda$ is given by Eq.(\[minVeff\]). A surprising feature of this case is that $\zeta$, Eq.(\[zeta-without-ferm-delta=0\]), disappears in the minimum: $$\zeta(\phi_{min})=0 \label{zeta-in-minVeff}$$
The components of the metric $g_{\mu\nu}$ in the ground state are as follows $$g_{00}|_{(ground\,state)}=\left(\frac{2V_2-b_gV_1}{b_g^3M^4}\right)^{1/2}=const,
\qquad g_{ii}|_{(ground\,state)}= -g_{00}|_{(ground\,state)}\cdot
e^{2\lambda t} \label{bgV1less2V2g00-gii}$$ with the respective behavior of the metrical volume measure $\sqrt{-g}\propto exp(3\lambda t)$. Hence the manifold volume measure in the ground state disappears $$\Phi|_{(ground\,state)}=0 \label{bgV1less2V2Phi}$$ in view of Eq.(\[zeta-in-minVeff\]).
Disappearance of the manifold volume measure $\Phi$ in the ground state may not allow to get the equation (\[varphi\]) by varying the $\varphi_a$ fields in the action (\[totaction\]). Therefore in the conservation law (\[covar-conserv-J-model2\]) one should use the current in the form $j^{\mu}=L_1B^{\mu}_a\varphi_a$ as we have noticed after Eq.(\[deltaS\]). Recall that $L_1$ is constituted by the terms of the Lagrangian in (\[totaction\]) coupled to the measure $\Phi$. However, after using the gravitational equation obtained by varying $g^{\mu\nu}$ in (\[totaction\]) and substituting the ground state value $\phi =\phi_{min}$ into $L_1$, we obtain $L_1=M^4$. Hence, the conservation law (\[covar-conserv-J-model2\]) in the ground state reads just $$\tilde{\nabla}_{\mu}J^{\mu}|_{(ground\,state)}=0.
\label{covar-conserv-J-model2-gs}$$
![The phase portrait (in the phase plane ($\phi$,$\dot{\phi})$) for the model with $\alpha =0.2$, $\delta
=0.1$, $V_{1}=10M^{4}$ and $V_{2}=9.9b_{g}M^{4}$. The region with $\rho >0$ is divided into two dynamically disconnected regions by the line $Q_{1}(\phi,\dot{\phi})=0$. To the left of this line $Q_{1}>0$ (the appropriate zone we call zone 1) and to the right $Q_{1}<0$. The $\rho
>0$ region to the right of the line $Q_{1}(\phi,\dot{\phi})=0$ is divided into two zones (zone 2 and zone 3) by the line $Q_2=0$ (the latter coincides with the line where $w=-1$). In zone 2 $w>-1$ but $c_s^2<0$. In zone 3 $w<-1$ and $c_s^2>0$. Phase curves started in zone 2 cross the line $w=-1$. All phase curves in zone 3 exhibit processes with super-accelerating expansion of the universe. Besides all the phase curves in zone 3 demonstrate dynamical attractor behavior to the line which asymptotically, as $\phi\to\infty$, approaches the straight line $\dot{\phi}=0$.[]{data-label="fig3"}](no-go-phase-test-1.eps){width="15.0cm" height="10.0cm"}
![For the model with $\alpha =0.2$, $\delta =0.1$, $V_{1}=10M^{4}$ and $V_{2}=9.9b_{g}M^{4}$: typical scalar factor dependence of $\phi$ (Fig.(a)) and of the energy density $\rho$, defined by Eq.(\[rho1\]), (Fig(b)) in the regime corresponding to the phase curves started in zone 2. Both graphs correspond to the initial conditions $\phi_{in}=M_{p}$, $\dot\phi_{in} =5.7M^2/\sqrt{b_g}$; $\rho$ increases approaching asymptotically the value $\frac{M^{4}}{b_{g}}e^{5.52}$[]{data-label="fig4"}](no-go-phi-rho.eps){width="16.0cm" height="7.0cm"}
![For the same model and with the same initial conditions as in Fig. 4: crossing the phantom divide $w=-1$ and changing sign of the total volume measure $(\Phi +b_{\phi}\sqrt{-g})$ in the scalar field $\phi$ kinetic term (in the underlying action (\[totaction\])) occur simultaneously.[]{data-label="fig5"}](w_muk.eps){width="10.0cm" height="8.0cm"}
Sign Indefiniteness of the Manifold Volume Measure as the Origin of a Phantom Dark Energy
=========================================================================================
We turn now to the non fine-tuned case of the model of Sec.VII applied to the spatially flat universe. We start from a short review of our recent results[@GK9] concerning qualitative structure of the appropriate dynamical system which consists of Eq.(\[phi1\]) and the equation $$\left(\frac{\dot{a}}{a}\right)^{2}=\frac{1}{3M_{p}^{2}}\rho
\label{FRW-eq2}$$ where the energy density $\rho$ is defined by Eq.(\[rho1\]). The case of the interest of this section is realized when the parameters of the model satisfy the condition $$(b_g+b_{\phi})V_1-2V_2<0 \label{cond-for-phase-structure-1}$$ In this case the phase plane has a very interesting structure presented in Fig.3. Recall that the functions $Q_1$, $Q_2$, $Q_3$ are defined by Eqs.(\[Q1\])-(\[Q3\]).
We are interested in the equation of state $w=p/\rho<-1$, where pressure $p$ and energy density $\rho$ are given by Eqs.(\[rho1\]) and (\[p1\]). The line indicated in Fig. 3 as “line $w=-1$” coincides with the line $Q_2(\phi,X)=0$ because $$w+1=\frac{X}{\rho}\cdot\frac{Q_2}{\left[b_g\left(M^4e^{-2\alpha\phi/M_p}+V_1\right)-V_2\right]}
\label{w+1}$$ Phase curves in zone 3 correspond to the cosmological solutions with the equation of state $w<-1$. In zone 2, $w>-1$ but this zone has no physical meaning since the squared sound speed of perturbations $$c_s^2=\frac{Q_2}{Q_1} \label{sound}$$ is negative in zone 3. But in zone 2, $c_s^2>0$. Some details of numerical solutions describing the cross of the phantom divide $w=-1$ and the super-accelerating expansion of the universe are presented in Figs. 4 and 5.
Note that the superaccelerating cosmological expansion is obtained here without introducing an explicit phantom scalar field into the underlying action (\[totaction\]). In Ref.[@GK9] we have discussed this effect from the point of view of the effective k-essence model realized in the Einstein frame when starting from the action (\[totaction\]). A deeper analysis of the same effect yields the conclusion that the true and profound [*origin of the appearance of an effective phantom dynamics in our model is sign-indefiniteness of the manifold volume measure $\Phi$*]{}. In fact, using the constraint (\[constraint2\]), Eqs.(\[w+1\]) and (\[ct\]) it is easy to show that $$\Phi
+b_{\phi}\sqrt{-g}=(w+1)\,\frac{\rho}{4X}\,\frac{[M^4e^{-2\alpha\phi/M_p}+V_1+\delta\cdot
b_gX]}{[b_g(M^4e^{-2\alpha\phi/M_p}+V_1)-V_2]}\, a^3,
\label{kinetic-measure}$$ where $a$ is the scale factor. The expression in the l.h.s of this equation is the total volume measure of the $\phi$ kinetic term in the underlying action (\[totaction\]): $$\int d^{4}x (\Phi
+b_{\phi}\sqrt{-g})\frac{1}{2}g^{\mu\nu}\phi_{,\mu}\phi_{,\nu}
\label{kinetic-total-measure}$$ The sign of this volume measure coincides with the sign of $w+1$ as well as with the sign of the function $Q_2$ (see Eq.(\[w+1\])). In Fig. 5 we present the result of numerical solution for the scale factor dependence of $w$ and $(\Phi +b_{\phi}\sqrt{-g})/a^3$. Thus [*crossing the phantom divide occurs when the total volume measure of the $\phi$ kinetic term in the underlying action changes sign from positive to negative for dynamical reasons*]{}. This dynamical effect appears here as a dynamically well-founded alternative to the usually postulated phantom kinetic term of a scalar field Lagrangian[@Phantom-usual].
Summary and Discussion
======================
Introducing the space-time manifold volume element (\[dV\]) and adding the appropriate degrees of freedom to a set of traditional variables (metric, connection, matter fields) we reveal that such a two measures theory (TMT) takes up a special position between alternative theories. First, the equations of motion can be rewritten in the Einstein frame (where the space-time becomes Riemannian) with [*the same Newtonian constant*]{} as in the underlying action (where the space-time is generically non-Riemannian). Second, the theory possesses remarkable features in what it concerns the CC problem. Third, the TMT model with spontaneously broken dilatation symmetry satisfies all existing tests of GR. There are other interesting results, for example a possibility of a dynamical protection from the initial singularity of the curvature.
In this paper we have studied the behavior of the manifold volume measure $\Phi$ and the metric tensor $g_{\mu\nu}$ (used in the underlying TMT action) in cosmological solutions for a number of scalar field models of dark energy. We have made a special accent on the [*sign indefiniteness of the manifold volume measure*]{} $\Phi$ that may yield interesting physical effects. An example of such type of effects we have seen in Sec.IX: the total volume measure of the dilaton scalar field kinetic term in the underlying action can change sign from positive to negative in the course of dynamical evolution of the late time universe. In the Einstein frame, this transition corresponds to [*the crossing of the phantom divide of the dark energy*]{}.
We have found out that in all studied models, the transition to the ground state is [*always accompanied by a certain degeneracy*]{} either in the metric (e.g., in $g_{00}$ or in all components) or in the manifold volume measure $\Phi$, or even in both of them. This result differs sharply from what was expected e.g. in Refs.[@Hawking1979]-[@Tseytlin1982] where degenerate metric solutions have been associated with high curvature and temperature phases. One should only take into account that degeneracy of $g_{\mu\nu}$ and/or $\Phi$ in the (transition to) ground state takes place only when one works with the set of variables of the underlying TMT action. In the Einstein frame, we deal with the effective picture where the measure $\Phi$ does not present at all and the metric tensor $\tilde{g}_{\mu\nu}$ (see Eqs.(\[gmunuEin\]) or (\[ct\])) has the same regularity properties as in GR. The regularity of $\tilde{g}_{\mu\nu}$ results from the singularity of the transformations (\[gmunuEin\]) or (\[ct\])): degeneracy of $g_{\mu\nu}$ in a discrete set of moments is compensated by a singularity of $\zeta$.
The CC problem
--------------
TMT provides two different possibilities for resolution of the CC problem: one which guarantees zero CC without fine tuning (see however the end of Sec.VI and Appendix B); another which allows an unexpected way to reach a tiny CC. Which of these possibilities is realized depends on the sign of the integration constant $sM^4$, $s=\pm 1$. We are going now to discuss these two issues.
### The case $\Lambda =0$ in TMT
This case is of a special interest for two reasons. First, as it was shown earlier[@GK9], the conditions of the Weinberg’s no-go theorem[@Weinberg1] fail and a transition to a zero CC state in TMT can be realized without fine tuning. This becomes possible for example if $V_1(\phi)>0$ and the integration constant $sM^4<0$. Second, as we have shown in Sec.V, in the course of transition to a zero CC state, $g_{\mu\nu}$ and $\Phi$ oscillate synchronously around zero and they cross zero each time $t_i$ ($i=1,2,3,...$) when the scalar field $\phi$ crosses the (zero) absolute minimum of the potential (\[Veff3\]) (or of the potential (\[eff-L-ala-Mukhanov-potential\]) for the model of Sec.VII with $V_1<0$, see [@GK9]).
One should recall that $\zeta(x)$ does not have its own dynamics: its values at the space-time point $x$ are determined directly and immediately by the local configuration of the matter fields and gravity through the algebraic constraint, which is nothing but a consistency condition of equations of motion. $\zeta(x)$ does not possess inertia and therefore it changes together and synchronously with changing matter and gravity fields. This notion is very important when trying to answer the natural question: can oscillations of $\zeta(x)$ be a source for particle creation? The answer is - no, it cannot. In fact, there is a coupling of $\zeta$ with fermions. But the structure of this coupling in the Einstein frame has very surprising features which we will shortly review in the next subsection. Here we are only formulating the conclusion: emergence of even a tiny amount of fermionic matter immediately yields a rearrangement of the vacuum[^11] in such a way that $\zeta$ instantly ceases the regime of oscillations and rapidly enters into a regime of monotonous approach to a nonzero constant. It is interesting to note that the latter effect may explain why the present day cosmological constant most likely is tiny but nonzero, in spite of the existence of a fine tuning free classical solution described a transition to the $\Lambda =0$ state.
An overall change of sign of $g_{\mu\nu}$ in the course of these oscillations means a change of the signature from $(+---)$ to $(-+++)$ and vice versa, while oscillations of the sign of $\Phi$ describe the change of orientation of the space-time manifold. The latter means that the arena of the gravitational dynamics should contain two space-time manifolds with opposite orientations. The discrete set of changes of the orientations happens in the form of a [*smooth*]{} dynamical process in the course of which the space-time passes [*the “degenerate” phase where both the metrical structure and the total 4D-volume measure disappear*]{}. The latter means also that the term “orientation of the space-time manifold” loses any sense at moments $t_i$ ($i=1,2,3,...$). We conclude therefore that two 4D differentiable manifolds with opposite orientations (described by means of a sign indefinite volume 4-form) equipped with connection and metrical structure still are not enough to describe the arena of the gravitational dynamics: the complete description of the space-time dynamics requires also the mentioned degenerate phase. This situation is somewhat similar to that discussed in Introduction: first-order formulation of GR where the degenerate phase with $g_{\mu\nu}= 0$ should be also added[@Hawking1979],[@D'Auria-Regge],[@Tseytlin1982]; see e.g. the recent discussion by Bañados[@Banados] where the limiting process $g_{\mu\nu}\to 0$ is analyzed.
A new interesting feature of ground states in TMT we have revealed in the present paper concerns the so-called global Einstein symmetry (GES), Eqs.(\[LES1\])-(\[linear-trans\]), which turns out generically to be explicitly broken in all models with non-trivial dynamics. The surprising result we have discovered here on the basis of a number of models is that [*the GES is restored in the course of transitions to the ground state*]{} in all models considered. Hence its subgroup of the sign inversions of $g_{\mu\nu}$ and $\Phi$, Eq.(\[reflection\]), is also restored. Therefore the oscillations of $g_{\mu\nu}$ and $\Phi$ around zero in the course of transition to a $\Lambda =0$ ground state provoke a wish to compare this [*dynamical effect*]{} with the attempts to solve the old CC problem developed in Refs.[@Erdem]-[@reflection2]. The main idea of these approaches is that the field theory or at least the ground state[@'t; @Hooft] should be invariant under transformations of a discrete symmetry. According to Refs.[@Erdem]-[@reflection2] it might be either an invariance under the metric reversal symmetry or under the space-time coordinate transformations with the imaginary unit $i$: $x^A\rightarrow ix^A$. In contrast with these approaches, in TMT there is no need to postulate such exotic enough symmetries. Nevertheless we have seen that sign inversions of $g_{\mu\nu}$ emerge as a dynamical effect in the course of the cosmological evolution and this effect has indeed a relation to the resolution of the old CC problem.
### The case of a tiny CC
In the scalar field models of dark energy, an interesting feature of TMT consists in a possibility to provide a small value of the CC. If in the model of Sec.VIII.B, the parameter $V_2<0$ and $|V_2|\gg b_gV_1$ then the CC can be very small without the need for $V_1$ and $V_2$ to be very small. For example, if $V_{1}$ is determined by the energy scale of electroweak symmetry breaking $V_{1}\sim
(10^{3}GeV)^{4}$ and $V_{2}$ is determined by the Planck scale $V_{2} \sim (10^{18}GeV)^{4}$ then $\Lambda_{1}\sim
(10^{-3}eV)^{4}$. Along with such a seesaw mechanism[@G1], [@seesaw], there exists another way to explain the smallness of the CC applicable in all types of scenarios discussed in Secs.VIII.B-VIII.D (see also Appendix A). As one can see from Eqs.(\[lambda\]), (\[a-phi-exact-bV1equal2V2\]) and (\[minVeff\]), the value of $\Lambda$ appears to be inverse proportional[^12] to the dimensionless parameter $b_g$ which characterizes the relative strength of the ’manifold’ and ’metrical’ parts of the gravitational action. If for example $V_{1}\sim (10^{3}GeV)^{4}$ then for getting $\Lambda_1\sim
(10^{-3}eV)^{4}$ one should assume that $b_{g}\sim 10^{60}$. Such a large value of $b_{g}$ (see Eq.(\[L1L2\])) permits to formulate [*a correspondence principle*]{}[@GK9] [*between TMT and regular (i.e. one-measure) field theories*]{}: when $\zeta/b_g\ll 1$ then one can neglect the gravitational term in $L_1$ with respect to that in $L_2$ (see Eq.(\[L1L2\]) or Eq.(\[S-model-scalar.f.\]) or Eq.(\[totaction\])). More detailed analysis shows that in such a case the manifold volume measure $\Phi =\zeta \sqrt{-g}$ has no a dynamical effect and TMT is reduced to GR. This happens e.g. in the model of Sec.VIII.C where the late time evolution proceeds in a quintessence-like manner: the energy density decreases to the cosmological constant, Eq.(\[Lambda-bV1equal2V2\]), and $\zeta\to 0$, Eq.(\[zeta-asympt-delta=0-1\]). Another example is the model of Sec.VIII.D where $\Phi = 0$ in the ground state, Eq.(\[zeta-in-minVeff\]), while $\sqrt{-g}$ is finite. However generically $\zeta/b_g$ is not small, as it happens for example in the quintessence-like scenario of the late time universe in the model of Sec.VIII.B (see Eq.(\[zeta-asympt-delta=0\])).
Possibilities for predictions of new physical effects
-----------------------------------------------------
### Short review of the TMT model with spontaneously broken dilatation symmetry in the presence of matter
It would be interesting to find out other possible physical manifestations of the sign indefiniteness of the manifold volume measure. In fact, the model with spontaneously broken dilatation symmetry studied in Secs.VII-IX and in Ref.[@GK9] allows extensions which include fermion and gauge fields[@GK5]-[@GK7] or, alternatively, dust as a phenomenological matter model[@GK10]. In the former case, for example, the constraint (\[constraint2\]) is modified to the following $$\frac{1}{(\zeta
+b)^{2}}\left\{(b-\zeta)\left[M^{4}e^{-2\alpha\phi/M_{p}}+
V_{1}\right]-2V_{2})\right\}= \frac{\mu(\zeta -\zeta_{1})(\zeta
-\zeta_{2})}{2(\zeta +k)^2(\zeta +b)^{1/2}}\bar{\Psi}\Psi,
\label{constraint3}$$ where $\Psi$ is the fermion field in the Einstein frame and for simplicity we have chosen $\delta =0$ (that is $b_{\phi}=b_g=b$); $\zeta_{1,2}$ are defined by $$\zeta_{1,2}=\frac{1}{2}\left[k-3h\pm\sqrt{(k-3h)^{2}+ 8b(k-h)
-4kh}\,\right].
\label{zeta12}$$ and the dimensionless parameters $k$ and $h$ appear in the underlying action in the total volume measures of the fermion kinetic term $$\int d^{4}x e^{\alpha\phi /M_{p}}(\Phi +k\sqrt{-g})
\frac{i}{2}\overline{\Psi}
\left(\gamma^{a}e_{a}^{\mu}\overrightarrow{\nabla}_{\mu}-
\overleftarrow{\nabla}_{\mu}\gamma^{a}e_{a}^{\mu}\right)\Psi
\label{k}$$ and the fermion mass term $$-\int d^{4}xe^{\frac{3}{2}\alpha\phi /M_{p}}
(\Phi +h\sqrt{-g})\mu\overline{\Psi}\Psi
\label{h}$$ respectively. Note that the fermion equation in the Einstein frame has a canonical form but the mass of the fermion turns out $\zeta$ dependent $$m(\zeta)= \frac{\mu(\zeta +h)}{(\zeta +k)(\zeta +b)^{1/2}}
\label{fermion_mass}$$
The constraint (\[constraint3\]) describes the local balance between the fermion energy density and the scalar field $\phi$ contribution to the dark energy density in the space-time region where the wave function of the primordial fermion is not equal to zero. By means of this balance the constraint determines the scalar $\zeta(x)$.
In the case of dust as a phenomenological matter model, the r.h.s. of the constraint (\[constraint3\]) looks $$\frac{\zeta -b_m +2b}{2\sqrt{\zeta +b}}\, m\, \tilde{n},
\label{rhs-dust}$$ where the dimensionless parameter $b_m$ appears in the total volume measure of the dust contribution to the underlying action $$S_{m}=\int (\Phi +b_{m}\sqrt{-g})L_m d^{4}x \label{dust}$$ $$L_m=-m\sum_{i}\int e^{\frac{1}{2}\alpha\phi/M_{p}}
\sqrt{g_{\alpha\beta}\frac{dx_i^{\alpha}}{d\lambda}\frac{dx_i^{\beta}}{d\lambda}}\,
\frac{\delta^4(x-x_i(\lambda))}{\sqrt{-g}}d\lambda \label{Lm}$$ and $m$ is the mass parameter.
The wonderful feature of these models in the Einstein frame consists of the exact coincidence of the following three quantities: a) the noncanonical (in comparison with GR) terms in the energy-momentum tensor; b) the effective coupling “constant” of the dilaton $\phi$ to the matter (up to the factor $\alpha/M_p$); c) the expressions in the r.h.s. of the above mentioned constraints (\[constraint3\]) and (\[rhs-dust\]) for fermionic matter and dust respectively. For [*matter in normal conditions*]{}, the local matter energy density (i.e. in the space-time region occupied by the matter) is many orders of magnitude larger than the vacuum energy density. Detailed analysis[@GK5]-[@GK7],[@GK10] shows that when the matter is in the normal conditions, the balance dictated by the constraint becomes possible if $\zeta$ with very high accuracy takes the constant values: $\zeta\approx\zeta_1$ or $\zeta\approx\zeta_2$ for fermions (and therefore the fermion masses become constant)) and $\zeta\approx b_m-2b$ for the dust. Then the mentioned three quantities simultaneously become extremely small. Besides for the matter in normal conditions the gravitational equations are reduced to the canonical GR equations. The practical disappearance of the dilaton-to-matter coupling “constant” for the matter in normal conditions which occurs without fine tuning of the parameters allows us to assert that in such type of models the fifth force problem is resolved[@GK7],[@GK10].
It does not mean however that matter does not interact with the dilaton at all. When the matter is in states different from normal, the effect of dilaton-to-matter coupling may yield new very interesting phenomena. One of such effects appears when the neutrino energy density decreases to the order of magnitude close to the vacuum energy density. The latter can happen due to spreading of the neutrino wave packet. Then the cold gas of uniformly distributed nonrelativistic neutrinos causes a reconstruction of the vacuum to a state with $\zeta\to
|k|$ and as a result the neutrino gas rapidly transmute into an exotic state called neutrino dark energy(see e.g. Ref.[@Nelson]). This effect was studied in details in Ref.[@GK7] where we have shown that transmutation from the pure scalar field dark energy to the neutrino dark energy regime is favorable from the energetic point of view.
### Prediction of strong gravity effect in high energy physics experiments
For the solutions $\zeta\approx\zeta_1$ or $\zeta\approx\zeta_2$ of the constraint (\[constraint3\]), the l.h.s. of the constraint has the order of magnitude close to the vacuum energy density. There exists however another solution if one allows a possibility that in the core of the support of the fermion wave function the local dark energy density may be much bigger than the vacuum energy density. Such a solution turns out to be possible as fermion density is very big and $\zeta$ becomes negative and close enough to the value $\zeta\approx -b$. Then the solution of the constraint (\[constraint3\]) looks[@GK5] $$\frac{1}{\sqrt{\zeta
+b}}\approx\left[\frac{\mu(b-h)}{4M^4b(b-k)}\bar{\Psi}\Psi
e^{2\alpha\phi/M_p}\right]^{1/3}.
\label{zeta-b}$$ In such a case, instead of constant masses, as it was for $\zeta\approx\zeta_{1,2}$, Eq.(\[fermion\_mass\]) results in the following fermion self-interaction term in the effective fermion Lagrangian $$L^{ferm}_{selfint}=3\left[\frac{1}{b}\left(\frac{\mu(b-h)}{4M(b-k)}\bar{\Psi}\Psi\right)^4
e^{2\alpha\phi/M_p}\right]^{1/3}.
\label{selfinter}$$
It is very interesting that the described effect is the direct consequence of the strong gravity. In fact, in the regime where $\zeta +b\ll 1$ the effective Newton constant in the gravitational term of underlying action(\[totaction\]) $$S_{grav}=-\int d^{4}x \sqrt{-g}\,\frac{\zeta +b}{\kappa b}
R(\Gamma ,g)e^{\alpha\phi /M_{p}}
\label{strong-grav}$$ becomes anomalously large. Recall that for simplicity we have chosen here $b_{\phi}=b_g=b$. But if one do not to imply this fine tuning then one can immediately see from Eqs.(\[gef1\])-(\[Veff2\]) that in the Einstein frame the regime of the strong gravity dictated by the dense fermion matter is manifested for the dilaton too.
The coupling constant in Eq.(\[selfinter\]) is dimensionless and depends exponentially of the dilaton $\phi$ if one can regard $\phi$ as a background field $\phi=\bar{\phi}$. But in a more general case Eq.(\[selfinter\]) may be treated as describing an anomalous dilaton-to-fermion interaction very much different from the discussed above case of interaction of the dilaton to the fermion matter in normal conditions where the coupling constant practically vanishes. Such an anomalous dilaton-to-fermion interaction should result in creation of quanta of the dilaton field in processes with very heavy fermions. The probability of these processes is of course proportional to the Newton constant $M_p^{-2}$. But the new effect consists of the fact that the effective coupling constant of the anomalous dilaton-to-fermion interaction is proportional to $e^{2\alpha\bar{\phi}/3M_p}$. If the dilaton is the scalar field responsible for the quintessential inflation type of the cosmological scenario[@Quint-ess] then one should expect an exponential amplification of the effective coupling of this interaction in the present day universe in comparison with the early universe. One can hope that the described effect of the strong gravity might be revealed in the LHC experiments in the form of missing energy due to the multiple production of quanta of the dilaton field (recall that coupling of the dilaton to fermions in normal conditions practically vanishes and therefore the dilaton will not be observed after being emitted).
### Some other possible effects
1\. [*Dark matter as effect of gravitational enhancement.*]{} In the case of dust as a phenomenological matter model, the constraint (\[constraint3\]) with the r.h.s. (\[rhs-dust\]) is the fifth degree algebraic equation with respect to $\sqrt{\zeta
+b}$. There are some indications that in a certain region of the parameters a solution of the constraint exists which could provide a very interesting effect of an amplification of the gravitational field of visible diluted galactic and intergalactic dust or/and neutrinos. Such an effect might imply that the dark matter is not a new sort of matter but it is just a result of a so far unknown enhancement of the gravitational field of low density states of usual matter.
2\. [*Dilaton to photon coupling.*]{} Astrophysical observations of few last years indicate anomalously large transparency of the Universe to gamma rays[@Nature1],[@Nature2]. It is hard to explain this astrophysical puzzle in the framework of extragalactic background light. Recently a natural mechanism was suggested by De Angelis, Mansutti and Roncadelli[@Angelis] in order to resolve this puzzle. The idea is to suppose that there exists a very light spin-zero boson coupled to the photon: $$L_{\phi\gamma}=-\frac{1}{4\mu}F_{\mu\nu}\tilde{F}_{\mu\nu}\phi.
\label{phi-to-photon}$$ where $\mu$ is a mass parameter. In the context of quintessential scenario such a coupling was studied by Carroll[@carroll]. Then $\gamma\longrightarrow\phi\longrightarrow\gamma$ oscillations emerge which explain[@Angelis] the observed transparency of the Universe to gamma rays in a natural way if mass of the spin-zero boson $m<10^{-10}eV$. The crucial feature of this boson is that no other coupling of this scalar to matter exists. In the standard quintessence models this feature seems to be a real problem. But in TMT, as we already mentioned (see also Refs.[@GK7],[@GK10]) the dilaton playing the role of quintessence field decouples from matter in normal conditions. At the same time its coupling to the photon in the form (\[phi-to-photon\]) is not suppressed.
3\. [*Creation of a universe in the laboratory.*]{} A theoretical attempt by Farhi, Guth and Guven to describe a creation of a universe in the laboratory[@Guth] runs across a need to allow vanishing and changing sign of $\sqrt{-g}$. In Ref.[@Guth], this need is naturally regarded as a pathology. If similar approach to the problem of creation of a universe in the laboratory could be formulated in the framework of TMT then instead of $\sqrt{-g}$ there should appear a linear combination of $\Phi$ and $\sqrt{-g}$ which, as we already know, is able to vanish and change sign. In recent paper[@GS] by Guendelman and Sakai a model of child universe production without initial singularities was studied. To provide the desirable absence of initial singularity a crucial point is that the energy momentum tensor of the domain wall should be dominated by a sort of phantom energy. A possible way to realize this idea is to apply the dynamical brane tension[@GKNP] obtained when using the modified volume measure similar to the signed measure $\Phi$ of the present paper. So it could be that applying the notions explored in the present paper one can obtain also a framework for formulating non singular child universe production.
4\. [*Unparticle physics.*]{} $\zeta$ dependence of the fermion mass, Eq.(\[fermion\_mass\]), together with the constraint (\[constraint3\]) can be treated as a $\bar{\Psi}\Psi$ dependence of the fermion mass. This means that in states different from the normal one, the fermion mass spectrum may be continuous, that allows to think of a possibility to establish relation with the idea of unparticle physics[@Georgi].
Note finally that for the matter in normal conditions the model does not impose essential constraints on the parameters of the model (such as $b_g$, $b_{\phi}$, $b_m$, $k$, $h$). But the appropriate constraints should appear when more progress in the study of the listed and another possible new effects will be achieved.
Acknowledgements
================
We acknowledge V. Goldstein, M. Lin, E. Nissimov and S. Pacheva for useful discussions of some mathematical subjects. We also thank M. Duff for explaining us his approach and M. Bañados for useful conversations. We are also grateful to the referee whose constructive remarks assisted us in the improvement of this paper.
The Ground State with Non Zero CC in Model I
=============================================
Let us consider the scalar field model I (Eqs.(\[S-model-scalar.f.\]) and (\[V12model\])) with $\delta =0$ where we now choose a positive integration constant ($s=+1$) and the parameters $V_2^{(0)}<0$, $b_g\mu_1^2>\mu_2^2$. Then the ground state is realized for $\phi =0$ and the vacuum energy is $$\Lambda =V_{eff}(0)=\frac{M^8}{4b_gM^4-V_2^{(0)}} \label{AppI}$$ In this ground state, both the measure $\Phi>0$ and all components of the metric tensor $g_{\mu\nu}$ are regular. Note that the presence of the free dimensionless parameter $b_g$ in the denominator allows again to reach a small vacuum energy by means of the correspondence principle discussed in item 2 of Sec.X.
Global Einstein Symmetry does not Guarantee Resolution of the CC Problem
========================================================================
In the model (\[S+Delta\]), the gravitational equations are modified to the following $$G_{\mu\nu}(\tilde{g})=\frac{\kappa}{2}\left[\phi_{,\mu}\phi_{,\nu}-\frac{1}{2}\tilde{g}_{\mu\nu}X
+\frac{b_g[sM^4+V_1(\phi)+2\lambda\zeta]-V_2(\phi)-\lambda\zeta^2}{(\zeta
+b_g)^2}\right] \label{grav-App}$$ while the form of the scalar field $\phi$ equation remains the same as in Eq.(\[phief\]). However the constraint is now very much differs from Eq.(\[constraint2-1\]): $$4\lambda\zeta^2 +[sM^4+V_1(\phi)-2b_g\lambda]\zeta +2V_2(\phi)
-b_g[sM^4+V_1(\phi)]=0\label{constr-App}$$ One can see from Eq.(\[grav-App\]) that $\zeta$-dependence emerges now in the numerator of the effective potential. Besides, it is evident that in contrast with what was in Sec.V, the regime with $\zeta\to\infty$ cannot be a solution of the constraint. It is evident that a zero minimum of the effective potential cannot be now reached without fine tuning. Thus although the second term in the action (\[S+Delta\]) is invariant under the GES, adding this term we loss the ability to resolve the old CC problem.
[99]{} Einstein A and Rosen N 1935 [*Phys. Rev.*]{} **48** 73
Hawking S W [*Nucl.Phys.*]{} 1978 B[*144*]{} 349; Hawking S W 1979 in [*Recent Developments in Gravitation*]{} ed M Levy and S Deser (New York; Plenum)
D’Auria R and Regge T 1982 [*Nucl. Phys.*]{} B **195** 308
Tseytlin A A 1982 [*J. Phys. A: Math. Gen.*]{} **15** L105
Horowitz Gary T 1991 Class Quantum Grav. **8** 587
Ashtekar A 1991 [*Lectures on Non-Perturbative Canonical Gravity*]{} (World Scientific)
Jacobson T and Smolin L 1988 [*Nucl. Phys.*]{} B **299** 295
Dray T, Manogue C A and Tucker R W 1991 [*Gen. Rel. Grav.*]{} [**23**]{} 967
Ellis G, Sumeruk A, Coule D and Hellaby C 1992 [*Class. Quantum Grav.*]{} **9** 1535
Elizalde E, Odintsov S and Romeo A 1994 [*Class. Quant. Grav.*]{} **11** 61
Dray T, Ellis G, Hellaby C and Corinne A. Manogue C A 1997 [*Gen. Rel. Grav.*]{} [**29**]{} 591
Dray T, Ellis G, Hellaby C 2001 [*Gen. Rel. Grav.*]{} [**33**]{} 1041
Borowiec A, Francaviglia M and Volovich I 2007. [*Int. J. Geom. Meth. Mod. Phys.*]{} **4** 647
Mars M, Senovilla Jose M M and Vera R 2008 [*Phys. Rev.*]{} D **77** 027501
Witten E 1988 [*Commun. Math. Phys.*]{} **117** 353\
Witten E 1988 [*Nucl. Phys.*]{} B **311** 46
Giddings S B 1991 Physics Letters B **268** 17
Bañados M 2007 [*Class. Quantum Grav.*]{}**24** 5911\
Bañados M 2008 [*Phys.Rev.*]{} **D** 77 123534
Cohn D L, [*Measure Theory*]{}, Birkhauser, Boston, 1993.
Taylor J G 1979 [*Phys. Rev.*]{} **19** 2336
Wilczek F 1998 [*Phys. Rev. Lett.*]{} **80** 4851
Mosna R A and Saa A 2005 [*J.Math.Phys.*]{} **46** 112502
Guendelman E I and Kaganovich A B [*Phys. Rev.*]{} 1996 D**53** 7020; [*Mod. Phys. Lett.*]{} 1997 A**12** 2421; [*Phys. Rev.*]{} D**55** 5970; [*Mod. Phys. Lett.*]{} 1997 A**12** 2421; [*Phys. Rev.*]{} 1997 D**56** 3548; [*Mod. Phys. Lett.*]{} 1998 A**13** 1583.
Guendelman E I and Kaganovich A B [*Phys. Rev.*]{} 1998 D**57** 7200).
Guendelman E I and Kaganovich A B [*Phys. Rev.*]{} 1999 **D60** 065004.
Guendelman E I 1999 [*Mod. Phys. Lett.*]{} A[**14**]{}, 1043; [*Class. Quant. Grav.*]{} 2000 [**17**]{} 361; gr-qc/9906025; [*Mod. Phys. Lett.*]{} 1999 A[**A4**]{}, 1397; gr-qc/9901067; hep-th/0106085; [*Found. Phys.*]{} 2001 [**31**]{} 1019;
Kaganovich A B 2001 [*Phys. Rev.*]{} [**D63**]{}, 025022.
Guendelman E I and Katz O 2003 [*Class. Quant. Grav.*]{} [**20**]{} 1715
Guendelman E I 1997 [*Phys. Lett.*]{} B[**412**]{} 42; Guendelman E I 2003 gr-qc/0303048; Guendelman E I and Spallucci E 2003 hep-th/0311102.
Guendelman E I and Kaganovich A B 2002 [*Int. J. Mod. Phys.*]{} A[**17**]{} 417.
Guendelman E I and Kaganovich A B 2002 [*Mod. Phys. Lett.*]{} A[**A7**]{} 1227 (2002).
Guendelman E I and Kaganovich A B 2004 hep-th/0411188; [*Int.J.Mod.Phys.*]{} 2006 A[**21**]{} 4373.
Guendelman E I and Kaganovich A B 2006 in [*Paris 2005, Albert Einstein’s century*]{}, AIP Conf.Proc. 2006 861 875, Paris; hep-th/0603229.
Guendelman E I and Kaganovich A B 2007 [*Phys.Rev.*]{} D[**75**]{} 083505.
Guendelman E I and Kaganovich A B 2008 [*Annals Phys.*]{}. [**323**]{} 866.
Comelli D arXiv:0704.1802 \[gr-qc\].
Weinberg S 1989 [*Rev. Mod. Phys.*]{} [**61**]{} 1
Unruh W G 1989 [*Phys. Rev.*]{} 1989 D[*40*]{} 1048.
Ng Y Jack and van Dam H 1991 [*J. Math. Phys.*]{} [**32**]{} 1337.
Chiba T, Okabe T and Yamaguchi M 2000 [*Phys.Rev.*]{} D[**62**]{} 023511; Armendariz-Picon C, Mukhanov V and Steinhardt P J 2000 [*Phys. Rev. Lett.*]{} [**85**]{} 4438; [*Phys. Rev.*]{} 2001 D[**63**]{} 103510; Chiba T 2002 [*Phys.Rev.*]{} D[**66**]{} 063514.
Armendariz-Picon C., Damour T and Mukhanov V F 1999 [ *Phys.Lett.*]{} B[**458**]{} 209.
Caldwell R R [*Phys.Lett.*]{} 2002 B[**545**]{} 23; Gibbons G W 2003, hep-th/0302199.
Erdem R 2005 [*Phys. Lett.*]{} B[**621**]{} 11; [*Phys. Lett.*]{} 2006 B[**639**]{} 348; [*J. Phys.*]{} 2007 A[**40**]{} 6945.
Nobbenhuis S 2006 [*Found. Phys.*]{} [**36**]{} 613; ’t Hooft G, Nobbenhuis S 2006 [*Class. Quant. Grav.*]{} [**23**]{} 3819.
Duff M J and Kalkkinen J. 2006 [*Nucl. Phys.*]{} B[**758**]{} 161; 2007 [*Nucl. Phys.*]{} B[**760**]{} 64.
Arkani-Hamed N, Hall L J, Kolda C F and Murayama H 2000 [*Phys.Rev.Lett.*]{} [**85**]{} 4434.
R. MacKenzie, F. Wilczek and A. Zee 1984 [*Phys. Rev. Lett.*]{} [**53**]{} 2203.
R Fardon, A. E. Nelson, N. Weiner 2004 [*JCAP*]{} [**0410**]{} 005.
P.J.E. Peebles and A. Vilenkin 1999 [*Phys.Rev*]{} D[**59**]{} 063505.
Uchiyama Y, Aharonian F, Tanaka T, Takahashi T and Maeda 2007 [*Nature*]{} [**449**]{} 576.
Mazin D, Raue M 2007 [*Astron.Astrophys.*]{} [**471**]{} 439.
De Angelis A, Mansutti O and Roncadelli M 2007 [*Phys.Rev.*]{} D[**76**]{} 121301.
Carroll S 1998 [*Phys. Rev. Lett.*]{} [**81**]{} 3067.
Farhi E, Guth A H and Guven J 1990 [*Nucl. Phys.*]{} B[**339**]{} 417.
Guendelman E I and Sakai N 2008 [*Phys. Rev.*]{} D[**77**]{} 125002.
Guendelman E I, Kaganovich A B, Nissimov E, Pacheva S 2002 [*Phys. Rev.*]{} D[**66**]{} 046003; “Impact of dynamical tensions in modified string and brane theories”, Presented at 5th International Workshop on Lie Theory and Its Applications in Physics, Varna, Bulgaria, 16-22 Jun 2003, H.D. Doebner and V. Dobrev Eds., World Scientific, 2004. Published in [*Varna 2003, Lie theory and its applications in physics V*]{} 241-250 e-Print: hep-th/0401083.
Georgi H 2007 [*Phys. Rev. Lett.*]{} [**98**]{} 221601.
[^1]: [email protected]
[^2]: [email protected]
[^3]: An opposite view on the role of the volume element has been studied by Wilczek[@Wilczek]. Another idea of modified volume element was studied in Ref.[@Mosna].
[^4]: Fore a more detailed discussion of the role of scalars $\varphi_a(x)$ in the TMT dynamics, see the end of Sec.II.
[^5]: See also Ref.[@GK9] where we study in detail a model with dilatation symmetry which also results in the $k$-essence type dynamics
[^6]: For the case $s=+1$ and the ground state with nonzero cosmological constant see Appendix A
[^7]: Since the metric in the Einstein frame $\tilde{g}_{\mu\nu}$ is diagonal, Eq.(\[FRW\]), it is clear from the transformation (\[gmunuEin\]) that $g_{\mu\nu}$ is also diagonal.
[^8]: Note however that these oscillations do not effect the sign of the metrical volume measure $g=det(g_{\mu\nu})$ used in the underlying action (\[S-model-scalar.f.\]). This notion is useful when comparing our model with an approach developed in Refs.[@Erdem]-[@reflection2]
[^9]: for the model of Sec.IV it means that in Eq.(\[S-model-scalar.f.\]) we take the limit $b_g\to 0$, $b_{\phi}\to 0$ and $V_2\to 0$)
[^10]: Note that the pure gravity model of Sec.III is invariant both under the LES and the GES.
[^11]: A possibility of a vacuum deformation in a different approach has been shown by MacKenzie, Wilczek and Zee[@Zee]
[^12]: In the pure gravity model, Sec.III, $\Lambda$ is proportional to $b_g$
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The quantization of the electromagnetic field has successfully paved the way for the development of the Standard Model of Particle Physics and has established the basis for quantum technologies. Gravity, however, continues to hold out against physicists’ efforts of including it into the framework of quantum theory. Experimental techniques in quantum optics have only recently reached the precision and maturity required for the investigation of quantum systems under the influence of gravitational fields. Here, we report on experiments in which a genuine quantum state of an entangled photon pair was exposed to a series of different accelerations. We measure an entanglement witness for $g$ values ranging from $30 mg$ to up to $30 g$ - under free-fall as well on a spinning centrifuge - and have thus derived an upper bound on the effects of uniform acceleration on photonic entanglement. Our work represents the first quantum optics experiment in which entanglement is systematically tested in geodesic motion as well as in accelerated reference frames with acceleration $a>>g = 9{,}81 \frac{m}{s^2}$.'
author:
- Matthias Fink
- 'Ana Rodriguez-Aramendia'
- Johannes Handsteiner
- Abdul Ziarkash
- Fabian Steinlechner
- Thomas Scheidl
- Ivette Fuentes
- Jacques Pienaar
- 'Tim C. Ralph'
- Rupert Ursin
bibliography:
- 'sample.bib'
title: Experimental test of photonic entanglement in accelerated reference frames
---
[^1]
\[sec:Quantum\]Quantum Theory vs. Relativity
============================================
Einstein’s Relativity Theory (RT) and quantum theory (QT) are both important pillars of modern physics and have been thoroughly tested by many high precision experiments. RT has proven its ability to describe the Universe on the largest scales. Conversely, QT was invented to describe the Universe at atomic and subatomic scales. All currently known physical effects can be described by either the one or the other theory to very high precision. Without the scientific insights of these two theories, a modern society would not have been imaginable, taking the global positioning system (GPS) or the importance of semiconductors as two prominent examples.
In a broader view, the unification of Einstein’s theory of relativity and quantum mechanics is a long-standing problem in contemporary physics. As long as the existing descriptions of nature remain confined to their own specialized regimes, they cannot contribute a unified theory that captures physics at the boundary between these regimes. Current models are therefore considered incomplete, and can at best be combined in a patchwork way, and only under special conditions. In natural science, where theory becomes uncertain, experimental guidance is needed. It is not unusual for the first steps into a new physical regime to turn up unexpected experimental results, as occurred in the historical black-body radiation [@Rubens:1901] and Michelson-Morley [@Michelson:1887] experiments.
Dimensional arguments suggest that quantum gravity must dominate close to the Planck scale, which is still far beyond current experimental capabilities. However, there remains the distinct possibility that quantum gravity phenomena already become significant at scales that are just becoming accessible. Quantum entangled systems subjected to high and low accelerations is one regime where new physical phenomena might be expected to arise. Also, experimental investigations involving hyper- or microgravity can cause unexpected changes to physical phenomena [@Dreyer:2010]. Exposing physical systems to such extreme conditions can aid in the understanding of that system, and lead to a deeper understanding of the physical processes themselves.
It has been conjectured e.g. by Penrose [@Penrose:1996ty] and Diosi [@diosi1987universal] that gravity may cause a quantum state to collapse. This is just one of many ways that researchers have tried to include gravity into the quantum theory paradigm [@Milburn:1991tn; @Adler:2000; @Ralph:2014; @Pikovski:2015; @Blencowe:2013]. Gravity and acceleration can have observable effects on quantum entanglement [@Bruschi2014a; @Bruschi2012; @Friis2012; @Alsing2012]. However, to date there has been no systematic experimental investigation of what happens to quantum entanglement under various gravitational fields and accelerations. With the onset of new technology, techniques now exist to transgress experimental barriers and begin to test quantum physics in regimes where gravity is important. However, experimental testing of quantum phenomena in these regimes is still sparse [@Amelino-Camelia2014]. To date, first experiments have been performed using only single quanta, e.g. in the pioneering work of Colella et al in 1975 [@Colella:1975; @Abele:2012] and correlated photon pairs in microgravity [@Tang:2016]. Just recently a remarkable one-shot experiment was carried out where an entangled photon source was exposed to explosive g-forces, but not during operation [@Tang:2016].
Following these lines we report on a series of experiments in which a provably quantum state [@Bell:1964] of photon pairs entangled in their polarization degree of freedom is exposed to uniform accelerations. The acceleration is imposed by its motion, in free-fall as well as on a centrifuge.
Apart from the inherent novelty of testing entanglement under different accelerated conditions, the results of the experiment can also be extrapolated to analogous hypothetical experiments under different gravitational fields. If, for example, the equivalence principle holds, then a local test of entanglement should not reveal the difference between uniform acceleration and a gravitational field [@Zych:2015]. The experiment therefore implies bounds on the amount of decoherence as well as potential unexpected unitary transformations caused by gravity, such as in the speculative models mentioned above. Our experiment takes a broad-brush approach by aiming to establish a bound on any arbitrary effect of this kind, without focusing on any specific model.
Firstly, a crate containing an entangled photon source and two single-photon polarization detection units was dropped from $12m$ in a drop tower to realize a microgravity environment, where we reached $30$ $mg$. Secondly, that crate was mounted on an arm of a rotational centrifuge and accelerated to as high as $30$ $g$ (comparable to the surface gravity of the sun) at a maximum angular speed of $9{.}9$ $rad$ $sec^{-1}$. We evaluated the impact of the acceleration on that genuine quantum system, by constantly monitoring an entanglement witness. Our results show (assuming the equivalence principle holds) that quantum entanglement is unaffected by ambient gravitational conditions to within the resolution of our test-system. The gravity experienced by the entangled state ranges from those on the orbiting ISS, Mars, Jupiter or close to the sun[^2]. This represents the first experimental effort exposing a genuine quantum system to micro- and hyper-gravitation and extends the experimental regime in which quantum effects can be said to exist in harmony with gravity. Future experiments could be designed to show that entanglement changes when the system undergoes non-uniform acceleration. Such experiments would require much higher accelerations or massive fields.
\[sec:Methods\]Entangled photon source
======================================
In order to guarantee the stability requirements for our intended drop-tower and centrifuge experiments, we designed a simple and very rigid entangled photon source capable of withstanding the high g-forces during the deceleration and acceleration phases (see Fig \[source\]) of the experimental runs. The source (see center and right part of Fig. \[source\]) is based on spontaneous parametric down conversion (SPDC) in a periodically poled Potassium Titanyl Phosphate (ppKTP) crystal in a degenerate collinear type-II quasi-phase matching configuration and post-selection on a beam splitter (BS) [@Kuklewicz:2004]. A continuous-wave pump laser at $405$ $nm$ is focused loosely into the crystal and creates pairs of photons at $810$ $nm$, with horizontally and vertically polarized signal and idler photons, respectively. The crystal is temperature stabilized to $39{.}6^{\circ}C \pm 0{.}1^{\circ}C$, to ensure degenerate phase matching [@Kiess:1993wn]. After being separated from the pump light by two dichoric mirrors and spectral filtered using an interference filter, the two photons are coupled into a sole polarization maintaining single-mode optical fiber (PMSMF). In order to monitor possible misalignment of the beam paths during acceleration, a fraction of the pump beam was imaged onto a CMOS camera that was placed behind one of the dichroic mirrors. The length of the PMSMF and an additional neodymium-doped yttrium orthovanadate (Nd:YVO4) crystal were chosen, such that the average longitudinal walk-off between the signal and idler photons due to the birefringence of the ppKTP crystal, was compensated. This ensures, that no information about the polarization of a photon can be obtained from the timing of the SPDC photons. The PMSMF also had the benefit of ensuring polarization stability even under high accelerations during an experimental run. The PMSMF guides the photons to a beam splitter, where they are split with $50\%$ probability. As a consequence, one obtains the maximally polarization entangled ${\vert \Psi^+ \rangle}=1/\sqrt{2}({\vert H_1V_2 \rangle}+{\vert V_1H_2 \rangle})$ state for the photon pairs that are split up at the BS in mode $1$ and $2$. Here, $H$ ($V$) denote the horizontal (vertical) polarization, where we define the linear polarization state relative to the base plate of the source, as it’s orientation changes during the experiment (Fig. \[towercentri\] right). Note that the orientation of this reference frame changes during the experiment relative to the laboratory frame.
A combination of a quarter-, half- and quarter-wave plate (QHQ) is used to compensate polarization dependent phase shifts caused by the reflection at the BS and allows setting different Bell states. For the presented experiments we set the QHQ such that we end up with an rotational invariant two photon state ${\vert \Psi^- \rangle}=1/\sqrt{2}({\vert H_1V_2 \rangle}-{\vert V_1H_2 \rangle})$. An additional motorized half-wave plate followed by a polarizing beam splitter (PBS) is inserted in each output mode of the 50/50 BS to analyze the polarization of the photons in any desired linear polarization measurement basis. Finally, the photons were detected using four passively quenched semiconductor avalanche photo diodes placed in every output mode of the two PBS. Colored glass and additional interference filters were placed in front of the detectors in order to minimize background counts from the remaining pump photons and stray background light. The electronic detector signals are recorded by a time-tagging module, allowing to post-select simultaneous detection events in the two output modes of the 50/50 BS. The optical setup together with the required electronic equipment was confined in a rigid 3-level crate (Fig. \[source\] left), constituting a fully functional and ultra-stable entangled-photon box. This facilitated flexible plug-and-play installation at the drop-tower and centrifuge support structures.
\[source\] {width="39.00000%"} {width="60.00000%"}
The Bell-state fidelity of the experimental state generated at various g values was assessed by measuring a fidelity witness $F_{\Psi^-}(\hat{\rho}_{exp}) = {\langle \Psi^- \vert} \hat{\rho}_{exp} {\vert \Psi^- \rangle}$ with $F_{\Psi^-}(\hat{\rho}_{exp}) \geq F_{\Psi^-}^M (\hat{\rho}_{exp}) = \frac{1}{2}(V_{HV}+V_{DA})$ [@Blinov:2004]. Here, $V_{HV}$ and $V_{DA}$ denote the visibilities measured in the horizontal/vertical (diagonal/anti-diagonal) basis $V=\frac{N_{12}+N_{21}-N_{11}-N_{22}}{N_{12}+N_{21}+N_{11}+N_{22}}$, where $N$ is the number of coincidence detection events between detectors 1 and 2 for each measurement setting (HV,DA). For a pump power of approximately $5$ $mW$ The setup provides $280$ $kcps$ detected single counts in total and $7$ $kcps$ coincident counts, yielding a visibility of $V_{HV}=97\%$ in HV-basis and $V_{DA}=96\%$ in the DA basis. Note that the visibility in the DA basis is lower, as it is affected by changes in the polarization basis measured as well as the partial distinguishably of the two photons (non-degenerate phase matching).
\[towercentri\] {width="29.00000%"} {width="69.00000%"}
\[sec:drop\]The drop tower experiment
=====================================
A $12$ $m$ high drop was used for the micro-g experiment which provides us with $1{.}4$ $sec$ of integration time (see Fig. \[towercentri\]). Power supply was provided by a battery built into the crate capable to keep the source and detectors alive for about $1{.}5$ $h$, a wireless network antenna provided data connection from the control room to the falling crate. Guide rails are used to keep the crate on track during flight, so as not to be displaced by Coriolis-force due to the earth’s rotation during the free-fall phase. Wind speed and the guide rails resulted in drag on the crate, which reaches $55$ $km/h$ of maximum speed, enabling us to measure down to $30 \pm 3$ $mg$ after the solenoid based release mechanism was opened. For a smooth deceleration, a high stack of foam mattresses was used, which still caused a saturation of our g-sensor at $16$ $g$.
We observed no lasting degradation in count rates or source visibility after impact on the mattresses, and could repeat the experiment many times without the need for re-alignment between measurement runs. In each successive experimental run we measured either in the HV- or in the DA-visibility shown in Fig \[g:drop\], which gives $\approx 400{,}000$ single counts in total stored as time-tags on the local computer for later evaluation. During the free-fall phase the visibility never dropped below $V_{DA}=96\% \pm 2\%$.
{width="80.00000%"}
\[sec:level1\]The centrifuge experiment
=======================================
For the hyper-gravity regime, we used a centrifuge that puts the crate in rotation around a fixed vertical axis, applying an outward oriented force perpendicular to the axis of variable spin (see Fig \[towercentri\] right). The centrifuge consists of two $3$ $m$ long arms with an articulated platform at the end (Fig. \[towercentri\] right), so that they swing outwards at increasing angular velocities of the centrifuge. The crate was mounted in one of the gondolas with a $37$ $kg$ counterweight on the other side. Power supply and data-connection of the crate was provided by means of a sliding contact at the axis of the centrifuge to the control room.
\[g:centri\] {width="80.00000%"}
Acceleration in the centrifuge experiment was varied by the angular velocity from $1$ $g$ to up to $30$ $g$ in approx. $5 g$ steps. The maximal tangential speed at the source layer is $174$ $km/h$. With respect to the stability of the optical setup, a movement of the UV-pump beam spot on the CCD camera of up to $2$ $pixel$ corresponding to $38$ $arcsec$ was monitored. This is well within the tolerances of the PMSM-fiber coupling lens system design. No drop in the count rates was observed. For each g value (set by the angular speed of the centrifuge) we evaluated the visibility for several minutes in two complementary basis states HV and DA (see Fig \[g:centri\]). Note that a small reduction of the DA visibility stems from the fact that the temperature of the crystal was not stable at high accelerations. This was due to high wind speed of about $174$ $km/h$ which effectively cooled the system as indicated by the temperature sensor of the oven.
\[sec:Discussion\]Discussion
============================
We have conducted a micro-gravity (drop-tower) and a hyper-gravity (centrifuge) experiment, in which we used a genuine quantum mechanical system to evaluate whether the amount of entanglement varies at different accelerations. We implemented an entangled photon source together with a detection system and the required electronics in one single crate. We contiguously tested the quantumness of the system by measuring a witness which imposes a bound on the miminmum Bell-state fidelity, from the measured visibility of polarization correlations in two mutually unbiased measurement basis (HV and DA).
Fig. \[g:total\] shows a summary graph displays the g-value vs. the entanglement fidelity measured for both conducted experiments in the falling tower and the centrifuge. A fidelity of above $96 \%$ at each acceleration level is achieved. Each data point was calculated using more than $10{,}000$ coincidence counts leading to an $3$ $\sigma$ error bar of $0{.}25 \% $. The measured fidelity is limited by numerous sources of systematic errors, such as birefringence-induced transformations of the entangled photon state, temperature-dependent spectral characteristics of the SPDC source [@Steinlechner:2014], as well as accidental coincidence counts [@Takesue:2010].
Our experiment can rule out any hypothetical variation of entanglement to within the precision of the current measurement apparatus. The resolution in our experiment was limited by the short experimental time as well as residual drifts of the crystal-temperature and alignment issues, which are not correlated with the acceleration itself. The error bars shown in the graphs are calculated considering Poissonian statistics, as well as systematical errors due to temperature fluctuations. Due to insufficient temperature stabilisation of the crystal ($\pm 0{.}1^{\circ}C$) and gusts of cold air in the centrifuge, an additional error of $\pm 0{.}6\%$ was observed in the DA-Visibility. This corresponds to an uncertainty of $\pm 0{.}3\%$ in the Fidelity, in addition to statistical errors. The total error is well within the average fidelity taking all acquired data into account.
\[g:total\] {width="80.00000%"}
In conclusion, we have used the techniques originally developed for quantum optics high precision experiments, to search for first experimental indications of unexpected relativistic and gravitational effects in quantum systems. Using the entangled system in the micro- and hyper-gravity condition experiments described above, we found that the phase and the entanglement of the state were not correlated to the acceleration of the system. These experiments therefore rule out any influence on the polarization entangled two-photon quantum state that could hypothetically cause a reduction in fidelity of more than $1{.}08\%$.
Our study tested photonic quantum entanglement in the case of flat space-time where the system undergoes uniform acceleration. Within this experiment we have shown that quantum entanglement should persist in a variety of gravitational settings ranging across the solar system.
Such conditions would also accompany a rocket launch into space and are also relevant to subsequent quantum optics experiments carried out in space, under low-g conditions. We believe it is the first time that such a technology involving an entangled photon source was functional under these conditions and shown to be capable of withstanding the stresses.
Our experiment demonstrates the extent to which state-of-the-art quantum hardware can be exposed to such harsh operational conditions and was intended to stimulate research on theories beyond the current paradigm which can be tested with the kind of experiments presented here. Our experimental platform represents a first test-bed that can readily be upgraded for measurements with higher precision, and higher-dimensional degrees of freedom, such as energy-time entanglement. We also envisage bringing such a system very close to zero-g conditions for several tens of seconds and reaching hyper-gravity environments of up to $150$ $g$ for many hours of operation.
We are especially thankful to Prof. Martin Tajmar and the Technical University Dresden (Germany) for their support in the drop-tower experiment. The centrifuge experiment was supported by Jens Schiefer and Clemens Greiner from AMST in Ranshofen (Austria). The Authors thank FFG-ALR (contract Nr. 844360), ESA (contract Nr. 4000112591/14/NL/US), FWF (P24621-N27) as well as the Austrian Academy of Sciences for their financial support.
[^1]: Correspondence and requests for materials should be addressed to Matthias Fink and Rupert Ursin.
[^2]: Of course, these environments would pose other challenges to entanglement, such as exposure to radiation and strong magnetic fields, which we do not address here.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The excursion set theory based on spherical or ellipsoidal gravitational collapse provides an elegant analytic framework for calculating the mass function and the large-scale bias of dark matter haloes. This theory assumes that the perturbed density field evolves stochastically with the smoothing scale and exhibits Markovian random walks in the presence of a density barrier. Here we derive an analytic expression for the halo bias in a new theoretical model that incorporates non-Markovian extension of the excursion set theory with a stochastic barrier. This model allows us to handle non-Markovian random walks and to calculate perturbativly these corrections to the standard Markovian predictions for the halo mass function and halo bias. Our model contains only two parameters: $\kappa$, which parameterizes the degree of non-Markovianity and whose exact value depends on the shape of the filter function used to smooth the density field, and $a$, which parameterizes the degree of stochasticity of the barrier. Appropriate choices of $\kappa$ and $a$ in our new model can lead to a closer match to both the halo mass function and halo bias in the latest $N$-body simulations than the standard excursion set theory.'
author:
- |
Chung-Pei Ma$^{1}$, Michele Maggiore$^{2}$, Antonio Riotto$^{3,4}$, Jun Zhang$^{5}$\
\
$^{1}$Department of Astronomy, University of California, Berkeley, CA 94720, USA\
$^{2}$Département de Physique Théorique, Université de Genève, CH-1211 Geneva, Switzerland\
$^{3}$CERN, PH-TH Division, CH-1211, Genève 23, Switzerland\
$^{4}$INFN, Sezione di Padova, Via Marzolo 8, I-35131 Padua, Italy\
$^{5}$Texas Cosmology Center, University of Texas, Austin, TX 78712 USA\
title: 'The Bias and Mass Function of Dark Matter Haloes in Non-Markovian Extension of the Excursion Set Theory'
---
\[firstpage\]
cosmology: theory
Introduction {#intro}
============
Dark matter haloes typically form at sites of high density peaks. The spatial distribution of dark matter haloes is therefore a biased tracer of the underlying mass distribution. A standard way to quantify this difference between haloes and mass is to use a halo bias parameter $b_h$, which can be defined as the ratio of the overdensity of haloes to mass, or as the square root of the ratio of the two-point correlation function (or power spectrum) of haloes to mass. Like the halo mass function, analytic expressions for the halo bias can be obtained from the excursion set theory [@Bond] based on the spherical gravitational collapse model [@CK89; @MW96]. In the excursion set theory, the density perturbation evolves stochastically with the smoothing scale, and the problem of computing the probability of halo formation is mapped into the first-passage time problem in the presence of a (constant) barrier. The approach to the clustering evolution is based on a generalization of the peak-background split scheme [@Bardeen], which basically consists in splitting the mass perturbations into a fine-grained (peak) component filtered on a scale $R$ and a coarse-grained (background) component filtered on a scale $R_0\gg R$. The underlying idea is to ascribe the collapse of objects on small scales to the high frequency modes of the density fields, while the action of large-scale structures of these non-linear condensations is due to a shift of the local background density.
Comparison with $N$-body simulations finds that the spherical collapse model underpredicts the halo bias for low mass halos [@jing98; @ST99]. The discrepancy reaches a factor of $\sim 2$ at $M\sim 0.01 M_*$, where $M_*$ is the characteristic nonlinear mass scale (defined by $\sigma(M_*)=1$ where $\sigma^2(M)$ is the variance of the density field in a volume of radius $R$ containing the mass $M$). [@SMT01] obtained an improved formula for the halo bias by using a moving barrier whose scale-dependent shape is motivated by the ellipsoidal gravitational collapse model. Compared to the spherical collapse model, this formula predicts a lower bias at the high mass end and a higher bias at the low mass end (see Fig. 1 below). The resulting bias is shown to be too high at the low mass end by $\sim 20$% compared with simulation results. Further modifications have been introduced that either used the functional form of [@ST99] or [@SMT01] with new fitting parameters (e.g., @tinker05), or proposed new fitting forms altogether (e.g., @SW04 [@pillepich10; @tinker10]).
Our goal in this paper is not to improve on the accuracy of the fits to the halo bias, but rather to gain deeper theoretical insight by deriving an analytical expression for the halo bias using a new model. This model modifies the excursion set theory by incorporating non-Markovian random walks in the presence of a stochastic barrier. It is based on a path integral formulation introduced in [@MR1; @MR2], which provides an analytic framework for calculating perturbatively the non-Markovian corrections to the standard version of the excursion set theory. Mathematically, the non-Markovianity in the theory is related to the choice of the filter function necessary to smooth out the density contrast. As soon as the filter function is different from a step (tophat) function in momentum space, the excursion of the smoothed density contrast is non-Markovian, namely, every step depends on the previous ones and the random walk acquires memory. As the computation of the bias parameter $b_h$ amounts to computing the first crossing rate with a non-trivial initial condition at a large, but not infinite, radius, the non-Markovianity makes the calculation much harder than the Markovian case.
Furthermore, the critical value for collapse in our model is itself assumed to be a stochastic variable, whose scatter reflects a number of complicated aspects of the underlying dynamics. The gravitational collapse of haloes is a complex dynamical phenomenon, and modeling it as spherical, or even as ellipsoidal, is a significant oversimplification. In addition, the very definition of what is a dark matter halo, both in simulations and observationally, is a non-trivial problem. [@MR2] proposed that some of the physical complications inherent to a realistic description of halo formation can be included in the excursion set theory framework, at least at an effective level, by taking into account that the critical value for collapse is itself a stochastic variable.
In Section 2 we review briefly the derivation for the halo mass function in the Markovian excursion set theory (Sec 2.1) and the path integral approach used to introduce non-Markovian terms (Sec 2.2) and stochastic barriers (Sec 2.3) into the theory. Section 3 is devoted to the discussion of halo bias, including a review of the standard derivation in the Markovian case (Sec 3.1), and a summary of our new derivation in the non-Markovian model (Sec 3.2 and 3.3). The details of how the halo bias is calculated from the conditional probability for two barrier crossings in the new model is provided in the Appendix. In Section 4 we compare the predictions for the halo bias and mass function in our new model with those from the Markovian model (both spherical and ellipsoidal collapse) and $N$-body simulations.
Non-Markovian Extension and Stochastic Barrier
==============================================
In this section we review the main points of the excursion set theory [@Bond] and then summarize how to introduce non-Markovian terms in the presence of a stochastic barrier [@MR1; @MR2].
Brief review of the excursion set theory
----------------------------------------
The basic variable is the smoothed density contrast, \[dfilter\] ([**x**]{},R) =d\^3x’ W(|[**x**]{}-[**x**]{}’|,R) ([**x**]{}’) , where $\delta({\bf x}) = \rho({\bf x})/\bar\rho - 1$ is the density contrast about the mean mass density $\bar\rho$ of the universe, $W(|{\bf
x}-{\bf x}'|,R)$ is the filter function, and $R$ is the smoothing scale. We are interested in the evolution of $\d({\bf x},R)$ with smoothing scale $R$ at a fixed point ${\bf x}$ in space, so we suppress the argument ${\bf
x}$ from this point on. It is also convenient to use, instead of $R$, the variance $S$ of the smoothed density field defined by $$S(R) \equiv \sigma^2(R) = \int \frac{d^3k}{(2\pi)^3}\, P(k) \tilde{W}^2(k, R) \,,
\label{S}$$ where $P(k)$ is the power spectrum of the matter density fluctuations in the cosmological model under consideration, and $\tilde{W}$ is the Fourier transform of the filter function $W$. A smoothing radius $R=\infty$ corresponds to $S=0$ and, in hierarchical models of structure formation such as the $\Lambda$CDM model, $S$ is a monotonically decreasing function of $R$. We can therefore use $S$ and $R$ interchangeably and denote our basic variable by $\d(S)$.
If the filter function is taken to be a tophat in momentum space, $\d(S)$ then satisfies a simple Langevin equation with $S$ playing the role of a “pseudo-time” [@Bond] \[Langevin1\] = (S) , where $\eta(S)$ represents a stochastic “pseudo-force” whose two-point correlation statistic obeys a Dirac-delta function: \[Langevin2\] (S\_1)(S\_2)=\_D (S\_1-S\_2) . It then follows that the function $\Pi(\d_0;\d;S)$, which gives the probability density of reaching a value $\d$ at “time” $S$ starting at a value $\d_0$ at $S=0$, satisfies the Fokker-Planck equation \[FPdS\] = . In [@Bond] this equation was supplemented by the boundary condition \[bc\] .(,S)|\_[=\_c]{}=0 to eliminate the trajectories that have reached the critical value $\d_c$ for collapse.
The corresponding solution of the Fokker-Planck equation is \[PiChandra\] (\_0;;S)= $$e^{-(\d-\d_0)^2/(2S)}- e^{-(2\d_c-\d_0-\d)^2/(2S)}$$ . The probability ${\cal F}(S) dS$ of first crossing the threshold density $\delta_c$ between “time” $S$ and $S+dS$ is then given by $${\cal F}(S) = -\int_{-\infty}^{\d_c}d\d\, \frac{\pa\Pi}{\pa S}
= \frac{1}{\sqrt{2\pi}}\, \frac{\delta_c}{S^{3/2}} e^{-\delta_c^2/(2S)}\,,
%\label{defcalF}
%\label{firstcrossT}
\label{calFmarkov}$$ where we have set $\d_0=0$. The number density of virialized objects with mass between $M$ and $M+dM$ is related to the first crossing probability between $S$ and $S+dS$ by $$\begin{aligned}
\frac{dn(M)}{dM} dM & = &\frac{\bar{\rho}}{M} {\cal F}(S) dS \nonumber \\
&=& \sqrt{\frac{2}{\pi}}\,\frac{\d_c}{\s}\, e^{-\d_c^2/(2\s^2)}
\, \frac{\bar{\rho}}{M^2} \frac{d\ln\s^{-1}}{d\ln M} dM \nonumber \\
&\equiv & f(\nu)\, \frac{\bar{\rho}}{M^2} \frac{d\ln\s^{-1}}{d\ln M} dM \, ,
\label{massfunction}\end{aligned}$$ where $\sigma=S^{1/2}$ is defined in equation (\[S\]) and $\nu \equiv
\delta_c/\sigma$. This expression reproduces the mass function of [@PS], including the correct overall normalization that had to be adjusted by hand in [@PS]. We will use the dimensionless function $f(\nu)$ defined in equation (\[massfunction\]) to denote the halo mass function below.
Non-Markovian extension
-----------------------
As already discussed in [@Bond], a difficulty of the excursion set approach in Sec 2.1 is that an unambiguous relation between the smoothing radius $R$ and the mass $M$ of the corresponding collapsed halo only exists when the filter function is a tophat in coordinate space: $M(R)=(4/3)\pi
R^3 \rho$. For all other filter functions (e.g., tophat in momentum space, Gaussian), it is impossible to associate a well-defined mass $M(R)$ (see also the recent review @zentner07). More importantly, $\d(S)$ obeys a Langevin equation with a Dirac delta noise as in when the filter function is a tophat in momentum space. Otherwise, the evolution of $\d$ with the smoothing scale becomes non-Markovian, and the distribution function $\Pi
(\d_0;\d;S)$ of the trajectories no longer obeys the Fokker-Planck equation, nor any local generalization of it. In this case, $\Pi
(\d_0;\d;S)$ obeys a complicated equation that is non-local with respect to the variable $S$ [@MR1].
To deal with this problem, [@MR1] proposed a “microscopic” approach, in which one computes the probability associated with each trajectory $\d(S)$, and sums over all relevant trajectories. As with any path integral formulation, it is convenient to discretize the time variable and to take the continuum limit at the end. Therefore we discretize the interval $[0,S]$ in steps $\D S=\eps$, so $S_k=k\eps$ with $k=1,\ldots n$, and $S_n\equiv S$, and a trajectory is defined by the collection of values $\{\delta_1,\ldots ,\delta_n\}$, such that $\delta(S_k)=\delta_k$. All trajectories start at a value $\d_0$ at time $S=0$.
The basic quantity in this approach is the probability density in the space of trajectories, defined as \[defW\] W(\_0;\_1,…,\_n;S\_n) \_D ((S\_1)-\_1)…\_D ((S\_n)-\_n)where $\d_D$ denotes the Dirac delta function. In terms of $W$ we define \[defPi\] \_ (\_0;\_n;S\_n) \_[-]{}\^[\_c]{} d\_1…d\_[n-1]{} W(\_0;\_1,…,\_n;S\_n) where $S_n=n\eps$, and $\Pi_{\eps} (\delta_0;\d;S)$ is the probability density of arriving at the “position" $\d$ in a “time” $S$, starting from $\delta_0$ at time $S_0=0$, through trajectories that never exceeded $\delta_c$. The problem of computing the distribution function of excursion set theory is therefore mapped into the computation of a path integral with a boundary at $\d=\d_c$.
The probability density $W$ can be computed in terms of the connected correlators of the theory. When the density field $\delta$ is a Gaussian random variable, only the two-point connected function is non-zero, and one finds \[WnNG0\] &&W(\_0;\_1,…,\_n;S\_n)=\
&& … e\^[ i\_[i=1]{}\^n\_i\_i -\_[i,j=1]{}\^n\_i\_j \_i\_j\_c]{} ,and $\delta_i\equiv\delta(S_i)$. We will restrict the discussion here to the Gaussian case since higher-order connected correlators must be included in the non-Gaussian case [@MR3; @MR4].
Consider first the case of a tophat filter in momentum space, so the evolution of $\d(S)$ is Markovian and obeys . Then one can show that the connected two-point correlator is given by (S\_i)(S\_j)\_c=[min]{}(S\_i,S\_j), and the integrals over $d\lambda_1,\ldots ,d\lambda_n$ in can be performed explicitly to give \[W\] W\^[ gm]{}(\_0;\_1,…,\_n;S\_n)= e\^[- \_[i=0]{}\^[n-1]{} (\_[i+1]{}-\_i)\^2]{} where the superscript “gm” stands for “Gaussian and Markovian.” Inserting this expression into it can be shown [@MR1] that, in the continuum limit, the corresponding distribution function $\Pi^{\rm
gm}_{\eps=0} (\delta_0;\delta;S)$ satisfies the Fokker-Planck equation (\[FPdS\]) as well as the boundary condition (\[bc\]), and therefore we recover the standard result (\[PiChandra\]) of excursion set theory.
The interesting case is to generalize the computation above to filter functions different from the conventional tophat in momentum space. The two-point correlator depends on the filter function. For a Gaussian filter and a tophat filter in coordinate space, for instance, we find \[rever\] (S\_i)(S\_j)= [ min]{}(S\_i,S\_j) + (S\_i,S\_j) , where $\D(S_i,S_j)=\D(S_j,S_i)$ and, for $S_i\leq S_j$, the function $\D(S_i,S_j)$ is well approximated by \[approxDelta\] (S\_i,S\_j) , with $\kappa\approx 0.35$ for a Gaussian filter and $\kappa\approx 0.44$ for a tophat filter in coordinate space. The parameter $\kappa$ gives a measure of the non-Markovianity of the stochastic process, and the computation of the distribution function $\Pi (\delta_0;\delta;S)$ can be performed order by order in $\kappa$. The technique necessary for evaluating the path integral in to first order in $\kappa$ has been developed in [@MR1], and will be further discussed below. To first order in the non-Markovian corrections, the resulting first-crossing rate becomes \[Ffinal\] [F]{}(S)= e\^[-\_c\^2/(2S)]{}+ $0,\frac{\delta_c^2}{2S}$ , where $\G(0,z)$ is the incomplete Gamma function. For $\kappa=0$ one recovers the Markovian result in equation (\[calFmarkov\]). The halo mass function is then obtained by substituting this expression for ${\cal
F}$ into equation (\[massfunction\]).
Stochastic barrier
------------------
The constant barrier $\d_c\simeq 1.686$ in the spherical collapse model is a significant oversimplification of the complex dynamics leading to halo formation and growth. Such a model can be improved in various ways. For instance, the excursion set theory results for the mass function have been shown to match more closely those from $N$-body simulations by considering a moving barrier whose shape is motivated by the ellipsoidal collapse model [@ST99; @SMT01; @ST02; @DSMR]. The equations are summarized in Table 1. The parameters $a, b, c$ are fixed by fit to $N$-body simulations, while $A$ is fixed by the normalization condition on the halo mass function. As already remarked in [@SMT01], these expressions can be obtained from a barrier shape that is virtually identical to the ellipsoidal collapse barrier, except for the factor of $a$, which is not a consequence of the ellipsoidal collapse model. In fact, the ellipsoidal collapse model reduces to the spherical collapse model in the large mass limit. In this limit the mass function is determined by the slope of the exponential factor, so even in an ellipsoidal collapse model we must have $a=1$, as in the spherical model. However, numerical simulations show that $a<1$ and its precise value also depends on the details of the algorithm used for identifying halos in the simulation, e.g., the link length in a friends-of-friends (FOF) halo finder, or the critical overdensity in a spherical density (SO) finder.
A physical understanding of the parameter $a$ is given by a second independent improvement of the spherical collapse model, the diffusing barrier model proposed in [@MR2]. These authors suggested that at least some of the physical complications inherent to a realistic description of halo formation, which involves a mixture of smooth accretion, violent encounters and fragmentations, can be included in the excursion set theory framework by assuming that the critical value for collapse is itself a stochastic variable, whose scatter reflects a number of complicated aspects of the underlying dynamics. In the simple example of a barrier performing a random walk with diffusion coefficient $D_B$ around the spherical collapse barrier, one finds indeed a mass function in which $\d_c$ is effectively replaced by $a^{1/2}\d_c$, with $a=1/(1+D_B)$, while at the same time $\kappa$ is replaced by $a\kappa$ (see Table 1).
[lccl]{} Model & Mass Function $f(\nu)$ & Halo Bias $b_h(\nu)$ & Parameters\
Spherical Collapse & $\sqrt{\frac{2}{\pi}} \nu \exp\left(-\frac{\nu^2}{2}\right)$ & $1 + \frac{\nu^2-1}{\delta_c}$ & $\delta_c=1.686$\
\
Ellipsoidal Collapse & $A \sqrt{\frac{2}{\pi}} \sqrt{a}\nu
\exp\left(-\frac{a\nu^2}{2}\right) \left[ 1+(a\nu^2)^q \right]$ & $1 + \frac{1}{\sqrt{a}\delta_c} \left[ \sqrt{a}(a\nu^2)
+ \sqrt{a} b (a\nu^2)^{1-c} \right. $ & $A=0.322, q=-0.3$\
&& $\left. -\frac{(a\nu^2)^c}{(a\nu^2)^c + b(1-c)(1-c/2)}\right]$ & $a=0.707, b=0.5, c=0.6$\
\
Non-Markovian & $\sqrt{\frac{2}{\pi}} \left[(1-\kappa)\nu
\exp\left(-\frac{\nu^2}{2}\right) + \kappa \frac{\nu}{2}
\Gamma\left(0,\frac{\nu^2}{2}\right) \right]$ & $1+ \frac{1}{\delta_c \left[ 1-\kappa+ \frac{\kappa}{2} e^{\nu^2/2}\Gamma(0,\nu^2/2) \right] }
\left\{ \nu^2-1 \right.$ & $\kappa=0$ for tophat-$k$ filter\
&& $ \left. +\frac{\kappa}{2} \left[2- \exp\left(\frac{\nu^2}{2}\right)\Gamma\left(
0,\frac{\nu^2}{2}\right) \right] \right\}$ & $\kappa=0.35$ for Gaussian\
&&& $\kappa=0.44$ for tophat-$x$\
\
Non-Markovian & $\kappa \rightarrow a\kappa, \quad \nu \rightarrow \sqrt{a}\nu$ & $\kappa \rightarrow a\kappa, \quad \nu \rightarrow \sqrt{a}\nu, \quad
\delta_c \rightarrow \sqrt{a} \delta_c $ & $a=\frac{1}{1+D_B}$\
$+$ Stochastic Barrier & && $D_B=$ diffusion coefficient\
Halo Bias
=========
We now apply the technique in Section 2 to the computation of the halo bias, including the non-Markovian corrections with stochastic barriers. We sketch here the main steps of the computations, leaving the details to the Appendix.
Conditional probability: the Markovian case
-------------------------------------------
To compute the bias, we need the probability of forming a halo of mass $M$, corresponding to a smoothing radius $R$, under the condition that the smoothed density contrast on a much larger scale $R_m$ has a specified value $\d_m=\delta(R_m)$. We use ${\cal F}(S_n|\d_m,S_m)$ to denote the conditional first-crossing rate. This is the rate at which trajectories first cross the barrier at $\d=\d_c$ at time $S_n$, under the condition that they passed through the point $\d=\d_m$ at an earlier time $S_m$. We also use the notation ${\cal F}(S_n|0)\equiv {\cal F}(S_n|\d_m=0,S_m=0)$, so ${\cal F}(S_n|0)$ is the first-crossing rate when the density approaches the cosmic mean value on very large scales.
The halo overdensity in Lagrangian space is given by (@Kaiser84 [@EFWD; @CK89; @MW96]; see also @zentner07 for a review) 1+\_[halo]{}\^L= . In a sufficiently large region, we have $S_m\ll S_n\equiv S$ and $\d_m\ll
\d_c$. Then, using the first crossing rate of excursion set theory and retaining only the term linear in $\d_m$, we obtain \_[halo]{}\^L=\_m , where $\nu =\d_c/\s$. After mapping to Eulerian space, one finds $\d_{\rm
halo} \approx 1 + \d_{\rm halo}^L$ in the limit of small overdensity $\d_m\simeq \d$ [@MW96], and \[bhmarkov\] b\_h()=1+ .
Non-Markovian corrections
-------------------------
We now use the path integral formalism discussed in Section 2.2 to compute the non-Markovian corrections to the halo bias. The relevant quantity for our purposes is the conditional probability $$\begin{aligned}
&&P(\delta_n,S_n| \delta_m,S_m)\equiv\label{Pbias}\\
&=&\frac{
\int_{-\infty}^{\delta_c}d\delta_1\cdots
\widehat{d\delta_{m}}\cdots d\delta_{n-1}
W\left(\delta_0=0;\delta_1,\ldots, \delta_n;S_n\right)
}{\int_{-\infty}^{\delta_c}d\delta_1\cdots d\delta_{m-1}
W\left(\delta_0=0;\delta_1,\cdots,\delta_m;S_m\right)}\, ,\nonumber\end{aligned}$$ where the hat over $d\delta_{m}$ means that $d\delta_{m}$ must be omitted from the list of integration variables. The numerator is a sum over all trajectories that start from $\d_0=0$ at $S=0$, have a given fixed value $\d_m$ at $S_m$, and a value $\d_n$ at $S_n$, while all other points of the trajectory, $\d_1, \ldots ,\d_{m-1},\d_{m+1},\ldots \d_{n-1}$ are integrated from $-\infty$ to $\d_c$ . The denominator gives the appropriate normalization to the conditional probability.
Similarly to , the conditional first-crossing rate ${\cal F}(S_n|\d_m,S_m)$ is obtained from the conditional probability $P (\delta_n,S_n| \d_m,S_m) $ using \[firstcrossTcond\] [F]{} (S\_n|\_m,S\_m) = -\_[-]{}\^[\_c]{}d\_n . In the Gaussian and Markovian case, the probability $W^{\rm gm}$ satisfies \[facto\] &&W\^[gm]{}(\_0;\_1,…,\_n; S\_n)= W\^[gm]{}(\_0;\_1,…, \_m;S\_m)\
&&W\^[gm]{}(\_m; \_[m+1]{}, …,\_n; S\_n-S\_m), and $P (\delta_n,S_n| \d_m,S_m) $ in becomes identical to the probability of arriving in $\d_n$ at time $S_n$, starting from $\d_m$ at time $S_m$, which is given by (with $\d_m$ identified with $\d_0$ at $S=S_n-S_m$), and we therefore recover the excursion set theory result.
We have computed the non-Markovian corrections to this result for the case of Gaussian fluctuations and a tophat filter in coordinate space, i.e. with the two-point function given in . The computation is quite involved, and we leave the details to the Appendix. Taking finally $S_m= 0$ and developing to first order in $\d_m\equiv \d_0$, which is the case relevant to the computation of the bias, for the conditional first crossing rate we find &&[F]{}(S|\_0,S\_0=0)= e\^[-\_c\^2/(2S)]{}\
&& { $1-\kappa+\frac{\kappa}{2} e^{\nu^2/2}\Gamma(0,\nu^2/2)
$.\[calFfinal\]\
&&.+$$(\nu^2-1)+\frac{\kappa}{2} \(2-
e^{\nu^2/2}\Gamma(0,\nu^2/2)\)$$} . From this we obtain the Lagrangian halo bias $b_h^L$ and the Eulerian halo bias $b_h$: \[bh\] b\_h () &=& 1 + b\_h\^L\
&=& 1 +\
&&{ (\^2-1)+ $$2- e^{\nu^2/2}\Gamma(0,\nu^2/2)$$ } .For $\kappa=0$ we recover the usual Markovian result in equation (\[bhmarkov\]). For $\nu\gg1$, corresponding to large masses, $e^{\nu^2/2}\Gamma(0,\nu^2/2)\rightarrow 2/\nu^2$, and the above expression simplifies to \[bhlargenu\] b\_h () & & 1 + $\frac{1+\frac{\kappa}{\nu^2} }{1-\kappa+\frac{\kappa}{\nu^2} } $\
& & , 1 .
Adding a stochastic barrier
---------------------------
In the presence of the stochastic barrier described in Section 2.3, we can easily modify the halo bias in equation (\[bh\]) using the substitution $\d_c \rightarrow a^{1/2}\d_c$ and $\kappa \rightarrow a\kappa$, where the parameter $a$ is related to the diffusion coefficient of the barrier. The Eulerian halo bias finally reads \[bhfinal\] b\_h()&=&1+\
&&{ (a\^2-1)+ $$2- e^{a\nu^2/2}\Gamma(0,a\nu^2/2)$$ } .We note that equation (\[bhfinal\]) raises the halo bias in the large $\nu$ (i.e. large halo mass) region compared to the bias in the ellipsoidal collapse model [@SMT01], and in fact get closer to the spherical result. For $\nu\gg1$, using again the asymptotic expression of the incomplete Gamma function and keeping only the leading term $\sim\nu^2$, our result reads b\_h() , 1 which differs from the asymptotic spherical collapse result by an overall factor of $a^{1/2}/(1-a\kappa)$.
Comparisons
===========
We now compare the predictions for the Eulerian halo bias $b_h(\nu)$ from our non-Markovian and stochastic barrier model with those from the standard excursion set theory as well as $N$-body simulations. We present the results for the halo mass function in parallel since as we have shown in Sections 2 and 3, an analytic theory for halo formation provides simultaneous predictions for the mass function and bias.
Our model contains two parameters: (1) $\kappa$, which parameterizes the degree of non-Markovianity and its exact value depends on the filter function used to smooth the density, e.g., $\kappa=0, 0.35, 0.45$ for tophat in momentum-space, Gaussian, and tophat in coordinate filter, respectively; (2) $a$, which parameterizes the stochasticity of the diffusing barrier with the diffusion coefficient $D_B$, where $a=1/(1+D_B)$. There is no a priori reason to favor one filter to another, nor is the choice of filters limited to the three functional forms given above. Furthermore, we recall that in [@MR1] the scaling $\kappa
\rightarrow a\,\kappa$ is obtained under the simplified assumption that the barrier makes a simple Brownian motion around the spherical collapse barrier; for more complicated stochastic motions of the barrier (and also for fluctuations around the ellipsoidal barrier), the rescaling of $\kappa$ might be different. For these reasons, we prefer in this work to treat both $a$ and $\kappa$ as free parameters and use simulations to calibrate their values.
For the simulations, we choose to compare with the fits to the latest $N$-body simulations [@tinker08; @tinker10]. These papers provide detailed discussions about the comparison within the different $N$-body results and numerical issues such as the dependence of the results on simulation resolution and halo definitions and finders.
Fig. \[fig:compare\_bias\] shows the results for the bias (left panels) and mass function (right panels) as a function of halo mass (as parameterized by $\nu=\delta_c/\sigma$). The bottom panels show the fractional difference between each model prediction and the fit to $N$-body results. For the spherical and ellipsoidal models, we use the standard parameters listed in Table 1. For our model, we plot the predictions using $a=0.818$ and $\kappa=0.23$, which provide a good match (within $\sim
20$%) to both the bias and mass function from $N$-body, and in particular to the mass function at high mass.
We note that since $\kappa$ is related to the two-point correlation function of the density field (see eq. 91 of [@MR1]), it can in principle depend on the cosmological model. As discussed in footnote 10 of [@MR1], however, the dependence of $\kappa$ on the cosmological parameters is extremely weak. Both $\kappa$ and $a$ can therefore be treated as universal parameters whose values can be calibrated with simulations.
Conclusions
===========
We derived an analytic expression (eq. \[bhfinal\] with $a=1$) for the halo bias in the non-Markovian extension of the excursion set theory. This new model is based on a path integral formulation introduced in [@MR1], which provides an analytic framework for handling the non-Markovian nature of the random walk and for calculating perturbatively the non-Markovian corrections to the standard version of the excursion set theory. The degree of non-Markovianity in our theory is parameterized by a single variable, $\kappa$, whose exact value depends on the shape of the filter function used to smooth the density field, e.g., $\kappa=0$ for a tophat filter in momentum space, $\kappa\approx 0.35$ for a Gaussian filter, and $\kappa \approx 0.44$ for a tophat filter in coordinate space.
As already discussed in [@Bond; @robertson09; @MR1], changing the filter function in the spherical collapse model from a tophat in momentum space to a tophat in coordinate space does not help alleviate the discrepancy in the mass function between the Press-Schechter model and $N$-body simulations. In another word, had we plotted the corresponding curves in Fig. 1 using $a=1$ (i.e. a constant barrier height as in the spherical collapse model) and $\kappa=0.44$ (for a tophat filter in coordinate space), the bias would be too high by up to $\sim 80$% at large $\nu$ compared to the $N$-body result, and the mass function would be too low by up to $\sim 80$% at large mass (see also Fig. 9 of @MR1). Using a Gaussian filter reduces $\kappa$ by only $\sim 20$% and has only a minor effect. Additional modifications to the theory beyond including non-Markovian corrections must therefore be introduced to match the $N$-body results.
We have explored one such modification by allowing the barrier height itself to be a stochastic variable (Sec. 2.3 and @MR2). This new ingredient introduces a second parameter $a$ in our theory, as summarized in Table 1. As the solid black curves in Fig. 1 illustrate, an appropriate choice of these two parameters for the non-Markovian correction and stochastic barrier (e.g., $\kappa=0.23$ and $a=0.818$) produces a good match to $N$-body results for [*both*]{} the halo mass function and bias, with fractional deviations being $\sim 20$% or less. In comparison, the ellipsoidal collapse model contains four fitting parameters ($a, b, c$ and $q$; see Table 1) and does a comparable job at matching $N$-body simulations (dot-dashed magenta curves in Fig. 1). Further improvement to the model presented in this paper can be obtained by computing the bias through the excursion set theory starting from the ellipsoidal model and including the effects of non-Markovianity. A step towards this computation has been taken recently in [@DSMR], where the first crossing rate has been computed using the path integral method for a generic barrier. As the next step, one could envisage to combine the diffusing barrier model with the ellipsoidal model, i.e. consider a barrier that fluctuates around an average value given by the ellipsoidal collapse model, rather than around the constant value provided by the spherical collapse model as done in this paper. We expect this combined model to be able to provide an even closer match to $N$-body results than the $\sim
20$% accuracy achieved by either model alone.
Acknowledgments {#acknowledgments .unnumbered}
===============
Support for CPM is provided in part by the Miller Institute for Basic Research in Science, University of California Berkeley. The work of MM is supported by the Fonds National Suisse. We thank Roman Scoccimarro for a useful comment.
Bardeen J.M., Bond J.R., Kaiser N. and Szalay A., 1986, ApJ 304, 15.
Bond J., Cole S., Efstathiou G., Kaiser N., 1991, ApJ, 379, 440.
Cole S. & Kaiser N., 1989, MNRAS, 237, 1127.
De Simone A., Maggiore M. and Riotto A., arXiv:1007.1903 \[astro-ph.CO\], submitted to MNRAS.
Efstathiou G., Frenk C. S., White S. D. M., Davis, M., 1988, MNRAS 235, 715.
Jing Y. P., 1998, ApJ, 503, L9
Kaiser N., 1984, ApJL, 284, L9
Maggiore, M. & Riotto, A., 2010, ApJ, 711, 907.
Maggiore, M. & Riotto, A., 2010, ApJ, 717, 515
Maggiore, M. & Riotto, A., 2010, ApJ, 717, 526
Maggiore, M. & Riotto, A., 2010, MNRAS, 405, 1244
Mo H. & White S., 1996, MNRAS, 282, 347.
Pillepich A., Porciani, C., Hahn, O., 2010, MNRAS, 402, 191
Press W. H. & Schechter P., 1974, ApJ, 187, 425.
Robertson, B., Kravtsov, A., Tinker, J., Zentner, A., 2009, ApJ, 696, 636
Seljak U., Warren M., 2004, MNRAS, 355, 129
Sheth R., Mo H., Tormen G., 2001, MNRAS, 323, 1
Sheth R. & Tormen G., 1999, MNRAS, 308, 119
Sheth R. & Tormen G., 2002, MNRAS, 329, 61
Tinker J., Weinberg, D., Zheng Z., Zehavi I., 2005, ApJ, 631, 41
Tinker J. et al. 2008, ApJ, 688, 709
Tinker J. et al. 2010, arXiv:1001.3162
Zentner A., 2007, Int. J. Mod. Phys. D, 16, 763.
Details of the computation
==========================
To perform the computation we use the technique discussed in detail in @MR1 (MR1). We consider first the numerator in . We start from , with the two-point function $\langle\d_i\d_j\rangle_c$ given in , and we expand to first order in $\kappa$ (recall that $\D_{ij}$ is proportional to $\kappa$). This gives $W$ in terms of $W^{\rm gm}$, \[WnNG\] &&W(\_0;…,\_n;S\_n)=\
&& e\^[ i\_[i=1]{}\^n\_i\_i -\_[i,j=1]{}\^n\_i\_j ([min]{}(S\_i,S\_j) + \_[ij]{})]{}\
&&W\^[gm]{} (\_0;…,\_n;S\_n)\
&&+\_[i,j=1]{}\^n\_[ij]{}\_i\_j W\^[gm]{} (\_0;…,\_n;S\_n) , where $\D_{ij}\equiv \D(S_i,S_j)$, $\pa_i\equiv \pa/\pa\d_i$, and we have used the identity \_ke\^[i\_[j=1]{}\^n\_j\_j]{}=-i\_k e\^[i\_[j=1]{}\^n\_j\_j]{} to transform the factor $-\D_{ij}\lambda_i\lambda_j$ coming from the expansion of the exponential into $\D_{ij}\pa_i\pa_j$. It is convenient to split the sum into various pieces $$\begin{aligned}
\frac{1}{2}\sum_{i,j=1}^{n}\Delta_{ij}\partial_i\partial_j&=&
\frac{1}{2}\sum_{i,j=1}^{m-1}\Delta_{ij}\partial_i\partial_j+
\sum_{i=1}^{m-1}\Delta_{im}\partial_i\partial_m\nn\\
&&\hspace*{-4mm}+
\frac{1}{2}\sum_{i,j=m+1}^{n-1}\Delta_{ij}\partial_i\partial_j+
\sum_{i=m+1}^{n-1}\Delta_{in}\partial_i\partial_n\nn\\
&&\hspace*{-4mm}+\sum_{i=1}^{m-1}\Delta_{in}\partial_i\partial_n+
\Delta_{mn}\partial_m\partial_n\nn\\
&&\hspace*{-4mm}+\sum_{i=1}^{m-1}\sum_{j=m+1}^{n-1}\Delta_{ij}\partial_i\partial_j+
\sum_{j=m+1}^{n-1}\Delta_{jm}\partial_j\partial_m\, .\nn\label{A3}\\\end{aligned}$$ Consider first the contribution from the first line of this expression. Using the factorization property (\[facto\]) of $W^{\rm gm}$, its contribution to the numerator in can be written as && \_[-]{}\^[\_c]{}d\_1d\_[m-1]{} d\_[m+1]{}d\_[n-1]{}\
&&$$\frac{1}{2}\sum_{i,j=1}^{m-1}\Delta_{ij}\partial_i\partial_j+
\sum_{i=1}^{m-1}\Delta_{im}\partial_i\partial_m$$\
&&W\^[gm]{}(\_0;…, \_m;S\_m) W\^[gm]{}(\_m; …,\_n; S\_n-S\_m)\
&=&\_[-]{}\^[\_c]{} d\_1d\_[m-1]{} $$\frac{1}{2}\sum_{i,j=1}^{m-1}\Delta_{ij}\partial_i\partial_j+
\sum_{i=1}^{m-1}\Delta_{im}\partial_i\partial_m$$\
&&W\^[gm]{}(\_0;…, \_m;S\_m)\
&&\_[-]{}\^[\_c]{}d\_[m+1]{}d\_[n-1]{} W\^[gm]{}(\_m; …,\_n; S\_n-S\_m)\
&+&\_[-]{}\^[\_c]{}d\_1d\_[m-1]{} \_[i=1]{}\^[m-1]{}\_[im]{}\_i W\^[gm]{}(\_0;…, \_m;S\_m)\
&&\_[-]{}\^[\_c]{} d\_[m+1]{}d\_[n-1]{} \_mW\^[gm]{}(\_m; …,\_n; S\_n-S\_m) .\[A4\]\
The first term is easily dealt by observing that &&\_[-]{}\^[\_c]{}d\_[m+1]{}d\_[n-1]{} W\^[gm]{}(\_m; …,\_n; S\_n-S\_m)\
&&= \^[gm]{}(\_m;\_n;S\_n-S\_m) . Combining this with the contribution coming from the zero-th order term $W^{\rm gm} (\delta_0;\ldots ,\delta_n;S_n) $ in and using again the factorization property (\[facto\]) of $W^{\rm gm}$, we therefore get &&\^[gm]{}(\_m;\_n;S\_n-S\_m) \_[-]{}\^[\_c]{}d\_1d\_[m-1]{}\
&&$$1+\frac{1}{2}\sum_{i,j=1}^{m-1}\Delta_{ij}\partial_i\partial_j+
\sum_{i=1}^{m-1}\Delta_{im}\partial_i\partial_m$$\
&&W\^[gm]{}(\_0;…, \_m;S\_m)\
&+&\_[-]{}\^[\_c]{}d\_1d\_[m-1]{} \_[i=1]{}\^[m-1]{}\_[im]{}\_i W\^[gm]{}(\_0;…, \_m;S\_m)\
&&\_[-]{}\^[\_c]{} d\_[m+1]{}d\_[n-1]{} \_mW\^[gm]{}(\_m; …,\_n; S\_n-S\_m) .We now observe that the term in brackets give just the expansion to $O(\kappa)$ of the denominator in . Therefore, to $O(\kappa)$, we can write P(\_n,S\_n| \_m,S\_m)&=&\^[gm]{}(\_m;\_n;S\_n-S\_m)\
&&+P\^[non-mark]{}(\_n,S\_n| \_m,S\_m) ,\[Ptot\] where \[PPnonm\] P\^[non-mark]{}(\_n,S\_n| \_m,S\_m)= , and $N_a,\ldots ,N_d$ are defined by N\_a&=&\_[-]{}\^[\_c]{}d\_1d\_[m-1]{} \_[i=1]{}\^[m-1]{}\_[im]{}\_i W\^[gm]{}(\_0;…, \_m;S\_m)\
&&\_[-]{}\^[\_c]{} d\_[m+1]{}d\_[n-1]{} \_mW\^[gm]{}(\_m; …,\_n; S\_n-S\_m) ,\
N\_b&=&\_[-]{}\^[\_c]{}d\_1d\_[m-1]{} d\_[m+1]{}d\_[n-1]{}\[Nb0\]\
&&$$\frac{1}{2}\sum_{i,j=m+1}^{n-1}\Delta_{ij}\partial_i\partial_j+
\sum_{i=m+1}^{n-1}\Delta_{in}\partial_i\partial_n$$\
&&W\^[gm]{}(\_0;…, \_m;S\_m) W\^[gm]{}(\_m; …,\_n; S\_n-S\_m) ,N\_c&=&\_[-]{}\^[\_c]{}d\_1d\_[m-1]{} d\_[m+1]{}d\_[n-1]{}\
&&$$\sum_{i=1}^{m-1}\Delta_{in}\partial_i\partial_n+
\Delta_{mn}\partial_m\partial_n$$\
&&W\^[gm]{}(\_0;…, \_m;S\_m) W\^[gm]{}(\_m; …,\_n; S\_n-S\_m) ,N\_d&=&\_[-]{}\^[\_c]{}d\_1d\_[m-1]{} d\_[m+1]{}d\_[n-1]{}\
&&$$\sum_{i=1}^{m-1}\sum_{j=m+1}^{n-1}\Delta_{ij}\partial_i\partial_j+
\sum_{j=m+1}^{n-1}\Delta_{jm}\partial_j\partial_m$$\
&&W\^[gm]{}(\_0;…, \_m;S\_m) W\^[gm]{}(\_m; …,\_n; S\_n-S\_m) .The contribution $N_a$ comes from , while $N_b,N_c$ and $N_d$ come from the second, third and fourth line in , respectively. Observe that in the denominator in we could replace $\Pi(\d_0;\d_m;S_m)$ by $\Pi^{\rm gm}(\d_0;\d_m;S_m)$, since the numerator is proportional to $\D_{ij}$ and therefore is already ${\cal O}(\kappa)$.
The contributions $N_a,\ldots, N_d$ can be computed using the techniques developed in MR1. The term $N_a$ is immediately obtained using eqs. (105) and (110) of MR1, and is given by N\_a&=& $\frac{2\d_c-\d_m}{\sqrt{2S_m}}$\
&&\_m\^[gm]{}(\_m;\_n;S\_n-S\_m) , where ${\rm Erfc}$ is the complementary error function. The term $N_b$ is given by \[Nb\] N\_b&=&\^[gm]{}(\_0;\_m;S\_m)\
&& ,where &&\^[b1]{} (\_m,S\_m;\_n,S\_n) \_[-]{}\^[\_c]{} d\_[m+1]{}d\_[n-1]{}\[Nb1\]\
&&\_[i=m+1]{}\^[n-1]{}\_[in]{}\_i\_n W\^[gm]{}(\_m; …,\_n; S\_n-S\_m) ,and &&\^[b2]{} (\_m,S\_m;\_n,S\_n) \_[-]{}\^[\_c]{} d\_[m+1]{}d\_[n-1]{}\[Nb2\]\
&&\_[i,j=m+1]{}\^[n-1]{}\_[ij]{}\_i\_j W\^[gm]{}(\_m; …,\_n; S\_n-S\_m) .The computation of $\Pi^{b1}$ and $\Pi^{b2}$ is quite similar to the computation of the terms called $\Pi^{\rm mem}$ and $\Pi^{\rm mem-mem}$ in MR1, and in the continuum limit $\eps\ra 0$ we get \[Pib1\] &&\^[b1]{}(\_m,S\_m;\_n,S\_n)=\_n\_[0]{} \_[S\_m]{}\^[S\_n]{}dS\_i\
&&(S\_i,S\_n)\_\^[gm]{}(\_m;\_c;S\_i-S\_m) \_\^[gm]{}(\_c;\_n;S\_n-S\_i)\
&=& (\_c-\_m) \_n{(\_c-\_n) \_[S\_m]{}\^[S\_n]{}dS\_i\
&&\
&&$$-\frac{(\d_c-\d_m)^2}{2(S_i-S_m)}-\frac{(\d_c-\d_n)^2}{2(S_n-S_i)}$$} \[finalb1\] and &&\^[b2]{}(\_m,S\_m;\_n,S\_n)=\_[0]{} \_[S\_m]{}\^[S\_n]{}dS\_i\_[S\_i]{}\^[S\_n]{}dS\_j\
&&(S\_i,S\_j)\_\^[gm]{}(\_m;\_c;S\_i-S\_m)\
&&\_\^[gm]{}(\_c;\_c;S\_j-S\_i) \_\^[gm]{}(\_c;\_n;S\_n-S\_j)\
&=& (\_c-\_m)(\_c-\_n)\
&&\_[S\_m]{}\^[S\_n]{}dS\_i e\^[-(\_c-\_m)\^2/\[2(S\_i-S\_m)\]]{}\
&&\_[S\_i]{}\^[S\_n]{}dS\_j . This can be rewritten as a total derivative with respect to $\d_n$, as &&\^[b2]{}(\_m,S\_m;\_n,S\_n)= (\_c-\_m)\_n\
&&\_[S\_m]{}\^[S\_n]{}dS\_i e\^[-(\_c-\_m)\^2/\[2(S\_i-S\_m)\]]{}\
&&\_[S\_i]{}\^[S\_n]{}dS\_j .\[finalb2\] The fact that both $\Pi^{b1}$ and $\Pi^{b2}$ can be written as a derivative with respect to $\d_n$ simplifies considerably the computation of the flux ${\cal F}(S)$, since we can integrate $\pa_n\equiv\pa/\pa\d_n$ by parts, and then we only need to evaluate the integrals in in $\d_n=\d_c$, which can be done analytically, as discussed in MR1.
The term $N_c$ is a total derivative with respect to $\pa_n$ of a quantity that vanishes in $\d_n=\d_c$ so, when inserted into , it gives a vanishing contribution to the first crossing rate. The most complicated term is $N_d$. Using the techniques developed in MR1, a rather long computation gives N\_d&=&\_n{ \_c (\_c-\_m) [Erfc]{}$\frac{2\d_c-\d_m}{\sqrt{2S_m}}$\_m I\[Nd\]\
&& + \^[gm]{}(\_0;\_m;S\_m)\_d} , where \_d&=& -\_m(\_c-\_m) I (\_m,\_n)+S\_m (\_c-\_m)\_m I (\_m,\_n)\
&&-S\_m I (\_m,\_n) , and I (\_m,\_n)&&\_[S\_m]{}\^[S\_n]{}dS\_j\
&&{- -} . Using , can be rewritten as P(\_n,S\_n| \_m,S\_m)=\^[gm]{}+\^[b1]{}+ \^[b2]{}+ . We can now compute the contribution to the flux from the various terms. The term $\Pi^{\rm gm}$ gives the zero-th order term, &&\^[gm]{}(S\_n|\_m,S\_m)=-\_[-]{}\^[\_c]{} d\_n \^[gm]{} (\_m;\_n;S\_n-S\_m)\
&&= e\^[-(\_c-\_m)\^2/\[2(S\_n-S\_m)\]]{} . The contribution of $\Pi^{b1}$ to the flux is zero since it is the derivative with respect to $\pa_n$ of a quantity that vanishes in $\d_n=\d_c$, and the same holds for $N_c$. The contribution of $\Pi^{b2}$ is &&[F]{}\^[b2]{}(S\_n|\_m,S\_m)\
&&= -\_[-]{}\^[\_c]{} d\_n \^[b2]{} (\_m,S\_m;\_n;S\_n)\
&&= - (\_c-\_m)\
&&\_[S\_m]{}\^[S\_n]{}dS\_i e\^[-(\_c-\_m)\^2/\[2(S\_i-S\_m)\]]{}\
&&\_[S\_i]{}\^[S\_n]{}dS\_j . The inner integral is elementary, \_[S\_i]{}\^[S\_n]{}dS\_j = , and we end up with &&[F]{}\^[b2]{}(S\_n|\_m,S\_m)= - ,which generalized eq. (118) of MR1 to $\d_m\neq 0$ and $S_m\neq 0$. For $S_m$ generic the integral cannot be performed analytically. However, for computing the bias we are actually interested in the limit $S_m\ra 0$ with $\d_m$ generic, and we see that in this case this contribution reduces to that computed in MR1, with the replacement $\d_c\ra\d_c-\d_m$.
The remaining contributions can be computed similarly. For the term $N_d$, again, rather than computing explicitly the derivative $\pa_n=\pa/\pa\d_n$ in , it is convenient to insert this expression directly into the first-crossing rate (\[firstcrossTcond\]), and use the fact that it is a total derivative with respect to $\pa_n$ to perform the integral over $d\d_n$. So, in the end, we only need I (\_m,\_n=\_c) &=& e\^[+(\_c-\_m)\^2/(2 S\_m)]{}\[Idndc\]\
&&$$(\d_c-\d_m)\sqrt{\frac{S_n}{2S_m(S_n-S_m)}}$$ .We can now put together all the terms and take the limit $S_m\ra 0$ (with $\d_c-\d_m>0$). In this limit the ${\rm Erfc}$ function in reduces to an exponential, so I (\_m,\_n=\_c) e\^[-(\_c-\_n)\^2/\[2(S\_n-S\_m)\]]{} . Denoting $\d_m=\d_0$ in this limit, we finally get (S|\_0,S\_m=0) &=& e\^[-(\_c-\_0)\^2/(2S)]{}\
&&+ $0,\frac{(\d_c-\d_0)^2}{2S}$\
&&- $$1- \frac{(\d_c-\d_0)^2}{S}$$ e\^[-(\_c-\_0)\^2/(2S)]{} .Expanding this result to first order in $\d_0$ we obtain .
\[lastpage\]
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
We introduce a robust estimator of the location parameter for the change-point in the mean based on the Wilcoxon statistic and establish its consistency for $L_1$ near epoch dependent processes. It is shown that the consistency rate depends on the magnitude of change. A simulation study is performed to evaluate finite sample properties of the Wilcoxon-type estimator in standard cases, as well as under heavy-tailed distributions and disturbances by outliers, and to compare it with a CUSUM-type estimator. It shows that the Wilcoxon-type estimator is equivalent to the CUSUM-type estimator in standard cases, but outperforms the CUSUM-type estimator in presence of heavy tails or outliers in the data.
KEYWORDS: Wilcoxon statistic; change-point estimator; near epoch dependence
author:
- 'Carina Gerstenberger$^*$'
title: 'Robust Estimation of Change-Point Location'
---
Introduction {#Introduction}
============
In many applications it can not be assumed that observed data have a constant mean over time. Therefore, extensive research has been done in testing for change-points in the mean, see e.g. @Giraitis.1996, @Csorgo.1997, @Ling.2007, and others. A number of papers deal with the problem of estimation of the change-point location. @Bai.1993 estimates the unknown location point for the break in the mean of a linear process by the method of least squares. @Antoch.1995 and @Csorgo.1997 established the consistency rates for CUSUM-type estimators for independent data, while @Csorgo.1997 considered weakly dependent variables. @Horvath.1997 established consistency of CUSUM-type estimators of location of change-point for strongly dependent variables. @Kokoszka.1998 [@Kokoszka.2000] discussed CUSUM-type estimators for dependent observations and ARCH models. In spite of numerous studies on testing for changes and estimating for change-points, however, just a few procedures are robust against outliers in the data. In a recent work @Dehling.2015 address the robustness problem of testing for change-points by introducing a Wilcoxon-type test which is applicable under short-range dependence (see also @Dehling.2013b for the long-range dependence case).
In this paper we suggest a robust Wilcoxon-type estimator for the change-point location based on the idea of @Dehling.2015 and applicable for $L_1$ near epoch dependent processes. The Wilcoxon change-point test statistic is defined as $$\label{Wilcoxon-TS}
W_{n}(k) = \sum_{i=1}^{k}\sum_{j=k+1}^{n}(1_{\{X_i\leq X_j\}}-1/2)$$ and counts how often an observation of the second part of the sample, $X_{k+1},\ldots,X_n$, exceeds an observation of the first part, $X_1,\ldots,X_k$. Assuming a change in mean happens at the time $k^*$, the absolute value of $W_{n}(k^*)$ is expected to be large. Hence, the Wilcoxon-type estimator for the location of the change-point, $$\label{bp_estimator_wilcoxon}
\hat{k} = \min\Big\{k:\max_{1\leq l < n}\Big| W_{n}(l) \Big|=\Big| W_{n}(k) \Big|\Big\},$$ can be defined as the smallest k for which the Wilcoxon test statistic $W_n(k)$ attains its maximum. Since the Wilcoxon test statistic is a rank-type statistic, outliers in the observed data can not affect the test statistic significantly. On the contrary, the CUSUM-type test statistic $$C_{n}(k) = \frac{1}{k}\sum_{i=1}^{k} X_i - \frac{1}{n}\sum_{i=1}^{n} X_i,$$ which compares the difference of the sample mean of the first $k$ observations and the sample mean over all observations, can be significantly disturbed by a single outlier.
The outline of the paper is as follows. In Section \[Main Results\] we discuss the consistency and the rates of the estimator $\hat{k}$ in (\[bp\_estimator\_wilcoxon\]). Section \[Simulation Study\] contains the simulation study. Section \[Useful Properties\] provides useful properties of the Wilcoxon test statistic and the proof of the main result. Sections \[Auxiliary Results\] and \[result\_literature\] contain some auxiliary results.
Definitions, assumptions and main results {#Main Results}
=========================================
Assume the random variables $X_1,\ldots,X_n$ follow the change-point model $$\label{model_bp}
X_i = \begin{cases}
Y_i + \mu, & 1\leq i \leq k^*\\
Y_i + \mu + \Delta_n, & k^* < i \leq n,
\end{cases}$$ where the process $(Y_j)$ is a stationary zero mean short-range dependent process, $k^*$ denotes the location of the unknown change-point and $\mu$ and $\mu + \Delta_n$ are the unknown means. We assume that $Y_1$ has a continuous distribution function $\operatorname{F}$ with bounded second derivative and that the distribution functions of $Y_1-Y_k$, $k\geq 1$ satisfy $$\label{Bed_gem_Verteilung}
\operatorname{P}(x\leq Y_1-Y_k\leq y) \leq C|y-x|,$$ for all $0\leq x\leq y \leq 1$, where $C$ does not depend on $k$ and $x,y$. We allow the magnitude of the change $\Delta_n$ vary with the sample size $n$.
\[assumption\_change\]
a) The change-point $k^*=[n\theta]$, $0<\theta<1,$ is proportional to the sample size $n$.
b) The magnitude of change $\Delta_n$ depends on the sample size $n$, and is such that $$\label{assumption_size_change_infty}
\Delta_n \rightarrow 0, \qquad n\Delta_n^2 \rightarrow \infty, \qquad n\rightarrow \infty.$$
Next we specify the assumptions on the underlying process $(Y_j)$. The following definition introduces the concept of an absolutely regular process which is also known as $\beta$-mixing.
A stationary process $\left(Z_j\right)_{j\in\mathbb{Z}}$ is called absolutely regular if $$\beta_k = \sup_{n\geq 1}\operatorname{E}\sup_{A\in \mathcal{F}_1^n} \left| \operatorname{P}\left(A|\mathcal{F}_{n+k}^{\infty}\right)-\operatorname{P}\left(A\right)\right| \rightarrow 0$$ as $k\rightarrow\infty$, where $\mathcal{F}_a^b$ is the $\sigma$-field generated by random variables $Z_a,\ldots,Z_b$.
The coefficients $\beta_k$ are called mixing coefficients. For further information about mixing conditions see @Bradley.2002. The concept of absolute regularity covers a wide range of processes. However, important processes like linear processes or AR processes might not be absolutely regular. To overcome this restriction, in this paper we discuss functionals of absolutely regular processes, i.e. instead of focusing on the absolute regular process $\left(Z_j\right)$ itself, we consider process $\left(Y_j\right)$ with $Y_j=f(Z_j,Z_{j-1},Z_{j-2},\ldots)$, where $f:\mathbb{R}^{\mathbb{Z}}\rightarrow\mathbb{R}$ is a measurable function. The following near epoch dependence condition ensures that $Y_j$ mainly depends on the near past of $\left(Z_j\right)$.
We say that stationary process $\left(Y_j\right)$ is $L_1$ near epoch dependent ($L_1$ NED) on a stationary process $\left(Z_j\right)$ with approximation constants $a_k$, $k\geq 0$, if conditional expectations $\operatorname{E}(Y_1|\mathcal{G}_{-k}^k)$, where $\mathcal{G}_{-k}^{k}$ is the $\sigma$-field generated by $Z_{-k},\ldots,Z_{k}$, have property $$\operatorname{E}\Big| Y_1 - \operatorname{E}(Y_1|\mathcal{G}_{-k}^k)\Big| \leq a_k, \qquad k=0,1,2,\ldots$$ and $a_k\rightarrow0$, $k\rightarrow\infty$.
Note that $L_1$ NED is a special case of more general $L_r$ near epoch dependence, where approximation constants are defined using $L_r$ norm: $\operatorname{E}\big| Y_1 - \operatorname{E}(Y_1|\mathcal{G}_{-k}^k)\big|^r \leq a_k$, $r\geq 1$. $L_r$ NED processes are also called $r$-approximating functionals. In testing problems considered in this paper we allow for heavy-tailed distributions. Hence, we deal with $L_1$ near epoch dependence, which assumes existence of only the first moment $\operatorname{E}|Y_1|$. The concept of near epoch dependence is applicable e.g. to GARCH(1,1) processes, see @Hansen.1991, and linear processes, see below. @borovkova.2001 provide additional examples and information about properties of $L_r$ near epoch dependent process.
Let $\left(Y_j\right)$ be a linear process, i.e. $Y_t=\sum_{j=0}^{\infty}\psi_j Z_{t-j}$, where $\left(Z_j\right)$ is white-noise process and the coefficients $\psi_j$, $j\geq 0$, are absolutely summable. Since $\left(Z_j\right)$ is stationary and $Z_{t-j}$ is $\mathcal{G}_{-k}^k$ measurable for $|t-j|\leq k$, we get $$\begin{gathered}
\operatorname{E}|Y_t-\operatorname{E}(Y_t|\mathcal{G}_{-k}^k)| \leq \sum_{j=k+1}^{\infty}|\psi_j|\operatorname{E}|Z_{t-j}-\operatorname{E}(Z_{t-j}|\mathcal{G}_{-k}^k)|\\
\leq 2\sum_{j=k+1}^{\infty}|\psi_j|\operatorname{E}|Z_{t-j}| = 2\operatorname{E}|Z_1|\sum_{j=k+1}^{\infty}|\psi_j|.\end{gathered}$$ Thus, the linear process $\left(Y_j\right)$ is $L_1$ NED on $\left(Z_j\right)$ with approximation constants $a_k=2\operatorname{E}|Z_1|\sum_{j=k+1}^{\infty}|\psi_j|$.
We will assume that the process $(Y_j)$ in (\[model\_bp\]) is $L_1$ near epoch dependent on some absolutely regular process $\left(Z_j\right)$. In addition, we impose the following condition on the decay of the mixing coefficients $\beta_k$ and approximation constants $a_k$: $$\label{condition_appr.const_regu.coeff}
\sum_{k=1}^{\infty}{k^2\left(\beta_k+\sqrt{a_k}\right)}< \infty.$$
The next theorem states the rates of consistency of the Wilcoxon-type change-point estimator $\hat{k}$ given in (\[bp\_estimator\_wilcoxon\]) and the estimator $\hat{\theta}=\hat{k}/n$ of the true location parameter $\theta$ for the change-point $k^*=[n\theta]$.
Let $X_1,\ldots, X_n$ follow the change-point model (\[model\_bp\]) and Assumption \[assumption\_change\] be satisfied. Assume that $\left(Y_j\right)$ is a stationary zero mean $L_1$ near epoch dependent process on some absolutely regular process $\left(Z_j\right)$ and (\[condition\_appr.const\_regu.coeff\]) holds. Then, $$\label{rate_consistency_k_hat}
\Big| \hat{k}-k^*\Big| = O_P\Big(\frac{1}{\Delta_n^2}\Big),$$ and $$\label{rate_consistency_theta}
\Big|\hat{\theta}-\theta\Big| =O_P\Big(\frac{1}{n\Delta_n^2}\Big).$$
The rate of consistency of $\hat{\theta}$ in (\[rate\_consistency\_theta\]) is given by $n\Delta_n^2$. The assumption $n\Delta_n^2\rightarrow\infty$ in (\[assumption\_size\_change\_infty\]) implies $\hat{k}-k^*=o_P(k^*)$ and yields consistency of the estimator: $\hat{\theta}\rightarrow_p\theta$. In particular, for $\Delta_n\geq n^{-1/2+\epsilon}$, $\epsilon>0$, the rate of consistency in (\[rate\_consistency\_theta\]) is $n^{2\epsilon}$: $\big|\hat{\theta}-\theta\big|= O_P\left(n^{-2\epsilon}\right)$.
The same consistency rate $n^{\epsilon}$ for the CUSUM-type change-point location estimator $\tilde{\theta}_C = \tilde{k}_C/n$, given by $$\label{bp_estimator_cusum}
\tilde{k}_C = \min\bigg\{ k: \max_{1\leq i \leq n}\bigg| \sum_{j=1}^{i}X_j - \frac{i}{n}\sum_{j=1}^{n}X_j \bigg| = \bigg| \sum_{j=1}^{k}X_j - \frac{k}{n}\sum_{j=1}^{n}X_j \bigg| \bigg\},$$ was established by @Antoch.1995 for independent data and by @Csorgo.1997 for weakly dependent data.
Simulation results {#Simulation Study}
==================
In this simulation study we compare the finite sample properties of the Wilcoxon-type change-point estimator $\hat{k}$, given in (\[bp\_estimator\_wilcoxon\]), with the CUSUM-type estimator $\tilde{k}_C$, given in (\[bp\_estimator\_cusum\]). We refer to the Wilcoxon-type change-point estimator by W and to the CUSUM-type estimator by C.
We generate the sample of random variables $X_1,\ldots,X_n$ using the model $$\label{model_bp_simu}
X_i = \begin{cases}
Y_i + \mu &, 1\leq i \leq k^*\\
Y_i + \mu + \Delta &, k^* < i \leq n
\end{cases}$$ where $Y_i = \rho Y_{i-1} + \epsilon_i$ is an AR(1) process. In our simulations we consider $\rho = 0.4$, which yields a moderate positive autocorrelation in $X_i$. The innovations $\epsilon_i$ are generated from a standard normal distribution and a Student’s t-distribution with 1 degree of freedom. We consider the time of change $k^*=[n\theta]$, $\theta = 0.25,0.5,0.75$, the magnitude of change $\Delta = 0.5,1,2$ and the sample sizes $n=50,100,200,500$. All simulation results are based on 10.000 replications. Note that we report estimation results not for $\hat{k}$ and $\tilde{k}_C$, but $\hat{\theta}=\hat{k}/n$ and $\tilde{\theta}_C=\tilde{k}_C/n$.
Figure \[figure.normal.hist\] contains the histogram based on the sample of 10.000 values of Wilcoxon-type estimator $\hat{\theta}$ and the CUSUM-type estimator $\tilde{\theta}_C$, for the model (\[model\_bp\_simu\]) with $\Delta=1$, $\theta=0.5$, $n=50$ and independent standard normal innovations $\epsilon_i$. Both estimation methods give very similar histograms.
Table \[table.normal\] reports the sample mean and the sample standard deviation based on 10.000 values of $\hat{\theta}$ and $\tilde{\theta}_C$ for other choices of parameters $\Delta$ and $\theta$. It shows that performance of both estimators improves when the sample size $n$ and the magnitude of change $\Delta$ are rising, and when the change happens in the middle of the sample. In general, Wilcoxon-type estimator performs in all experiments as good as the CUSUM-type estimator.
Figure \[figure.t1.hist\] shows the histogram based on 10.000 values of $\hat{\theta}$ and $\tilde{\theta}_C$, for the model (\[model\_bp\_simu\]) with $t_1$-distributed heavy-tailed iid innovations $\epsilon_i$, $\Delta=1$, $\theta=0.5$ and $n=500$. For heavy-tailed innovations $\epsilon_i$, both estimators deviate from the true value of the parameter $\theta$ more significantly than under normal innovations. Nevertheless, the Wilcoxon-type estimator seems to outperform the CUSUM-type estimator.
Figure \[figure.outliers.hist\] shows the histogram based on 10.000 values for $\hat{\theta}$ and $\tilde{\theta}_C$ when the data $X_1,\ldots,X_n$ is generated by (\[model\_bp\_simu\]) with $\Delta =1$, $\theta=0.5$, $n=200$ and $\epsilon_i \sim \operatorname{NIID}(0,1)$ and contains outliers. The outliers are introduced by multiplying observations $X_{[0.2n]}$, $X_{[0.3n]}$, $X_{[0.6n]}$ and $X_{[0.8n]}$ by the constant $M=50$. The histogram shows that the Wilcoxon-type estimator is rarely affected by the outliers, whereas the CUSUM-type estimator suffers large distortions.
Table \[table.normal.t1.outliers\] reports the sample mean and the sample standard deviation based on 10.000 values of $\hat{\theta}$ and $\tilde{\theta}_C$ for $\Delta=1$ and $\theta=0.5$ for sample size $n=50,100,200,500$ in the case of the normal, normal with outliers and $t_1$-distributed innovations. Figures \[figure.normal.hist\], \[figure.t1.hist\] and \[figure.outliers.hist\] presents results for $n=50,200,500$.
In general, we conclude that the Wilcoxon-type change-point location estimator performs equally well as the CUSUM-type change-point estimator in standard situations, but outperforms the CUSUM-type estimator in presence of heavy tails and outliers.
[0.5]{} ![Histogram based on 10.000 values for the Wilcoxon-type estimator $\hat{\theta}$ and the CUSUM-type estimator $\tilde{\theta}_C$. $X_i$ follows the model (\[model\_bp\_simu\]) with $\Delta=1$, $\theta=0.5$, $n=50$ and normal innovations $\epsilon_i \sim \operatorname{NIID}(0,1)$.[]{data-label="figure.normal.hist"}](C_50_04_1_05_norm.pdf "fig:"){width="7cm" height="5cm"}
[0.5]{} ![Histogram based on 10.000 values for the Wilcoxon-type estimator $\hat{\theta}$ and the CUSUM-type estimator $\tilde{\theta}_C$. $X_i$ follows the model (\[model\_bp\_simu\]) with $\Delta=1$, $\theta=0.5$, $n=50$ and normal innovations $\epsilon_i \sim \operatorname{NIID}(0,1)$.[]{data-label="figure.normal.hist"}](W_50_04_1_05_norm.pdf "fig:"){width="7cm" height="5cm"}
---------- ---------- ------ ------ ------ ------ ------ ------ ------ ------ ------
$\Delta$ $\theta$
0.5 0.25 mean 0.46 0.46 0.43 0.44 0.40 0.40 0.34 0.34
sd 0.21 0.21 0.20 0.20 0.18 0.18 0.13 0.13
0.50 mean 0.50 0.50 0.50 0.50 0.50 0.50 0.50 0.50
sd 0.18 0.18 0.16 0.16 0.13 0.13 0.08 0.08
0.75 mean 0.54 0.54 0.57 0.56 0.61 0.61 0.66 0.66
sd 0.20 0.20 0.20 0.20 0.18 0.18 0.13 0.13
1 0.25 mean 0.39 0.39 0.35 0.35 0.31 0.31 0.28 0.28
sd 0.18 0.18 0.14 0.14 0.10 0.10 0.05 0.06
0.50 mean 0.50 0.50 0.50 0.50 0.50 0.50 0.50 0.50
sd 0.12 0.12 0.09 0.09 0.05 0.05 0.02 0.02
0.75 mean 0.61 0.60 0.65 0.65 0.69 0.69 0.72 0.72
sd 0.17 0.17 0.15 0.15 0.10 0.10 0.05 0.06
2 0.25 mean 0.30 0.31 0.28 0.29 0.27 0.28 0.26 0.26
sd 0.10 0.10 0.06 0.07 0.04 0.04 0.02 0.02
0.50 mean 0.50 0.50 0.50 0.50 0.50 0.50 0.50 0.50
sd 0.05 0.05 0.03 0.03 0.02 0.01 0.01 0.01
0.75 mean 0.69 0.68 0.72 0.71 0.73 0.73 0.74 0.74
sd 0.09 0.10 0.06 0.07 0.04 0.04 0.02 0.02
---------- ---------- ------ ------ ------ ------ ------ ------ ------ ------ ------
: Sample mean and the sample standard deviation based on 10.000 values of $\hat{\theta}$ and $\tilde{\theta}_C$. $X_i$ follows the model (\[model\_bp\_simu\]) with normal innovations $\epsilon_i \sim \operatorname{NIID}(0,1)$.[]{data-label="table.normal"}
[0.5]{} ![Histogram of CUSUM-type estimator $\tilde{\theta}_C$ and Wilcoxon-type estimator $\hat{\theta}$ based on 10.000 values of $\tilde{\theta}_C$ and $\hat{\theta}$ for the model (\[model\_bp\_simu\]) with iid $t_1$-distributed innovations, $\Delta=1$, $\theta=0.5$ and $n=500$.[]{data-label="figure.t1.hist"}](C_500_04_1_05_t1.pdf "fig:"){width="7cm" height="5cm"}
[0.5]{} ![Histogram of CUSUM-type estimator $\tilde{\theta}_C$ and Wilcoxon-type estimator $\hat{\theta}$ based on 10.000 values of $\tilde{\theta}_C$ and $\hat{\theta}$ for the model (\[model\_bp\_simu\]) with iid $t_1$-distributed innovations, $\Delta=1$, $\theta=0.5$ and $n=500$.[]{data-label="figure.t1.hist"}](W_500_04_1_05_t1.pdf "fig:"){width="7cm" height="5cm"}
[0.5]{} ![Histogram based on 10.000 values of $\tilde{\theta}_C$ and $\hat{\theta}$ for the model (\[model\_bp\_simu\]) with normal innovations $\epsilon_i\sim\operatorname{NIID}(0,1)$, $\Delta=1$, $\theta=0.5$, $n=200$ and outliers.[]{data-label="figure.outliers.hist"}](C_200_04_1_05_o50.pdf "fig:"){width="7cm" height="5cm"}
[0.5]{} ![Histogram based on 10.000 values of $\tilde{\theta}_C$ and $\hat{\theta}$ for the model (\[model\_bp\_simu\]) with normal innovations $\epsilon_i\sim\operatorname{NIID}(0,1)$, $\Delta=1$, $\theta=0.5$, $n=200$ and outliers.[]{data-label="figure.outliers.hist"}](W_200_04_1_05_o50.pdf "fig:"){width="7cm" height="5cm"}
------------- ------ ------ ------ ------ ------ ------ ------ ------ ------
Innovations
normal mean 0.50 0.50 0.50 0.50 0.50 0.50 0.50 0.50
sd 0.12 0.12 0.09 0.09 0.05 0.05 0.02 0.02
$t_1$ mean 0.52 0.50 0.51 0.50 0.51 0.50 0.50 0.50
sd 0.23 0.20 0.24 0.19 0.24 0.17 0.25 0.14
normal with mean 0.50 0.49 0.50 0.50 0.50 0.50 0.51 0.50
outliers sd 0.17 0.13 0.16 0.09 0.15 0.06 0.09 0.02
------------- ------ ------ ------ ------ ------ ------ ------ ------ ------
: Sample mean and the sample standard deviation of $\hat{\theta}$ and $\tilde{\theta}_C$ based on 10.000 replications for the normal, normal with outliers and $t_1$-distributed innovations, $\Delta=1$ and $\theta=0.5$.[]{data-label="table.normal.t1.outliers"}
Useful properties of the Wilcoxon test statistic and proof of {#Useful Properties}
==============================================================
This section presents some useful properties of the Wilcoxon test statistic and the proof of .
Throughout the paper without loss of generality, we assume that $\mu = 0$ and $\Delta_n>0$. We let $C$ denote a generic non-negative constant, which may vary from time to time. The notation $a_n \sim b_n$ means that two sequences $a_n$ and $b_n$ of real numbers have property $a_n/b_n\rightarrow c$, as $n\rightarrow \infty$, where $c\neq 0$ is a constant. $\|g\|_{\infty}=\sup_{x}|g(x)|$ stands for the supremum norm of function $g$. By $\xrightarrow{d}$ we denote the convergence in distribution, by $\rightarrow_p$ the convergence in probability and by $\overset{d}{=}$ we denote equality in distribution.
U-statistics and Hoeffding decomposition {#Sub_U-statistics and Hoeffding decomposition}
----------------------------------------
The Wilcoxon test statistic $W_n(k)$ in (\[Wilcoxon-TS\]) under the change-point model (\[model\_bp\]) can be decomposed into two terms $$\begin{aligned}
&W_{n}(k) =\sum_{i=1}^{k}\sum_{j=k+1}^{n}( 1_{\{X_i \leq X_j\}}-1/2 )\nonumber\\
&= \begin{cases}
\sum_{i=1}^{k}\sum_{j=k+1}^{n}( 1_{\{Y_i \leq Y_j\}}-1/2 ) + \sum_{i=1}^{k}\sum_{j=k^*+1}^{n} 1_{\{Y_j < Y_i \leq Y_j+\Delta_n\}}, &1\leq k \leq k^*\\
\sum_{i=1}^{k}\sum_{j=k+1}^{n}( 1_{\{Y_i \leq Y_j\}}-1/2 ) + \sum_{i=1}^{k^*}\sum_{j=k+1}^{n} 1_{\{Y_j < Y_i \leq Y_j+\Delta_n\}}, &k^* < k \leq n,
\end{cases}\nonumber\\
&= \begin{cases}
U_n(k) + U_n(k,k^*), & 1\leq k \leq k^*\\
U_n(k) + U_n(k^*,k), & k^* < k \leq n, \label{decompose_Wn}
\end{cases}\end{aligned}$$ where $$\begin{aligned}
U_n(k) &= \sum_{i=1}^{k}\sum_{j=k+1}^{n}( 1_{\{Y_i \leq Y_j\}}-1/2 ), &1\leq k \leq n,\label{decompose_Un}\\
U_n(k,k^*) &= \sum_{i=1}^{k}\sum_{j=k^*+1}^{n} 1_{\{Y_j < Y_i \leq Y_j+\Delta_n\}}, &1\leq k \leq k^*,\label{decompose_Un_k1}\\
U_n(k^*,k) &= \sum_{i=1}^{k^*}\sum_{j=k+1}^{n} 1_{\{Y_j < Y_i \leq Y_j+\Delta_n\}}, &k^* < k \leq n.\label{decompose_Un_k2}\end{aligned}$$
The first term $U_n(k)$ depends only on the underlying process $(Y_j)$, while the terms $U_n(k,k^*)$ and $U_n(k^*,k)$ depend in addition on the change-point time $k^*$ and the magnitude $\Delta_n$ of the change in the mean.
The term $U_n(k)$ can be written as a second order U-statistic $$U_n\left(k\right) = \sum_{i=1}^{k}\sum_{j=k+1}^{n}\left(h\left(Y_i,Y_j\right)-\Theta\right), \qquad 1\leq k \leq n,$$ with the kernel function $h\left(x,y\right) = 1_{\left\{x\leq y\right\}}$ and the constant $\Theta = \operatorname{E}h\left(Y_1',Y_2'\right) = 1/2$, where $Y_1'$ and $Y_2'$ are independent copies of $Y_1$.
We apply to $U_n\left(k\right)$ Hoeffding’s decomposition of U-statistics established by @Hoeffding.1948. It allows to write the kernel function as the sum $$\label{hoeffding_decomposition_h}
h\left(x,y\right) = \Theta + h_1\left(x\right) + h_2\left(y\right) + g\left(x,y\right),$$ where $$\begin{aligned}
h_1\left(x\right) &= \operatorname{E}h\left(x,Y_2'\right) - \Theta = 1/2 - \operatorname{F}\left(x\right),\qquad
h_2\left(y\right)= \operatorname{E}h\left(Y_1',y\right) - \Theta = \operatorname{F}\left(y\right) - 1/2,\\
g\left(x,y\right) &= h\left(x,y\right) - h_1\left(x\right) - h_2\left(y\right)-\Theta.\end{aligned}$$ By definition of $h_1$ and $h_2$, $\operatorname{E}h_1(Y_1)=0$ and $\operatorname{E}h_2(Y_1)=0$. Hence, $\operatorname{E}g(x,Y_1)=\operatorname{E}g(Y_1,y)=0$, i.e. $g(x,y)$ is a degenerate kernel.
The term $U_n(k,k^*)$ in (\[decompose\_Un\_k1\]) (and $U_n(k^*,k)$ in (\[decompose\_Un\_k2\])) can be written as a U-statistic $$U_n(k,k^*) = \sum_{i=1}^{k}\sum_{j=k^*+1}^{n}h_n(Y_i,Y_j),\qquad 1\leq k \leq k^*,$$ with the kernel $h_n(x,y) = h(x,y+\Delta_n)-h(x,y) = 1_{\{y<x\leq y+\Delta_n\}}$. The Hoeffding decomposition allows to write the kernel as $$\label{hoeffding_decomposition_hn}
h_n\left(x,y\right) = \Theta_{\Delta_n} + h_{1,n}\left(x\right) + h_{2,n}\left(y\right) + g_n\left(x,y\right),$$ with $\Theta_{\Delta_n} = \operatorname{E}1_{\{Y_2'\leq Y_1' \leq Y_2'+\Delta_n\}},$ $$\begin{aligned}
h_{1,n}\left(x\right) &= \operatorname{E}h_n\left(x,Y_2'\right) - \Theta_{\Delta_n} = \operatorname{F}\left(x\right) - \operatorname{F}\left(x-\Delta_n\right) - \Theta_{\Delta_n}, \\
h_{2,n}\left(y\right)&= \operatorname{E}h_n\left(Y_1',y\right) - \Theta_{\Delta_n} = \operatorname{F}\left(y+\Delta_n\right) - \operatorname{F}\left(y\right) - \Theta_{\Delta_n},\\
g_n\left(x,y\right) &= h_n\left(x,y\right) - h_{1,n}\left(x\right) - h_{2,n}\left(y\right)-\Theta_{\Delta_n}.\end{aligned}$$ By assumption the distribution function $\operatorname{F}$ of $Y_1$ has bounded probability density $f$ and bounded second derivative. This allows to specify the asymptotic behaviour of $\Theta_{\Delta_n}$, as $n\rightarrow\infty,$ $$\begin{aligned}
\Theta_{\Delta_n} &= \operatorname{E}1_{\{Y_2' < Y_1' \leq Y_2'+\Delta_n\}} = \operatorname{P}\left(Y_2' < Y_1' \leq Y_2'+\Delta_n\right)\nonumber\\
&= \int_{\mathbb{R}}\left(\operatorname{F}\left(y+\Delta_n\right)-\operatorname{F}\left(y\right)\right)dF\left(y\right) = \Delta_n \bigg(\int_{\mathbb{R}}f^2\left(y\right)dy + o(1)\bigg). \label{remark_theta_delta_n}\end{aligned}$$ Note that $\operatorname{E}h_{1,n}(Y_1)=0$ and $\operatorname{E}h_{2,n}(Y_1)=0$. Therefore, $g_n(x,y)$ is a degenerate kernel, i.e. $\operatorname{E}g_n(x,Y_1)=\operatorname{E}g_n(Y_1,y)=0$. Furthermore, $\|h_{1,n} \|_{\infty}\rightarrow 0$, as $n\rightarrow \infty$, since $$\label{abschaetzung_h1n_kleiner_delta}
|h_{1,n}(x)| \leq |\operatorname{F}(x)-\operatorname{F}(x-\Delta_n)-\Theta_{\Delta_n}|\leq C\Delta_n + \Theta_{\Delta_n} \leq C\Delta_n,$$ where $C>0$ is a constant and $\Delta_n\rightarrow 0$, as $n\rightarrow\infty$.
1-continuity property of kernel functions and {#Sub_1-continuity_h_hn}
----------------------------------------------
Asymptotic properties of near epoch dependent processes $(Y_j)$ introduced in Section \[Main Results\] are well investigated in the literature, see e.g. @borovkova.2001. In the context of change-point estimation we are interested in asymptotic properties of the variables $h(Y_i,Y_j)$, where $h(x,y)=1_{\{x\leq y\}}$ is the Wilcoxon kernel, and also in properties of the terms $h_1(Y_j)$ and $h_{1,n}(Y_j)$ of the Hoeffding decomposition of the kernels in (\[hoeffding\_decomposition\_h\]) and (\[hoeffding\_decomposition\_hn\]). We will need to show that the variables $(h(Y_i,Y_j))$, $(h_1(Y_j))$ and $(h_{1,n}(Y_j))$ retain some properties of $(Y_j)$. To derive them, we will use the fact that the kernels $h$ in (\[hoeffding\_decomposition\_h\]) and $h_n$ in (\[hoeffding\_decomposition\_hn\]) satisfy the $1$-continuity condition introduced by @borovkova.2001.
We say that the kernel $h\left(x,y\right)$ is $1$-continuous with respect to a distribution of a stationary process $(Y_j)$ if there exists a function $\phi(\epsilon)\geq 0$, $\epsilon\geq 0$ such that $\phi\left(\epsilon\right)\rightarrow 0$, $\epsilon\rightarrow 0$, and for all $\epsilon > 0$ and $k\geq 1$ $$\begin{aligned}
\operatorname{E}\left(\left| h\left(Y_1,Y_k\right)-h\left(Y_1',Y_k\right) \right| 1_{\left\{\left|Y_1-Y_1'\right|\leq \epsilon\right\}}\right) & \leq \phi\left(\epsilon\right),\label{1-cont_11}\\
\operatorname{E}\left(\left| h\left(Y_k,Y_1\right)-h\left(Y_k,Y_1'\right) \right| 1_{\left\{\left|Y_1-Y_1'\right|\leq \epsilon\right\}}\right) & \leq \phi\left(\epsilon\right),\nonumber\end{aligned}$$ and $$\begin{aligned}
\operatorname{E}\left(\left| h\left(Y_1,Y_2'\right)-h\left(Y_1',Y_2'\right) \right| 1_{\left\{\left|Y_1-Y_1'\right|\leq \epsilon\right\}}\right) & \leq \phi\left(\epsilon\right),\label{1-cont_22}\\
\operatorname{E}\left(\left| h\left(Y_2',Y_1\right)-h\left(Y_2',Y_1'\right) \right| 1_{\left\{\left|Y_1-Y_1'\right|\leq \epsilon\right\}}\right) & \leq \phi\left(\epsilon\right),\nonumber\end{aligned}$$ where $Y_2'$ is an independent copy of $Y_1$ and $Y_1'$ is any random variable that has the same distribution as $Y_1$.
For a univariate function $g(x)$ we define the $1$-continuity property as follows.
The function $g\left(x\right)$ is $1$-continuous with respect to a distribution of a stationary process $(Y_j)$ if there exists a function $\phi(\epsilon)\geq 0$, $\epsilon\geq 0$ such that $\phi\left(\epsilon\right)\rightarrow 0$, $\epsilon\rightarrow 0$, and for all $\epsilon > 0$ $$\begin{aligned}
\label{1-cont_3}
\operatorname{E}\left(\left| g\left(Y_1\right)-g\left(Y_1'\right) \right| 1_{\left\{\left|Y_1-Y_1'\right|\leq \epsilon\right\}}\right) & \leq \phi\left(\epsilon\right),\end{aligned}$$ where $Y_1'$ is any random variable that has the same distribution as $Y_1$.
below establishes the $1$-continuity of functions $h(x,y)=1_{\{x\leq y\}}$ and $h_n(x,y)=1_{\{y<x\leq y+\Delta_n\}}$, $n\geq 1$. For $h_n$, $n\geq 1$ we assume that (\[1-cont\_11\]) and (\[1-cont\_22\]) hold with the same $\phi(\epsilon)$ for all $n\geq 1$. We start the proof by showing the $1$-continuity of the more general kernel function $h(x,y;t)=1_{\{x-y\leq t\}}$.
Let $(Y_j)$ be a stationary process, $Y_1$ have distribution function $F$ which has bounded first and second derivative and $Y_1-Y_k$, $k\geq 1$ satisfy (\[Bed\_gem\_Verteilung\]). Then the function $h(x,y;t)=1_{\{x-y\leq t\}}$ is $1$-continuous with respect to the distribution function of $(Y_j)$.
The proof is similar to the proof of $1$-continuity of the kernel function $h(x,y;t)=1_{\{|x-y|\leq t\}}$ given in Example 2.2 of @borovkova.2001.
Note that $1_{\{Y_1- Y_k\leq t\}}-1_{\{Y_1'- Y_k\leq t\}}=0$ if $Y_1-Y_k\leq t$ and $Y_1'-Y_k\leq t$; or $Y_1-Y_k> t$ and $Y_1'-Y_k> t$. The difference is not zero if $Y_1-Y_k\leq t$ and $Y_1'-Y_k> t$; or $Y_1-Y_k> t$ and $Y_1'-Y_k\leq t$. Let $|Y_1-Y_1'|\leq \epsilon$, where $\epsilon>0$. Then $Y_1-Y_k<t-\epsilon$ implies $Y_1'-Y_k<t$, and $Y_1-Y_k>t+\epsilon$ implies $Y_1'-Y_k>t$.
Hence, $|1_{\{Y_1- Y_k\leq t\}}-1_{\{Y_1'- Y_k\leq t\}}|1_{\left\{\left|Y_1-Y_1'\right|\leq \epsilon\right\}}\leq 1_{\left\{t-\epsilon\leq Y_1-Y_k\leq t+\epsilon\right\}}$. Therefore, $$\label{ineq_1_continuous}
\operatorname{E}\left(\left|1_{\{Y_1- Y_k\leq t\}}-1_{\{Y_1'- Y_k\leq t\}}\right|1_{\left\{\left|Y_1-Y_1'\right|\leq \epsilon\right\}}\right) \leq \operatorname{P}\left(t-\epsilon\leq Y_1-Y_k\leq t+\epsilon\right)\leq C_1\epsilon,$$ because of assumption (\[Bed\_gem\_Verteilung\]). Similar argument yields $$\begin{aligned}
\operatorname{E}\left(\left|1_{\{Y_k- Y_1\leq t\}}-1_{\{Y_k- Y_1'\leq t\}}\right|1_{\left\{\left|Y_1-Y_1'\right|\leq \epsilon\right\}}\right)&\leq \operatorname{P}\left(t-\epsilon\leq Y_1-Y_k\leq t+\epsilon\right)\leq C_1\epsilon, \\
\operatorname{E}\left(\left|1_{\{Y_1- Y_2'\leq t\}}-1_{\{Y_1'- Y_2'\leq t\}}\right|1_{\left\{\left|Y_1-Y_1'\right|\leq \epsilon\right\}}\right) &\leq \operatorname{P}(t-\epsilon \leq Y_1-Y_2'\leq t+\epsilon)\leq C_2\epsilon,\\
\operatorname{E}\left(\left|1_{\{Y_2'- Y_1\leq t\}}-1_{\{Y_2'- Y_1'\leq t\}}\right|1_{\left\{\left|Y_1-Y_1'\right|\leq \epsilon\right\}}\right) &\leq \operatorname{P}(t-\epsilon \leq Y_1-Y_2'\leq t+\epsilon)\leq C_2\epsilon,\end{aligned}$$ where $Y_2'$ is an independent copy of $Y_1$, noting that by the mean value theorem and $|d\operatorname{F}(y)/dy|\leq C$, $$\begin{aligned}
\operatorname{P}(t-\epsilon\leq Y_1-Y_2'\leq t+\epsilon) &= \int_{\mathbb{R}}\left(\operatorname{F}\left(y+t+\epsilon\right)-\operatorname{F}\left(y+t-\epsilon\right)\right)dF\left(y\right)\\
&\leq C\epsilon\int_{\mathbb{R}}f(y)dy = C_2\epsilon.\end{aligned}$$ These bounds imply (\[1-cont\_11\]) and (\[1-cont\_22\]) with $\phi(\epsilon)=C\epsilon$, where $C$ does not depend on $t$. This completes the proof.
Assume that assumptions of are satisfied. Then,
(i) Function $h(x,y)=1_{\{x\leq y\}}$ is $1$-continuous with respect to the distribution function of $(Y_j)$.
(ii) Function $h_n(x,y)=1_{\{y<x\leq y+\Delta_n\}}$ is $1$-continuous with respect to the distribution function of $(Y_j)$.
*(i)* follows from , noting that $1_{\{x\leq y\}}=h(x,y;0)$.
*(ii)* We need to verify (\[1-cont\_11\]) and (\[1-cont\_22\]). Write $h_n(x,y)=h(x,y)-h(x,y+\Delta_n)=1_{\{x\leq y\}}-1_{\{x\leq y+\Delta_n\}}$. Then by (\[ineq\_1\_continuous\]), $$\begin{gathered}
\operatorname{E}\big(|h_n(Y_1,Y_k)-h_n(Y_1',Y_k)|1_{\left\{\left|Y_1-Y_1'\right|\leq \epsilon\right\}}\big)
\leq \operatorname{E}\big(|1_{\{Y_1\leq Y_k\}}-1_{\{Y_1'\leq Y_k\}}|1_{\left\{\left|Y_1-Y_1'\right|\leq \epsilon\right\}}\big)
\\ + \operatorname{E}\big(|1_{\{Y_1\leq Y_k+\Delta_n\}}-1_{\{Y_1'\leq Y_k+\Delta_n\}}|1_{\left\{\left|Y_1-Y_1'\right|\leq \epsilon\right\}}\big)
\leq C\epsilon.\end{gathered}$$ Similar argument yields $\operatorname{E}\big(|h_n(Y_k,Y_1)-h_n(Y_k,Y_1')|1_{\left\{\left|Y_1-Y_1'\right|\leq \epsilon\right\}}\big) \leq C\epsilon,$ $$\begin{aligned}
\operatorname{E}\big(|h_n(Y_1,Y_2')-h_n(Y_1',Y_2')|1_{\left\{\left|Y_1-Y_1'\right|\leq \epsilon\right\}}\big) &\leq C\epsilon,\\
\operatorname{E}\big(|h_n(Y_2',Y_1)-h_n(Y_2',Y_1')|1_{\left\{\left|Y_1-Y_1'\right|\leq \epsilon\right\}}\big)&\leq C\epsilon.\end{aligned}$$ Hence, (\[1-cont\_11\]) and (\[1-cont\_22\]) hold with $\phi(\epsilon)=C\epsilon$.
Note that condition (\[Bed\_gem\_Verteilung\]) is satisfied if variables $(Y_1,Y_k)$, $k\geq 1$, have joint probability densities that are bounded by the same constant $C$ for all $k$. If the joint density does not exist, for examples of verification of condition (\[Bed\_gem\_Verteilung\]) see pages 4315, 4316 of @borovkova.2001.
Lemma 2.15 of @borovkova.2001 yields that if a general function $h(x,y)$ is $1$-continuous, i.e. satisfies (\[1-cont\_11\]) and (\[1-cont\_22\]) with function $\phi(\epsilon)$ then $\operatorname{E}h\left(x,Y_2'\right)$, where $Y_2'$ is an independent copy of $Y_1$, is also $1$-continuous and satisfies the condition in (\[1-cont\_3\]) with the same function $\phi(\epsilon)$. Hence, $h_i(x)$ and $h_{i,n}(x)$, $i=1,2$ are $1$-continuous and satisfy the condition in (\[1-cont\_3\]) with $\phi(\epsilon)=C\epsilon$.
Next we turn to $1$-continuity property of $g(x,y)$. By Hoeffding decomposition (\[hoeffding\_decomposition\_h\]), $g(x,y)=h(x,y)-\Theta-h_1(x)-h_2(y)$. Since $h(x,y)$, $h_1(x)$ and $h_2(x)$ in (\[hoeffding\_decomposition\_h\]) are $1$-continuous and satisfy (\[1-cont\_11\]), (\[1-cont\_22\]) and (\[1-cont\_3\]) with the same function $\phi(\epsilon)=C\epsilon$, then $g(x,y)$ is also $1$-continuous with function $\phi(\epsilon)=C\epsilon$. Indeed, $$\begin{gathered}
\operatorname{E}\left(|g(Y_1,Y_k)-g(Y_1',Y_k)|1_{\left\{\left|Y_1-Y_1'\right|\leq \epsilon\right\}} \right)\\
\leq \operatorname{E}\left(|h(Y_1,Y_k)-h(Y_1',Y_k)|1_{\left\{\left|Y_1-Y_1'\right|\leq \epsilon\right\}} \right) + \operatorname{E}\left(|h_1(Y_1)-h_1(Y_1')|1_{\left\{\left|Y_1-Y_1'\right|\leq \epsilon\right\}} \right)
\leq 2\phi(\epsilon)\end{gathered}$$ and similarly, $\operatorname{E}\big(|g(Y_k,Y_1)-g(Y_k,Y_1')|1_{\left\{\left|Y_1-Y_1'\right|\leq \epsilon\right\}} \big)\leq 2\phi(\epsilon)$.
Using the same argument, it follows that the function $g_n(x,y)=h_n(x,y)-\Theta_{\Delta_n}-h_{1,n}(x)-h_{2,n}(x)$ in the Hoeffding decomposition (\[hoeffding\_decomposition\_hn\]) is also $1$-continuous and satisfies (\[1-cont\_11\]), (\[1-cont\_22\]) with $\phi(\epsilon)=C\epsilon$.
NED property of and {#Properties of the kernel}
--------------------
In Proposition 2.11 of @borovkova.2001 it is shown that if $(Y_j)$ is $L_1$ NED on a stationary absolutely regular process $(Z_j)$ with approximation constants $a_k$ and $g(x)$ is $1$-continuous with function $\phi$, then $(g(Y_j))$ is also $L_1$ NED on $(Z_j)$ with approximation constants $\phi\left(\sqrt{2a_k}\right)+2\sqrt{2a_k}||g||_{\infty}$.
Thus, the processes $(h_1(Y_j))$ and $(h_2(Y_j))$ in (\[hoeffding\_decomposition\_h\]) and $(h_{1,n}(Y_j))$ and $(h_{2,n}(Y_j))$ in (\[hoeffding\_decomposition\_hn\]) are $L_1$ NED processes with approximation constants $a'_k=C\sqrt{a_k}$.
Corollary 3.2 of @wooldridge.1988 provides a functional central limit theorem for partial sum process $\sum_{i=1}^k \tilde{Y}_i$, $k\geq 1$, where $(\tilde{Y}_j)$ is $L_2$ NED on a strongly mixing process $(\tilde{Z}_j)$. To apply this result to $(h_1(Y_j))$ which is $L_1$ NED on $(Z_j)$ with approximation constants $a_k'$, we need to show that $(h_1(Y_j))$ is also $L_2$ NED process. Note that the variables $\eta_k := h_1(Y_1)-\operatorname{E}(h_1(Y_1)|\mathcal{G}_{-k}^k)$ have property $$\begin{gathered}
\operatorname{E}\eta_k^2 = \operatorname{E}\big(\eta_k^2 1_{\big\{|\eta_k|\leq {a_k'}^{-\frac{1}{2}}\big\}}\big)+ \operatorname{E}\big(\eta_k^2 1_{\big\{|\eta_k|> {a_k'}^{-\frac{1}{2}}\big\}}\big)\\
\leq {a_k'}^{-\frac{1}{2}}\operatorname{E}|\eta_k| + \sqrt{a_k'}\operatorname{E}|\eta_k|^{4}\leq \sqrt{a_k'} + \sqrt{a_k'} C =: a_k''.\end{gathered}$$ The last inequality holds, because by of $L_1$ near epoch dependence, $\operatorname{E}|h_1(Y_1)-\operatorname{E}(h_1(Y_1)|\mathcal{G}_{-k}^k)|\leq a_k'$ and because $|h_1(Y_1)|\leq 1/2$. Therefore the process $(h_1(Y_j))$ is $L_2$ NED on $(Z_j)$ with approximation constant $a_k''$. Since absolute regular process $(Z_j)$ is strongly mixing process, from Corollary 3.2 of @wooldridge.1988, we obtain $$\bigg( \frac{1}{n^{1/2}}\sum_{i=1}^{\left[nt\right]}h_1\left(Y_i\right) \bigg)_{0\leq t \leq 1} \xrightarrow{d} \left( \sigma W\left(t\right) \right)_{0\leq t \leq 1},$$ where $W\left(t\right)$ is a Brownian motion and $\sigma^2 =\sum_{k=-\infty}^{\infty}\operatorname{Cov}(h_1(Y_0),h_1(Y_k))$.
Since $h_2(x) = -h_1(x)$, all properties of $(h_1(Y_j))$ remain valid also for $(h_2(Y_j))$.
Proof of {#Proof}
---------
First we show consistency property $|k^*-\hat{k}|=o_P(k^*)$ of the estimate $\hat{k}=\operatorname{argmax}_{1\leq k\leq n}|W_n(k)|$. To prove it, we verify that for any $\epsilon>0$, $$\label{zz_consistency_1}
\lim_{n\rightarrow \infty}\operatorname{P}\left( |k^*-\hat{k}| \leq \epsilon k^*\right) = 1.$$ This means that the estimated value $\hat{k}$ with probability tending to $1$ is in a neighbourhood of the true value $k^*$: $$\operatorname{P}\big(\hat{k}\in[k^*(1-\epsilon),k^*(1+\epsilon)]\big) \rightarrow 1.$$ We will show that as $n\rightarrow\infty$, $$\label{zz_consistency_2}
\operatorname{P}\Big( \max_{k:|k-k^*|\geq \epsilon k^*}|W_n(k)| < |W_n(k^*)| \Big) \rightarrow 1.$$ Since $|W_n(k^*)|\leq \max_{k:|k^*-k|\leq \epsilon k^*}|W_n(k)|$, this proves (\[zz\_consistency\_1\]).
By (\[decompose\_Wn\]), $$W_{n}(k)= \begin{cases}
U_n(k) + U_n(k,k^*),\qquad 1\leq k \leq k^*\\
U_n(k) + U_n(k^*,k),\qquad k^* < k \leq n.
\end{cases}$$
implies $\max_{1\leq k \leq n} \left|U_n(k)\right| = O_P\left(n^{3/2}\right)$ and below yields $$\begin{aligned}
\max_{1\leq k \leq k^*}\left| U_n(k,k^*)- k\left(n-k^*\right)\Theta_{\Delta_n}\right| &= o_P\left(n^{3/2}\right),\\
\max_{k^*\leq k \leq n}\left|U_n(k^*,k)-k^*\left(n-k\right)\Theta_{\Delta_n}\right| &= o_P\left(n^{3/2}\right).\end{aligned}$$ Hence, $$\begin{aligned}
W_n(k^*) &=k^*(n-k^*)\Theta_{\Delta_n} + (U_n(k^*,k^*)-k^*(n-k^*)\Theta_{\Delta_n})+U_n(k^*)\\ &= k^*(n-k^*)\Theta_{\Delta_n} + O_P(n^{3/2}),\\
\max_{1\leq k \leq k^*(1-\epsilon)} |W_n(k)| &\leq (1-\epsilon)k^*(n-k^*)\Theta_{\Delta_n} + O_P(n^{3/2}),\\
\max_{(1+\epsilon)k^*\leq k \leq n} |W_n(k)| &\leq k^*(n-(1+\epsilon)k^*)\Theta_{\Delta_n} + O_P(n^{3/2}).\end{aligned}$$ Thus, $$|W_n(k^*)|-\max_{k:|k^*-k|\geq \epsilon k^*} |W_n(k)| \geq \epsilon \delta_n + O_P(n^{3/2}),$$ where $\delta_n = k^*\min(n-k^*,{k^*})\Theta_{\Delta_n}$.
By definition $k^*=[n\theta]\sim n\theta$, and by (\[remark\_theta\_delta\_n\]) and (\[assumption\_size\_change\_infty\]), $\sqrt{n}\Theta_{\Delta_n}\sim c\sqrt{n}\Delta_n\rightarrow\infty$. Hence, $\delta_n^{-1}=o(n^{-3/2})$ and $\epsilon \delta_n + O_P(n^{3/2}) = \epsilon\delta_n(1+O_P(n^{3/2}\delta_n^{-1}))=\epsilon\delta_n(1+o_P(1))$ which proves (\[zz\_consistency\_2\]).
Next we establish the rate of convergence in (\[rate\_consistency\_k\_hat\]), $k^*-k=O_P(1/\Delta_n^2)$. Set $a(n)=\frac{M}{\Delta_n^2}$. Then for fixed $M>0$, $a(n)\rightarrow\infty$, as $n\rightarrow\infty$. We will verify that $$\lim_{n\rightarrow \infty}\operatorname{P}\left( |k^*-\hat{k}| \leq a(n)\right) \rightarrow 1, \qquad \text{as} \; M\rightarrow\infty,$$ which implies (\[rate\_consistency\_k\_hat\]). As in (\[zz\_consistency\_2\]), we prove this by showing $$\label{zz_consistency_3}
\lim_{n\rightarrow \infty}\operatorname{P}\Big( \max_{k:|k-k^*|\geq a(n)}|W_n(k)| < |W_n(k^*)| \Big) \rightarrow 1, \qquad \text{as} \; M\rightarrow\infty.$$
Define $V_k := W_{n}^2\left(k\right)-W_{n}^2\left(k^*\right)$. If $|W_{n}\left(k\right)|$ attains its maximum at $k'$, it is easy to see that $V_k$ attains its maximum at the same $k'$. Hence, $ \hat{k}= \min \{ k: \max_{1\leq l \leq n} |W_n(l)|=|W_n(k)| \} = \min \{ k: \max_{1\leq l \leq n} V_l = V_k\}$. Thus, instead of (\[zz\_consistency\_3\]) it remains to show that $$\label{zz_consistency_4}
\lim_{n\rightarrow \infty}\operatorname{P}\Big( \max_{k:|k-k^*|\geq a(n)}V_k < 0 \Big) \rightarrow 1, \qquad M\rightarrow\infty.$$
Define $\tilde{k} := \min \{ k: |k-k^*|\leq \epsilon k^*; V_k = \max_{n\alpha\leq l \leq n\beta} V_l\}$. Since by (\[zz\_consistency\_1\]) $\hat{k}$ is a consistent estimator of $k^*$, it holds $\lim_{n\rightarrow\infty}\operatorname{P}(\hat{k}=\tilde{k})=1$.
So, in the proof of (\[zz\_consistency\_4\]) it suffices to consider $\max$ over $k$, such that $|k-k^*|\leq \epsilon k^*$, $|k-k^*|\geq a(n)$, which corresponds to $(1-\epsilon)k^*\leq k \leq k^*-a(n)$ and $k^*+a(n)<k\leq (1+\epsilon)k^*$.
Let us start with $(1-\epsilon)k^*\leq k \leq k^*-a(n)$. Since $k^*-k>0$, relation (\[zz\_consistency\_4\]) holds for such $k$, if $$\label{zz_consistency_5}
\lim_{n\rightarrow \infty}\operatorname{P}\Big( \max_{(1-\epsilon)k^*\leq k \leq k^*-a(n)}\frac{V_k}{(n(k^*-k))^2} < 0 \Big) \rightarrow 1, \qquad M\rightarrow\infty.$$
Note that $$\begin{aligned}
\label{zerlegung_v_k}
\frac{-V_k}{(n(k^*-k))^2} &= \frac{W_n^2(k^*)-W_n^2(k)}{(n(k^*-k))^2}\nonumber\\
&= -\bigg(\frac{W_n(k^*)-W_n(k)}{n(k^*-k)}\bigg)^2+2\frac{W_n(k^*)-W_n(k)}{n(k^*-k)}\frac{W_n(k^*)}{n(k^*-k)}.\end{aligned}$$
By (\[decompose\_Wn\]), $W_n(k)=U_n(k)+U_n(k,k^*)$. Then, $$\frac{W_n(k^*)-W_n(k)}{n(k^*-k)} = \frac{n-k^*}{n}\Theta_{\Delta_n} + \delta_{1,k} + \delta_{2,k},$$ where $$\delta_{1,k}= \frac{U_n(k^*)-U_n(k)}{n(k^*-k)}, \qquad \delta_{2,k}=\frac{U_n(k^*,k^*)-U_n(k,k^*)}{n(k^*-k)}-\frac{n-k^*}{n}\Theta_{\Delta_n}.$$ Observe that by (\[remark\_theta\_delta\_n\]), $\Theta_{\Delta_n}\sim c_{*}\Delta_n$, $c_{*}>0$, and $k^*/n\rightarrow\theta$. Therefore, $(n-k^*)/n\Theta_{\Delta_n}\sim c_0\Delta_n$, where $c_0=(1-\theta)c_{*}$. Moreover, $\max_{1\leq k \leq k^*-a(n)}|\delta_{i,k}|=o_P(\Delta_n)$, $i=1,2$, by (\[zz\_gewichtetes\_maximum\]) and (\[zz\_gewichtetes\_maximum\_unkk\]) of . Hence, $$\label{abschaetzung_1}
\frac{W_n(k^*)-W_n(k)}{n(k^*-k)}=c_0\Delta_n(1+o_P(1)), \qquad \bigg(\frac{W_n(k^*)-W_n(k)}{n(k^*-k)}\bigg)^2=c_0^2\Delta_n^2(1+o_P(1)).$$ In turn, $$\frac{W_n(k^*)}{n(k^*-k)}=\frac{U_n(k^*,k^*)}{n(k^*-k)}+\frac{U_n(k^*)}{n(k^*-k)}$$ and $$\frac{U_n(k^*,k^*)}{n} = \frac{k^*(n-k^*)\Theta_{\Delta_n}}{n}+ \frac{\delta_{3,k}}{n},$$ where $\delta_{3,k}=U_n(k^*,k^*)-k^*(n-k^*)\Theta_{\Delta_n}$. By , $\max_{1\leq k\leq k^*}\delta_{3,k}/n=o_P(n^{1/2})$. Since $$\frac{k^*(n-k^*)\Theta_{\Delta_n}}{n}\sim k^*c_0\Delta_n \sim \theta c_0 n\Delta_n$$ and $\sqrt{n}=o(n\Delta_n)$, this implies $$\frac{U_n(k^*,k^*)}{n} = k^*c_0\Delta_n(1+o_P(1)).$$ Next, by below, $U_n(k^*)=O_P(n^{3/2})$, and hence, $U_n(k^*)/n=O_P(n^{1/2})$. Therefore, $W_n(k^*)/n= k^*c_0\Delta_n(1+o_P(1))$. Hence, for $(1-\epsilon)k^*\leq k\leq k^*-a(n)$, $$\label{abschaetzung_2}
\frac{W_n(k^*)}{n(k^*-k)}=\frac{k^*c_0\Delta_n(1+o_P(1))}{k^*-k} \geq \frac{1}{\epsilon}c_0\Delta_n(1+o_P(1)).$$ Using (\[abschaetzung\_1\]) and (\[abschaetzung\_2\]) in (\[zerlegung\_v\_k\]), it follows $$\begin{aligned}
-\frac{V_k}{(n(k^*-k))^2} \geq \frac{2}{\epsilon}c_0^2\Delta_n^2(1+o_P(1))-c_0^2\Delta_n^2(1+o_P(1)) \geq \Big(\frac{2}{\epsilon}-1\Big)(c_0\Delta_n)^2(1+o_P(1)) >0.\end{aligned}$$ This proves (\[zz\_consistency\_5\]). Similar argument yields $$\lim_{n\rightarrow\infty}\operatorname{P}\Big( \max_{k^*+a(n) \leq k \leq k^*(1+\epsilon)} V_k < 0 \Big) \rightarrow 1, \qquad M\rightarrow\infty,$$ which completes the proof of (\[zz\_consistency\_4\]) and the theorem. $\hfill \Box $
Auxiliary results {#Auxiliary Results}
=================
This section contains auxiliary results used in the proof of .
We establish asymptotic properties of the quantities $U_n(k)$, $U_n(k,k^*)$ and $U_n(k^*,k)$ defined in (\[decompose\_Un\])-(\[decompose\_Un\_k2\]) and appearing in the decomposition (\[decompose\_Wn\]) of $W_n(k)$.
The following lemma derives a Hájek-Rényi type inequality for $L_1$ NED random variables.
Let $\left(Y_j\right)$ be a stationary $L_1$ near epoch dependent process on some absolutely regular process $\left(Z_j\right)$, satisfying (\[condition\_appr.const\_regu.coeff\]). Assume that $\operatorname{E}Y_j=0$ and $|Y_j|\leq K\leq \infty$ a.s. for some $K\geq 0$. Then, for all fixed $\epsilon>0$, for all $1\leq m \leq n$, $$\label{hajek_renyi}
\operatorname{P}\bigg(\max_{m\leq k \leq n} \frac{1}{k}\bigg|\sum_{i=1}^{k}Y_i\bigg| >\epsilon\bigg) \leq \frac{1}{\epsilon^2}\frac{C}{\sqrt{m}},$$ where $C>0$ does not depend on $m$, $n$, $\epsilon$.
To prove (\[hajek\_renyi\]), we use the Hájek-Rényi type inequality of established in @Kokoszka.2000, $$\begin{gathered}
\label{inequality_KL}
\epsilon^2 \operatorname{P}\bigg(\max_{m\leq k \leq n} \frac{1}{k}\bigg|\sum_{i=1}^{k}Y_i\bigg| >\epsilon\bigg) \leq \frac{1}{m^2}\operatorname{E}\bigg(\sum_{i=1}^{m}Y_i\bigg)^2 + \sum_{k=m}^{n}\Big|\frac{1}{(k+1)^2}-\frac{1}{k^2}\Big|\operatorname{E}\bigg(\sum_{i=1}^{k}Y_i\bigg)^2\\
+ 2\sum_{k=m}^{n}\frac{1}{(k+1)^2}\operatorname{E}\bigg(\left|Y_{k+1}\right|\bigg|\sum_{j=1}^{k}Y_j\bigg|\bigg) + \sum_{k=m}^{n}\frac{1}{(k+1)^2}\operatorname{E}Y_{k+1}^2.\end{gathered}$$ First we bound $\operatorname{E}\big(\sum_{i=1}^{k}Y_i\big)^2$. Under assumptions of this lemma, by below, for $i,j\geq 0$ $$\label{abschaetzung_covariance}
\left| \operatorname{Cov}\left(Y_i,Y_{i+j}\right)\right|=\left| \operatorname{E}\left(Y_iY_{i+j}\right)\right| \leq 4Ka_{\lfloor \frac{j}{3}\rfloor} + 2K^2\beta_{\lfloor \frac{j}{3}\rfloor} \leq C(a_{\lfloor \frac{j}{3}\rfloor}+\beta_{\lfloor \frac{j}{3}\rfloor}).$$ By stationarity of $(Y_j)$, $$\left| \operatorname{E}\left(Y_iY_{j}\right)\right| = \left| \operatorname{Cov}\left(Y_i,Y_{j}\right)\right|= \left| \operatorname{Cov}\left(Y_0,Y_{|i-j|}\right)\right|.$$ Hence, $$\begin{gathered}
\operatorname{E}\Big(\sum_{i=1}^{k}Y_i\Big)^2 = \sum_{i,j=1}^{k}\operatorname{E}\left(Y_iY_j\right)\leq \sum_{i,j=1}^{k}\left| \operatorname{Cov}\left(Y_0,Y_{|i-j|}\right)\right|\\
\leq C\sum_{i,j=1}^{k}(a_{\lfloor \frac{|i-j|}{3}\rfloor}+\beta_{\lfloor \frac{|i-j|}{3}\rfloor})\leq C\sum_{i=1}^{k}\sum_{k=0}^{\infty}(a_{\lfloor \frac{k}{3}\rfloor}+\beta_{\lfloor \frac{k}{3}\rfloor})\leq Ck,\end{gathered}$$ by (\[abschaetzung\_covariance\]) and (\[condition\_appr.const\_regu.coeff\]). Since $|Y_j|\leq K$, then $$\operatorname{E}\Big(\left|Y_{k+1}\right|\Big|\sum_{j=1}^{k}Y_j\Big|\Big) \leq K\operatorname{E}\Big(\Big|\sum_{j=1}^{k}Y_j\Big|\Big)\leq K\Big( \operatorname{E}\Big(\sum_{i=1}^{k}Y_i\Big)^2 \Big)^{1/2}\leq C\sqrt{k}.$$ Using these bounds in (\[inequality\_KL\]) together with $$\frac{1}{(k+1)^2}\leq \frac{1}{k^2}, \qquad \Big|\frac{1}{(k+1)^2}-\frac{1}{k^2}\Big| \leq \frac{1+2k}{(k+1)^2k^2}\leq \frac{4}{k^3},$$ we obtain (\[hajek\_renyi\]): $$\epsilon^2 \operatorname{P}\bigg(\max_{m\leq k \leq n} \frac{1}{k}\bigg|\sum_{i=1}^{k}Y_i\bigg| >\epsilon\bigg) \leq C\bigg[ \frac{1}{m}+\sum_{k=m}^{n}\frac{1}{k^2} + \sum_{k=m}^{n}\frac{1}{k^{3/2}}\bigg] \leq \frac{C}{\sqrt{m}}.$$
The next lemma establishes asymptotic bounds of the sums $$\label{s_k}
S_k^{(1)} = \sum_{i=1}^{k}h_{1,n}\left(Y_i\right), \qquad S_k^{(2)} =\sum_{j=1}^{k}h_{2,n}\left(Y_j\right).$$
Assume that $\left(Y_j\right)$ is a stationary zero mean $L_1$ near epoch dependent process on some absolutely regular process $\left(Z_j\right)$ and (\[condition\_appr.const\_regu.coeff\]) holds. Furthermore, let Assumption \[assumption\_change\] be satisfied and $S_k^{(i)}$, $i=1,2$, be as in (\[s\_k\]). Then $$\label{abschaetzung_h1}
\max_{1\leq k \leq n}n^{-1/2}\Big| S_k^{(i)}\Big| = o_P\left(1\right), \qquad i=1,2.$$
To show (\[abschaetzung\_h1\]) for $i=1$, we will use the inequality given in . Define $S_k = \sum_{i=1}^{k}n^{-1/2} h_{1,n}\left(Y_i\right)$, $k\geq 1$, and set $S_0=0$. We need to evaluate $\operatorname{E}( S_l-S_k )^4 $ for $1\leq k < l \leq n$. Note that $$\operatorname{E}\left( S_l-S_k \right)^4
= n^{-2}\operatorname{E}\bigg|\sum_{i=k+1}^{l} h_{1,n}\left(Y_i\right)\bigg|^4 = n^{-2}\operatorname{E}\bigg|\sum_{i=1}^{l-k} h_{1,n}\left(Y_i\right)\bigg|^4,$$ where the last equality holds because $(h_{1,n}(Y_j))$ is a stationary process. Since $(h_{1,n}(Y_j))$ is $L_1$ NED on an absolutely regular process, see Section \[Properties of the kernel\], $\operatorname{E}h_{1,n}(Y_0)=0$ and $|h_{1,n}(x)|\leq C\Delta_n$ by (\[abschaetzung\_h1n\_kleiner\_delta\]), then by and the comment below $$\operatorname{E}\bigg|\sum_{i=1}^{l-k}h_{1,n}(Y_i)\bigg|^4 \leq C(l-k)^2\Delta_n^2,$$ where $C$ does not depend on $l$, $k$ or $n$. Thus, $$\operatorname{P}(|S_l-S_k|\geq \lambda) \leq \frac{1}{\lambda^4}\operatorname{E}|S_l-S_k|^4 \leq \frac{C(l-k)^2\Delta_n^2}{\lambda^4n^2} = \frac{1}{\lambda^4}\bigg(\sum_{i=k+1}^{l}u_{n,i}\bigg)^2,$$ where $u_{n,i} = C^{1/2} \Delta_n n^{-1}$. Hence, $S_j$ satisfies assumption (\[condition\_prop\_billingsley\]) of with $\beta=4$, $\alpha=2$. Therefore, by (\[result\_prop\_billingsley\]), for any fixed $\epsilon >0$, as $n\rightarrow\infty$, $$\begin{aligned}
\operatorname{P}\bigg(\max_{1\leq k \leq n}n^{-1/2} \Big|S_k^{(1)}\Big| \geq \epsilon\bigg) \leq \frac{K}{\epsilon^4}\bigg(\sum_{i=1}^{n} u_{n,i} \bigg)^2 = \frac{KC\Delta_n^2}{\epsilon^4} \rightarrow 0,\end{aligned}$$ since $\Delta_n\rightarrow 0$. The proof of (\[abschaetzung\_h1\]) for $i=2$ follows using a similar argument as in the proof for $i=1$.
Assume that $\left(Y_j\right)$ is $L_1$ near epoch dependent process on some absolutely regular process $\left(Z_j\right)$ and (\[condition\_appr.const\_regu.coeff\]) holds. Furthermore, let Assumption \[assumption\_change\] be satisfied. Then $$\label{Lemma_abschaetzung_1}
\max_{1\leq k \leq k^*}n^{-3/2}\bigg| U_n(k,k^*) -k\left(n-k^*\right)\Theta_{\Delta_n} \bigg| = o_P\left(1\right)$$ and $$\label{Lemma_abschaetzung_2}
\max_{k^*\leq k \leq n}n^{-3/2}\bigg| U_n(k^*,k) -k^*\left(n-k\right)\Theta_{\Delta_n} \bigg| = o_P\left(1\right),$$ where $\Theta_{\Delta_n}$ is the same as in (\[remark\_theta\_delta\_n\]).
By the Hoeffding decomposition (\[hoeffding\_decomposition\_hn\]), $$h_n\left(x,y\right) - \Theta_{\Delta_n} = h_{1,n}\left(x\right) + h_{2,n}\left(y\right) + g_n\left(x,y\right).$$ Hence, $$\begin{gathered}
U_n(k,k^*)-k(n-k^*)\Theta_{\Delta_n} = \sum_{i=1}^{k} \sum_{j=k^*+1}^{n} \big(h_{1,n}\left(Y_i\right) + h_{2,n}\left(Y_j\right) + g_n\left(Y_i,Y_j\right)\big)\\
= (n-k^*)\sum_{i=1}^{k}h_{1,n}\left(Y_i\right) + k \sum_{j=k^*+1}^{n}h_{2,n}\left(Y_j\right) + \sum_{i=1}^{k} \sum_{j=k^*+1}^{n}g_n\left(Y_i,Y_j\right).\end{gathered}$$ Denote $$\begin{aligned}
U_n^{(g)}(k,k^*)&=\sum_{i=1}^{k} \sum_{j=k^*+1}^{n}g_n\left(Y_i,Y_j\right), &U_n^{(g)}(k^*,k)&=\sum_{i=1}^{k^*} \sum_{j=k+1}^{n}g_n\left(Y_i,Y_j\right).\label{U_n_g}\end{aligned}$$ Since $|n-k^*|\leq n$, $k^*\leq n$ and $\sum_{i=k^*+1}^{n}h_{2,n}(Y_j)=S_n^{(2)}-S_{k^*}^{(2)}$, then $$\begin{aligned}
\Big|U_n(k,k^*)-k(n-k^*)\Theta_{\Delta_n}\Big| &\leq n\Big(\big|S_k^{(1)}\big|+\big|S_{n}^{(2)}\big|+\big|S_{k^*}^{(2)}\big|\Big)+\Big|U_n^{(g)}(k,k^*)\Big|,\\
\Big|U_n(k^*,k)-k^*(n-k)\Theta_{\Delta_n}\Big| &\leq n\Big(\big|S_{k^*}^{(1)}\big|+\big|S_{n}^{(2)}\big|+\big|S_{k}^{(2)}\big|\Big)+\Big|U_n^{(g)}(k^*,k)\Big|,\end{aligned}$$ where $S_{k}^{(i)}$, $i=1,2$ are defined in (\[s\_k\]). Therefore, $$\begin{gathered}
\label{abschaetzung_Lemma_abschaetzung_1}
\max_{1\leq k \leq k^*}n^{-3/2}\Big|U_n(k,k^*)-k(n-k^*)\Theta_{\Delta_n}\Big|\\
\leq \max_{1\leq k \leq n}n^{-1/2}\Big(\big|S_k^{(1)}\big|+\big|S_{n}^{(2)}\big|+\big|S_{k^*}^{(2)}\big|\Big) + \max_{1\leq k \leq k^*}n^{-3/2}\Big|U_n^{(g)}(k,k^*)\Big|.\end{gathered}$$ The degenerate kernel $g_n$ is bounded and $1$-continuous, see Subsections \[Sub\_U-statistics and Hoeffding decomposition\] and \[Sub\_1-continuity\_h\_hn\]. Thus, by below, $$\begin{gathered}
\label{abschaetzung_gn_1}
\max_{1\leq k \leq k^*}n^{-3/2}\bigg| U_n^{(g)}(k,k^*)\bigg|\\
\leq \max_{1\leq k \leq n}n^{-3/2}\bigg|\sum_{i=1}^{k}\sum_{j=k+1}^{n}g_n(Y_i,Y_j)\bigg|+\max_{1\leq k \leq k^*}n^{-3/2}\bigg|\sum_{i=1}^{k}\sum_{j=k+1}^{k^*}g_n(Y_i,Y_j)\bigg|
=o_P(1).\end{gathered}$$ Similar argument implies $$\max_{k^*< k \leq n}n^{-3/2}\bigg| U_n^{(g)}(k^*,k)\bigg| =o_P(1).$$ Using in (\[abschaetzung\_Lemma\_abschaetzung\_1\]) the bounds (\[abschaetzung\_gn\_1\]) and (\[abschaetzung\_h1\]) of we obtain $$\max_{1\leq k \leq k^*}n^{-3/2}\Big|U_n(k,k^*)-k(n-k^*)\Theta_{\Delta_n}\Big| = o_P(1)$$ which proves (\[Lemma\_abschaetzung\_1\]). The proof of (\[Lemma\_abschaetzung\_2\]) follows using similar argument.
Denote $$\label{tilde_U_n_g}
\tilde{U}_n^{(g)}(k) = \sum_{i=1}^{k}\sum_{j=k+1}^{n}g(Y_i,Y_j).$$
Assume that $\left(Y_j\right)$ is $L_1$ near epoch dependent process on some absolutely regular process $\left(Z_j\right)$ and (\[condition\_appr.const\_regu.coeff\]) holds. Furthermore, let Assumption \[assumption\_change\] be satisfied and let $a(n) = M/\Delta_n^2$, $M>0$, and $\tilde{U}_n^{(g)}(k)$, $U_n^{(g)}(k,k^*)$ and $U_n^{(g)}(k^*,k)$ are defined as in (\[tilde\_U\_n\_g\]), (\[U\_n\_g\]). Then there exists $C>0$ such that for any $\epsilon>0$, $$\label{hajek_renyi_g_1}
\operatorname{P}\bigg(\max_{k:|k-k^*|\geq a(n)}\Big|\frac{\tilde{U}_n^{(g)}(k)}{k^*-k}\Big|>\epsilon\bigg) \leq \frac{C}{\epsilon^2}\Big(\frac{n^2}{a(n)}+\frac{1}{n}\Big),$$ $$\label{hajek_renyi_g_2}
\operatorname{P}\bigg(\max_{1\leq k \leq k^*-a(n)}\Big|\frac{U_n^{(g)}(k,k^*)}{k^*-k}\Big|>\epsilon\bigg) \leq \frac{C}{\epsilon^2}\Big(\frac{n^2}{a(n)}+\frac{1}{n}\Big),$$ $$\operatorname{P}\bigg(\max_{k^*+a(n)\leq k \leq n}\Big|\frac{U_n^{(g)}(k^*,k)}{k-k^*}\Big|>\epsilon\bigg) \leq \frac{C}{\epsilon^2}\Big(\frac{n^2}{a(n)}+\frac{1}{n}\Big),$$ where $C$ does not depend on $\epsilon$, $n$ and $a(n)$.
Recall $\{k:|k-k^*|\geq a(n)\}=\{k\leq k^*-a(n)\}\cup\{k\geq k^*+a(n)\}$. We consider only the case $\max_{1\leq k \leq k^*-a(n)}$ since the proof for $\max_{k^*+a(n)\leq k\leq n}$ is similar.
*Proof of (\[hajek\_renyi\_g\_1\]).* Define $R_k = \tilde{U}_n^{(g)}(k)-\tilde{U}_n^{(g)}(k-1)$, $k\geq 1$, $\tilde{U}_n^{(g)}(0)=0$ and $R_0=0$. Then $\tilde{U}_n^{(g)}(k)=\sum_{i=1}^k R_i$. Inequality (\[inequality\_hr\]) of , applied to the random variables $R_i$ with $c_k=1/(k^*-k)$ yields $$\begin{gathered}
\label{zz_hajek_renyi_g}
\rho_n := \epsilon^2 \operatorname{P}\bigg(\max_{1\leq k \leq k^*-a(n)} \frac{1}{k^*-k}\bigg|\sum_{i=1}^{k}R_i\bigg| >\epsilon\bigg)\\
\leq \frac{1}{(k^*-1)^2}\operatorname{E}R_1^2 + \sum_{k=1}^{k^*-a(n)}\bigg|\frac{1}{(k^*-k-1)^2}-\frac{1}{(k^*-k)^2}\bigg|\operatorname{E}\bigg(\sum_{i=1}^{k}R_i\bigg)^2\\
+ 2\sum_{k=1}^{k^*-a(n)}\frac{1}{(k^*-k-1)^2}\operatorname{E}\bigg(\left|R_{k+1}\right|\bigg|\sum_{j=1}^{k}R_j\bigg|\bigg) + \sum_{k=1}^{k^*-a(n)}\frac{1}{(k^*-k-1)^2}\operatorname{E}R_{k+1}^2.\end{gathered}$$ In Subsections \[Sub\_U-statistics and Hoeffding decomposition\] and \[Sub\_1-continuity\_h\_hn\], we showed that kernel function $g(x,y)$ is bounded and $1$-continuous. Therefore, by below $$\label{hr_g_1}
\operatorname{E}\Big[ (\tilde{U}_n^{(g)}(k))^2\Big]=\operatorname{E}\bigg(\sum_{i=1}^{k}R_i\bigg)^2 \leq Ck(n-k), \qquad k=1,\ldots,n.$$ also yields $$\label{hr_g_2}
\operatorname{E}R_{k+1}^2 = \operatorname{E}\left(\tilde{U}_n^{(g)}(k+1)-\tilde{U}_n^{(g)}(k)\right)^2 \leq n^3C\frac{(k+1)-k}{n^2} = Cn, \qquad k=1,\ldots,n.$$ Then, $$\operatorname{E}\bigg(\left|R_{k+1}\right|\bigg|\sum_{j=1}^{k}R_j\bigg|\bigg) \leq \bigg(\operatorname{E}R_{k+1}^2 \bigg)^{1/2}\bigg(\operatorname{E}\Big(\sum_{j=1}^{k}R_j\Big)^2 \bigg)^{1/2}\leq C\sqrt{n}\sqrt{k(n-k)}.$$ From (\[zz\_hajek\_renyi\_g\]), (\[hr\_g\_1\]) and (\[hr\_g\_2\]), using $\frac{1}{(k^*-k-1)^2}-\frac{1}{(k^*-k)^2}\leq \frac{2}{(k^*-k-1)^3}$, we obtain $$\rho_n \leq C\bigg[ \frac{n-1}{(k^*-1)^2} + \sum_{k=1}^{k^*-a(n)}\bigg\{\frac{k(n-k)}{(k^*-k-1)^3}
+ \frac{\sqrt{n}\sqrt{k(n-k)}+n}{(k^*-k-1)^2}\bigg\}\bigg] .$$ Noting that $\sqrt{k(n-k)}\leq n$, $(k^*-k-1)^{-3}\leq (k^*-k-1)^{-2}$, it follows $$\rho_n \leq C\bigg( \frac{1}{n}+ \sum_{k=1}^{k^*-a(n)}\frac{n^2}{(k^*-k-1)^2}\bigg) \leq C\Big(\frac{1}{n}+\frac{n^2}{a(n)}\Big).$$ *Proof of (\[hajek\_renyi\_g\_2\]).* It follows a similar line to the proof of (\[hajek\_renyi\_g\_1\]). Denote $\tilde{R}_k=U_n^{(g)}(k,k^*)-U_n^{(g)}(k-1,k^*)$. We verified in Subsections \[Sub\_U-statistics and Hoeffding decomposition\] and \[Sub\_1-continuity\_h\_hn\] that function $g_n(x,y)$ is bounded and $1$-continuous. Therefore, by below, $$\operatorname{E}\Big[(U_n^{(g)}(k,k^*))^2\Big]=\operatorname{E}\bigg(\sum_{i=1}^{k}\tilde{R}_i\bigg)^2 \leq Ck(n-k^*), \qquad k=1,\ldots,k^*$$ and $$\begin{gathered}
\operatorname{E}\tilde{R}_{k+1}^2 = \operatorname{E}\left(U_n^{(g)}(k+1,k^*)-U_n^{(g)}(k,k^*)\right)^2\\
= \operatorname{E}\Big(\sum_{j=k^*+1}^{n}g_n(Y_{k+1},Y_j)\Big)^2 \leq C(n-k^*), \qquad k=1,\ldots,k^*.\end{gathered}$$ Combining both bounds, we obtain $$\operatorname{E}\bigg(\left|\tilde{R}_{k+1}\right|\bigg|\sum_{j=1}^{k}\tilde{R}_j\bigg|\bigg) \leq \bigg(\operatorname{E}\tilde{R}_{k+1}^2 \bigg)^{1/2}\bigg(\operatorname{E}\Big(\sum_{j=1}^{k}\tilde{R}_j\Big)^2 \bigg)^{1/2}\leq C(n-k^*)\sqrt{k}.$$ Using the same argument as in the proof of (\[hajek\_renyi\_g\_1\]), we obtain $$\operatorname{P}\bigg(\max_{1\leq k \leq k^*-a(n)}\frac{1}{k^*-k}\Big|U_n^{(g)}(k,k^*)\Big|>\epsilon\bigg) \leq \frac{C}{\epsilon^2}\Big(\frac{1}{n}+\frac{n^2}{a(n)}\Big).$$ This completes proof of (\[hajek\_renyi\_g\_2\]) and the lemma.
Assume that $\left(Y_j\right)$ is a stationary zero mean $L_1$ near epoch dependent process on some absolutely regular process $\left(Z_j\right)$ and (\[condition\_appr.const\_regu.coeff\]) holds. Furthermore, let Assumption \[assumption\_change\] be satisfied and let $a(n) = M/\Delta_n^2$, $M>0$. Then, as $n\rightarrow\infty$, $M\rightarrow\infty$,
*(i)* For any $\epsilon>0$, $$\label{zz_gewichtetes_maximum}
\operatorname{P}\bigg(\frac{1}{n\Delta_n}\max_{k:|k-k^*|\geq a(n)}\bigg|\frac{U_n(k^*)-U_n(k)}{k^*-k}\bigg|>\epsilon\bigg)\rightarrow 0.$$ *(ii)* For any $\epsilon>0$, $$\label{zz_gewichtetes_maximum_unkk}
\operatorname{P}\bigg(\frac{1}{n\Delta_n}\max_{1\leq k \leq k^*-a(n)}\bigg|\frac{U_n(k^*,k^*)-U_n(k,k^*)}{k^*-k}-(n-k^*)\Theta_{\Delta_n}\bigg|>\epsilon\bigg)\rightarrow 0,$$ and $$\operatorname{P}\bigg(\frac{1}{n\Delta_n}\max_{k^*+a(n)\leq k \leq n}\bigg|\frac{U_n(k^*,k^*)-U_n(k^*,k)}{k-k^*}-k^*\Theta_{\Delta_n}\bigg|>\epsilon\bigg)\rightarrow 0,$$ where $\Theta_{\Delta_n}$ is the same as in (\[remark\_theta\_delta\_n\]).
Notice that $\{k:|k-k^*|\geq a(n)\}=\{k\leq k^*-a(n)\}\cup\{k\geq k^*+a(n)\}$. We will prove relations (\[zz\_gewichtetes\_maximum\]) and (\[zz\_gewichtetes\_maximum\_unkk\]) for $\max_{1\leq k \leq k^*-a(n)}$. The proof for $\max_{k^*+a(n)\leq k\leq n}$ is similar.
Notice, that $a(n)=\frac{Mn}{\Delta_n^2n}=o(Mn)$ since $n\Delta_n^2\rightarrow\infty$ by assumption (\[assumption\_size\_change\_infty\]). Therefore, for a fixed $M$, $a(n)=o(k^*)$ and $k^*-a(n)>1$ as $n\rightarrow\infty$.
*(i)* Denote $$\tilde{S}_k = \sum_{i=1}^k h_1(Y_i).$$
By Hoeffding’s decomposition (\[hoeffding\_decomposition\_h\]), for $k\leq k^*$, and using $h_1(x)=-h_2(x)$, it follows $$\begin{gathered}
U_n(k) = \sum_{i=1}^{k}\sum_{j=k+1}^{n} \big(h_1(Y_i)+ h_2(Y_j) + g(Y_i,Y_j)\big)\\
= (n-k)\sum_{i=1}^{k}h_1(Y_i) - k \sum_{j=k+1}^{n}h_1(Y_j) + \sum_{i=1}^{k}\sum_{j=k+1}^{n}g(Y_i,Y_j) = n\tilde{S}_k- k\tilde{S}_n + \tilde{U}_n^{(g)}(k),\end{gathered}$$ where $\tilde{U}_n^{(g)}(k)$ is defined in (\[tilde\_U\_n\_g\]). Hence, $$\Big|U_n(k^*)-U_n(k)\Big| = \Big|n\big(\tilde{S}_{k^*}-\tilde{S}_k\big)-(k^*-k)\tilde{S}_n + \tilde{U}_n^{(g)}(k^*)-\tilde{U}_n^{(g)}(k)\Big|.$$ Therefore, for $1\leq k\leq k^*-a(n)$, $$\begin{aligned}
\frac{1}{n\Delta_n}\Big|\frac{U_n(k^*)-U_n(k)}{k^*-k}&\Big|\\
&\leq \frac{1}{\Delta_n}\frac{\big|\tilde{S}_{k^*}-\tilde{S}_k\big|}{k^*-k}
+ \frac{1}{\Delta_n}\frac{\big|\tilde{S}_n\big|}{n} + \frac{\big|\tilde{U}_n^{(g)}(k^*)\big|}{n\Delta_na(n)} + \frac{1}{n\Delta_n}\frac{\big|\tilde{U}_n^{(g)}(k)\big|}{k^*-k}\\
&=: \rho_k^{(1)}+\rho_k^{(2)}+\rho_k^{(3)}+\rho_k^{(4)}.\end{aligned}$$
It suffices to show that for any $\epsilon>0$, as $n\rightarrow\infty$, for $l=1,\ldots,4$, $$\label{zz_gewichtetes_maximum_2}
\operatorname{P}\Big( \max_{1\leq k \leq k^*-a(n)} \rho_k^{(l)} >\epsilon \Big) \rightarrow 0, \quad M\rightarrow \infty,$$ which proves (\[zz\_gewichtetes\_maximum\]) for $\max_{1\leq k\leq k^*-a(n)}$.
For $l=1$, stationarity of the process $(h_1(Y_j))$ yields $$\Big\{ \big|\tilde{S}_{k^*}-\tilde{S}_{k}\big|=\big| \sum_{i=k+1}^{k^*}h_1(Y_i) \big|,\; 1\leq k\leq k^*-a(n) \Big\} \overset{d}{=} \Big\{ \big|\tilde{S}_{k^*-k}\big|,\; 1\leq k\leq k^*-a(n) \Big\}.$$ Therefore, $$\max_{1\leq k\leq k^*-a(n)}\rho_k^{(1)} \overset{d}{=} \frac{1}{\Delta_n}\max_{k\geq 1:\; k^*-k\geq a(n)}\frac{\big|\tilde{S}_{k^*-k}\big|}{k^*-k} \overset{d}{=} \frac{1}{\Delta_n}\max_{a(n)\leq j\leq n}\frac{\big|\tilde{S}_{j}\big|}{j}.$$ Since $(h_1(Y_j))$ is $L_1$ NED on an absolutely regular process $(Z_j)$, $\operatorname{E}h_1(Y_1)=0$ and $|h_1(x)|\leq 1/2$, then by , $$\label{hri}
\max_{a(n)\leq k \leq n}\frac{1}{k}\big|\tilde{S}_{k}\big| = O_P\bigg(\frac{1}{\sqrt{a(n)}}\bigg).$$ Thus, $$\max_{1\leq k\leq k^*-a(n)}\rho_k^{(1)} = O_P\bigg(\frac{1}{\Delta_n\sqrt{a(n)}}\bigg) = O_P\bigg(\frac{1}{\sqrt{M}}\bigg) = o_P(1), \quad \text{as} \; M\rightarrow\infty,$$ which proves (\[zz\_gewichtetes\_maximum\_2\]) for $l=1$.
For $l=2$, by (\[hri\]), $|\tilde{S}_n|/n = O_P(n^{-1/2})$. Thus, $$\rho_k^{(2)} = \frac{1}{\Delta_n}\frac{\tilde{S}_n}{n} = O_P\bigg(\frac{1}{\Delta_n\sqrt{n}}\bigg) = o_P(1),$$ since $\Delta_n\sqrt{n}\rightarrow\infty$ by (\[assumption\_size\_change\_infty\]), which proves (\[zz\_gewichtetes\_maximum\_2\]) for $l=2$.
To show (\[zz\_gewichtetes\_maximum\_2\]) for $l=3$, recall that $g(x,y)\leq 3/2$ is $1$-continuous, see Subsection \[Properties of the kernel\]. Therefore, by , $$\operatorname{E}\bigg( \frac{\tilde{U}_n^{(g)}(k^*)}{\sqrt{k^*(n-k^*)}} \bigg)^2 \leq C,$$ which implies that $$\frac{\tilde{U}_n^{(g)}(k^*)}{\sqrt{k^*(n-k^*)}} = O_P(1).$$ Thus, $$\begin{gathered}
\rho_k^{(3)} = \frac{\big|\tilde{U}_n^{(g)}(k^*) \big|}{n\Delta_na(n)} = O_P\bigg( \frac{\sqrt{k^*(n-k^*)}}{n\Delta_na(n)} \bigg)\\ = O_P\bigg( \frac{1}{\Delta_na(n)} \bigg) = O_P\bigg( \frac{\Delta_n}{M} \bigg) = O_P\bigg( \frac{1}{M} \bigg) = o_P(1),\quad \text{as} \; M\rightarrow\infty,\end{gathered}$$ which proves (\[zz\_gewichtetes\_maximum\_2\]) for $l=3$.
Finally, for $l=4$, by , $$\begin{aligned}
\operatorname{P}\Big( \max_{1\leq k \leq k^*-a(n)} \rho_k^{(4)} > \epsilon\Big) &= \operatorname{P}\Big( \max_{1\leq k \leq k^*-a(n)} \frac{\tilde{U}_n^{(g)}(k)}{k^*-k} > \epsilon n\Delta_n\Big)\\
&\leq \frac{C}{(\epsilon n\Delta_n)^2}\Big(\frac{n^2}{a(n)}+\frac{1}{n}\Big) = \frac{C}{\epsilon^2}\Big(\frac{1}{M}+\frac{1}{n^3\Delta_n^2}\Big) \rightarrow 0,\quad \text{as} \; M\rightarrow\infty, \end{aligned}$$ which proves (\[zz\_gewichtetes\_maximum\_2\]) for $l=4$ and completes the proof of *(i)*.
*(ii)* Let $S_k^{(1)}$, $S_k^{(2)}$ and $U_n^{(g)}(k,k^*)$ be defined as in (\[s\_k\]) and (\[U\_n\_g\]). By Hoeffding’s decomposition (\[hoeffding\_decomposition\_hn\]), for $k\leq k^*$, $$\begin{gathered}
U_n(k,k^*)-k(n-k^*)\Theta_{\Delta_n} = \sum_{i=1}^{k}\sum_{j=k^*+1}^{n} \big(h_{1,n}(Y_i)+ h_{2,n}(Y_j) + g_n(Y_i,Y_j)\big)\\
= (n-k^*)\sum_{i=1}^{k}h_{1,n}(Y_i) + k \sum_{j=k^*+1}^{n}h_{2,n}(Y_j) + \sum_{i=1}^{k}\sum_{j=k^*+1}^{n}g_n(Y_i,Y_j)\\
= (n-k^*)S_k^{(1)}+k(S_n^{(2)}-S_{k^*}^{(2)}) + U_n^{(g)}(k,k^*).\end{gathered}$$ Hence, $$\begin{gathered}
\big|U_n(k^*,k^*)-U_n(k,k^*)-(k^*-k)(n-k^*)\Theta_{\Delta_n}\big|\\
=\big|(n-k^*)(S_{k^*}^{(1)}-S_k^{(1)})+(k^*-k)(S_n^{(2)}-S_{k^*}^{(2)}) + U_n^{(g)}(k^*,k^*)-U_n^{(g)}(k,k^*)\big|.\end{gathered}$$ Therefore, for $1\leq k\leq k^*-a(n)$, $$\begin{aligned}
\frac{1}{n\Delta_n}&\Big|\frac{U_n(k^*,k^*)-U_n(k,k^*)}{k^*-k}-(n-k^*)\Theta_{\Delta_n}\Big|\\
&\leq \frac{1}{\Delta_n}\frac{\big|S_{k^*}^{(1)}-S_k^{(1)}\big|}{k^*-k}
+ \frac{1}{\Delta_n}\frac{\big|S_n^{(2)}-S_{k^*}^{(2)}\big|}{n} + \frac{\big|U_n^{(g)}(k^*,k^*)\big|}{n\Delta_na(n)} + \frac{1}{n\Delta_n}\frac{\big|U_n^{(g)}(k,k^*)\big|}{k^*-k}\\
&=: \nu_k^{(1)}+\nu_k^{(2)}+\nu_k^{(3)}+\nu_k^{(4)}.\end{aligned}$$ It suffices to show that for any $\epsilon>0$, as $n\rightarrow\infty$, for $l=1,\ldots,4$, $$\label{zz_gewichtetes_maximum_3}
\operatorname{P}\Big( \max_{1\leq k \leq k^*-a(n)} \nu_k^{(l)} >\epsilon \Big) \rightarrow 0, \quad M\rightarrow \infty,$$ which proves (\[zz\_gewichtetes\_maximum\_unkk\]) for $\max_{1\leq k \leq k^*-a(n)}$. The process $(h_{1,n}(Y_j))$ is stationary and $L_1$ NED on an absolutely regular process, see Section \[Properties of the kernel\]. Furthermore, it has zero mean and $|h_{1,n}|\leq C\Delta_n$ by (\[abschaetzung\_h1n\_kleiner\_delta\]). Hence, by the same argument as for $\rho_k^{(1)}$, using , it follows $$\max_{1\leq k \leq k^*-a(n)}\nu_k^{(1)} \overset{d}{=} \frac{1}{\Delta_n}\max_{a(n)\leq j\leq n}\frac{|S_j^{(1)}|}{j}=O_P\Big(\frac{1}{\Delta_n\sqrt{a(n)}}\Big)=O_P\Big(\frac{1}{\sqrt{M}}\Big)=o_P(1),$$ as $M\rightarrow\infty$.
yields $\max_{1\leq k\leq n}n^{-1/2}|S_k^{(2)}|=o_P(1)$. Therefore, $$\max_{1\leq k \leq k^*-a(n)}\nu_k^{(2)} \leq \frac{1}{\sqrt{n}\Delta_n}\big(n^{-1/2}|S_n^{(2)}|+n^{-1/2}|S_{k^*}^{(2)}|\big)= o_P(1),$$ since $\sqrt{n}\Delta_n\rightarrow\infty$.
We showed in Subsections \[Sub\_U-statistics and Hoeffding decomposition\] and \[Sub\_1-continuity\_h\_hn\] that the function $g_n(x,y)$ is bounded and $1$-continuous. Hence, by , $$\operatorname{E}\bigg( \frac{{U}_n^{(g)}(k^*,k^*)}{\sqrt{k^*(n-k^*)}} \bigg)^2 \leq C.$$ Therefore, the claim $\max_{1\leq k \leq k^*-a(n)}\nu_k^{(3)}=o_P(1)$ follows using the same argument as in the proof of (\[zz\_gewichtetes\_maximum\_2\]) for $l=3$.
By , $$\begin{gathered}
\operatorname{P}\Big( \max_{1\leq k \leq k^*-a(n)} \nu_k^{(4)} > \epsilon\Big) = \operatorname{P}\Big( \max_{1\leq k \leq k^*-a(n)} \frac{|{U}_n^{(g)}(k,k^*)|}{k^*-k} > \epsilon n\Delta_n\Big)\\
\leq \frac{C}{(\epsilon n\Delta_n)^2}\Big(\frac{n^2}{a(n)}+\frac{1}{n}\Big) = \frac{C}{\epsilon^2}\Big(\frac{1}{M}+\frac{1}{n^3\Delta_n^2}\Big) \rightarrow 0,\quad \text{as} \; M\rightarrow\infty, \end{gathered}$$ which proves (\[zz\_gewichtetes\_maximum\_3\]) for $l=4$. This completes the proof of (\[zz\_gewichtetes\_maximum\_unkk\]) and the lemma.
Auxiliary results from the literature {#result_literature}
=====================================
This section contains results from the literature used in the proofs of this paper.
states a correlation and a moment inequality for $L_1$ NED random variables, established by @borovkova.2001.
(Lemma 2.18 and 2.24, @borovkova.2001) Let $(Y_j)$ be $L_1$ near epoch dependent on an absolutely regular, stationary process with mixing coefficients $\beta_k$ and approximation constants $a_k$, and such that $|Y_0|\leq K\leq \infty$ a.s. Then, for all $i,k\geq 0$, $$\left| \operatorname{Cov}(Y_i,Y_{i+k})\right| \leq 4 Ka_{\lfloor \frac{k}{3} \rfloor} + 2 K^2 \beta_{\lfloor \frac{k}{3} \rfloor}.$$ In addition, if $\sum_{k=0}^{\infty}k^2 (a_k+\beta_k)<\infty$, then there exists $C>0$ such that for all $n\geq 1$ $$\label{lemma_2.24_result}
\operatorname{E}\bigg(\sum_{i=1}^n \big(Y_i-\operatorname{E}Y_i\big)\bigg)^4 \leq Cn^2.$$
The proof of Lemma 2.24 in @borovkova.2001 shows that (\[lemma\_2.24\_result\]) holds with $C=C_0K^2$, where $C_0>0$ does not depend on $K$ and $n$.
In Theorem 3 of @Dehling.2015 the asymptotic distribution of the Wilcoxon test statistic for $L_1$ NED random process is obtained. We use this result to show the consistency of the Wilcoxon-type estimator $\hat{k}$.
(Theorem 3, @Dehling.2015) Assume that $\left(Y_j\right)$ is stationary and $L_1$ near epoch dependent process on some absolutely regular process $\left(Z_j\right)$ and (\[condition\_appr.const\_regu.coeff\]) holds. Then, $$\frac{1}{n^{3/2}}\max_{1\leq k < n}\bigg| \sum_{i=1}^k\sum_{j=k+1}^n \left(1_{\left\{ Y_i \leq Y_j \right\}}-1/2\right) \bigg| \overset{d}{\rightarrow} \sigma \sup_{0\leq \tau \leq 1}\left| B\left(\tau\right) \right|,$$ where $\left(B\left(\tau\right)\right)_{0\leq \tau \leq 1}$ is the standard Brownian bridge process, $$\sigma^2 = \sum_{k=-\infty}^\infty \operatorname{Cov}\left(F\left(Y_k\right),F\left(Y_0\right)\right),$$ and $F$ denotes the distribution function of $Y_j$.
We use the following results from @Dehling.2015 to handle the degenerate part $g(x,y)$ of the Hoeffding decomposition (\[hoeffding\_decomposition\_h\]).
(Proposition 1, @Dehling.2015) Let $(Y_j)$ be stationary and $L_1$ near epoch dependent on an absolutely regular process with mixing coefficients $\beta_k$ and approximation constants $a_k$ satisfying $$\sum_{k=1}^{\infty}k\left(\beta_k+\sqrt{a_k}+\phi(a_k)\right) < \infty,$$ with $\phi(\epsilon)$ as in . If $g(x,y)$ is a 1-continuous bounded degenerate kernel, then, as $n\rightarrow\infty$, $$\frac{1}{n^{3/2}}\max_{1\leq k\leq n}\bigg| \sum_{i=1}^{k}\sum_{j=k+1}^{n}g\left(Y_i,Y_j\right)\bigg|\rightarrow_p 0.$$
(Lemma 1 and 2, @Dehling.2015) Under assumptions of there exists $C>0$ such that for all $1\leq m \leq k \leq n$, $n\geq 2$, $$\operatorname{E}\bigg(\sum_{i=1}^{k}\sum_{j=k+1}^{n}g(Y_i,Y_j) \bigg)^2 \leq Ck(n-k),$$ $$\operatorname{E}\bigg(n^{-3}\bigg|\sum_{i=1}^{k}\sum_{j=k+1}^{n}g(Y_i,Y_j)-\sum_{i=1}^{m}\sum_{j=m+1}^{n}g(Y_i,Y_j)\bigg|^2\bigg) \leq C\frac{k-m}{n^2}.$$
In our proofs we use the maximal inequality of @Billingsley.1999, which is valid for stationary/non-stationary and independent/dependent random variables $\xi_i$.
(Theorem 10.2, @Billingsley.1999) Let $\xi_1,\ldots,\xi_n$ be random variables and $S_k= \sum_{i=1}^{k}\xi_k$, $k\geq 1$, $S_0=0$ denotes the partial sum. Suppose that there exist $\alpha > 1$, $\beta>0$ and non-negative numbers $u_{n,1},\ldots,u_{n,n}$ such that $$\label{condition_prop_billingsley}
\operatorname{P}\bigg(\left|S_j-S_i\right|\geq \lambda\bigg) \leq \frac{1}{\lambda^{\beta}}\bigg(\sum_{l=i+1}^{j}u_{n,l}\bigg)^{\alpha},$$ for $\lambda>0$, $0\leq i \leq j \leq n$. Then for all $\lambda>0$, $n\geq 2$, $$\label{result_prop_billingsley}
\operatorname{P}\left(\max_{1\leq k \leq n}\left|S_k\right| \geq \lambda\right) \leq \frac{K}{\lambda^{\beta}}\bigg(\sum_{l=1}^{n}u_{n,l}\bigg)^{\alpha},$$ where $K>0$ depends only on $\alpha$ and $\beta$.
By the Markov inequality, (\[condition\_prop\_billingsley\]) is satisfied if $$\operatorname{E}\left|S_j-S_i\right|^{\beta} \leq \bigg(\sum_{l=i+1}^{j}u_{n,l}\bigg)^{\alpha}.$$
In the proof of we use a Hájek-Rényi type inequality established by @Kokoszka.2000.
(Theorem 4.1, @Kokoszka.2000) Let $X_1,..., X_n$ be any random variables with finite second moments and $c_1, ..., c_n$ be any non-negative constants. Then $$\begin{aligned}
\epsilon^2\operatorname{P}\bigg(\max_{m\leq k\leq n}c_k\bigg|\sum_{i=1}^{k}X_i\bigg|>\epsilon\bigg) &\leq c_m^2\sum_{i,j=1}^m\operatorname{E}\left(X_iX_j\right) + \sum_{k=m}^{n-1}\left|c_{k+1}^2-c_k^2\right|\sum_{i,j=1}^k\operatorname{E}\left(X_iX_j\right)\nonumber\\
&+ 2\sum_{k=m}^{n-1}c_{k+1}^2\operatorname{E}\bigg(\left|X_{k+1}\right|\bigg|\sum_{j=1}^{k}X_j\bigg|\bigg) + \sum_{k=m}^{n-1}c_{k+1}^2\operatorname{E}X_{k+1}^2.\label{inequality_hr}\end{aligned}$$
Acknowledgement {#acknowledgement .unnumbered}
===============
The author would like to thank Herold Dehling, Liudas Giraitis and Isabel Garcia for valuable discussions. The research was supported by the Collaborative Research Centre 823 *Statistical modelling of nonlinear dynamic processes* and the Konrad-Adenauer-Stiftung.
[998]{}
Antoch, J., Hu[š]{}kov[á]{}, M. and Veraverbeke, N. (1995). Change-point problem and bootstrap. *J. Nonparametr. Stat.* **5**, 123-144.
Bai, J. (1994). Least squares estimation of a shift in linear processes. *J. Time Series Anal.* **15**, 453-472.
Billingsley, P. (1999). *Convergence of Probability Measures*, 2nd ed. Wiley, New York.
Borovkova, S., Burton, R. and Dehling, H. (2001). Limit theorems for functionals of mixing processes with applications to [U]{}-statistics and dimension estimation. *Trans. Amer. Math. Soc.* **353**, 4261-4318.
Bradley, R.C. (2002). *Introduction to Strong Mixing Conditions*. Kendrick Press, Heber City.
Cs[ö]{}rg[ö]{}, M. and Horv[á]{}th, L. (1997). *Limit Theorems in Change-Point Analysis*. Wiley, New York.
Dehling, H., Fried, R., Garcia Arboleda, I. and Wendler, M. (2015). Change-point detection under dependence based on two-sample U-statistics. In: Dawson, D., Kulik, R., Jaye, M. O., Szyszkowicz, B., Zhao, Y.(Eds.) *Asymptotic laws and methods in stochastics. Fields Institute Communication* **76**, 195-220.
Dehling, H., Rooch, A. and Taqqu, M. S. (2013). Non-parametric change-point tests for long-range dependent data. *Scand. J. Stat.* **40**, 153-173.
Giraitis, L., Leipus, R. and Surgailis, D. (1996). The change-point problem for dependent observations. *J. Statist. Plann. Inference* **53**, 297-310.
Hansen, B. E. (1991). [GARCH]{}(1,1) processes are near epoch dependent. *Econom. Lett.* **36**, 181-186.
Hoeffding, W. (1948). A class of statistics with asymptotically normal distribution. *Ann. Math. Stat.* **19**, 293-325.
Horv[á]{}th, L. and Kokoszka, P. (1997). The effect of long-range dependence on change-point estimators. *J. Statist. Plann. Inference* **64**, 57-81.
Kokoszka, P. and Leipus, R. (1998). Change-point in the mean of dependent observations. *Statist. Probab. Lett.* **40**, 385-393.
Kokoszka, P. and Leipus, R. (2000). Change-point estimation in [ARCH]{} models. *Bernoulli* **6**, 513-539.
Ling, S. (2007). Testing for change points in time series models and limiting theorems for [NED]{} sequences. *Ann. Statist.* **35**, 1213-1237.
Wooldridge, J. M. and White, H. (1988). Some invariance principles and central limit theorems for dependent heterogeneous processes. *Econometric Theory* **4**, 210–230.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'State-of-the-art methods for sensing weak AC fields are only efficient in the low frequency domain ($<10$MHz). The inefficiency of sensing high frequency signals is due to the lack of ability to use dynamical decoupling. In this paper we show that dynamical decoupling can be incorporated into high frequency sensing schemes and by this we demonstrate that the high sensitivity achieved for low frequency can be extended to the whole spectrum. While our scheme is general and suitable to a variety of atomic and solid-state systems, we experimentally demonstrate it with the nitrogen-vacancy center in diamond. For a diamond with natural abundance of $^{13}$C we achieve coherence times up to 1.43ms resulting in a smallest detectable magnetic field strength of 4nT at 1.6GHz. Attributed to the inherent nature of our scheme, we observe an additional increase in coherence time due to the signal itself.'
author:
- 'A. Stark'
- 'N. Aharon'
- 'T. Unden'
- 'D. Louzon'
- 'A. Huck'
- 'A. Retzker'
- 'U.L. Andersen'
- 'F. Jelezko'
title: 'Narrow-bandwidth sensing of high-frequency fields with continuous dynamical decoupling'
---
Improving the sensitivity of high frequency sensing schemes is of great significance, especially for classical fields sensing [@chipaux_wide_2015; @kolkowitz_probing_2015; @shao_diamond_2016], detection of electron spins in solids [@sushkov_all-optical_2014; @hall_detection_2016] and nuclear magnetic resonance spectroscopy [@kimmich_field-cycling_2004]. The common method to detect high frequency field components is based on relaxation measurements, where the signal induces an observable effect on the lifetime, T$_{1}$, of the probe system [@sushkov_all-optical_2014; @schmid-lorch_relaxometry_2015; @hall_detection_2016]. Nevertheless, the sensitivity of this method is limited by the pure dephasing time T$_{2}^{*}$ of the probe system. Pulsed dynamical decoupling [@hahn_spin_1950; @carr_effects_1954; @meiboom_modified_1958] can substantially increase the coherence time [@viola_dynamical_1998; @biercuk_optimized_2009; @du_preserving_2009; @lange_universal_2010; @ryan_robust_2010; @naydenov_dynamical_2011; @wang_comparison_2012; @bar-gill_solid-state_2013]. In order to carry out sensing with a decoupling scheme, the frequency of the decoupling pulses has to be matched with the frequency of the target field [@taylor_high-sensitivity_2008; @kotler_single-ion_2011]. This largely restricts the approach to low frequencies, as the repetitive application of pulses is limited by the maximum available power per pulse [@gordon_optimal_2008]. The same power restrictions are present for very rapid and composite pulse sequences aimed to decrease both external and controller noise [@khodjasteh_fault-tolerant_2005; @uhrig_keeping_2007; @souza_robust_2011; @yang_preserving_2011; @farfurnik_optimizing_2015]. With continuous dynamical decoupling (CDD) [@fanchini_continuously_2007; @bermudez_electron-mediated_2011; @bermudez_robust_2012; @cai_long-lived_2012; @xu_coherence-protected_2012; @golter_protecting_2014; @rabl_strong_2009; @gordon_optimal_2008; @clausen_bath-optimized_2010; @laucht_dressed_2017; @cai_robust_2012; @cohen_multi-qubit_2015; @teissier_hybrid_2017] robustness to external and controller noise can be attained, especially for multi-level systems [@baumgart_ultrasensitive_2016; @timoney_quantum_2011; @aharon_general_2013; @aharon_fully_2016]. However, the significance of CDD for sensing high frequency fields remained elusive. Indeed, it was unclear whether it is possible to incorporate such a protection into the metrology task of sensing frequencies in the GHz domain. The first step towards this goal was done recently by integrating CDD in the sensing of high frequency fields with three level systems [@aharon_fully_2016]. In this article, we propose, analyze and experimentally demonstrate for the first time a sensing scheme that is capable of probing high frequency signals with a coherence time, T$_2$, limited sensitivity. Unlike relaxation measurements comprising a bandwidth $\propto 1/\text{T}_2^*$, determined by the pure dephasing time, T$_2^*$, of the sensor (up to the MHz range), our protocol overcomes the imposed limitation by protecting the addressed two-level system (TLS) with an adapted concatenated CDD approach. We use and adjust it such that high frequency sensing becomes feasible even for not phase-matched signals. As a result, the proposed scheme is generic and works for many atomic or solid state TLS, in which the energy gap matches the frequency of the signal under interrogation. A remarkable feature of our scheme is the fact that the signal to be probed also works partially as a decoupling drive and thus improves further the sensitivity of the sensor.
![Schematic representation of our setup: (a) The NV center probes an external signal while it is being manipulated by the control fields. (b) Schematic representation of the sequence applied in this work. (c) The protected TLS: The bare system, $H_0$, is subjected to strong environmental noise $\delta B$. Applying a strong drive, $\Omega_1$, opens a protected gap, now subjected mainly to drive fluctuations $\delta \Omega_1$. A second drive, $\Omega_2$, is then applied to protect the TLS, $H_I$, from these fluctuations, resulting in a TLS, $H_{II}$, on resonance with the signal, $g''=g/4$, with noise mainly from the second weak drive $\delta \Omega_2 \ll \delta \Omega_1$. []{data-label="fig:scheme-ddrive"}](fig1.pdf){width="\columnwidth"}
We demonstrate the performance of continuous dynamical decoupling by applying it to a nitrogen-vacancy (NV) center in diamond with natural abundance of $^{13}$C (cf. Fig. \[fig:scheme-ddrive\]a). Here, we utilize two of its ground sub-levels as the TLS. The states of the NV center can be read out and initialized by a 532nm laser, which reveals spin dependent fluorescence between the two levels [@jelezko_observation_2004; @jelezko_observation_2004-1; @balasubramanian_nanoscale_2008]. The system can be manipulated by driving it with microwave fields. We show that by using a concatenation of two drives, an improvement in coherence time of the sensor by more than one order of magnitude is achieved. Taking into account the effect of an external signal, $g$, on the sensor during a concatenation of two drive fields, we obtain an improvement in bandwidth for high frequency sensing by three orders of magnitude in comparison to the relaxometry approach. Moreover, we report on the measurement of a weak high frequency signal with strength $g$, which relates to a smallest detectable magnetic field amplitude of $\delta B_{\text{min}}\approx 4$nT.
The sensing scheme {#the-sensing-scheme .unnumbered}
==================
The basic idea of utilizing concatenated continuous driving to create a robust qubit is illustrated in Fig. \[fig:scheme-ddrive\]b. The concatenation of two phase-matched driving fields results in a robust qubit [@cai_robust_2012; @aharon_fully_2016]. In what follows we show that such a robust qubit can be utilized as a sensor for frequencies in the range of the qubit’s energy separation and hence, dynamical decoupling can be integrated into the sensing task.
By the concatenated driving, the qubit is prepared in a state that allows for strong coherent coupling to the high-frequency signal to be probed (corresponding to the last TLS in Fig. \[fig:scheme-ddrive\]c). In the total Hamiltonian, $H$, we consider the concatenation of two driving fields of strength (the Rabi frequency) $\Omega_1$ and $\Omega_2$, respectively. The Hamiltonians of the TLS, $H_0$, the protecting driving fields, $H_{\Omega_1}, H_{\Omega_2}$ and the signal, $H_s$, are given by $$\begin{aligned}
% \nonumber to remove numbering (before each equation)
H &=& H_0 + H_{\Omega_1} + H_{\Omega_2} + H_s = \frac{\omega_{0}}{2}\sigma_{z}+\Omega_{1}\sigma_{x}\cos\left(\omega_{0}t\right) \nonumber\\
&+& \Omega_{2}\sigma_{y}\cos\left(\omega_{0}t\right)\cos\left(\Omega_{1}t\right) + g \sigma_{x}\cos\left(\omega_{s}t+\varphi\right),
\label{eq:double-drive-hamiltonian}\end{aligned}$$ where $\omega_{0}$ is the energy gap of the bare states ($\hbar=1$), $\omega_{s}$ is the frequency of the signal, and $g$ is the signal strength which we want to determine. We tune the system, i.e., $\omega_{0}$, $\Omega_{1}$, and $\Omega_{2}$, such that $\omega_{s}=\omega_{0}+\Omega_1+\Omega_2/2$.
It is an important feature that phase matching between the signal and the control is not required, which means that the signal phase $\varphi$ can be unknown and moreover, it may vary between experimental runs. In addition, we make the assumption that $\omega_{0} \gg \Omega_1 \gg \Omega_2 \gg g$. Moving to the interaction picture (IP) with respect to $H_{0}=\frac{\omega_{0}}{2}\sigma_{z}$ and making the rotating-wave-approximation, we obtain $$\begin{aligned}
H_{I} &=& \frac{\Omega_1}{2}\sigma_{x}+ \frac{\Omega_2}{2}\sigma_{y}\cos\left(\Omega_{1}t\right)\nonumber\\ &+& \frac{g}{2}\left(\sigma_{+}e^{-i\left(\left(\Omega_1+\Omega_2/2\right)t+\varphi\right)}+\sigma_{-}e^{+i\left(\left(\Omega_1+\Omega_2/2\right)t+\varphi\right)}\right).\end{aligned}$$ This picture incorporates the effect of $\Omega_1$ onto a TLS and express the new system in eigenstates of $\sigma_x$, the $\ket{\pm}$ (dressed) states, which separates the contributions from $\Omega_2$ and $g$. For a large enough drive $\Omega_1$, the $\ket{\pm}$ eigenstates are decoupled (in first order) from magnetic noise, $\delta B \sigma_z$, because $\bra{\pm}\sigma_z\ket{\pm}=0$. However, power fluctuations $\delta \Omega_1$ of $\Omega_1$ limit the coherence time of the dressed states. The resulting IP is illustrated in the second TLS in Fig. \[fig:scheme-ddrive\]c. We continue by moving to a second IP with respect to $H_{01}=\frac{\Omega_{1}}{2}\sigma_{x}$, which leads to $$H_{II} =\frac{\Omega_2}{4}\sigma_{y}+ \frac{g}{4}\left(-i\sigma_{+}e^{-i\left(\frac{\Omega_2}{2} t+\varphi\right)}+i\sigma_{-}e^{+i\left(\frac{\Omega_2}{2} t+\varphi\right)}\right).
\label{eq:h2}$$ Once again, we incorporate $\Omega_2$ into the dressed states, so that solely the contribution of the signal $g$ becomes obvious, which is depicted in the last TLS of Fig. \[fig:scheme-ddrive\]c. The second drive, $\Omega_2$, which is larger than $\delta \Omega_1$, creates effectively doubly-dressed states (the $\sigma_y$ eigenstates). These doubly-dressed states are immune to power fluctuations of $\Omega_1$ and hence prolong the coherence time (see Supplementary Sec. II for more details). Moving to the third IP with respect to $H_{02}=\frac{\Omega_{2}}{4}\sigma_{y}$ results in $$H_{III}=\frac{g}{8}\left(\sigma_{+}e^{-i\varphi}+\sigma_{-}e^{+i\varphi}\right),$$ where we can clearly see that the signal $g$ induces rotations in the robust qubit subspace (either with $\sigma_+$ or $\sigma_-$). These rotations are obtained for any value of an arbitrary phase $\varphi$ and the bandwidth ($\propto 1/\text{T}_2$) is now limited by the coherence time, T$_2$, of the sensor. Hence, if a given TLS exhibits the possibility of manipulating it via drive fields $H_{\Omega_1}$ and $H_{\Omega_2}$, we can achieve a high frequency sensor in the range of $\omega_0$.
![Measurements of an external signal of strength $g$. (a) In a single drive approach with $\Omega_1/2\pi=3.002$MHz a signal $g'=g/2$ is recorded. (b) By the application of two drive fields with $\Omega_1/2\pi=3.363$MHz and $\Omega_2/2\pi=505$kHz, we record a signal $g''=g/4$ and increase the coherence time of the sensor by one order of magnitude with respect to the case of $g=0$.[]{data-label="fig:result-longrabi"}](fig2.pdf){width="\columnwidth"}
By this, we overcome the low frequency limit that is common to state-of-the-art pulsed dynamical decoupling sensing methods. In addition, we present an analog pulsed version of our scheme, where the pulsing rate is much lower than the frequency of the signal (see Supplementary Sec. IX). However it is not a direct measurement of the signal, but based on a signal demodulation approach. Compared to the pulsed schemes, continuous dynamical decoupling does not suffer from being susceptible to higher harmonics of the decoupling window appearing naturally from the periodic character of the pulsed sequence [@loretz_spurious_2015]. Eventually, less power per unit time is used in the continuous scheme leading to a smaller overall noise contribution from the drive.
Results {#results .unnumbered}
=======
![Comparison of the smallest measurable magnetic field change $\delta B_{\text{min}} = (2\pi \delta g_{\text{min}})/ \gamma_{\text{NV}} $ as a function of total measurement time. To show the total improvement, we obtain $\sigma(t)$ at $\tau = \text{T}_{2}^{\Omega_1}\approx 60$s in the single drive case and $\sigma(t)$ at $\tau=\text{T}_{2}^{\Omega_1, \Omega_2, g}\approx 1.43$ms in the double drive case. Note, that for both data traces a signal was always present, $g/2\pi=26.9$kHz and $g/2\pi=69.2$kHz in the single and in the double drive, respectively. But only in the double drive the coherence time prolonging effect of $g$ was included into the choice of $\tau$ for equation (i.e. the measurement was performed at $\tau=\text{T}_{2}^{\Omega_1, \Omega_2, g}$ instead at $\tau=\text{T}_{2}^{\Omega_1, \Omega_2}\approx 393$s). []{data-label="fig:result-sensitivity"}](fig3.pdf){width="\columnwidth"}
After determining the optimal drive parameters, $\Omega_1$ and $\Omega_2$, for the concatenated sensing sequence, and thereby maximize the coherence times, T$_{2}^{\Omega_1}$ and T$_{2}^{\Omega_1, \Omega_2}$, respectively, of the sensor (see Supplementary Sec. VI), we apply an external high frequency signal (according to $H_s$ in equation ) tuned to one of the four appearing energy gaps $\omega_s$ of the doubly-dressed states. In these energy gaps an effective population transfer can occur between the states of the robust TLS, evidenced by signal induced Rabi oscillations at a rate $g''=g'/2 = g/4$ in the double drive case (see Supplementary Sec. II).
The measurements take place in the laboratory frame, i.e. all three contributions $\Omega_1, \Omega_2$ and $g$ to the population dynamics of the TLS will be visible. In order to see solely the effect of $g$ on the TLS, we alter the modulation of the second drive in equation to $\cos(\Omega_1 t + \pi/2)$. This does not change the performance of the scheme, but only changes the axis of rotation to $\sigma_z$ for the second drive $\Omega_2$. Since the readout laser is effectively projecting the population in the $\sigma_z$ eigenbasis, we can make the Rabi rotations of $\Omega_2$ invisible to the readout. To remove the effect of $\Omega_1$ in the data, we can simply sample the measurement at multiple times of $\tau_{\Omega_1} =2 \pi/\Omega_1$, i.e. we measure at times $t=N \tau_{\Omega_1} $ ($N \in \mathbb{N}$). This procedure reveals directly $g''$($g'$) as the signal induces Rabi oscillations of the robust qubit under double (single) drive (Fig. \[fig:result-longrabi\]). Alternatively, we could have applied at the end of the drive a correction pulse in order to complete the full $\Omega_1$ and $\Omega_2$ rotations so that just the effect of $g''$ remains.
Without a signal, $g$, we achieve coherence times of T$_{2}^{\Omega_1} \approx60$s with a single drive ($\Omega_2 = 0$, compare Fig. S3 in Sec. VI.A in Supplementary) and T$_{2}^{\Omega_1, \Omega_2} \approx 393$ s with a double drive (compare Fig. S4 in Sec. VI.B in Supplementary). The results for long and slow Rabi oscillations induced by an external signal, $g$, under single and double drive ($\Omega_2/\Omega_1 \approx 0.15$) are shown in Fig. \[fig:result-longrabi\]. These illustrate a significant increase of the coherence time of the sensor by two orders of magnitude, from T$_{2}^{\Omega_1}\approx 60$ s to a lifetime limited coherence time of (T$_1/2 \approx$) T$_{2}^{\Omega_1,\Omega_2, g} \approx 1.43$ms. It should be noted that the signal itself can be considered as an additional drive (cf. equation \[eq:h2\]), correcting external errors $\delta \Omega$ of the previous drive and thereby prolonging the coherence time even further. Consequently, we can improve the bandwidth for high frequency sensing by almost three orders of magnitude from $\sim 900\,$kHz (for T$_2^* \approx 1.1$ s) to $\sim 700$ Hz (for a T$_{2}^{\Omega_1,\Omega_2, g} \approx 1.43$ms). Moreover, in the Supplementary Sec. III we discuss an improved version of our scheme which has the potential to push the coherence time of the sensor further towards the lifetime limit.
To benchmark the double drive scheme against a standard single drive approach, we determine the smallest magnetic field which can be sensed after an accumulation time $t$. The smallest measurable signal $S$ is eventually bounded by the smallest measurable magnetic field change $\delta B_{\text{min}}$, which is found to be $$\delta B_{\text{min}}(t,\tau) = \frac{\delta S}{\text{max} \left| \frac{\partial S}{\partial B} \right|} = \frac{1}{ \gamma_{\text{NV}}} \frac{\sigma(t)}{\alpha \tau C} \, .
\label{eq:min-b-field}$$ Here, $\gamma_{\text{NV}}/2 \pi=28.8\,$GHz T$^{-1}$ is the gyromagnetic ratio of the NV defect, $\sigma (t)$ is the standard deviation of the measured normalized fluorescence counts after time $t$, $\alpha$ accounts for a different phase accumulation rate depending on the decoupling scheme and $C$ is the contrast of the signal (see Supplementary Sec. V for detailed derivation). Since the photon counting is shot noise limited, we have $\sigma (t) = 1/\sqrt{N_{ph} \cdot N}$, with $N_{\text{ph}}$ being the number of photons measured in $\tau$ and $N=t/\tau$ is the number of sequence repetitions. With this, equation will transform in the commonly known form [@itano_quantum_1993; @budker_optical_2007] with some measurement dependent constants.
![Projected sensitivity of the single drive scheme (from Fig. S5 in the Supplementary) as a function of signal strength $g$, where the magenta dashed line indicates the sensitivity of the sensor if no signal is applied. The figure illustrates that an external signal has a non-linear effect on the sensitivity of the sensor, which has to be taken into account in the sensitivity estimation.[]{data-label="fig:projected-sensitivity"}](fig4.pdf){width="\columnwidth"}
We recorded $\sigma(t)$ as a function of time and use this to determine $\delta B_{\text{min}}$. The results of this measurement for both the single and double drive are summarized in Fig. \[fig:result-sensitivity\]. The sensitivity can be obtained by $\eta (\tau) = \delta B_{\text{min}}(t, \tau) \sqrt{t} $, which is optimal in the vicinity of the coherence time of the sensor, $\tau \approx \text{T}_{2}$. With our system, we achieve a sensitivity of $\eta_{\Omega_1, \Omega_{2}, g} \lesssim 1$ T Hz$^{-0.5}$ in the double drive case at $\sim 1.6$GHz, which should be compared to $\eta_{\Omega_1, g} \lesssim 20$ T Hz$^{-0.5}$ for a single drive approach. Both traces in Fig. \[fig:result-sensitivity\] were recorded while a signal $g$ was applied. Apart from the mere fact, that the number of driving fields are different, the specific choice for $\tau$ will also determine the magnitude of the smallest measurable magnetic field change $\delta B_{\text{min}}$. Obtaining the coherence time without a signal, $g$, (which are T$_{2}^{\Omega_1}$ and T$_{2}^{\Omega_1, \Omega_2}$), is a common practice in the field, but will not result in a correct choice of $\tau$ for the sensitivity measurement and also for equation , since the signal has an impact on the sensor’s sensitivity. However, if $\delta B_{\text{min}}$ shall be evaluated correctly, then the non-linearity, i.e. the coherence time prolonging effect of the signal, has to be taken into account. Otherwise an even worse $\delta B_{min}$ will be measured as it is exemplarily shown for the single drive case in Fig. \[fig:result-sensitivity\], where $\delta B_{min}$ was evaluated and measured under the naive assumption that the signal has no effect on the coherence time of the sensor (i.e. we measure at $\tau = \text{T}_2^{\Omega_1}$ and not at $\tau = \text{T}_2^{\Omega_1,g}$). This effect was included in the double drive case.
To examine the signal protection effect more in detail, the coherence time of the sensor is measured as a function of signal strength, $g$, in a single drive configuration (see also Supplementary Sec. VII). From these measurements we project the sensitivity associated with a specific signal strength (Fig. \[fig:projected-sensitivity\]), assuming $\sigma(t)$ is unchanged for the same repetition $N$. This is a reasonable assumption given that the only difference between measurements is the signal strength, $g$, and sequence length, $\tau$.
Eventually, this phenomenon, which seems to be an inherent part of this continuous scheme, can be used to further increase the performance of the sensor by fine tuning the controlled parameters (static bias field $B_{\text{bias}}$ and thereby changing $\omega_0$, $\Omega_1$ and $\Omega_2$) to match the signal frequency, $\omega_s$, and strength, $g$ (see Supplementary Sec. III for further discussions).
Conclusion {#conclusion .unnumbered}
==========
We have demonstrated for the first time that dynamical decoupling can be used in the context of sensing high frequency fields. In contrast to state-of-the-art pulsed dynamical decoupling protocols, we can show that continuous dynamical decoupling can be simultaneously integrated into the sensing task. By utilizing a NV center in diamond we have demonstrated by pure concatenation of two drives a coherence time of $\sim 393$s which constitutes an improvement of more than two orders of magnitude over T$_{2}^{*}$, and an increase of resolution from the MHz to a few kHz. The application of this method for wireless communication [@baylis_solving_2014] could have a transformative effect due to the high resolution of the protocol. Since the protocol is applicable to a variety of solid-state, molecular, and atomic systems, we believe that it has a great potential to have a significant impact on many fields and tasks that involve high frequency sensing (up to frequencies in the THz range). Eventually, this method could also be used to improve the coupling to quantum systems [@cai_long-lived_2012].
Acknowledgements {#acknowledgements .unnumbered}
----------------
The experiments presented here were supported by the Qudi Software Suite [@binder_qudi_2017]. A. S., A. H. and U. L. A. acknowledge funding from the Innovation Foundation Denmark through the project EXMAD and the Qubiz center, and the Danish Research Council via the Sapere Aude project (DIMS). T. U. and F.J. acknowledge the Volkswagenstiftung. A. R. acknowledges the support of the Israel Science Foundation (grant no. 1500/13), the support of the European commission, EU Project DIADEMS. This project has received funding from the European Union’s Horizon 2020 research and innovation program under grant agreement No 667192 Hyperdiamond and Research Cooperation Program and DIP program (FO 703/2-1).
[50]{}ifxundefined \[1\][ ifx[\#1]{} ]{}ifnum \[1\][ \#1firstoftwo secondoftwo ]{}ifx \[1\][ \#1firstoftwo secondoftwo ]{}““\#1””@noop \[0\][secondoftwo]{}sanitize@url \[0\][‘\
12‘\$12 ‘&12‘\#12‘12‘\_12‘%12]{}@startlink\[1\]@endlink\[0\]@bib@innerbibempty [****, ()](\doibase
10.1063/1.4936758) [****, ()](\doibase 10.1126/science.aaa4298) [****, ()](\doibase
10.1103/PhysRevApplied.6.064008) [****, ()](\doibase 10.1021/nl502988n) [****, ()](\doibase
10.1038/ncomms10211) [****, ()](\doibase 10.1016/j.pnmrs.2004.03.002) [****, ()](\doibase
10.1021/acs.nanolett.5b00679) [****, ()](\doibase 10.1103/PhysRev.80.580) [****, ()](\doibase 10.1103/PhysRev.94.630) [****, ()](\doibase 10.1063/1.1716296) [****, ()](\doibase 10.1103/PhysRevA.58.2733) [****, ()](\doibase
10.1038/nature07951) [****, ()](\doibase 10.1038/nature08470) [****, ()](\doibase
10.1126/science.1192739) [****, ()](\doibase 10.1103/PhysRevLett.105.200402) [****, ()](\doibase 10.1103/PhysRevB.83.081201) [****, ()](\doibase
10.1103/PhysRevB.85.155204) [****, ()](\doibase
10.1038/ncomms2771) [****, ()](\doibase 10.1038/nphys1075) [****, ()](\doibase
10.1038/nature10010) [****, ()](\doibase 10.1103/PhysRevLett.101.010403) [****, ()](\doibase 10.1103/PhysRevLett.95.180501) [****, ()](\doibase 10.1103/PhysRevLett.98.100504) [****, ()](\doibase 10.1103/PhysRevLett.106.240501) [****, ()](\doibase 10.1007/s11467-010-0113-8) [****, ()](\doibase
10.1103/PhysRevB.92.060301) [****, ()](\doibase 10.1103/PhysRevA.75.022329) [****, ()](\doibase 10.1103/PhysRevLett.107.150503) [****, ()](\doibase 10.1103/PhysRevA.85.040302) [****, ()](\doibase
10.1088/1367-2630/14/9/093030) [****, ()](\doibase 10.1103/PhysRevLett.109.070502) [****, ()](\doibase 10.1103/PhysRevLett.113.237601) [****, ()](\doibase
10.1103/PhysRevB.79.041302) [****, ()](\doibase 10.1103/PhysRevLett.104.040401) [****, ()](\doibase 10.1038/nnano.2016.178) [****, ()](\doibase
10.1088/1367-2630/14/11/113023) [****, ()](\doibase 10.1088/1367-2630/17/4/043008) [****, ()](\doibase 10.1088/2040-8986/aa5f62) [****, ()](\doibase
10.1103/PhysRevLett.116.240801) [****, ()](\doibase
10.1038/nature10319) [****, ()](\doibase 10.1103/PhysRevLett.111.230507) [****, ()](\doibase 10.1088/1367-2630/aa4fd3) [****, ()](\doibase
10.1103/PhysRevLett.92.076401) [****, ()](\doibase
10.1103/PhysRevLett.93.130501) [****, ()](\doibase
10.1038/nature07278) [****, ()](\doibase
10.1103/PhysRevX.5.021009) [****, ()](\doibase 10.1103/PhysRevA.47.3554) [****, ()](\doibase 10.1038/nphys566) [****, ()](\doibase
10.1109/MMM.2014.2321253) [****, ()](\doibase
10.1016/j.softx.2017.02.001)
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We report on an experimental and theoretical study of the high-frequency mixing properties of ion-irradiated [$\mbox{YBa}_2$$\mbox{Cu}_3$$\mbox{O}_7$[$\mbox{ } $]{}]{}Josephson junctions embedded in THz antennas. We investigated the influence of the local oscillator power and frequency on the device performances. The experimental data are compared with theoretical predictions of the general three-port model for mixers, in which the junction is described by the resistively shunted junction model. A good agreement is obtained for the conversion efficiency in different frequency ranges, spanning above and below the characteristic frequencies $f_c$ of the junctions.'
author:
- 'M. Malnou$^1$, C. Feuillet-Palma$^1$, C. Ulysse$^2$, G. Faini$^2$, P. Febvre$^3$, M. Sirena$^4$,L. Olanier$^1$, J. Lesueur$^1$, N. Bergeal$^1$'
title: Josephson mixers for terahertz detection
---
The THz region of the electromagnetic spectrum, which covers the range from 0.3 to 10 THz is a frontier area for research in many fields, including physics, astronomy, chemistry, material science and biology. However, so far, this range is hardly exploitable because of the limited number of suitable sources and detectors [@ferguson; @tonouchi]. Indeed, THz frequencies lie between the frequency range of electronics and photonics where the existing technologies cannot be simply extended. An important challenge for practical applications is heterodyne detection, that is needed to translate the THz frequency window of interest to lower frequencies, for which semiconductor electronics can process the signals. This detection technique has the advantage to combine high sensitivity with high frequency resolution. It involves detecting a signal at frequency $f_s$ by non-linear mixing with a continuous wave reference signal produced by a local oscillator (LO) at frequency $f_{\mathrm{LO}}$. The output signal has a low frequency component at the intermediate frequency (IF) $f_{\mathrm{IF}}=|f_{\mathrm{LO}}-f_s|$ which contains the information carried by the original signal. Metal-semiconductor Schottky diodes are often employed as mixing elements in heterodyne receivers, offering ease of use and a wide coverage of the THz frequency region [@crowe; @yasui]. Their main drawbacks are the limited sensitivity and the need for a high LO power. Superconducting hot electron bolometers (HEB) made of Niobium or Niobium Nitride (NbN) are able to operate at low noise with very good frequency coverage but require cooling to 4K [@gershenzon; @zmuidzinas; @zhang]. So far, the most sensitive frequency-mixing elements are the low temperature superconductor-insulator-superconductor (SIS ) Niobium tunnel junctions [@graauw; @mears]. However, these junctions are intrinsically limited in frequency by the gap energy of Nb ($\sim$800 GHz ) and operate only at low temperature (4 K). More recently, NbN SIS junctions with higher gap energy have been developed to extend the frequency range.\
![ Illustration of the fabrication process steps: (a) 70-nm thick [$\mbox{YBa}_2$$\mbox{Cu}_3$$\mbox{O}_7$[$\mbox{ } $]{}]{}film grown on sapphire covered by an in situ 250-nm gold layer; (b) spiral antenna in the CPW transmission line defined in an ma-N negative e-beam resist; (c) 500-eV Ar ion-beam-etching of the gold layer (d) gold antenna in the CPW transmission line on [$\mbox{YBa}_2$$\mbox{Cu}_3$$\mbox{O}_7$[$\mbox{ } $]{}]{}; (e) high-dose 70-keV oxygen ion irradiation to create insulating regions in exposed [$\mbox{YBa}_2$$\mbox{Cu}_3$$\mbox{O}_7$[$\mbox{ } $]{}]{}. A 2-$\mu$m wide channel in the center of the antenna is protected by a ma-N resist mask; (f) patterned superconducting and insulating [$\mbox{YBa}_2$$\mbox{Cu}_3$$\mbox{O}_7$[$\mbox{ } $]{}]{}regions; (g) low-dose 110 keV oxygen ion irradiation of the Josephson junction patterned as a 20-nm-wide slit in PMMA photoresist; and (h) device after resist cleaning.](figure1){width="8.8cm"}
An alternative to these devices consist in using high-temperature superconducting (HTS) receivers. In addition to the obvious advantage of a much higher operating temperature, their higher energy gap results in a cut-off frequency of several THz. Hence, it is important to develop HTS devices and related heterodyne mixer technology for applications in the THz range. However, SIS tunnel junction technology is not available with these materials and it is not possible to directly adapt the low-T$_c$ superconductor technology. In this context, the interest for non-tunnel Josephson junction mixers has been renewed because they can be fabricated by various methods with high-T$_c$ materials. Unlike SIS mixers, whose operation is based on the quasiparticle non-linearity near the gap energy, Josephson mixers use the non-linearity of the Cooper pair current. First realizations of mixers with high-T$_c$ superconductors which were mainly based on grain-boundary or ramp edge junctions [@chen; @scherbel; @harnack; @tarasov], produced promising results but the development was slowed down by the difficulty to build a junction technology sufficiently reliable to fabricate complex devices. In recent years, a new approach based on ion irradiation has been developed to fabricate Josephson junctions with high temperature superconductors. This method has been used to produce reproducible junctions [@kahlmann; @bergeal], SQUIDs [@bergealsq] and large-scale integrated Josephson circuits [@cybart1; @cybart]. Here, we present an analysis of the high-frequency mixing properties of Josephson junctions made by ion irradiation. The experimental data are compared with theoretical predictions form the general three-port model.\
**I. Fabrication of the Josephson mixer.**\
HTS thin films are structured at the nanometer scale by combining e-beam lithography with ion irradiation. This technique relies on the extreme sensitivity of HTS to defects, owing to the d-wave symmetry of their order parameter. Disorder induced in the material by irradiation reduces the superconducting transition temperature and increases the resistivity because of enhanced scattering. Beyond a critical defect density, a superconductor-to-insulator transition takes place, a phenomenon that can be used to selectively insulate some regions of a superconducting film [@bergealjap]. Figure 1 describes the different steps of the fabrication process. Starting from a commercial 70-nm-thick [$\mbox{YBa}_2$$\mbox{Cu}_3$$\mbox{O}_7$[$\mbox{ } $]{}]{}film ($T_c$ = 86 K) [@ceraco] grown on sapphire covered by an in-situ 250-nm gold layer (Fig.1a), a three steps fabrication process is performed. The spiral antenna embedded in a 50$\Omega$ co-planar waweguide (CPW) transmission line is first defined in the gold layer through a ma-N e-beam resist patterning followed by a 500-eV Ar$^+$ Ion Beam Etching (IBE) (Fig.1b,c,d). Then a 2-$\mu m$ wide channel located at the center of the antenna is patterned in a ma-N e-beam resist, followed by a 70-keV oxygen ion irradiation at a dose of 2$ \times$10$^{15}$ at/cm$^2$ (Fig.1e). This process ensures that the regions of the film which are not protected either by the resist or by the gold layer become deeply insulating (Fig. 1f). No HTS material is removed and the superconducting parts, including the antenna, the CPW line and the 2-$\mu m$ wide channel, remain embedded in the insulating film preventing degradation. Finally, the junction is defined at the center of the superconducting channel by irradiating through a 20-nm wide slit patterned in a PMMA resist with 100 keV oxygen ions (Fig.1g,h). A fluence of 3$ \times$10$^{13}$ at/cm$^2$ is used to lower the $T_c$ in the region underneath the slit. The parameters of the junctions such as the normal resistance $R_n$, the critical current $I_c$ and the operating temperature can be engineered simply by modifying the width of the slit, the fluence of irradiation and the ion energy.\
Josephson behaviour, in particular the Fraunhofer pattern of the critical current under magnetic field and Shapiro steps under microwave irradiation has been reported previously in this type of junctions [@lesueur]. One main advantage of this technique is that the process is by nature highly scalable, with no design constraint. It is therefore particularly suitable in creating THz devices that include Josephson junctions embedded in their circuitry [@malnou]. In this device, the self-complementary spiral wideband antenna \[80GHz-6THz\] has an impedance $Z=\frac{Z_0}{2}\sqrt{\frac{2}{\epsilon_r+1}}\approx$ 80 $\Omega$ where $Z_0\approx$ 377$\Omega$ is the vacuum impedance and $\epsilon_r\approx$ 10 is the dielectric constant of the sapphire substrate. The 50-$\Omega$ CPW transmission line is directly connected to the junction to readout the intermediate frequency signal.\
{width="8.5cm"}
**II. Experimental set-up**\
The experimental set-up is shown in Fig. 2. The back side of the sapphire substrate is placed in contact with a silicon hyper-hemispheric lens located at the focal point of a parabolic mirror exposed to external signals through the window of the cryostat. The junction is connected to contact pads for dc biasing and to a microwave transmission line. A cryogenic HEMT amplifier operating in the 4-8GHz band amplifies the output signal at the intermediate frequency before further amplification at room temperature. An isolator is placed in the chain to minimize the back-action of the amplifier on the Josephson mixer. The local oscillator is combined with the signal through a beam splitter. Mixing experiments were performed in five different frequency ranges centered on 20, 70, 140, 280 and 420 GHz. At 20 GHz, signals are provided by microwave generators whereas for the higher frequencies, signals are provided by Gunn diodes emitting at 70GHz coupled to a set of frequency doublers and triplers.\
**III. dc and ac response of the junction**\
The resistance of the junction as a function of temperature measured at very low current bias, reveals the existence of two characteristic temperatures in our device, namely $T_{c}$ and $T_J$ (fig. 3a). The highest transition at $T_{c}$ = 84 K refers to the superconducting transition of the non-irradiated regions of sample, which corresponds to the transition temperature of the unprocessed [$\mbox{YBa}_2$$\mbox{Cu}_3$$\mbox{O}_7$[$\mbox{ } $]{}]{}film [@bergeal]. The second transition at the lower temperature $T_{J}$ = 66 K corresponds to the occurrence of a Josephson coupling between the two electrodes, strong enough for the critical current to resist thermal fluctuations. Below $T_{J}$, the critical current $I_c$ grows quadratically when lowering temperature, as expected from Josephson coupling by the proximity effect [@degennes]. A third characteristic temperature $T'_c$ is also observed when the barrier itself becomes superconducting. Its existence is inherent to the irradiation fabrication technique which lowers the $T_c$ of the material in the region below the slit. To retrieve this temperature, we measured the R(T) curve while illuminating the junction with a sufficiently high-power RF signal to suppress the Josephson supercurrent (Fig. 3a). The temperature at which the resistance reaches zero defines $T'_c$. The Josephson regime therefore lies between $T'_c$=45K and $T_{J}$ = 66 K.\
{width="8.5cm"}
Junctions have non-hysteretic current-voltage characteristics with an upward curvature in the dissipation branch at low voltage and no sharp feature at the gap voltage (Fig. 3b). In this low capacitance regime, the electrical behavior of the junction is expected to be well described by the resistively shunted junction (RSJ) model [@mccumber], which considers a Josephson element in parallel with a resistance $R_n$. The voltage across the junction is given by the expressions : $$\begin{aligned}
V&=&R_n\big[I_{\mathrm{dc}}+\delta I_{n}-I_{c}sin\phi\big]\\
\label{RS}
V&=&\frac{\hbar}{2e}\frac{d\phi}{dt}
\label{RSJbis}\end{aligned}$$
where $I_{dc}$ is the bias dc current and $\phi$ the superconducting phase difference across the junction. Here $\delta I_{n}$ is an additive Gaussian white noise of variance $\sigma^2=\frac{\hbar\Gamma}{eI_cR_n\Delta t}$, where $\Gamma=2ek_BT/\hbar I_c$ is the ratio of the thermal energy to the Josephson energy and $\Delta t$ is the time step chosen for the numerical integration of equations (1) and (2). From the determination of $I_c$ and $R_n$, we extract the characteristic frequency $f_c = (2e/h)I_cR_n$ of the mixer. As seen in Fig. 3c, $f_c$ displays a dome as a function of the temperature with a maximum value $f_c^{\mathrm{opt}}$ of 85 GHz at 55 K. Note that $f_c$ is not a cut-off frequency and that mixing can be performed up to frequencies corresponding to several times the value of $f_c$ at the cost of a reduced conversion efficiency [@malnou]. However, for optimal operation, it is desirable to have $f_c$ larger than the frequencies of the incoming signals $f_{\mathrm{LO}}$ and $f_{s}$ as the resulting ac current would then interact mainly with the Josephson non-linear inductive element.\
![a, b and c) Current-voltage characteristics of the junction (open circles) measured at different temperatures 53K, 58K and 62K respectively. Dashed lines correspond to the curve under strong microwave radiation and red solid lines correspond to a fit with using the RSJ model (\[RSJ2\]) in which the non-linear resistance is introduced. d–f) Current voltage characteristics of the junction (open circles) measured at T=58K under LO radiation at 20 GHz, 70 GHz and 140 GHz. Curves are fitted using the RSJ model (\[RSJ2\]).](figure4){width="9cm"}
{width="9cm"}
In the Josephson regime, the dissipation branch at large bias reveals that the resistance increases with voltage. The origin of this non-linear behavior of the resistance stems from the non-uniform distribution of defects in the barrier resulting from the irradiation process [@katz]. In Figure 4 we show that the non linear resistance, in particular at low bias, can be retrieved by suppressing the Josephson supercurrent with high-power microwave radiation (dashed lines). The behaviour of the I(V) is then well described by the RSJ equations (1) and (2) provided we enter the non-linear normal resistance $R_n(I_{dc})$ in the model (Fig. 4).\
Current-voltage characteristics measured at T= 58K upon LO illumination are shown in Fig. 4(d)–(f) for three different frequencies, 20 GHz, 70 GHz and 140 GHz. Shapiro steps at the quantized voltage $V_n=n\frac{\hbar}{2e}f_{\mathrm{LO}}$ can be clearly observed [@shapiro]. To analyze these features, we added a LO current term into equation (1) $$\begin{aligned}
V=R_n(I_{\mathrm{dc}})\big[I_{\mathrm{dc}}+I_{\mathrm{LO}}\cos(2\pi f_{\mathrm{LO}}t)+\delta I_{n}-I_{c}sin\phi\big]
\label{RSJ2}\end{aligned}$$ A good agreement with the experimental data is obtained as can be seen in Fig. 4(d)–(f). The junction response to 20-GHz LO illumination has also been measured at different temperatures. Fig. 5 shows the differential resistance of the junction $\frac{dV}{dI}$ as a function of bias current and power radiation in the Josephson regime (Fig. 5(a) and (b)) and below (Fig. 5(c) and (d)). For strong LO power, several Shapiro steps can be seen as well as their modulation with LO power. In particular, the critical current (n=0) can be fully suppressed by the application of the correct amount of LO power. However, below $T'_c$ the modulation of the critical current is no longer complete (Fig. 5(d)), indicating that the dynamics of the junction deviates from a pure Josephson one. A crossover towards a flux flow regime is then observed although some features of the Josephson effect remain observable.\
{width="9cm"}
**IV. High frequency mixing**\
The junction is illuminated with a strong LO signal at frequency $f_{\mathrm{LO}}$ and a much weaker test signal at frequency $f_{\mathrm{s}}$. These conditions guarantee that the IF signal is produced by a first order mixing mechanism between the signal and the LO. Figure 6 shows the output power measured at the intermediate frequency $f_{\mathrm{IF}}=|f_{\mathrm{LO}}-f_{s}|$=6 GHz as a function of the dc voltage $V$ across the junction for the different ranges of frequency. At 20 GHz, 70GHz and 140 GHz, the power of the LO has been set to reduce the critical current to approximately half its value, as it corresponds to an optimal operation point for mixer performances (see part VI). The IF output power $P_{\mathrm{IF}}$ displays strong modulations whose period is given by the quantized voltage $\Delta V=\frac{\hbar}{2e} f_{\mathrm{LO}}$ between two Shapiro steps. Two mixing regimes can be identified. For $f_{\mathrm{LO}}$=20GHz (Fig. 6(a)), $P_{\mathrm{IF}}$ is maximum at voltages corresponding to the exact center between two Shapiro steps (see arrow). We will show in part V that such a behavior is obtained when $f_{\mathrm{LO}}<f_c^{\mathrm{max}}$. For $f_{\mathrm{LO}}$=140GHz (Fig. 6(c)), $P_{\mathrm{IF}}$ has two maxima close to the Shapiro steps (see arrows), separated by a dip. This corresponds to the condition $f_{\mathrm{LO}}>f_c^{\mathrm{max}}$ . In the intermediate situation where $f_{\mathrm{LO}}\approx f_c^{\mathrm{max}}$, $P_{\mathrm{IF}}$ is approximately flat at the center of the steps (Fig. 6(b)). Measurements performed at higher frequencies, $f_{\mathrm{LO}}$=280 and 410 GHz (Fig. 6(d) and 6(e)) indicate that the junction responds in the lower part of the THz range. However, in these cases the power of the LO source was not sufficient to reach optimal bias conditions. Mixing at frequencies higher than 410 GHz was not investigated in this study.\
The output power $P_{\mathrm{IF}}$ at the intermediate frequency was measured as a function of the signal power for the three main ranges of frequency. After calibration of the IF output line, the conversion efficiency $\eta=\frac{P_{\mathrm{IF}}}{P_{\mathrm{s}}}$ was calculated and plotted as a function of the signal power $P_{\mathrm{s}}$. The mixer displays a linear dynamical range of constant conversion efficiency of more than 55 dB at 20 GHz and 30 dB at 140 GHz (Fig. 6(f)). For strong signal power, the amplitude of the modulation of the IF signal decreases and the mixer saturates. In this situation, the signal power can no longer be considered to be small compared to the LO power and second-order mixing processes take place.\
{width="8cm"}
**V. Three-port model\
**
Following the pioneering work of Taur [@taur], we used the three-port model to calculate the performance of the mixer. It describes the linear response for a small signal by solving the non-linear response of the mixer under the LO illumination. When the junction is driven by a strong LO signal, the relation between small currents $I$ and voltages $V$ is linear in the frequency domain [@likharev] : $$\tilde{V}(f)=\sum_{k=-\infty}^{^{+\infty}}Z_{k}(f-kf_\mathrm{LO})\tilde{I}(f-kf_\mathrm{LO})\label{eq:vtild}$$
Let us now consider the case when a signal of frequency $f_{\mathrm{s}}$ close to $f_{\mathrm{LO}}$ is shined onto the junction. At first order in (\[eq:vtild\]), there are only three frequencies of interest each containing a term at the intermediate frequency $f_{\mathrm{IF}}=|f_{s}-f_{\mathrm{LO}}|$ : the lower side band frequency $f_{\mathrm{LSB}}=f_{\mathrm{LO}}-f_{\mathrm{IF}}$, the intermediate frequency itself $f_{\mathrm{IF}}$ and the upper side band one $f_{\mathrm{USB}}=f_{\mathrm{IF}}+f_{\mathrm{LO}}$. Limiting ourselves to these three frequencies, (\[eq:vtild\]) can be written as a matrix equation: $$\begin{aligned}
\left(\begin{array}{c}
\tilde{V}_{\mathrm{USB}}\\
\tilde{V}_{\mathrm{IF}}\\
\tilde{V}_{\mathrm{LSB}}^{*}
\end{array}\right)=\left[\begin{array}{ccc}
Z_{uu} & Z_{u0} & Z_{ul}\\
Z_{0u} & Z_{00} & Z_{0l}\\
Z_{lu} & Z_{l0} & Z_{ll}
\end{array}\right]\left(\begin{array}{c}
\tilde{I}_{\mathrm{USB}}\\
\tilde{I}_{\mathrm{IF}}\\
\tilde{I}^{*}_{\mathrm{LSB}}
\end{array}\right)\label{eq:3ports}\end{aligned}$$
where $u$, $l$, and $0$ stand for USB, LSB, and IF respectively. $\tilde{\tilde{Z}}$ is the impedance matrix of the mixer which characterizes in particular its ability to down-convert at $f_{IF}$ the information at $f_{\mathrm{USB}}$ ( or $f_{\mathrm{LSB}}$. Each of its elements $Z_{ij}$ is simply the ratio of the voltage $\tilde{V}_j $ at frequency $f_j$ to the current $\tilde{I}_i$ injected at frequency $f_i$. For a low-Q mixer, the symmetric properties of the 3-port matrix imply $Z_{uu}=Z_{ll}^{*}$, $Z_{lu}=Z_{ul}^{*}$, $Z_{u0}=Z_{l0}^{*}$ and $Z_{0u}=Z_{0l}$. A general mixer theory provides the following expression for the matrix elements [@torrey; @taur; @schoelkopf]
$$\begin{aligned}
Z_{uu}&=&\frac{1}{2}\Big[\frac{\partial\tilde{V}(f_{\mathrm{LO}})}{\partial I_{\mathrm{LO}}}+\frac{\tilde{V}(f_{\mathrm{LO}})}{I_{\mathrm{LO}}}\Big]\,(\mathrm{RF}\,\mathrm{impedance})\\
Z_{u0}&=&\frac{\partial\tilde{V}(f_{\mathrm{LO}})}{\partial I_{dc}}\,(\mathrm{up}\,\mathrm{conversion})\\
Z_{ul}&=&\frac{1}{2}\Big[\frac{\partial\tilde{V}(f_{\mathrm{LO}})}{\partial I_{\mathrm{LO}}}-\frac{\tilde{V}(f_{\mathrm{LO}})}{I_{\mathrm{LO}}}\Big]\,(\mathrm{image}\,\mathrm{conversion})\\
Z_{0u}&=&\frac{1}{2}\frac{\partial V_{dc}}{\partial I_{\mathrm{LO}}}\,(\mathrm{down}\,\mathrm{conversion})\\
Z_{00}&=&\frac{\partial V_{dc}}{\partial I_{dc}}\,(\mathrm{dc}\,\mathrm{impedance})\end{aligned}$$
To determine the conversion efficiency of the mixer, we introduce in Fig. 7 the external part of the circuit which is described by the diagonal impedance matrix $\tilde{\tilde{Z}}_{\mathrm{ext}}$ whose elements $Z_u$, $Z_l$ and $Z_0$ are connected to the mixer inputs. Here $Z_u$ and $Z_l$ represent the impedance of the spiral antenna (80 $\Omega$) at USB and LSB frequencies respectively and are taken to be identical. $Z_0$ is the 50-$\Omega$ impedance of the IF microwave readout line. Assuming that the signal $V_{\mathrm{s}}$ incoming on the antenna is at the USB frequency, the equation for the circuit shown in Fig. 7 is [@taur] $$\begin{aligned}
\left(\begin{array}{c}
\tilde{V}_{\mathrm{USB}}\\
\tilde{V}_{\mathrm{IF}}\\
\tilde{V}_{\mathrm{LSB}}^{*}
\end{array}\right)+\left[\begin{array}{ccc}
Z_{u} & 0 & 0\\
0 & Z_{0} & 0\\
0 & 0 & Z_{l}
\end{array}\right]\left(\begin{array}{c}
\tilde{I}_{\mathrm{USB}}\\
\tilde{I}_{\mathrm{IF}}\\
\tilde{I}^{*}_{\mathrm{LSB}}
\end{array}\right)=\left(\begin{array}{c}
V_{\mathrm{s}}\\
0\\
0
\end{array}\right)
\label{ext}\end{aligned}$$
We therefore obtain a relation between the currents at different frequencies and the input signal
$$\begin{aligned}
\left(\begin{array}{c}
\tilde{I}_{\mathrm{USB}}\\
\tilde{I}_{\mathrm{IF}}\\
\tilde{I}^{*}_{\mathrm{LSB}}
\end{array}\right)=\tilde{\tilde{Y}}
\left(\begin{array}{c}
V_{\mathrm{s}}\\
0\\
0
\end{array}\right)
\label{tot}\end{aligned}$$
where $\tilde{\tilde{Y}}=(\tilde{\tilde{Z}}+\tilde{\tilde{Z}}_{\mathrm{ext}})^{-1}$ is the admittance matrix.
We define the conversion efficiency as the ratio of the IF power $P_{IF}=\frac{1}{2}Z_0|\tilde{I}_{\mathrm{IF}}|^2$ dissipated in the impedance $Z_0$ to the available signal power $P_{s}=\frac{\left|V_{\mathrm{s}}\right|^{2}}{8Z_{u}}$ on the antenna impedance $Z_u$, $$\begin{aligned}
\eta_{c}=\frac{P_{IF}}{P_{\mathrm{s}}}=4Z_0Z_{u}\left|Y_{u0}\right|^{2}
\label{conversion}
\end{aligned}$$
where $Y_{u}^{0}$ is a non-diagonal matrix element of the admitance matrix $\tilde{\tilde{Y}}$ with the same convention as in (\[eq:3ports\])
In the limit—well satisfied experimentally— where $|Z_{ul}|,|Z_{u0}|\ll|Z_{u}|$ and $|Z_{0u}|\ll|Z_{0}|$, the conversion efficiency takes the simple form
$$\begin{aligned}
\eta_{c}=4\frac{Z_{u}}{|Z_{u}+Z_{uu}|^2}\times\frac{Z_{0}}{|Z_{0}+Z_{00}|^2}\times Z_{0u}^2
\label{conv_simple}
\end{aligned}$$
The first two factors correspond to the matching impedance conditions at the USB and IF frequencies. The conversion is optimal when the antenna impedance $Z_{u}$ matches the RF impedance of the junction $Z_{uu}$ and when the readout line impedance $Z_{0}$ matches the dc impedance of the junction. The last factor $Z_{0u}^2$ represents the ability of the junction to down convert the USB signal to the intermediate frequency. Within this approximation, the performance of the device mainly depends on three elements of the $\tilde{\tilde{Z}}$ matrix : (i) the RF impedance at the USB (or LSB) frequency $Z_{uu}$ (=$Z_{ll}^{*}$), (ii) the dc impedance $Z_{00}$, and (iii) the down-conversion impedance $Z_{0u}$ (=$Z_{0l}^{*}$).\
{width="9cm"}
To derive the impedance matrix $\tilde{\tilde{Z}}$, the RSJ equation (\[RSJ2\]), including the non-linear resistance $R_n(I_{dc})$ is first solved numerically in the time domain. The biasing condition in terms of LO power is chosen to reproduce the experimental conditions of Fig. 6. The matrix elements $Z_{ij}$ are then calculated according to expressions (6), (9) and (10) and plotted in Fig. 8 as a function of normalized voltage for the different LO frequencies. The impedances $Z_{00}$ and $Z_{0u}$ reproduce the shape of the output power $P_{\mathrm{IF}}$ of Fig. 6. For $f_{LO}$=20GHz the mixer should be dc biased halfway between the Shapiro steps whereas for $f_{\mathrm{LO}}$=140 GHz it should be biased close to the steps. The impedance $Z_{0u}$ and therefore the ability of the junction to down-convert decreases significantly when the LO frequency is increased. Figure 8(d) shows that the theoretical calculations of the conversion efficiency obtained from (\[conv\_simple\]) are in good agreement with experimental data. A crossover from the first regime of mixing $f_\mathrm{LO}<f_c$ to the second regime $f_\mathrm{LO}>f_c$ is observed. At T=53K, the noise parameter $\Gamma$=0.07 is much lower than 1 which guarantees that the Josephson non-linearity is not smeared out by the noise.\
The conversion efficiency takes a maximum value of 0.1% at 20GHz and decreases to 0.01% at 140GHz. An improvement of the mixer performances requires optimizing the three factors of expression (\[conv\_simple\]). In particular, the impedance mismatch resulting from the low values of $Z_{uu}$ and $Z_{00} $ compared with $Z_u$ and $Z_0$ respectively, leads to a significant deterioration of $\eta_c$. In practice, the matrix elements are determined by two parameters, the normal resistance $R_n$ of the junction and its characteristic frequency $f_c$ (i.e. the $I_cR_n$ product), through the RSJ equation. As $R_n$ is the only impedance entering this equation, all the matrix elements are directly proportional to it. This resistance needs to be increased significantly to improve the impedance matching. This can be done by decreasing both the width and the thickness of the junction and by increasing the ion irradiation fluence. Finally, impedance matching elements both between the antenna and the junction and between the readout line and the junction could also be added at a cost of reduced bandwidth.\
The value of $f_c$ influences all the matrix elements, but affects mainly the down-conversion one $Z_{0u}$. Assuming for simplification that $Z_{uu}\sim Z_{00}\sim R_n$, the amount of LO current necessary to reduce the critical current to zero is $\Delta I_{LO}\approx\frac{hf_\mathrm{LO}}{2eR_n}$ [@grimes]. We thus obtain the dependence of the $Z_{0u}$ element with the ratio $f_c$/$f_\mathrm{LO}$ $$Z_{0u}=\frac{1}{2}\frac{\partial V_{dc}}{\partial I_{\mathrm{LO}}}\sim\frac{1}{2}\frac{\Delta (R_nI_c)}{\Delta I_\mathrm{LO}}\sim \frac{1}{2}R_n\frac{f_c}{f_\mathrm{LO}}$$ From this expression, we see that it is desirable to fabricate junctions with high $f_c$ values, i.e. high $I_cR_n$ product. The junction presented in this study has a characteristic frequency which is lower than the ones usually reported in grain-boundary junctions[@chen; @scherbel; @harnack; @tarasov]. However, several developments can be made to optimize the $I_cR_n$ product [@sirena; @sirenaJAP2007] in our junctions. In particular, a higher irradiation fluence combined with an annealing of the sample should lead to a significant improvement [@sirenaAPL2007; @sirenaJAP2009]. For $f_\mathrm{{LO}}>>f_{c}$ the signal and the LO ac current interact weakly with the inductive Josephson element. As a result, a large part of the IF power is generated by mixing on the non-linear resistance. As can be seen in Fig. 6(d) and (e), this produces a continuous background on top of which, Josephson mixing can still be distinguished.\
{width="9cm"}
**VI. Influence of the LO power\
**
In a practical heterodyne receiver application, the LO power necessary to optimally bias the mixer is a critical parameter and must satisfy two important requirements : (i) it has to be as low as possible to minimize the power consumption and to be easily driven by available sources in the frequency range of interest and (ii) its variations and fluctuations must not modify significantly the performance of the mixer. For a Josephson mixer, the dependence of the conversion efficiency with the LO power is mainly determined by the characteristic frequency. Additionally, it is generally expected that the conversion should be greatest for a LO power corresponding to a suppression by approximately 50$\%$ of the critical current. However, a careful analysis of this point has never been done, and the mixer should in principle operate for a range of LO power. Figure 8(a) and (b) show the behavior of the output power $P_\mathrm{{IF}}$ as a function of voltage across the junction for different values of the LO power received by the junction, for $f_\mathrm{{LO}}$= 20 GHz and 140 GHz. The conversion efficiency taken at $2e/h\times V=f_\mathrm{{LO}}/2$ is plotted as a function of $P_\mathrm{{LO}}$ (Fig. 8(c)). For $f_\mathrm{{LO}}<f_c$, $\eta$ is constant on more than one decade and decreases at strong LO power. $P_\mathrm{{LO}}$ as low as 20 pW at $f_{LO}$= 20 GHz and 100 pW at $f_{LO}$= 70 GHz are sufficient to drive optimally the mixer whereas at 140 GHz, 10 nW of power are required. It is clear that the conversion efficiency does not depend critically on the LO power as long as $f_\mathrm{{LO}}< f_c$; otherwise, as can been seen at 140 GHz, $\eta$ is optimal for a given LO power which corresponds approximately to a suppression by 50$\%$ of the critical current (Fig. 8(d)).
In conclusion, we have demonstrated the mixing operation of ion-irradiated [$\mbox{YBa}_2$$\mbox{Cu}_3$$\mbox{O}_7$[$\mbox{ } $]{}]{}Josephson junctions up to 420 GHz at temperature higher than 50K. The performances of the mixer were studied as a function of LO frequency and LO power. For LO frequencies lower or close to the characteristic frequency of the junction, a conversion efficiency in the range of 0.02-0.1% was obtained for a LO power lower than 1 nW. A detailed analysis of the mixer within the framework of the general three-port model and the RSJ model was proposed. A good agreement between experimental data and numerical simulation was obtained.\
The authors thank T. Ditchi, E. Geron, S. Hole and J. Lucas for useful discussions, and Y. Legall for his help with the ion irradiations. This work was supported by the ANR ASTRID program, the Emergence program Contract of Ville de Paris and by the Région Ile-de-France in the framework of CNano IdF and Sesame program. CNano IdF is the nanoscience competence center of the Paris Region, supported by CNRS, CEA, MESR and Région Ile-de-France.\
M. Tonouchi, Nature Photonics **1**, 97-105 (2007). B. Ferguson, X.-C. Zhang, Nature Mater. **1**, 26Ð33 (2002). T. Yasui, Nishimura, A., Suzuki, T., Nakayama, K. Okajima, S. Detection system operating at up to 7 THz using quasioptics and Schottky barrier diodes. Rev. Sci. Instr. 77, 066102 (2006). T. W. Crowe et et al. Proc. IEEE **80**, 1827-1841 (1992). E. M. Gershenzon, G. N. Goltsman, I. G. Gogidze, Y. P. Gousev, A. I. ElantÕev, B. S. Karasik, and A. D. Semenov, Sov. Phys. Superconductivity. **3**, 1582 (1990). J. Zmuidzinas and P. L. Richards Proc. IEEE **92**, 1597 (2004). W. Zhang, Appl. Phys. Lett. **96**, 111113 (2010). T. De Graauw, F. P. Helmich, T.G. Phillips, J. Stutzki, E. Caux, N. D. Whyborn, P. Dieleman, P. R. Roelfsema, H.Aarts, R. Assendorp et al., A&A **518**, L6 (2010). C. A. Mears, Q. Hu, P. L. Richards, A. H. Worsham, D. E. Prober and A. V. Räisänen, Appl. Phys. Lett. **57**, 2487-2489 (1990). J. Chen, H. Myoren, K. Nakajima, T. Yamashita, and P. H. Wu, Appl. Phys. Lett. **71**, 707 (1997). M. Tarasov, E. Stepantsov, D. Golubev, Z. Ivanov, T. Claeson, O. Harnack, M. Darula, S. Beuven, H. Kohlstedt, IEEE Trans. Appl. Supercond. **9**, 3761-3764, (1999). J. Scherbel, M. Darula, O. Harnack, M. Siegel, IEEE Trans. Appl. Supercon. **12**, 1828 (2002). O. Harnack, M. Darula, S. Beuven, and H. Kohlstedt. Appl. Phys. Lett., **76**, 1764 (2000). F. Kahlmann, A. Engelhardt, J. Schubert, W. Zander, C. Buchal, J. Hollkott, Appl. Phys. Lett. **73**, 2354-2356 (1998). N. Bergeal, X. Grison, J. Lesueur, G. Faini, M. Aprili, J.P. Contour, App. Phys. Lett. **87**, 102502 (2005). N. Bergeal, J. Lesueur, G. Faini, M. Aprili, and J.-P. Contour, Appl. Phys. Lett. **89**, 112515 (2006). S.A. Cybart, S. M. Wu, S. M. Anton, I. Siddiqi, J. Clarke, R. C. Dynes, App. Phys. Lett **89**, 112515 (2006). S. A. Cybart, S. M. Anton, S. M. Wu, J. Clarke, R. C. Dynes, Nano Lett. **9**, 3581 (2009). N. Bergeal, J. Lesueur, M. Sirena, G. Faini, M. Aprili, J-P. Contour, B. Leridon, J. App. Phys **102**, 083903 (2007). Ceraco ceramic coating GmbH. J. Lesueur et al. IEEE Trans. Appl. Supercond. **17**, 963–966 (2007). M. Malnou, A. Luo, T. Wolf, Y. Wang, C. Feuillet-Palma, C. Ulysse, G. Faini, P. Febvre, M. Sirena, J. Lesueur and N. Bergeal Appl. Phys. Lett. **101**, 233505 (2012). P. G. de Gennes and E. Guyon, Phys. Lett. **3**, 168 (1963). D. E. McCumber, J. App. Phys. **39**, 3113 (1968). A. S. Katz, S. I. Woods and R. C. Dynes, J. Appl. Phys. **87**, 2978–2983 (2000). S. Shapiro, Phys. Rev. Lett., **11**, 80 (1963). C. C. Grimes and S. Shapiro, Phys. Rev. **169**, 397 (1968). Y. Taur, IEEE T. Electron. Dev., **27**, 1921 (1980). K. K. Likharev and V. K. Semenov : Jetp Lett. **15**, 442 (1972) H. C. Torrey and C. A. Whitmer. Crystal Rectifiers, McGraw-Hill Book Company, Inc. (1948) R.J. Schoelkopf. PhD thesis, California Institute of Technology, Pasadena, (1995). M. Sirena, X. Fabreges, N. Bergeal, J. Lesueur, G. Faini, R. Bernard, J. Briatico, App. Phys. Lett **91** 262508 (2007). M. Sirena, S. Matzen, N. Bergeal, J. Lesueur, G. Faini, R. Bernard, J. Briatico, D. G. Crete, J. P. Contour, J. App. Phys. **101**, 123925 (2007). M. Sirena, S. Matzen, N. Bergeal, J. Lesueur, G. Faini, R. Bernard, J. Briatico, D. G. Crete, J. P. Contour, App. Phys. Lett. **91**, 142506 (2007). M. Sirena, S. Matzen, N. Bergeal, J. Lesueur, G. Faini, R. Bernard, J. Briatico, D. G. Crete, J. App. Phys. **105**, 023910 (2009).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Machine learning models are known to perpetuate the biases present in the data, but oftentimes these biases aren’t known until after the models are deployed. We present the Visual Bias Extraction (ViBE) Tool that assists in the investigation of a visual dataset, surfacing potential dataset biases along three dimensions: (1) object-based, (2) gender-based, and (3) geography-based. Object-based biases relate to things like size, context, or diversity of object representation in the dataset; gender-based metrics aim to reveal the stereotypical portrayal of people of different genders within the dataset, with future iterations of our tool extending the analysis to additional axes of identity; geography-based analysis considers the representation of different geographic locations. Our tool is designed to shed light on the dataset along these three axes, allowing both dataset creators and users to gain a better understanding of what exactly is portrayed in their dataset. The responsibility then lies with the tool user to determine which of the revealed biases may be problematic, taking into account the cultural and historical context, as this is difficult to determine automatically. Nevertheless, the tool also provides actionable insights that may be helpful for mitigating the revealed concerns. Overall, our work allows for the machine learning bias problem to be addressed early in the pipeline at the dataset stage. ViBE is available at <https://github.com/princetonvisualai/vibe-tool>.'
author:
- Angelina Wang
- Arvind Narayanan
- Olga Russakovsky
title: 'ViBE: A Tool for Measuring and Mitigating Bias in Image Datasets'
---
Introduction
============
Related Work
============
ViBE Tool Overview
==================
Object-Based Analysis
=====================
Gender-Based Analysis
=====================
Geography-Based Analysis
========================
Conclusions
===========
Acknowledgments
===============
Appendices {#appendices .unnumbered}
==========
[10]{} \[1\][`#1`]{} \[1\][https://doi.org/\#1]{}
Amazon rekognition <https://aws.amazon.com/rekognition/>
A [Y]{}ahoo [F]{}lickr grand challenge on tag and caption prediction (2016), <https://multimediacommons.wordpress.com/tag-caption-prediction-challenge/>
United [N]{}ations statistics division - methodology (2019), <https://unstats.un.org/unsd/methodology/m49/>
Alwassel, H., Heilbron, F.C., Escorcia, V., Ghanem, B.: Diagnosing error in temporal action detectors. European Conference on Computer Vision (ECCV) (2018)
Bellamy, R.K.E., Dey, K., Hend, M., Hoffman, S.C., Houde, S., Kannan, K., Lohia, P., Martino, J., Mehta, S., Mojsilovic, A., Nagar, S., Ramamurthy, K.N., Richards, J., Saha, D., Sattigeri, P., Singh, M., Varshney, K.R., Zhang, Y.: [AI]{} [F]{}airness 360: An extensible toolkit for detecting, understanding, and mitigating unwanted algorithmic bias. arXiv:1810.01943 (2018)
Berg, A.C., Berg, T.L., III, H.D., Dodge, J., Goyal, A., Han, X., Mensch, A., Mitchell, M., Sood, A., Stratos, K., Yamaguchi, K.: Understanding and predicting importance in images. Conference on Computer Vision and Pattern Recognition (CVPR) (2012)
Bolukbasi, T., Chang, K.W., Zou, J., Saligrama, V., Kalai, A.: Man is to computer programmer as woman is to homemaker? debiasing word embedding. Advances in Neural Information Processing Systems (NeurIPS) (2016)
Brown, C.: Archives and recordkeeping: Theory into practices. Facet Publishing (2014)
Buda, M., Maki, A., Mazurowski, M.A.: A systematic study of the class imbalance problem in convolutional neural networks. arXiv:1710.05381 (2017)
Buolamwini, J., Gebru, T.: Gender shades: Intersectional accuracy disparities in commercial gender classification. ACM Conference on Fairness, Accountability, Transparency (FAccT) (2018)
Burns, K., Hendricks, L.A., Saenko, K., Darrell, T., Rohrbach, A.: Women also snowboard: Overcoming bias in captioning models. European Conference on Computer Vision (ECCV) (2018)
Caliskan, A., Bryson, J.J., Narayanan, A.: Semantics derived automatically from language corpora contain humanlike biases. Science **356**(6334), 183–186 (2017)
Choi, M.J., Torralba, A., Willsky, A.S.: Context models and out-of-context objects. Pattern Recognition Letters p. 853–862 (2012)
Chouldechova, A.: Fair prediction with disparate impact: A study of bias in recidivism prediction instruments. Big Data (2017)
Deng, J., Dong, W., Socher, R., Li, L.J., Li, K., Fei-Fei, L.: [ImageNet: A Large-Scale Hierarchical Image Database]{}. In: CVPR09 (2009)
DeVries, T., Misra, I., Wang, C., van der Maaten, L.: Does object recognition work for everyone? Conference on Computer Vision and Pattern Recognition workshops (CVPRW) (2019)
Dwork, C., Hardt, M., Pitassi, T., Reingold, O., Zemel, R.: Fairness through awareness. Proceedings of the 3rd Innovations in Theoretical Computer Science Conference (2012)
Dwork, C., Immorlica, N., Kalai, A.T., Leiserson, M.: Decoupled classifiers for fair and efficient machine learning. arXiv:1707.06613 (2017)
Everingham, M., Gool, L.V., Williams, C.K.I., Winn, J., Zisserman, A.: The pascal visual object classes (voc) challenge. International Journal of Computer Vision (IJCV) (2010)
Fei-Fei, L., Fergus, R., Perona, P.: Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. IEEE CVPR Workshop of Generative Model Based Vision (2004)
Gajane, P., Pechenizkiy, M.: On formalizing fairness in prediction with machine learning. arXiv:1710.03184 (2017)
Galleguillos, C., Rabinovich, A., Belongie, S.: Object categorization using co-occurrence, location and appearance. Conference on Computer Vision and Pattern Recognition (CVPR) (2008)
Gebru, T., Morgenstern, J., Vecchione, B., Vaughan, J.W., Wallach, H., III, H.D., Crawford, K.: Datasheets for datasets. ACM Conference on Fairness, Accountability, Transparency (FAccT) (2018)
Hardt, M., Price, E., Srebro, N.: Equality of opportunity in supervised learning. Advances in Neural Information Processing Systems (NeurIPS) (2016)
He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning for image recognition. European Conference on Computer Vision (ECCV) (2016)
Hoiem, D., Chodpathumwan, Y., Dai, Q.: Diagnosing error in object detectors. European Conference on Computer Vision (ECCV) (2012)
Holland, S., Hosny, A., Newman, S., Joseph, J., Chmielinski, K.: The dataset nutrition label: A framework to drive higher data quality standards. arXiv:1805.03677 (2018)
Hua, J., Xiong, Z., Lowey, J., Suh, E., Dougherty, E.R.: Optimal number of features as a function of sample size for various classification rules. Bioinformatics **21**, 1509–1515 (2005)
Idelbayev, Y.: (2019), <https://github.com/akamaster/pytorch_resnet_cifar10>
Jain, A.K., Waller, W.: On the optimal number of features in the classification of multivariate gaussian data. Pattern Recognition **10**, 365–374 (1978)
Jo, E.S., Gebru, T.: Lessons from archives: Strategies for collecting sociocultural data in machine learning. ACM Conference on Fairness, Accountability, Transparency (FAccT) (2020)
Joulin, A., Grave, E., Bojanowski, P., Douze, M., J[é]{}gou, H., Mikolov, T.: Fasttext.zip: Compressing text classification models. arXiv preprint arXiv:1612.03651 (2016)
Joulin, A., Grave, E., Bojanowski, P., Mikolov, T.: Bag of tricks for efficient text classification. arXiv preprint arXiv:1607.01759 (2016)
Kay, M., Matuszek, C., Munson, S.A.: Unequal representation and gender stereotypes in image search results for occupations. Human Factors in Computing Systems pp. 3819–3828 (2015)
Khosla, A., Zhou, T., Malisiewicz, T., Efros, A.A., Torralba, A.: Undoing the damage of dataset bias. European Conference on Computer Vision (ECCV) (2012)
Kilbertus, N., Rojas-Carulla, M., Parascandolo, G., Hardt, M., Janzing, D., Schölkopf, B.: Avoiding discrimination through causal reasoning. Advances in Neural Information Processing Systems (NeurIPS) (2017)
Kleinberg, J., Mullainathan, S., Raghavan, M.: Inherent trade-offs in the fair determination of risk scores. Proceedings of Innovations in Theoretical Computer Science (ITCS) (2017)
Krasin, I., Duerig, T., Alldrin, N., Ferrari, V., Abu-El-Haija, S., Kuznetsova, A., Rom, H., Uijlings, J., Popov, S., Veit, A., Belongie, S., Gomes, V., Gupta, A., Sun, C., Chechik, G., Cai, D., Feng, Z., Narayanan, D., Murphy, K.: Openimages: A public dataset for large-scale multi-label and multi-class image classification. Dataset available from https://github.com/openimages (2017)
Krishna, R., Zhu, Y., Groth, O., Johnson, J., Hata, K., Kravitz, J., Chen, S., Kalanditis, Y., Li, L.J., Shamma, D.A., Bernstein, M., Fei-Fei, L.: Visual genome: Connecting language and vision using crowdsourced dense image annotations (2016)
Krizhevsky, A.: Learning multiple layers of features from tiny images. Technical Report (2009)
Krizhevsky, A., Sutskever, I., Hinton, G.E.: Imagenet classification with deep convolutional neural networks. Advances in Neural Information Processing Systems (NeurIPS) pp. 1097–1105 (2012)
Lin, T.Y., Maire, M., Belongie, S., Bourdev, L., Girshick, R., Hays, J., Perona, P., Ramanan, D., Zitnick, C.L., Dollar, P.: Microsoft [COCO]{}: Common objects in context. European Conference on Computer Vision (ECCV) (2014)
Liu, X.Y., Wu, J., Zhou, Z.H.: Exploratory undersampling for class-imbalance learning. IEEE Transactions on Systems, Man, and Cybernetics (2009)
Mehrabi, N., Morstatter, F., Saxena, N., Lerman, K., Galstyan, A.: A survey on bias and fairness in machine learning. arXiv:1908.09635 (2019)
Moulton, J.: The myth of the neutral ’man’. Sexist Language: A Modern Philosophical Analysis pp. 100–116 (1981)
Oksuz, K., Cam, B.C., Kalkan, S., Akbas, E.: [Imbalance Problems in Object Detection: A Review]{}. arXiv e-prints p. arXiv:1909.00169 (Aug 2019)
Oliva, A., Torralba, A.: The role of context in object recognition. Trends in Cognitive Sciences (2007)
Ouyang, W., Wang, X., Zhang, C., Yang, X.: Factors in finetuning deep model for object detection with long-tail distribution. Conference on Computer Vision and Pattern Recognition (CVPR) (2016)
Pleiss, G., Raghavan, M., Wu, F., Kleinberg, J., Weinberger, K.Q.: On fairness and calibration. Advances in Neural Information Processing Systems (NeurIPS) (2017)
Rosenfeld, A., Zemel, R., Tsotsos, J.K.: The elephant in the room. arXiv:1808.03305 (2018)
Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M., Berg, A.C., Fei-Fei, L.: [ImageNet Large Scale Visual Recognition Challenge]{}. International Journal of Computer Vision (IJCV) **115**(3), 211–252 (2015).
Salakhutdinov, R., Torralba, A., Tenenbaum, J.: Learning to share visual appearance for multiclass object detection. Conference on Computer Vision and Pattern Recognition (CVPR) (2011)
Shankar, S., Halpern, Y., Breck, E., Atwood, J., Wilson, J., Sculley, D.: No classification without representation: Assessing geodiversity issues in open datasets for the developing world. NeurIPS workshop: Machine Learning for the Developing World (2017)
Sigurdsson, G.A., Russakovsky, O., Gupta, A.: What actions are needed for understanding human actions in videos? International Conference on Computer Vision (ICCV) (2017)
Swinger, N., De-Arteaga, M., IV, N.H., Leiserson, M., Kalai, A.: What are the biases in my word embedding? Proceedings of the AAAI/ACM Conference on Artificial Intelligence, Ethics, and Society (AIES) (2019)
Thomee, B., Shamma, D.A., Friedland, G., Elizalde, B., Ni, K., Poland, D., Borth, D., Li, L.J.: Yfcc100m: The new data in multimedia research. Communications of the ACM (2016)
Tommasi, T., Patricia, N., Caputo, B., Tuytelaars, T.: A deeper look at dataset bias. German Conference on Pattern Recognition (2015)
Torralba, A., Efros, A.A.: Unbiased look at dataset bias. Conference on Computer Vision and Pattern Recognition (CVPR) (2011)
Torralba, A., Fergus, R., Freeman, W.T.: 80 million tiny images: a large dataset for non-parametric object and scene recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence **30**(11), 1958–1970 (2008)
Wang, T., Zhao, J., Yatskar, M., Chang, K.W., Ordonez, V.: Balanced datasets are not enough: estimating and mitigating gender bias in deep image representations. International Conference on Computer Vision (ICCV) (2019)
Wang, Z., Qinami, K., Karakozis, Y., Genova, K., Nair, P., Hata, K., Russakovsky, O.: Towards fairness in visual recognition: Effective strategies for bias mitigation. Conference on Computer Vision and Pattern Recognition (CVPR) (2020)
Wilson, B., Hoffman, J., Morgenstern, J.: Predictive inequity in object detection. arXiv:1902.11097 (2019)
Xiao, J., Hays, J., Ehinger, K.A., Oliva, A., Torralba, A.: Sun database: Large-scale scene recognition from abbey to zoo. Conference on Computer Vision and Pattern Recognition (CVPR) (2010)
Yang, J., Price, B., Cohen, S., Yang, M.H.: Context driven scene parsing with attention to rare classes. Conference on Computer Vision and Pattern Recognition (CVPR) (2014)
Yang, K., Qinami, K., Fei-Fei, L., Deng, J., Russakovsky, O.: Towards fairer datasets: Filtering and balancing the distribution of the people subtree in the imagenet hierarchy. ACM Conference on Fairness, Accountability, Transparency (FAccT) (2020)
Yang, K., Russakovsky, O., Deng, J.: Spatialsense: An adversarially crowdsourced benchmark for spatial relation recognition. International Conference on Computer Vision (ICCV) (2019)
Yao, Y., Zhang, J., Shen, F., Hua, X., Xu, J., Tang, Z.: Exploiting web images for dataset construction: A domain robust approach. IEEE Transactions on Multimedia pp. 1771–1784 (2017)
Zhang, B.H., Lemoine, B., Mitchell, M.: Mitigating unwanted biases with adversarial learning. Proceedings of the 2018 AAAI/ACM Conference on AI, Ethics, and Society (2018)
Zhao, J., Wang, T., Yatskar, M., Ordonez, V., Chang, K.W.: Men also like shopping: Reducing gender bias amplification using corpus-level constraints. Proceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP) (2017)
Zhou, B., Lapedriza, A., Khosla, A., Oliva, A., Torralba, A.: Places: A 10 million image database for scene recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence (2017)
Zhu, X., Anguelov, D., Ramanan, D.: Capturing long-tail distributions of object subcategories. Conference on Computer Vision and Pattern Recognition (CVPR) (2014)
| {
"pile_set_name": "ArXiv"
} |
---
address: 'Department of Mathematics, University of California, Riverside, CA 92521'
author:
- Mario Maican
title: '[On Vector Bundles of Finite Order]{}'
---
Introduction {#introduction .unnumbered}
============
Bundles of finite order were first introduced in [@griffiths] and were studied systematically in [@griffiths-cornalba]. Griffiths’ “working hypothesis” is that on an affine variety (smooth, over $\C$ and satisfying some extra conditions, cf. §1) every holomorphic vector bundle has a finite order structure. This he calls the “Oka principle with growth conditions” by analogy with the Oka-Grauert principle which states that on a Stein manifold every topological vector bundle has a unique holomorphic structure.
Griffiths and Cornalba prove the Oka principle for line bundles only. See (4.4) for the precise statement. For bundles of rank greater than 1 their theory is somewhat insufficient. To understand why we review here the four definitions they gave for what should mean that a holomorphic vector bundle $E$ on $X$ has finite order:
1. $E$ has a Hermitian metric whose holomorphic bisectional curvature has polynomial growth, see (4.1);
2. there is a holomorphic map $f$ of finite order from $X$ to some Grassmannian such that $E\isom f^* \mathcal{U}$, where $\mathcal{U}$ is the universal bundle on the Grassmannian;
3. $E$ has Schubert cycles of finite order;
4. $E$ has transition matrices of finite order relative to a suitable covering of $X$ with punctured polycylinders.
We refer to section 2 for the notions of holomorphic map of finite order, resp. analytic subset of finite order. The polycylinders from (IV) are as in §1. In the case of a line bundle $L$ the above conditions take a simpler form:
1. $L$ has a Hermitian metric whose first Chern form has polynomial growth;
2. there is a holomorphic map of finite order $f:X\lra \P^N$ such that $L\isom f^* \O_{\P^N}(1)$;
3. $L=[Z]$ with $Z\subset X$ an analytic divisor of finite order;
4. $L$ has transition functions of finite order relative to a suitable covering of $X$.
As shown in [@griffiths-cornalba] the four conditions from above are equivalent in the case of a line bundle. The implication “(I)$\implies$(II)” is true for any rank and follows from the vanishing theorem (4.6).
The implication “(II)$\implies$(III)” fails in general because of the so-called “unsolvability of the Bezôut problem in codimension greater than 1”. In a form tailored to our purposes, the Bezôut problem asks whether intersection of sets of finite order is of finite order, too. It has negative answer because of the counter-example from [@cornalba-shiffman]. Shiffman and Cornalba give two analytic sets of order zero in $\C^2$ whose intersection has infinite order.
The implication “(III)$\implies$(IV)” cannot be carried out if the rank is greater than 1 because of technical reasons: when $k\ge 2$ one does not know whether an analytic set of codimension $k$ in $\C^n$ can be given as the zero-set of $k$ functions of finite order. The case $k=1$ has positive answer, cf. [@skoda]. In the case $k>1$ Skoda showed that one can find $n+1$ defining functions of finite order, cf. [@skoda_2], but this is of little help.
In view of these difficulties we propose the following approach: one should give the conditions for finite order growth not in terms of $E$ but in terms of the hyperplane bundle $\O_{\P(E^*)}(1)$. This approach is inspired from classical complex geometry: one knows that a holomorphic vector bundle $E$ over a compact complex manifold is ample (in the sense that the zero-section of its dual can be collapsed to a point, or in the sense that Cartan’s Theorem A or Theorem B hold) if and only if $\O_{\P(E^*)}(1)$ is ample. Denoting $\L= \O_{\P(E)}(1)$ we propose the following definition for what should mean that the bundle $E^*$ have finite order:
1. $E$ has a Finsler metric of finite order, cf. (5.5).
One is forced to think about Finsler metrics because giving a Finsler metric on $E$ is, roughly speaking, equivalent to giving a Hermitian metric on $\L$. For Finsler metrics one has the notion of holomorphic bisectional curvature, cf. (3.21), and horizontal holomorphic bisectional curvature, cf. (3.22). We define finite order Finsler metrics by imposing estimates on the curvature in a very similar manner to the Hermitian case.
Once we are given a Finsler metric on $E$ there is a natural way of doing Nevanlinna theory on $\P(E)$ which we explain in section 7. Thus we can formulate a definition analogous to (II):
1. there is a holomorphic map of finite order to some projective space $f:\P(E)\lra \P^N$ such that $f^* \O_{\P^N}(1)\isom \L$.
The crucial step in the proof of “(I)$\implies$(II)” is a vanishing theorem for the sheaf of sections of finite order of $E$. At (5.12) we prove the corresponding statement in a Finsler context:
\
[**Theorem 1:**]{} *Let $X$ be a special affine variety and $E$ a holomorphic vector bundle on $X$ equipped with a Finsler metric of finite order. Then, for any $q\ge 1$, the $q^{\text{th}}$ cohomology of $\O_{\fo}(E^*)$ vanishes.*
\
As a consequence we get at (7.10) the implication “(I’)$\implies$(II’)”:
\
[**Theorem 2:**]{} *Let $X$ be a special affine variety of dimension $n$. Let $E$ be a holomorphic vector bundle of rank $r+1$ on $X$ equipped with a Finsler metric of finite order. Then, for any integer $N > 2(n+r)$, there is a holomorphic immersion $f:\P(E) \lra \P^N$ of finite order satisfying $f^* \O_{\P^N}(1) \isom \L$.*\
In §1 we introduce the spaces on which we are working. They come equipped with complete Kähler metrics of finite volume and bounded Ricci curvature. In §2 we present just the amount of Nevanlinna theory that we need. In §3 we define Finsler metrics and, following [@cao-wong], we compute the curvature of $\L$. In §4 we recall the main results of Griffiths and Cornalba on Hermitian metrics of finite order. We begin to make the transition to Finsler metrics by giving an equivalent definition for sections of finite order, cf. (4.12). §5 and §7 contain our two main results as stated above. They should be understood as a very small step in attempting to elucidate the Oka principle with growth conditions. In §6 we apply the classical theory of embeddings of Stein manifolds to show that the map from theorem 2 can be made to be an immersion. Notice that we do not claim that the image of $f$ in $\P^N$ is closed and, in fact, it is not.\
\
Preliminaries
=============
We begin by introducing the spaces on which we will be working. Let $\overline{X}$ be a complex projective manifold of dimension $n$. Let $D_1,\ldots, D_\n$ be effective smooth ample divisors on $\Xb$. We put $D= D_1+\ldots +D_\n$. We assume that $D$ has simple normal crossings. This means that around each point $x\in \Xb$ there is a coordinate chart $(z_1,\ldots, z_n)$ in which $D =\{ z_1 \cdots
z_k =0\}$ for some $0\le k\le n$. We put $X= \Xb \setminus D$ and call $X$ a *special affine variety*. Note that, by Hironaka’s resolution of singularities, a smooth affine variety always admits a compactification such that the divisor at infinity have simple normal crossings. Thus the word “special” refers to our requirement that each $D_i$ be ample.
Let $\D = \{ z\in \C,\ |z|\le 1\}$ be the unit disc and $\D^* =
\{ z\in \C,\ 0< |z| \le 1 \}$ be the punctured unit disc. A *k-fold punctured polycylinder* is of the form $P^* = (\D^*)^k
\times \D^{n-k},\ \ 0\le k \le n$. The compact manifold $\Xb$ can be covered with finitely many polycylinders of the form $P= \D^n$. They can be chosen in such a manner that the intersections $P^* =P\cap X$ be k-fold punctured polycylinders. If $k=0$ then we have a polycylinder that is entirely included in $X$. If $k>0$ we call $P^*$ a *neighbourhood at infinity*.
On the punctured unit disc we have the Poincaré metric $$\begin{aligned}
\o_{\D^*}= \frac{\i}{\p}\cdot \frac{dz \wedge d \bar{z}}{|z|^2
(\text{log} \frac{|z|^2}{c})^2} = - dd^c \text{log}(\text{log}
\frac{|z|^2}{c})^2 ,\ \ c>1,\end{aligned}$$ which has constant negative Gauss curvature, is complete and has finite volume. By the Poincaré metric $\o_{P^*}$ on a punctured polycylinder we mean the product of Poincaré metrics on the punctured components and Euclidean metrics on the non-punctured components. For a positive (1,1)-form $\f$ on $X$ we write $$\begin{aligned}
\tag{1.1}
\f \sim \o_{P^*}\end{aligned}$$ if on each polycylinder at infinity $\f$ is equivalent to the Poincaré metric.
Using the assumption that $D_i$ are ample Griffiths and Cornalba construct an exhaustive function $\t$ on $X$ with the following properties:
1. $\t$ is strictly plurisubharmonic;
2. the Levi-form $\f = dd^c \t$ induces a complete metric on $X$;
3. the Ricci curvature of this metric is bounded: $$\begin{aligned}
\tag{1.2}
|\text{Ric}(\f)| = O(\f);\end{aligned}$$
4. $\f$ has finite volume.
The contruction of $\t$ goes as follows: Consider the line bundles $[D_i]$ on $\Xb$ associated to $D_i$. They are ample, so on each of them we can choose a Hermitian metric with positive Chern form. We choose global sections $\s_i$ of $[D_i]$ whose zero-sets coincide with $D_i$ and which have norm less than 1 at every point. We put $\s= \s_1 \tensor \ldots \tensor \s_\n$, which is a section of $[D]= [D_1]\tensor \ldots \tensor [D_\n]$, and define $$\begin{aligned}
\tag{1.3}
\t=c\ \text{log} \frac{1}{|\s |^2} - \
\text{log} \{ (\text{log}|\s_1 |^2)^2 \cdots
(\text{log}|\s_\n |^2)^2 \}.\end{aligned}$$ Notice that $dd^c$log$\frac{1}{|\s |^2} = c_1 ([D])$ is a positive (1,1)-form on $\Xb$, while\
$dd^c$ log(log$ |\s_i |^2)^2$ is the Poincaré metric in the vicinity of $D_i$. Thus, choosing $c$ large enough, we can make sure that the Levi-form $\f =dd^c \t$ be positive, i.e. (i) holds. In addition $\f \sim \o_{P^*}$ which tells us that (ii) and (iv) hold. The remaining property (iii) was proven in [@griffiths-cornalba] and amounts to the fact mentioned above that $\o_{\D^*}$ has bounded curvature. We shall also make use of the exhaustive function $\r = e^{\t/2}$.
From now on $X$ will be a special affine variety. We fix a compactification $\Xb$ and an exhaustive function $\t$ as above. We assume that $X$ is embedded in some $\C^m$ and $\Xb$ is the closure of $X$ in $\P^m$. Such an embedding exists because the line bundle $[D]$ associated to the divisor at infinity is ample. In addition, we may assume that $[D]$ is the restriction of $\O_{\P^m}(1)$ to $X$. We choose affine coordinates $z=(z_1,\ldots, z_m)$ on $\C^m$ and homogeneous coordinates $(z_0;\ldots; z_m)$ on $\P^m$. As global section of $[D]$ in (1.3) we can take $\s = z_0$. We have $|\s |^2 = (1+||z||^2)^{-1}$ and $\r \sim
(1+||z||^2)^{c/2}$ outside a compact subset of $X$. Hence $$\begin{aligned}
\tag{1.4}
\r \sim ||z||^c\end{aligned}$$ outside a compact subset of $X$.
Next we introduce a semipositive (1,1)-form $\psi$ on $X$ which is suitable for doing Nevanlinna theory. We fix a projection $\pp : X\lra \C^n $ onto an n-dimensional subspace of $\C^m$ such that $\pp$ is a branched covering. We put $$\begin{aligned}
\t'= \text{log} ||\pp (x)||^2,\ \ \r'=e^{\t'/2},\ \ \psi = dd^c \t'.\end{aligned}$$ We have $\psi \ge 0,\ \psi^{n-1}\neq 0$ and $\psi^n =0$. Thus $\psi$ is not positive definite. We notice that outside a compact subset we have $$\begin{aligned}
\tag{1.5}
\t \sim \t'.\end{aligned}$$
We finish this section with several estimates from [@maican] that we will need later: $$\begin{aligned}
\tag{1.6}
\psi \le \r^c \f,\end{aligned}$$ $$\begin{aligned}
\tag{1.7}
d\t' \wedge d^c \t' \le \r^c \f,\end{aligned}$$ $$\begin{aligned}
\tag{1.8}
d\t \wedge d^c \t \le \r^c \f\end{aligned}$$ for some positive constant $c$.\
\
Nevanlinna Theory on Special Affine Varieties
=============================================
In this section we will recall several standard notions from Nevanlinna theory: the characteristic function measuring the growth of a holomorphic map from a special affine variety $X$ to projective space, the counting function measuring the growth of an analytic subset of $X$. We will then formulate basic relationships between these functions. As general reference we point out the elegantly written [@shabat].
Relative to the exhaustive function $\r' = e^{\t'/2}$ introduced in §1 we put $$\begin{aligned}
X[r]= \{ x\in X,\ \r'(x) \le r\},\quad X<r>=\{x\in X,\ \r'(x)= r\}.\end{aligned}$$ By Sard’s theorem the sets $X<r>$ are smooth for all $r$ outside a set of measure zero. In the sequel, each time we integrate over the set $X<r>$, it will tacitly be assumed that the latter is smooth.
Given a holomorphic map $f:X\lra \P^N$ we define its *characteristic function* $$\begin{aligned}
T_f (r,s) = \int_s^r \frac{dt}{t} \int_{X[t]} f^* \o \wedge \psi^{n-1}\end{aligned}$$ and its *higher characteristic functions* $$\begin{aligned}
T_f^{(k)} (r,s) = \int_s^r \frac{dt}{t} \int_{X[t]} f^* \o^k \wedge \psi^{n-k},\end{aligned}$$ where $r>s>0$ are real numbers, $1\le k \le n$ and $\o$ is the Fubini-Study form on $\P^N$. Given an analytic subset $Z\subset X$ of pure dimension $k$ we define its *counting function* $$\begin{aligned}
N_Z (r,s) = \int_s^r \frac{dt}{t} \int_{Z[t]} \psi^k,\end{aligned}$$ where $Z[t]=Z\cap X[t]$. In this definition and hereafter $Z$ does not have to be reduced, in other words its components may have multiplicities. Note that if the image of $Z$ under the projection $\pp :X\lra \C^n$ does not contain the origin then, the quantity $N_Z(r)=
N_Z (r,0)$ is well-defined. We call it “counting function” because in the case $k=0$, i.e. when $Z$ is a sequence of points, $N_Z (r)$ equals the logarithmic average of the number of points of $Z$ inside the ball $X[r]$. Given a global section $\s$ of $\O_{\P^N}(1)$ we define the *proximity function* of $f$ to the zero-set of $\s$ by $$\begin{aligned}
m_\s (r) =\int_{X<r>} \text{log} \frac{1}{|\s \circ f|} d^c \t' \wedge
\psi^{n-1}.\end{aligned}$$ Here $|\cdot |$ is the canonical norm of $\O_{\P^N}(1)$ given at a point $[v]$ by $|\s |_{[v]} =\frac{|<\s, v>|}{|v|}$. Notice that $d^c \t' \wedge
\psi^{n-1}$ is a volume form on $X<r>$. We call $m_\s (r)$ “proximity function” because it takes large values whenever the image of $X<r>$ under $f$ is close to the zero-set of $\s$. By choosing $\s$ to have norm less than 1 at all points we can arrange that $m_\s (r)$ be non-negative.\
[**(2.1) First Main Theorem:**]{} *Let $f:X\lra \P^N$ be a holomorphic map. Let $\s$ be a global section of $\O_{\P^N}(1)$ and $Z= f^* \{ \s =0\}$. Assume that the image of $f$ is not contained in $\{ \s = 0 \}$. Then, for $r>s>0$, we have* $$\begin{aligned}
N_Z(r,s)+m_\s(r)-m_\s(s)= T_f(r,s).\end{aligned}$$
[**(2.2) Corollary**]{} (Nevanlinna’s Inequality): *Fix $s>0$. Then, under the hypotheses of the previous theorem, we have for $r>s$* $$\begin{aligned}
N_Z(r,s)\le T_f(r,s) +O(1).\end{aligned}$$ *Here $O(1)$ is a constant that may depend on $s$.*
\
Thus the growth of the preimage of a hyperplane is bounded by the growth of $f$. The corresponding statement for planes of codimension greater than 1 is false: Cornalba and Shiffman give in [@cornalba-shiffman] an example of a holomorphic map of order zero from $\C^2$ to $\C^2$ for which the preimage of the origin has infinite order. This falls into the circle of ideas known as “Bezôut problem”. More about the Bezôut problem we said in the introduction. We now state a converse to (2.2) in an “average” sense. To explain it we introduce the Grassmannian $G(N,k)$ of planes of codimension $k$ in $\P^N$. There is a unique measure $\m$ on $G(N,k)$ which is invariant under the action of the unitary group $U(n+1)$ and which is so normalized that the measure of the total space be 1.
\
[**(2.3) Crofton Formula:**]{} *Let $f:X\lra \P^N$ be a holomorphic map which is non-degenerate in the sense that the preimage of a plane $P$ of codimension $k$ in $\P^N$ is an analytic subset of codimension $k$ in $X$. Then, for $r>s>0$, we have* $$\begin{aligned}
T_f^{(k)} (r,s)= \int_{P\in G(N,k)} N_{f^* P} (r,s) d\m (P).\end{aligned}$$
\
[**(2.4) Definition:**]{} Let $f:X\lra \P^N$ be a holomorphic map. We say that $f$ *has finite order* if there is $\l \ge 0$ such that $$\begin{aligned}
T_f (r,s)= O(r^\l).\end{aligned}$$ Likewise, we say that an analytic subset $Z\subset X$ of pure dimension $k$ *has finite order* if there is $\l \ge 0$ such that $$\begin{aligned}
N_Z (r,s)= O(r^\l).\end{aligned}$$ By an abuse of language we will say that $\l$ *is the order of* $f$ or of $Z$ if the above estimates hold. Technically, we would have to say that “$Z$ has order at most $\l$” but we want to avoid obstinate repetitions.
\
[**(2.5) Remark:**]{} Assume that $f$ has finite order and is linearly non-degenerate. Then (2.2) tells us that for any hyperplane $H\subset
\P^N$ the preimage $f^* H$ has finite order. Conversely, assume that the preimages $f^* H$ have finite order in a uniform fashion, i.e. there are $r_o, \k, \l \ge 0$ such that for $r\ge r_o$ and all $H$ we have $N_{f^* H} (r,s)\le \k r^\l$. Then, by (2.3), also $f$ has finite order.\
The order of growth of an analytic subset $Z \subset X$ depends on the embedding $X\subset \C^m$ and the projection $\pp :X\lra \C^n$ that we fixed in §1. However, the notion of $Z$ having finite order is intrinsic. We see this as follows: assume that $Z$ has pure dimension $k$ and consider the counting function of $Z$ computed using $\f$ instead of $\psi$: $$\begin{aligned}
\hat{N}_Z (r,s) = \int_s^r \frac{dt}{t} \int_{Z_{\r \le t}} \f^k.\end{aligned}$$ Using (1.5), (1.7) and the technique from [@griffiths-carlson] we can show that $N_Z (r,s) = O(r^\l)$ if and only if there is $\m$ depending only on $\l$ such that $\hat{N}_Z (r,s)= O(r^\m)$. See [@maican] for the details. Now, if $\t_1$ and $\t_2$ are constructed starting from two different embeddings of $X$ then it is an easy matter to see that $\t_1 \sim \t_2$ and $dd^c \t_1 \sim dd^c \t_2$. Thus $\hat{N}_1$ and $\hat{N}_2$ have polynomial growth at the same time.
By (2.3) the same considerations apply to holomorphic maps $f:X\lra \P^N$. The order of growth of $f$ is not intrinsic however, the notion of $f$ having finite order does not depend on any of the choices made.
Finally, we would like to mention the sheaf of holomorphic functions of finite order. First note that a holomorphic function $f:X\lra \C$ can be regarded as a map to $\P^1$ by sending $x$ to $(1;f(x))$. Its characteristic function takes the form $$\begin{aligned}
T_f (r,s)= \int_s^r \frac{dt}{t} \int_{X[t]} dd^c \text{log}(1+|f|^2)
\wedge \psi^{n-1}.\end{aligned}$$ Let us introduce one last growth function, the *maximum modulus function* $$\begin{aligned}
M_f (r)= \text{log max}\{ |f(x)|,\ \r(x)=r\}.\end{aligned}$$ Using the fact that log$(1+|f|^2)$ is plurisubharmonic one can show that $T_f (r,s)$ has roughly the same growth as $M_f (r)$. This point of view allows us to localize the notion of finite order function:
\
[**(2.6) Definition:**]{} We define the *sheaf $\O_\l$ of germs of holomorphic functions of order $\l$ on $X$* as follows: $\O_\l$ is a sheaf on $\Xb$; for each open set $U\subset \Xb$ the space of sections $\O_\l (U)$ consists of those holomorphic functions on $U\cap X$ with the property that around each point at infinity $x\in (\Xb \setminus
X)\cap U$ there is a neighbourhood $W\subset U$ and a constant $\k\ge 0$ such that the estimate $|f(w)|\le$ exp$(\k r^\l)$ with $w\in W,\
r=\r (w)$ holds.
Similarly we define the *sheaf $\O_{\fo}$ of germs of holomorphic functions of finite order on $X$* by the requirement that the estimate $|f(w)|\le$ exp$(\k r^\l)$ hold with $\k, \l$ depending on the point at infinity $x$.
It is easily seen that $\O_\l$ and $\O_{\fo}$ are sheaves on $\Xb$. In fact they are $\O_{\Xb}$-modules. Their restrictions to $X$ coincide with the sheaf $\O_X$ of germs of holomorphic functions on $X$. By the compactness of the divisor at infinity it is also clear that the spaces of global sections $\O_\l (\Xb)$ and $\O_{\fo} (\Xb)$ coincide with the space of holomorphic functions of order $\l$ on $X$, respectively the space of holomorphic functions of finite order on $X$.
These sheaves were studied in [@griffiths-cornalba] and [@wong-mulflur]. Griffiths and Cornalba showed that $\O_\l$ and $\O_{\fo}$ are acyclic, i.e. their cohomology groups, apart from $H^0$, vanish. Wong et al. proved that $\O_{\fo}$ is flat over $\O_{\Xb}$ and as a corollary obtained the acyclicity of $\O_{\fo}$.\
Finsler metrics and the geometry of $\P(E)$
===========================================
Let $X$ be a special affine variety of dimension $n$. Let $E$ be a holomorphic vector bundle of rank $r+1$ on $X$. We begin by recalling some facts about Hermitian metrics on $E$. We will then introduce Finsler metrics and see how the notion of holomorphic bisectional curvature generalizes from the Hermitian case to the Finsler case. Most of our calculations are taken from [@cao-wong]. We also refer to [@abate].
We will denote by $\O_X, \A_X, \A_X^k, \A_X^{p,q}$ the sheaves of holomorphic functions, of smooth $\C$-valued functions, of smooth $\C$-valued k-forms and of (p,q)-forms on $X$. We will denote by $\O_X(E), \A_X(E), \A_X^k(E), \A_X^{p,q}(E)$ the corresponding sheaves of $E$-valued functions or forms.
Let $U\subset X$ be an open coordinate set with coordinates $(z^1,\ldots,z^n)$. We assume that $E$ is trivial over $U$ and we choose a holomorphic frame $\{ e_0,\ldots,e_r \}$ for $E$ over $U$. Relative to this frame vectors $v$ of $E$ can be written $$\begin{aligned}
v= \sum_{i=0}^r v^i e_i\end{aligned}$$ and $E$-valued (p,q)-forms can be written $$\begin{aligned}
u= \sum_{i=0}^r \sum_{|I|=p,|J|=q} u^i_{IJ} dz^I \wedge
d\zb^J \tensor e_i\end{aligned}$$ where $dz^I= dz^{i_1}\wedge \ldots \wedge dz^{i_p}$ and $J,I=\{ i_1,\ldots,i_p \}$ are increasing multiindices.
Let $h$ be a Hermitian metric on $E$. We represent it by a Hermitian matrix $$\begin{aligned}
h=(h_{i\jb})_{0\le i,j\le r},\quad h_{i\jb}= <e_i, e_j>_h.\end{aligned}$$ Associated to it there is the Chern connection $\nabla :\A(E)
\lra \A^1 (E)$ which can be represented by a matrix of 1-forms $$\begin{aligned}
(\th_i^j)_{0\le i,j\le r},\quad \nabla e_i = \sum_{j=0}^r
\th_i^j \tensor e_j.\end{aligned}$$ The Chistoffel symbols of the first kind $\G^j_{ik}$ are given by the relations $$\begin{aligned}
\th_i^j = \sum_{k=1}^n \G^j_{ik} dz^k.\end{aligned}$$ We have $$\begin{aligned}
\tag{3.1}
\th = (\de h)\cdot h^{-1},\quad \G^j_{ik}=\sum_{s=0}^r
\frac{\de h_{i\sb}}{\de z^k} \cdot h^{\sb j}.\end{aligned}$$ Here $(h^{\sb j})_{s,j}$ is the inverse of the matrix $h$. The connection $\nabla$ induces unique $\C$-linear maps $\nabla :\A^k(E) \lra \A^{k+1}(E)$ by enforcing the Leibnitz rule: $\nabla (u \tensor v)= du \tensor v + (-1)^{|u|}u
\wedge \nabla v$. The composition $\nabla^2 :\A(E) \lra \A^2(E)$ is called the *curvature* of $\nabla$ and has the remarkable property that it is a tensor, i.e. it is a morphism of $\A$-modules. $\nabla^2$ can be represented by a matrix of (1,1)-forms $$\begin{aligned}
(\Th_i^j)_{0\le i,j\le r},\quad \nabla^2 e_i = \sum_{j=0}^r
\Th_i^j \tensor e_j.\end{aligned}$$ The Chistoffel symbols of the second kind $K^j_{ik\lb}$ are given by the relations $$\begin{aligned}
\Th_i^j = \sum_{k,l=1}^n K^j_{ik\lb} dz^k \wedge d\zb^l.\end{aligned}$$ We have $$\begin{aligned}
\Th = d\th - \th \wedge \th = \deb \th,\end{aligned}$$ $$\begin{aligned}
\tag{3.2}
K^j_{ik\bar{l}}= -\sum_{s=0}^r \frac{\de^2 h_{i\sb}}{\de z^k \de \zb^l}
h^{\sb j}+ \sum_{p,q,s=0}^r \frac{\de h_{i\bar{q}}}{\de z^k}\cdot
\frac{\de h_{p\bar{s}}}{\de \zb^l} h^{\bar{q}p} h^{\bar{s}j}.\end{aligned}$$ Given $v\in E_x$ we construct the (1,1)-form $$\begin{aligned}
\Th (v)=\frac{\i}{2\pi}
\frac{<\nabla^2 v, v >_h}{||v||^2_h}= \frac{\i}{2\pi} \frac{1}{||v||^2_h}
K^j_{i k \bar{l}} v^i h_{j \bar{s}} \bar{v}^s dz^k \wedge d \bar{z}^l.\end{aligned}$$ In the case of a line bundle $L$ the form $\Th (v)$ is nothing but the first Chern form of the metric defined by $$\begin{aligned}
c_1 (L,h)= - dd^c \ \text{log}\ h.\end{aligned}$$ Given $\x \in$ T$_x X$ and $v\in E_x$ we define the *holomorphic bisectional curvature of $h$ along $\x$ and $v$* by $$\begin{aligned}
\tag{3.3}
k_x (\x,v)= \frac{<\nabla^2_{\x,\xb} v,v>_h}{||\x ||^2_{\f}
||v||^2_h} = \frac{1}{||\x ||^2_{\f} ||v||^2_h}
\sum_{i,j=0}^r \sum_{k,l=1}^n K^j_{ik\lb}\x^k\xb^l v^i h_{j\sb}\vb^s.\end{aligned}$$
[**(3.4) Definition:**]{} Let $X$ be a complex manifold of dimension $n$ and $E$ a holomorphic vector bundle of rank $r+1\ge 2$ on $X$. A *Finsler metric* $h$ on $E$ is a function $$\begin{aligned}
h:E\lra [0,\infty)\end{aligned}$$ satisfying the following conditions:
1. $h$ is continuous on $E$ and smooth on the complement of the zero-section;
2. $h(\l v)= |\l |h(v)$ for $v\in E,\ \l \in \C$;
3. $h(v)>0$ if $v$ is non-zero;
4. $h_{| E_x \setminus \{ 0\}}$ is a strictly plurisubharmonic function for all $x\in X$.\
The norm associated to a Hermitian metric is thus a particular case of Finsler metric. We will see at (3.11) that a Finsler metric $h$ comes from a Hermitian metric if and only if $h^2$ is of class $\mathcal{C}^2$ on $E$.
Let $\P(E)$ be the fiber bundle with fibers $\P(E_x)=$projective space of lines through the origin in $E_x,\ \ x\in X$. We denote by $\pi :\P(E)\lra X$ the projection onto the base. The tautological line bundle $\L^{-1}= \O_{\P(E)}(-1)$ is the subbundle of $\p^*E$ whose fiber at $(x,[v])$ consists of the line generated by $v$ inside $E_x$. The bundle space of $\L^{-1}$ is the blow-up of $E$ along the zero-section; let $\b :|\L^{-1}|\lra E$ be the blowing-up map. Outside the zero-sections $\b$ is an isomorphism. To give $h$ satisfying properties (i), (ii) and (iii) from above is equivalent to giving a Hermitian metric $\htil^{-1}$ on $\L^{-1}$ via the correspondence $\htil^{-1}= h\circ \b$.
In the sequel we will denote $E_o =E\setminus \{$zero-section}. We denote $p:E\lra X$ the projection onto the base and by $q:E_o \lra \P(E)$ the quotient map. Thus $$\xymatrix {
E_o \ar[rr]^{q}
\ar[dr]^{p} & & \P (E) \ar[dl]^{\p} \\
& X
}$$ is a commutative diagram. We will consider the function $G=h^2$ which is continuous on $E$ and smooth on $E_o$.
As before, let $U$ be a coordinate set which trivializes $E$. On $p^{-1}U$ we have coordinates $(z^1,\ldots,z^n,v^0,\ldots,v^r)$. We will consider the following smooth functions on $p^{-1}U\cap E_o$: $$\begin{aligned}
\tag{3.5}
G_{i\jb}=\frac{\de^2 G}{\de v^i \de \vb^j},\quad 0\le i,j\le r.\end{aligned}$$ By property (ii) for all $\l\in \C$ we have $$\begin{aligned}
\tag{3.6}
G(z,\l v)= |\l |^2 G(z,v).\end{aligned}$$ Differentiating with respect to $v^i$ and $\vb^j$ we obtain $$\begin{aligned}
\tag{3.7}
G_{i\jb}(z,\l v)= G_{i\jb} (z,v).\end{aligned}$$ Differentiating this relation with respect to $\l$ we obtain $$\begin{aligned}
\tag{3.8}
\sum_{p=0}^r v^p \frac{\de G_{i\jb}}{\de v^p} = 0 =
\sum_{q=0}^r \vb^q \frac{\de G_{i\jb}}{\de \vb^q}.\end{aligned}$$ Differentiating (3.6) with respect to $\l$ and $\bar{\l}$ we obtain $$\begin{aligned}
\tag{3.9}
G(z,v)= \sum_{i,j=0}^r G_{i\jb} (z,\l v) v^i \vb^j\end{aligned}$$ which, in view of (3.7), takes the form $$\begin{aligned}
\tag{3.10}
G(z,v)= \sum_{i,j=0}^r G_{i\jb} (z,v) v^i \vb^j.\end{aligned}$$
If $h$ were a Hermitian metric with matrix $(h_{i\jb})$ we would have $G= \sum_{i,j=0}^r h_{i\jb} v^i \vb^j$ and $G_{i\jb}(z,v)=
h_{i\jb}(z)$. Thus $h$ comes from a Hermitian metric if and only if the functions $G_{i\jb}$ are constant along the fibers. In fact, more can be said:
\
[**(3.11) Remark:**]{} $h$ comes from a Hermitian metric on $E$ if and only if the function $G=h^2$ is of class $\mathcal{C}^2$ on $E$. Indeed, if this were the case, we could take limit as $\l$ tends to 0 in (3.9) and we would obtain $G(z,v)= \sum_{i,j=0}^r G_{i\jb}(z,0)
v^i \vb^j$. Differentiating again with respect to $v^i$ and $\vb^j$ we would get $G_{i\jb}(z,v)=G_{i\jb}(z,0)$, q.e.d.\
Let us denote by $p_*:$T$E\lra $T$X$ the differential of $p$. The kernel of $p_*$ is a holomorphic subbundle $\V\subset $T$E$ called the *vertical tangent bundle*. Notice that, relative to coordinates $(z^1,\ldots,z^n,v^0,\ldots,v^r)$ on $p^{-1}U$, a frame of $\V$ is given by $\{ \frac{\de}{\de v^0},\ldots,
\frac{\de}{\de v^r}\}$. The transformation rule for this frame is the same as for the frame $\{ e_0,\ldots, e_r\}$. This shows that there is a canonical isomorphism $\V \isom p^*E$ identifying $\frac{\de}{\de v^i}$ with $p^*e_i$. A vector $V$ of $\V$ can be written $$\begin{aligned}
V=\sum_{i=0}^r V^i \frac{\de}{\de v^i}.\end{aligned}$$ A section of $\V$ of particular interest is the *position vector field* $$\begin{aligned}
P(z,v)=\sum_{i=0}^r v^i \frac{\de}{\de v^i}.\end{aligned}$$ Notice that $q^* \L^{-1}$ is the subbundle of $p^*E$ generated by $P$. Property (iv) tells us that the matrix $(G_{i\jb})_{0\le i,j\le r}$ is positive definite at each point of $E_o$. This allows us to define a Hermitian metric on $\V_{|E_o}$ by putting $$\begin{aligned}
<V,W>_\V =\sum_{i,j=0}^r G_{i\jb}(z,v)V^i \overline{W}^j.\end{aligned}$$ This metric is nothing but pull-back of $h$ in the case when $h$ is Hermitian. Notice that, by virtue of (3.10), we have $G=||P||^2_\V$. Also notice that, by virtue of (3.9), $<,>_V$ is constant on the fibers of $q$ hence it descends to a Hermitian metric of $\p ^*E$.
The Chern connection $\nabla^\V$ and its curvature make sense just as for any Hermitian metric. $\nabla^\V$ has connection matrix $$\begin{aligned}
\th_i^j =\sum_{k=1}^n \G^j_{ik} dz^k + \sum_{p=0}^r \g^j_{ip} dv^p\end{aligned}$$ where $$\begin{aligned}
\tag{3.12}
\G_{ik}^j=\sum_{s=0}^r \frac{\de G_{i\sb}}{\de z^k}\cdot G^{\sb j},
\quad \g^j_{ip}=\sum_{s=0}^r \frac{\de G_{i\sb}}{\de v^p}\cdot G^{\sb j}.\end{aligned}$$ The curvature matrix $\Th$ has now horizontal, vertical and mixed components: $$\begin{aligned}
\Th_i^j & = & \sum_{k,l=1}^n K^j_{ik\bar{l}}\ dz^k\wedge d\zb^l +
\sum_{p,q=0}^r \k^j_{ip\bar{q}}\ dv^p \wedge d\vb^q \\
& & + \sum_{k=1}^n \sum_{q=0}^r \m^j_{ik\bar{q}}\ dz^k
\wedge d\vb^q + \sum_{l=1}^n
\sum_{p=0}^r \n^j_{ip\bar{l}}\ dv^p \wedge d\zb^l,\end{aligned}$$ where $$\begin{aligned}
\tag{3.13}
K^j_{ik\bar{l}}= -\sum_{s=0}^r \frac{\de^2 G_{i\sb}}{\de z^k \de \zb^l}
G^{\sb j}+ \sum_{p,q,s=0}^r \frac{\de G_{i\bar{q}}}{\de z^k}\cdot
\frac{\de G_{p\bar{s}}}{\de \zb^l} G^{\bar{q}p} G^{\bar{s}j},\end{aligned}$$ $$\begin{aligned}
\k^j_{ip\qb}= -\frac{\de \g^j_{ip}}{\de \vb^q},\ \m^j_{ik\qb}
=-\frac{\de \G^j_{ik}}{\de \vb^q},\ \n^j_{ip\lb}=
-\frac{\de \g^j_{ip}}{\de \zb^l}.\end{aligned}$$ Using (3.8) we obtain the following relations: $$\begin{aligned}
\tag{3.14}
\sum_{i=0}^r \g^j_{ip} v^i=0,\end{aligned}$$ $$\begin{aligned}
\tag{3.15}
\sum_{i=0}^r \n^j_{ip\lb}v^i= \sum_{i=0}^r \k^j_{ip\qb}v^i=
\sum_{j,s=0}^r \m^j_{ik\qb} G_{j\sb} \vb^s =0.\end{aligned}$$
For any tangent vector $\z \in$ T$E_o$ of the form $$\begin{aligned}
\z =\sum_{k=1}^n a^k \frac{\de}{\de z^k} +
\sum_{i=0}^r b^i \frac{\de}{\de v^i}\end{aligned}$$ we have $$\begin{aligned}
\nabla_\z^\V P & = & \sum_{i=0}^r (b^i +\sum_{j=0}^r \sum_{k=1}^n
\G^i_{jk} v^j a^k + \sum_{j,p=0}^r \g^i_{jp} v^j b^p)
\frac{\de}{\de v^i}\\
& = & \sum_{i=0}^r (b^i +\sum_{j=0}^r \sum_{k=1}^n
\G^i_{jk} v^j a^k )
\frac{\de}{\de v^i} \quad \quad \text{by (3.14)}.\end{aligned}$$ This calculation shows that the linear map of bundles $$\begin{aligned}
\nabla^\V P:\text{T}E_o \lra \V,\quad \z \lra \nabla_\z^\V P\end{aligned}$$ is a surjection. Thus the kernel of $\nabla^\V P$ is a smooth subbundle $\H \subset $ T$E_o$ which we call the *horizontal tangent bundle*. We have the smooth decomposition $$\begin{aligned}
\text{T}E_o =\H \oplus \V\end{aligned}$$ and at any point $(x,v)\in E_o$ the differential $p_*:\H_{(x,v)}
\lra $ T$_x X$ is an isomorphism. Under this isomorphism a vector $\x \in $ T$_x X$ corresponds to a vector $\x^\H \in \H_{(x,v)}$ called the *horizontal lift of* $\x$. The vertical tangent fields together with the horizontal lifts of the tangent fields $\de_k =\de /\de z^k$ give a smooth frame of T$E_o$: $$\begin{aligned}
\tag{3.16}
\left\{
\begin{array}{ll}
\de / \de v^i, & 0\le i\le r, \\
\de_k^\H = \de / \de z^k - \sum_{i,j=0}^r \G^i_{jk} v^j
\de / \de v^i, & 1\le k\le n.
\end{array}
\right.\end{aligned}$$ The dual basis is $$\begin{aligned}
\tag{3.17}
\left\{
\begin{array}{ll}
dz^k, & 1\le k\le n,\\
\z^i = dv^i + \sum_{k=1}^n \sum_{j=0}^r \G^i_{jk} v^j dz^k,
& 0\le i\le r.
\end{array}
\right.\end{aligned}$$
Just as in the Hermitian case, for any point $(x,v)\in E_o$ there is a holomorphic frame of $E$ on a neighbourhood of $x$ with respect to which we have $$\begin{aligned}
\tag{3.18}
G_{i\jb} (x,v)= \d_{ij},\quad \frac{\de G_{ij}}{\de z^k}(x,v)=
0,\quad 0\le i,j\le r,\ 1\le k\le n.\end{aligned}$$ Here $\d_{ij}$ is the Kronecker symbol. We call such a frame *normal at* $(x,v)$. Relative to such a frame we have $\de_k^\H =\de /\de z^k,\ \z^i=dv^i$ at $(x,v)$.
We claim that for any non-zero vector $v \in E_x$ the following expression gives us a well-defined (1,1)-form on T$_x X$: $$\begin{aligned}
\tag{3.19}
<K(\cdot,\cdot)v,v>_\V = \sum_{i,j,p=0}^r \sum_{k,l=1}^n
K^p_{ik\lb} v^i G_{p\jb} \vb^j dz^k\wedge d \zb^l.\end{aligned}$$ To see this we will show that for any $\x_1,\x_2 \in$ T$_x X$ we have $$\begin{aligned}
\tag{3.20}
<K(\x_1,\bar{\x}_2)v,v>_\V= <\nabla^2_{\x_1^\H,\bar{\x}_2^\H}
P,P>_\V\end{aligned}$$ where $\nabla^2$ is the curvature of $\nabla^\V$. Indeed, adopting Einstein’s summation convention, we have $$\begin{aligned}
<\nabla^2 P,P>_\V & = & K^j_{ik\lb} v^i G_{j\sb} \vb^s
d z^k \wedge d\zb^l + \k^j_{ip\qb} v^i G_{j\sb} \vb^s dv^p
\wedge d\vb^q \\
& & +\m^j_{ik\qb} v^i G_{j\sb}\vb^s dz^k\wedge d\vb^q +
\n^j_{ip\lb} v^i G_{j\sb}\vb^s dv^p\wedge d\zb^l.\end{aligned}$$ By (3.15) the last three terms on the right-hand-side vanish, yielding (3.20). As a byproduct we get the relation $$\begin{aligned}
\Th(P)= \frac{\i}{2\pi}\cdot \frac{1}{||P||^2_\V}
\sum_{i,j,s=0}^r K^j_{ik\lb} v^i G_{j\sb} \vb^s dz^k \wedge
d\zb^l.\end{aligned}$$
In order to define the holomorphic bisectional curvature of $h$ we need to first put a metric on T$E_o$. This is done by making the decomposition T$E_o=\H\oplus \V$ an orthogonal decomposition. Namely, any tangent vector $\z \in $ T$E_o$ can be written $\z= \x^\H + V$ and we put $||\z||^2=||\x||^2_\f + ||v||^2_\V$. We then define the *holomorphic bisectional curvature of $h$ along $\z$ and $v$* just as at (3.3): $$\begin{aligned}
\tag{3.21}
k(\z,v)= \frac{<\nabla^2_{\z,\bar{\z}}v,v>_\V}{||\z ||^2 ||v||^2_h}
\qquad \qquad \qquad \qquad \qquad \qquad \quad\end{aligned}$$ $$\begin{aligned}
\qquad \qquad \quad \quad \
& = & \frac{1}{||\z ||^2 ||v||^2_h} \sum_{i,j,s=0}^r
\Big\{ \sum_{k,l=1}^n K^j_{ik\lb} a^k \ab^l +
\sum_{p,q=0}^r \k^j_{ip\qb} b^p \bb^q \\
& & \qquad \quad + \sum_{k=1}^n \sum_{q=0}^r \m^j_{ik\qb} a^k \bb^q +
\sum_{l=1}^n \sum_{p=0}^r \n^j_{ip\lb} b^p \ab^l
\Big\} v^i G_{j\sb}\vb^s.\end{aligned}$$ The knowledge that (3.19) is well defined allows us to introduce the notion of *horizontal holomorphic bisectional curvature of $h$ along $\x\in $ T$_x X$ and $v\in E_x$*: $$\begin{aligned}
\tag{3.22}
k_x (\x,v)= k_{(x,v)}(\x^\H,v)=
\frac{<\nabla^2_{\x^\H,\bar{\x}^\H}P,P>_\V}
{||\x ||^2_\f ||v||^2_h} \end{aligned}$$ $$\begin{aligned}
\qquad \qquad \quad \ \ = \frac{1}{||\x ||^2_\f ||v||^2_h}
\sum_{i,j,s=0}^r \sum_{k,l=1}^n
K^j_{ik\lb} \x^k \bar{\x}^l v^i G_{j\sb} \vb^s.\end{aligned}$$
In the remaining part of this section we will compute $c_1 (\L,\htil)$. Recall that, under the isomorphism $\V \isom p^*E$, we can identify $q^* \L^{-1}$ with the sub-bundle of $\V$ spanned by $P$. Relation (3.10) tells us that the pull-back metric $q^* \htil^{-1}$ is nothing but the induced metric from $\V$. Therefore we have $$\begin{aligned}
q^* c_1(\L,\htil)= dd^c \text{log} ||P||^2_\V.\end{aligned}$$ We claim that $$\begin{aligned}
dd^c \text{log} ||P||^2_\V & = &
\frac{\i}{2\pi}\cdot \frac{1}{G^2} \Big\{ \sum_{i,j=0}^r G G_{i\jb}
\z^i \wedge \bar{\z}^j -
\sum_{i,j,p,q=0}^r G_{i\qb} G_{p\jb} \vb^q v^p
\z^i \wedge \bar{\z}^j \Big\}\end{aligned}$$ $$\begin{aligned}
\tag{3.23}
\quad \quad -\frac{\i}{2\pi} \cdot \frac{1}{G}
\sum_{i,j,s=0}^r \sum_{k,l=1}^r
K^j_{ik\lb} v^i G_{j\sb}\vb^s dz^k \wedge d\zb^l.\end{aligned}$$ The part involving terms $\z^i \wedge \bar{\z}^j$ will be called the *vertical component of* $c_1 (\L,\htil)$ and will be denoted $c_1 (\L,\htil)^\V$. The part involving terms $dz^k \wedge d\zb^l$ will be called the *horizontal component of* $c_1 (\L,\htil)$ and will be denoted $c_1 (\L,\htil)^\H$. Notice that $c_1 (\L,\htil)^\V$ is semi-positive definite and its restriction to the fibers of $\P(E)$ is positive-definite. Indeed, relative to a normal frame for $E$ at $(x,v)$ we have $$\begin{aligned}
c_1 (\L,\htil)^\V = \frac{\i}{2\pi}\cdot \frac{1}{||v||^2}
\sum_{i=0}^r dv^i \wedge d\vb^i
- \frac{\i}{2\pi}\cdot \frac{1}{||v||^4}
\sum_{i,j=0}^r \vb^i v^j dv^i \wedge d\vb^j\end{aligned}$$ with $||v||^2= \sum_{i=0}^r |v_i |^2$. The above is nothing but the Fubini-Study form on projective space which is known to be positive-definite. Notice that $$\begin{aligned}
c_1 (\L,\htil)^\H = - \Th (P).\end{aligned}$$ In normal coordinates at $(x,v)$ we have $$\begin{aligned}
\tag{3.24}
c_1 (\L,\htil)= \o - \Th(P)\end{aligned}$$ where $\o$ is the Fubini-Study form on the first component of $$\begin{aligned}
\text{T}_{(x,[v])}\P(E)= \text{T}_{[v]} \P(E_x)
\oplus \text{T}_x X.\end{aligned}$$ Thus $\L$ equipped with $\htil$ is a positive line bundle if and only if the horizontal holomorphic bisectional curvature of the Finsler metric $h$ is negative! This has an important consequence: assume $X$ to be compact; then $E^*$ is ample if and only if there is a Finsler metric (with only properties (i), (ii) and (iii)) on $E$ having negative horizontal holomorphic bisectional curvature. We will not need this fact, but it is worth mentioning because it illustrates the philosophy we advertized in the introduction: geometric properties of $E^*$ should be defined by means of $\O_{\P(E)}(1)$.\
\
Hermitian metrics of finite order
=================================
Let $E$ be a holomorphic vector bundle on the special affine variety $X$. We recall from [@griffiths-cornalba] the notion of Hermitian metric on $E$ of finite order. It is defined by means of an estimate on the holomorphic bisectional curvature. The theorem of Griffiths and Cornalba (4.4) states that a holomorphic line bundle on $X$ admits a unique finite order structure. In fact, as we explained in the introduction, Griffiths and Cornalba gave four definitions for the notion of finite order vector bundle which are equivalent in the case of line bundles. One of the crucial steps in the proof of the four equivalences is their vanishing theorem (4.6) for the sheaf of sections of finite order.
\
[**(4.1) Definition:**]{} A Hermitian holomorphic vector bundle $(E,h)$ on $X$ is said to *have order $\l$* if its holomorphic bisectional curvature is of order $\r^\l$: there is $\k >0$ such that $$\begin{aligned}
|k_x (\x,v)|\le \k \r^\l\end{aligned}$$ for all $x\in X,\ \x \in $ T$_x X,\ v\in E_x$.
Notice that, adopting the notation following (3.2), the above estimate is equivalent to the inequality $$\begin{aligned}
|\Th (v)|\le \k \r^\l \f.\end{aligned}$$ Also notice that a line bundle $(L,h)$ has finite order if and only if $$\begin{aligned}
|c_1 (L,h)| \le \k \r^\l \f.\end{aligned}$$
[**(4.2) Remark:**]{} The above definition is preserved under standard operations with vector bundles: direct sum, tensor product, dualization, symmetric power, exterior power etc. In particular, if $E_1$ and $E_2$ are Hermitian vector bundles of finite order then so is ${\mathcal Hom}(E_1, E_2)
\isom E_1^* \tensor E_2$.\
Sections of finite order of a Hermitian vector bundle are defined the same way as functions: for a global section $\s$ of $E$ we introduce the *maximum modulus function* $$\begin{aligned}
M_\s (r)= \text{log max}\{ |\s(x)|_h,\ \r (x)\le r\}.\end{aligned}$$ We say that $\s$ *has order $\l$ in the sup-norm sense* if $M_\s (r)=O(r^\l)$. In view of the previous remark this gives us a notion of *finite order morphism* between two Hermitian vector bundles. The composite of morphisms of finite order is of finite order, too. Thus we have constructed the category of *Hermitian holomorphic vector bundles of finite order*. In this category a vector bundle of rank $r$ is trivial, i.e. it is isomorphic to the trivial vector bundle $\underline{\C}^r$, if and only if it posesses $r$ global sections of finite order which generate the fiber at every point. We will denote by Vect$^r_{\fo}(X)$ the set of all Hermitian holomorphic vector bundles of rank $r$ on $X$ modulo isomorphisms in the finite order category. Vect$^1_{\fo}(X)$ together with the tensor product forms a group which we call the *Picard group of finite order line bundles on $X$* and denote Pic$_{\fo}(X)$. As we explained in the introduction, the motivation for our work is the following:
\
[**(4.3) Question:**]{} Is the canonical map Vect$^r_{\fo}(X)
\lra $ Vect$^r_{\text{hol}} (X)$ a bijection?\
The question was answered in the affirmative for the case of line bundles.
\
[**(4.4) Theorem**]{} (Griffiths and Cornalba [@griffiths-cornalba]): *The canonical map $$\begin{aligned}
\text{Pic}_{\fo} (X)\lra \text{Pic}_{\text{hol}}(X)\end{aligned}$$ is an isomorphism of groups. In other words, every holomorphic line bundle on $X$ admits a metric of finite order and a Hermitian holomorphic line bundle of finite order is trivial in the holomorphic category if and only if it is trivial in the finite order category.*
\
[**(4.5) Definition:**]{} Let $(E,h)$ be a holomorphic Hermitian vector bundle. A global section $\s$ of $E$ is said to *have order $\l$* if there is $\k \ge 0$ such that $$\begin{aligned}
\int_X |\s |^2_h \cdot e^{-\k \r^\l} \Phi \ < \ \infty.\end{aligned}$$ Here $\Phi =\f^n$ is the volume form of $\f$. We define the *sheaf $\O_\l (E)$ of germs of holomorphic sections of $E$ of order $\l$* as follows: $\O_\l (E)$ is a sheaf on $\Xb$; for each open set $U\subset \Xb$ the space of sections $\O_\l (U,E)$ consists of those holomorphic sections $\s$ of $E$ defined on $U \cap X$ with the property that around each point at infinity $x\in (\Xb \setminus X)\cap
U$ there is a neighbourhood $W\subset U$ and a constant $\k >0$, depending on $x$ and $\s$, such that $$\begin{aligned}
\int_{W\cap X} |\s |^2_h \cdot e^{-\k \r^\l} \Phi \ < \ \infty.\end{aligned}$$ Similarly, we define the *sheaf $\O_{\fo}(E)$ of germs of holomorphic sections of $E$ of finite order* by the same estimate as above with the additional requirement that $\l$ depend on $\s$ and on the point at infinity $x$.
It is easily seen that $\O_\l (E)$ and $\O_{\fo}(E)$ are modules over $\O_\l$, respectively over $\O_{\fo}$. Their restrictions to $X$ coincide with the sheaf $\O (E)$ of germs of sections of $E$ because in (4.5) there are no conditions at the points $x\in X$ away from infinity. By the compactness of the divisor at infinity it is also transparent that the spaces of global sections $\O_\l (\Xb,E)$ and $\O_{\fo} (\Xb,E)$ are nothing but the spaces of global sections of $E$ of order $\l$, respectively of finite order.\
Let us now assume that $E$ has order $\l$. Then a global section $\s$ of $E$ has finite order in the sup-norm sense if and only if it has finite order in the sense of (4.5). One direction is immediate: if $M_\s (r)= O(r^\l)$ then $|\s |_h \le e^{\k \r^\l}$ for some $\k \ge 0$ and the integral $$\begin{aligned}
\int_X |\s |^2_h \cdot e^{-2 \k \r^\l} \Phi\end{aligned}$$ is finite because $\Phi$ has finite volume. For the other direction we refer to [@griffiths-cornalba].
\
[**(4.6) Theorem:**]{} *Let $E$ be a Hermitian holomorphic vector bundle of order $\l$ on a special affine variety $X$. Then, for all $q\ge 1$, we have* $$\begin{aligned}
H^q (\Xb, \O_\l(E))=0,\quad H^q (\Xb, \O_{\fo}(E))=0.\end{aligned}$$
We now prepare to move from the Hermitian to the Finsler case. We begin by translating definition (4.5) “upstairs” on $\P (E)$. There is a canonical isomorphism of vector bundles $\pi_* (\L) \isom E^*$ which associates to a section $\s$ of $E^*$ the section $\st$ of $\L$ defined by $$\begin{aligned}
\tag{4.7}
\st_{(x,[v])} (t v)= \s (tv),\end{aligned}$$ for all $x\in X,\ v\in E_x \setminus \{ 0\},\ t\in \C$. Recall that the Hermitian metric $h$ on $E$ (which in particular is a Finsler metric) induces a metric $\htil$ on $\L$. Denote by $h^*$ the dual metric on $E^*$. It happens that $$\begin{aligned}
\tag{4.8}
|\s (x)|_{h^*}= \text{max} \{ |\st (x,[v])|_{\htil},\ \
v\in E_x \setminus \{ 0\} \}.\end{aligned}$$
Assume that $E$ has a Hermitian metric $h$ of order $\l$. We claim that there is a positive constant $\k$ such that the expression $$\begin{aligned}
\tag{4.9}
\ft = c_1 (\L,\htil) + dd^c (\k \r^\l)\end{aligned}$$ gives a positive (1,1)-form on $\P(E)$. We check this at a point $(x,[v])$. Since positivity is preserved under change of coordinates we can work with a frame of $E$ which is normal at $x$. By our hypothesis (4.1) we have $|\Th (v)|\le \k \r^\l \f$. We also have $$\begin{aligned}
\frac{\k \l}{2} \r^\l \f =
\frac{\k \l}{2} \r^\l dd^c \t & \le & \frac{\k \l}{2} \r^\l dd^c \t
+ \frac{\k \l^2}{4} \r^\l d\t \wedge d^c\t \\
& = & \k \ dd^c e^{\l \t /2} \\
& = & dd^c (\k \r^\l) \\
& \le & \frac{\k \l}{2} \r^\l dd^c \t
+ \frac{\k \l^2}{4} \r^{\l+c}\f \quad \quad \qquad
\text{by (1.8)}. \end{aligned}$$ From this calculation we get the estimate $$\begin{aligned}
\tag{4.10}
\frac{\k \l}{2}\r^\l \f \le -\Th(v) + dd^c (\k \r^\l)
\le 2\k \r^{\l +c}\f.\end{aligned}$$ Combining (3.24) with (4.10) we get the following estimate at $(x,[v])$: $$\begin{aligned}
\tag{4.11}
\o + \frac{\k \l}{2} \r^\l \f \le \ft \le \o + 2 \k \r^{\l +c}\f.\end{aligned}$$ This insures the positivity of $\ft$. In the sequel $\P(E)$ will be considered equipped with the Kähler metric induced by $\ft$. Since $\r^\l \circ \pi$ is a plurisubharmonic exhaustive function, we can arrange, possibly by choosing a larger $\k$, that this metric on $\P(E)$ be complete.
\
[**(4.12) Claim:**]{} *Assume that $E$ is a holomorphic Hermitian vector bundle of order $\l$. Let $U\subset X$ be an open subset and $\s$ a section of $E^*$ over $U$. Then there exists $\k_1 \ge 0$ such that* $$\begin{aligned}
\int_U |\s |^2_{h^*} \cdot e^{-\k_1 \r^\l} \Phi \ < \ \infty\end{aligned}$$ *if and only if there exists $\k_2 \ge 0$ such that* $$\begin{aligned}
\int_{\pi^{-1}(U)} |\st |^2_{\htil}\cdot e^{-\k_2 \r^\l}\Phit
\ < \ \infty.\end{aligned}$$
*Proof:* Let us denote by $\ft^\V$ and $\ft^\H$ the vertical, respectively the horizontal part of $\ft$, as defined at the end of §3. From (4.10) we have $$\begin{aligned}
\tag{4.13}
\frac{\k \l}{2} \r^\l \f \le \ft^\H \le 2\k \r^{\l +c}\f\end{aligned}$$ where, by an abuse of notation, we write $\f$ instead of $\pi^* \f$. We have $$\begin{aligned}
\Phit = \ft^{n+r} = {n+r \choose n}
(\ft^\H)^n \wedge (\ft^\V)^r\end{aligned}$$ which, combined with (4.13), gives $$\begin{aligned}
\tag{4.14}
C^{-1} \r^{n\l} \f^n \wedge (\ft^\V)^r \le
\Phit \le
C \r^{n\l +nc} \f^n \wedge (\ft^\V)^r\end{aligned}$$ for a fixed positive constant $C$. This estimate is useful because one can apply Fubini’s theorem to $|\st |^2_{\htil} \cdot e^{-\k \r^\l} \f^n \wedge (\ft^\V)^r$. More precisely, one can integrate this form first vertically along the fibers of $\P(E)$ and then horizontally along $X$. This finishes the proof of the claim because there is a positive constant $A$ depending only on $r$ such that for all $x\in X$ $$\begin{aligned}
\tag{4.15}
\int_{[v]\in \P(E_x)} |\st (x,[v])|^2_{\htil} (\ft_{|\P(E_x)})^r
= A |\s (x)|^2_{h^*}.\end{aligned}$$\
Finsler Metrics of Finite Order. The Vanishing Theorem
======================================================
We explained in the introduction the difficulty one encounters in trying to generalize (4.4) and the four equivalences to bundles of rank greater than 1. It seems to us that the correct antidote is to translate the definition of finite order “upstairs” on $\P(E)$. So we define Finsler metrics of finite order by a very similar estimate on the holomorphic bisectional curvature. We then define sections of finite order and we prove that they span at every point, cf. (5.23). This we achieve by means of the vanishing theorem (5.12) which generalizes (4.6).
Traditionally, vanishing theorems are given either on Stein manifolds or on compact Kähler manifolds. Our case here is a hybrid: we will have to solve the $\deb$-equation on $\P(E)$ whose base is Stein (because it is affine) while its fibers are compact.
The original proof of Kodaira’s Vanishing Theorem makes use of Hodge theory: one knows that on a compact Kähler manifold cohomology classes can be represented by harmonic forms and one argues, using the a priori estimate, that such forms do not exist. For non-compact manifolds this argument does not work because we do not know if the Hodge representation theorem holds. Instead we will use a very potent technique developed by Hörmander in [@hormander-estimates] involving some rudiments of functional analysis. We synthetize this technique in proposition (5.3) and theorem (5.4) from below. We refer to [@hormander], [@hormander-estimates], [@andreotti-vesentini] and [@kodaira].
Let $Y$ be a complex manifold of dimension $m$ equipped with a positive Kähler form $\o$ which induces a complete metric. Relative to local holomorphic coordinates $(z^1,\ldots,z^m)$ we write $$\begin{aligned}
\o =\i \sum_{i,j=1}^m g_{i \jb}\ dz^i \wedge d\zb^j.\end{aligned}$$ The condition that $\o$ be positive definite means that the matrix $(g_{i\jb})_{1\le i,j\le m}$ is Hermitian and positive-definite. The condition that $\o$ be Kähler means that $d \o =0$. This is a very natural condition to consider because it is equivalent to saying that the complex structure on the real tangent bundle of $Y$ is compatible with the Levi-Civita connection associated to the induced Riemannian metric on the real tangent bundle of $Y$.
Let Ric$(\o)$ be the Ricci curvature of $\o$. It is a (1,1)-form given in local coordinates by $$\begin{aligned}
\text{Ric}(\o)= dd^c \text{log(det}(g_{i \jb})).\end{aligned}$$ It is nothing but the first Chern form of the canonical line bundle $K_Y := \wedge^m $ T$^* Y$ equipped with the metric induced by $\o$.
Recall from §3 that the first Chern form of a holomorphic line bundle $L$ equipped with a Hermitian metric $h$ is given by $$\begin{aligned}
c_1 (\L,h)= -dd^c \text{log}(h).\end{aligned}$$
Let us choose an orthonormal frame $\{ \x_1, \ldots , \x_m \}$ for the tangent space at a point $y\in Y$ and a unitary vector $e\in L_y$. Relative to the dual frame $\{ d\x^1, \ldots, d\x^m \}$ for T$^*_y Y$ we write $$\begin{aligned}
\text{Ric}(\o)= \i \sum_{i,j=1}^m
R_{i \jb}\ d\x^i \wedge d\xb^j,\\
c_1(\L)= \i \sum_{i,j=1}^m K_{i \jb}\ d\x^i \wedge d\xb^j.\end{aligned}$$ For an $L$-valued (0,q)-form $$\begin{aligned}
u= \sum_{|I|=q} u_I \ d\xb^I \tensor e\end{aligned}$$ we define the pointwise operators $$\begin{aligned}
<Ru,u>= q \sum_{i,j=1}^m \sum_{|I|= q-1} R_{i \jb}\ u_{iI} \bar{u}_{jI},\end{aligned}$$ $$\begin{aligned}
<Ku,u>= q \sum_{i,j=1}^m \sum_{|I|= q-1} K_{i \jb}\ u_{iI} \bar{u}_{jI},\end{aligned}$$ and their integrated versions $$\begin{aligned}
(Ru,u)= \int_Y <Ru,u> dV,\end{aligned}$$ $$\begin{aligned}
(Ku,u)= \int_Y <Ku,u> dV.\end{aligned}$$ Here $u$ is assumed to have compact support and $dV=\o^m$ is the volume form of $\o$. In the sequel we will denote by $\DD^{p,q}(Y,L)$ the space of smooth $L$-valued (p,q)-forms with compact support. The Hermitian inner products on T$^*Y$ and on $L$ induce a Hermitian inner product $<,>$ on $\wedge^{p,q}$T$^*Y \tensor L$. Given $u,v\in \DD^{p,q}(Y,L)$ we put $$\begin{aligned}
(u,v)= \int_Y <u,v>\ dV.\end{aligned}$$ Clearly $(\cdot,\cdot)$ defines a Hermitian inner product on $\DD^{p,q}(Y,L)$. Its associate norm is given by $||u||^2 = (u,u)$. The operator $$\begin{aligned}
\deb : \DD^{p,q}(Y,L) \lra \DD^{p,q+1}(Y,L)\end{aligned}$$ has a formal adjoint $$\begin{aligned}
\del : \DD^{p,q+1}(Y,L) \lra \DD^{p,q}(Y,L)\end{aligned}$$ given by the condition $(\deb u, v)=(u,\del v)$ for all $u\in \DD^{p,q}(Y,L),\ v\in \DD^{p,q+1}(Y,L)$. We mention in passing that $\del =-\star D' \star$ where $\star$ is the Hodge-$\star$ operator while $D'$ is the (1,0)-component of the Chern connection of $L$.
Our key ingredient towards proving vanishing theorems is the following inequality which we state in a particular case that is of interest to us. See [@kodaira] p. 124 for the full statement and p. 126 for the statement from below.
\
[**(5.1) Weitzenböck Inequality:**]{} *Let $Y$ be a Kähler manifold. Let $L$ be a holomorphic Hermitian line bundle on $Y$ and $u$ a smooth $L$-valued (0,q)-form with compact support. Then we have:* $$\begin{aligned}
||\deb u||^2 +||\del u||^2 \ge (Ku,u) - (Ru,u).\end{aligned}$$
\
[**(5.2) Definition:**]{} We say that $L$ is *(p,q)-elliptic* if there is a positive constant $\e$, called *ellipticity constant*, such that for all $u\in \DD^{p,q}(Y,L)$ we have the estimate $$\begin{aligned}
||\deb u||^2 +||\del u||^2 \ \ge \ \e ||u||^2.\end{aligned}$$
\
[**(5.3) Proposition:**]{} *Assume $L$ is (p,q)-elliptic and that the metric on $Y$ is complete and Kähler. Then for any $u\in $ L$^{p,q}(Y,L)$ with $\deb u=0$ there is $v\in $ L$^{p,q-1}(Y,L)$ with $\deb v=u$. In addition $v$ is smooth.*\
Here $\text{L}^{p,q}(Y,L)$ is the completion of $\DD^{p,q}(Y,L)$ relative to the norm $||u||^2$. Notice here the crucial assumption that $Y$ be complete. An equivalent condition is that any ball in $Y$ be relatively compact. We need this for a density argument to make sure that (5.1) holds not only for smooth forms but also for square-integrable forms. See [@andreotti-vesentini], p. 92, lemma 4.
\
[**(5.4) Theorem:**]{} *Let $Y$ be a complex manifold equipped with a complete Kähler metric. Let $L$ be a holomorphic line bundle on $Y$ equipped with a Hermitian metric $h$. We denote by $\o$ the Kähler form of $Y$. Assume that there is a positive constant $\e$ such that* $$\begin{aligned}
c_1 (L,h) - \text{Ric}(\o) \ \ge \ \e \ \o.\end{aligned}$$ *Let $q$ be a positive integer. Then for any $u\in $ L$^{0,q}(Y,L)$ with $\deb u=0$ there is $v\in $ L$^{0,q-1}(Y,L)$ with $\deb v=u$. In addition $v$ is smooth.*
\
*Proof:* By hypothesis for any $u\in \DD^{0,q}(Y,L)$ we have $$\begin{aligned}
(Ku,u)-(Ru,u) \ \ge \ q\e ||u||^2.\end{aligned}$$ Combining this with the Weitzenböck Inequality we obtain $$\begin{aligned}
||\deb u||^2 +||\del u||^2 \ \ge \ q\e ||u||^2.\end{aligned}$$ Hence $L$ is (0,q)-elliptic with ellipticity constant $q\e$. The theorem now follows from (5.3).
\
[**(5.5) Definition:**]{} Let $E$ be a holomorphic vector bundle on $X$ and $h$ a Finsler metric on $E$. We say that *$h$ has order $\l$* if its holomorphic bisectional curvature is of order $\r^\l$: there is $\k>0$ such that $$\begin{aligned}
|k (\z,v)|\le \k \r^\l\end{aligned}$$ for all $y\in E_o,\ \z \in \text{T}_y E_o,\ v\in E_x$ where $p (y)=x$.
Notice that this definition encompasses (4.1). For a line bundle a Finsler metric is the same thing as a Hermition metric, so (5.5) is relevant only when the rank of $E$ is at least 2, which we will assume henceforth.
Notice that if we take $\z$ to be the horizontal lift of a tangent vector from $X$ and $v$ to be the position vector field we obtain that the horizontal holomorphic bisectional curvature is also of order $\r^\l$: $$\begin{aligned}
\tag{5.6}
| k_x (\x,v)|\le \k \r^\l\end{aligned}$$ for all $x\in X,\ \x \in \text{T}_x X,\ v\in E_x$. Adopting the notation below (3.20) the above estimate is equivalent to the inequality $$\begin{aligned}
|\Th (P)|\le \k \r^\l \f.\end{aligned}$$ In view of (3.24) and using the same argument as in the Hermitian case we can find a positive constant $\k$ such that the expression $$\begin{aligned}
\tag{5.7}
\ft = c_1 (\L,\htil)+dd^c (\k \r^\l)\end{aligned}$$ defines a Kähler form on $\P(E)$ which induces a complete metric. Estimate (4.13) also holds: $$\begin{aligned}
\tag{5.8}
\frac{\k \l}{2} \r^\l \f \le \ft^\H \le 2\k \r^{\l+c}\f.\end{aligned}$$ Moreover, the constant $\k$ can be so chosen that $$\begin{aligned}
|c_1 (G_{i\jb})| \le \k \r^\l \ft.\end{aligned}$$ This is so because an estimate on the holomorphic bisectional curvature of a vector bundle $E$ gives an estimate of the curvature of the induced metric on the determinant of $E$. From the proof of (5.12) it transpires that the Ricci curvature of $\P(E)$ is, in essence, equal to $c_1 (G_{i\jb})$ plus some other terms that are controllable. Thus, roughly speaking, the above estimate says that the Ricci curvature of $\P(E)$ has polynomial growth. Taking horizontal and vertical parts we obtain the following inequalities: $$\begin{aligned}
\tag{5.9}
|c_1 (G_{i\jb})^\H|\le \k \r^\l \f,\end{aligned}$$ $$\begin{aligned}
\tag{5.10}
|c_1(G_{i\jb})^\V|\le \k \r^\l c_1 (\L,\htil)^\V.\end{aligned}$$ In fact, (5.6), (5.9) and (5.10) are all that we will need in the sequel. It is not clear to us if these estimates imply (5.5), in other words it seems to us that (5.5) is stronger than (5.6), (5.9) and (5.10). We could have chosen the three latter estimates as our definition of finite order Finsler metric, but then the similarity with the Hermitian case would have been obscured.
Our Finsler metric is on $E$ but we will be concerned with and we will prove a vanishing theorem for the sections of order $\l$ of $E^*$. We define the latter in the spirit of (4.12), by using the one-to-one correspondence (4.7) between sections of $E^*$ and sections of $\L$. Before doing that we fix a Hermitian metric $g$ of finite order on det$(E)$. Such a metric exists by (4.4). Under the isomorphism $$\begin{aligned}
\L \isom \L \tensor \p^* (\text{det}(E^*))\tensor
\p^* (\text{det}(E))\end{aligned}$$ we consider the following metric on $\L$: $$\begin{aligned}
\ltil = \htil \cdot \text{det}(G_{i\jb})^{-1} \cdot \p^*(g).\end{aligned}$$
[**(5.11) Definition:**]{} Let $E$ be a holomorphic vector bundle on $X$ equipped with a Finsler metric of order $\l$. A global section $\s$ of $E^*$ is said to *have order* $\l$ if there is $\k >0$ such that $$\begin{aligned}
\int_{\P(E)} |\st |^2_{\ltil} \cdot e^{-\k \r^\l} \Phit
\ < \ \infty.\end{aligned}$$
We define the *sheaf $\O_\l (E^*)$ of germs of holomorphic sections of $E^*$ of order $\l$* as follows: $\O_\l (E^*)$ is a sheaf on $\Xb$; for each open set $U\subset \Xb$ the space of sections $\O_\l (U,E^*)$ consists of those holomorphic sections $\s$ of $E^*$ defined on $U \cap X$ with the property that around each point at infinity $x\in (\Xb \setminus X)\cap
U$ there is a neighbourhood $W\subset U$ and a constant $\k >0$, depending on $X$ and $\s$, such that $$\begin{aligned}
\int_{\p^{-1}(W\cap X)}
|\st |^2_{\ltil} \cdot e^{-\k \r^\l} \Phit \ < \ \infty.\end{aligned}$$
Similarly, we define the *sheaf $\O_{\fo}(E^*)$ of germs of holomorphic sections of $E^*$ of finite order* by the same estimate as above with the additional requirement that $\l$ depend on $\s$ and on the point at infinity $x$.
It is easily seen that $\O_\l (E^*)$ and $\O_{\fo}(E^*)$ are modules over $\O_\l$, respectively over $\O_{\fo}$. Their restrictions to $X$ coincide with the sheaf $\O (E^*)$ of germs of sections of $E^*$ because in (5.11) there are no conditions at the points $x\in X$ away from infinity. By the compactness of the divisor at infinity it is also transparent that the spaces of global sections $\O_\l (\Xb,E^*)$ and $\O_{\fo} (\Xb,E^*)$ are the space of global sections of $E^*$ of order $\l$, respectively of finite order which is independent of the choice of compactification $\Xb$. Finally, the use of $\ltil$ instead of $\htil$ may seem awkward but, in doing so, we do not deviate from our aim of studying finite order objects. Our choice of metric is dictated by technical reasons which will become clear in the proof of the next theorem. So here is the main result of this section:
\
[**(5.12) Theorem:**]{} *Let $X$ be a special affine variety and $E$ a holomorphic vector bundle on $X$ equipped with a Finsler metric $h$ of order $\l$. Then there is $\m \ge \l$ such that for all $q\ge 1$, we have* $$\begin{aligned}
H^q (\Xb, \O_\m(E^*))=0,\quad H^q (\Xb, \O_{\fo}(E^*))=0.\end{aligned}$$
*Proof:* Let $\m$ be the largest between $\l$ and the order of $g$. For conciseness of notation we write $Y$ instead of $\P(E)$. It is known, see for instance [@sommese], that we have the isomorphism $$\begin{aligned}
K_Y \isom \L^{-r-1} \tensor \p^*(\text{det}(E^*))
\tensor \p^*(K_X).\end{aligned}$$ On the canonical line bundle $K_X$ we have the metric $k$ induced by $\f$ as follows: if, relative to a local coordinates system $(z^1,\ldots,z^n)$, we can write $$\begin{aligned}
\f = \i \sum_{i,j=1}^n \f_{i\jb}\ dz^i \wedge d\zb^j\end{aligned}$$ then, relative to the frame $dz^1 \wedge \ldots \wedge dz^n$ of $K_X$ we have $$\begin{aligned}
k= \text{det}(\f_{i\jb})^{-1}.\end{aligned}$$ Notice that $c_1 (K_X,k) = \text{Ric}(\f)$. Likewise, let $\kt$ be the metric on $K_Y$ induced by $\ft$. Using the above isomorphism we put another metric on $K_Y$ by $$\begin{aligned}
\kt'= \htil^{-r-1} \cdot \text{det}(G_{i\jb})^{-1} \cdot
\p^*(k).\end{aligned}$$ We claim that $\kt$ and $\kt'$ are almost equivalent: there is $C >0$ such that $$\begin{aligned}
\tag{5.13}
C\ \r^{-n\l -nc} \ \kt' \le \kt \le
C\ \r^{-n\l} \ \kt'.\end{aligned}$$ We check this at a point $(x,[v])$. We may assume that the frame of $E$ is normal at $(x,v)$. Taking into account that det$(G_{i\jb})=1$ at $(x,[v])$ and taking determinants in (4.11) we obtain $$\begin{aligned}
\text{det}(\o) \text{det}(G_{i\jb}) 2^{-n} \k^n \r^{n\l}
\text{det}(\f)
\le \text{det}(\ft) \le
\text{det}(\o) \text{det}(G_{i\jb}) 2^{n} \k^n \r^{n\l +nc}
\text{det}(\f).\end{aligned}$$ Recall that $\o$ is the Fubini-Study form on T$_{[v]}\P^r$. It is known that the metric det$(\o)$ induced by $\o$ on $\wedge^r \text{T}\P^r \isom \O_{\P^r}(r+1)$ coincides with the canonical metric of $\O_{\P^r}(r+1)$. Therefore, under the identification $\L_{(x,[v])}^{\tensor (r+1)}\isom
\wedge^r \text{T}_{(x,[v])}\P(E_x)$, we have $\htil^{r+1} =
\text{det}(\o)$. Taking inverse in the above inequalities we obtain (5.13).
Under the isomorphism $\L \isom \L \tensor K_Y^{-1}
\tensor K_Y$ we put a metric $\ltil'$ on $\L$ by $$\begin{aligned}
\ltil'=\ltil \cdot (\kt')^{-1} \cdot \kt.\end{aligned}$$ By (5.13) the two metrics $\ltil$ and $\ltil'$ on $\L$ are almost equivalent: $$\begin{aligned}
C\ \r^{n\l} \ \ltil' \ \le \ \ltil \le C\ \r^{n\l +nc} \ \ltil'.\end{aligned}$$ Thus, in definition (5.11) we can replace $\ltil$ by $\ltil'$. The latter has the advantage that its Chern form can be bounded from below. Indeed, denoting $\L_\k$ the line bundle $\L$ equipped with the metric $\ltil_\k =\ltil' \cdot e^{-\k\r^\m}$, we have: $$\begin{aligned}
c_1 (\L_\k)- \text{Ric}(\ft)
& = & c_1 (\L,\ltil')+ dd^c (\k \r^\m)- c_1 (K_Y, \kt) \\
& = & c_1 (\L,\ltil) - c_1 (K_Y,\kt') +dd^c (\k \r^\m) \\
& = & c_1 (\L,\htil) - c_1 (\p^*(\text{det}(E)), \text{det}(G_{i\jb}))
+ c_1 (\text{det}(E),g) \\
& & + (r+1) c_1 (\L,\htil) + c_1 (\p^*(\text{det}(E)),
\text{det}(G_{i\jb})) \\
& & - \p^* \text{Ric}(\f)+dd^c(\k \r^\m) \\
& = & (r+2) c_1 (\L,\htil) + c_1 (\text{det}(E),g) \\
& & - \p^* \text{Ric}(\f)+dd^c(\k \r^\m).\end{aligned}$$ From (1.2) and from the hypothesis that $g$ have order $\m$ we conclude that there is $\k_o >0$ such that for all $\k \ge \k_o$ we have $$\begin{aligned}
\tag{5.14}
c_1 (\L_\k)-\text{Ric}(\ft) \ \ge \ \ft.\end{aligned}$$
The abstract de Rham theorem tells us that the cohomology of a sheaf can be computed by taking an acyclic resolution. We recall that a sheaf is said to be *acyclic* if all its cohomology groups, beside $H^0$, vanish. Therefore, in order to show that the higher cohomology of $\O_\m (E^*)$ vanishes, we will construct an acyclic resolution of this sheaf which is exact at the level of global sections. We define the sheaves $\A_\m^{0,q}(E^*)$ as follows: at each $x\in X$ the stalk $\A_\m^{0,q}(E^*)_x$ is the space of germs of smooth $\L$-valued (0,q)-forms defined on $\p^{-1}(U)$, where $U$ is an open neighbourhood of $x$ in $X$. If $x\in \Xb \setminus X$ is a point at infinity, the stalk $\A_\m^{0,q}(E^*)_x$ is the space of germs of smooth $\L$-valued (0,q)-forms defined on $\p^{-1}(U\cap X)$, with $U$ some open neighbourhood of $x$ in $\Xb$, such that both $u$ and $\deb u$ have order $\m$ in the L$^2$-sense. This means that there is $\k >0$ such that $$\begin{aligned}
\int_{\p^{-1}(U\cap X)} |u|^2 \cdot e^{-\k \r^\m} \Phit
\ < \ \infty, \quad \quad
\int_{\p^{-1}(U\cap X)} |\deb u|^2 \cdot e^{-\k \r^\m} \Phit
\ < \ \infty.\end{aligned}$$ Here $|\cdot |$ is taken with respect to $\ltil'$. Clearly $\A_\m^{0,q}(E^*),\quad 0 \le q \le n+r$, are sheaves on $\Xb$. In fact, they are modules over the sheaf $\A_{\Xb}$ of smooth $\C$-valued functions on $\Xb$. As such they are soft, because any $\A_{\Xb}$-module is a soft sheaf. This is due to the fact that we can find smooth partitions of the unity on $\Xb$. We recall that a sheaf is said to be *soft* if any section over a closed subset can be extended to a global section. One knows that soft sheaves are acyclic. Notice that we have a complex $$\begin{aligned}
\tag{5.15} 0 \lra \O_\m (E^*) \lra \A_\m^{0,0}(E^*)
\stackrel{\deb}{\lra} \A_\m^{0,1} (E^*)
\stackrel{\deb}{\lra} \ldots \stackrel{\deb}{\lra}
\A_\m^{0,q} (E^*) \stackrel{\deb}{\lra} \ldots\end{aligned}$$ which is clearly exact at the level of $\A_\m^{0,0}(E^*)$. Thus the theorem will be proven once we manage to establish that
1. the complex (5.15) is exact,
2. the complex (5.15) is exact at the level of global sections.
We begin with the latter. Choose $u\in \A_\m^{0,q} (\Xb,E^*)$ such that $\deb u=0$. There is $\k \ge \k_o$ so large that $u$ be square integrable with respect to the metric of $\L_\k$. But (5.14) tells us that $\L_\k$ satisfies the hypotheses of theorem (5.4). We can find a smooth $v\in \text{L}^{0,q-1}(Y,\L_\k)$ such that $\deb v=u$. By definition $v\in \A_\m^{0,q-1}(\Xb,E^*)$. This proves (ii).
We now turn to (i). Exactness of a complex of sheaves means exactness at the level of stalks. Fix an arbitrary point $x\in \Xb$ and a germ $u_x \in \A_\m^{0,q}(E^*)_x$ represented by some $u\in \A_\m^{0,q} (U,E^*)$. Here $U$ is a small open neighbourhood of $x$ in $\Xb$. Our aim is to find $v\in \A_\m^{0,q} (W,E^*)$, defined over a possibly smaller neighbourhood $W$ of $x$, such that $\deb v=u$ on $\p^{-1}(W\cap X)$.
Let us choose a Stein neighbourhood of $x$, say an open polycylinder $P\subset U$. After possibly shrinking $P$ we may assume that there is $\k \ge \k_o$ such that $u$ is square integrable relative to $\ltil_\k$, i.e. $u\in \text{L}^{0,q}(\p^{-1}(P\cap X), \L_\k)$. We will fix this $\k$ for the remainder of this proof. However, we cannot yet use (5.4) because the metric $\ft$ on $\p^{-1}(U\cap X)$ is not complete. We correct this by adding a horizontal term to $\ft$.
Let $\chi$ be a strictly plurisubharmonic exhaustive function of $P$. Such a function exists because $P$ is Stein. We put $Z=\p^{-1}(P\cap X)$. We claim that the (1,1)-form $$\begin{aligned}
\ft_Z = \ft + dd^c (\chi \circ \p)\end{aligned}$$ determines a complete metric on $Z$. To justify this we need to show that any ball $B$ in $Z$ is relatively compact. Here $B$ is a ball relative to the geodesic distance induced by $\ft_Z$. Since $\ft_Z$ dominates $\ft$ it is clear that $B$ is included in a ball relative to the geodesic distance induced by $\ft$. Thus $\p (B)$ stays away from the divisor at infinity $\Xb \setminus X$. Also $\p (B)$ is included in a ball relative to the geodesic distance induced by $dd^c\chi$ on $P$. But $dd^c \chi$ induces a complete metric on $P$ because $\chi$ is exhaustive. Thus $\p (B)$ stays away from the boundary of $P$. We conclude that the closure of $B$ in $Z$ is compact, which justifies the claim.
Let $\kt_Z$ be the metric on $K_Z$ induced by $\ft_Z$. It is easy to see that there are smooth functions $\a, \b :P\lra (0,\infty)$ such that on $P\cap X$ we have $$\begin{aligned}
\begin{array}{rclcl}
\a \r^\l \f & \le & \frac{\k \l}{2} \r^\l \f + dd^c \chi,\\
& & 2 \k \r^{\l +c}\f + dd^c \chi & \le &
\b \r^{\l +c}\f.
\end{array}\end{aligned}$$ From this and (4.10) we get $$\begin{aligned}
\a \r^\l \f \ \le \
-\Th(v) + dd^c(\k \r^\l) + dd^c \chi
\ \le \ \b \r^{\l +c}\f.\end{aligned}$$ From this we obtain the analogue of (4.11): $$\begin{aligned}
\o + \a \r^\l \f \ \le \ \ft_Z \ \le \ \o + \b \r^{\l +c}\f.\end{aligned}$$ This can be used to get the analogue of (5.13): $$\begin{aligned}
\tag{5.16}
\b^{-n} \r^{-n\l -nc} \kt' \le \kt_Z \le
\a^{-n} \r^{-n\l} \kt'.\end{aligned}$$
Under the isomorphism $\L_{|Z}\isom \L_{|Z} \tensor K_Z^{-1}
\tensor K_Z$ we put a metric $\ltil_Z'$ on $\L_{|Z}$ by $$\begin{aligned}
\ltil_Z' = \ltil \cdot (\kt')^{-1} \cdot \kt_Z.\end{aligned}$$ By (5.16) the two metrics $\ltil$ and $\ltil_Z'$ on $\L_Z$ satisfy $$\begin{aligned}
\tag{5.17}
\a^n \r^{n\l} \ltil'_Z \ \le \ \ltil \ \le \ \b^n \r^{n\l +nc}
\ltil_Z'.\end{aligned}$$ The metric $\ltil_Z'$ has the advantage that we can make it “satisfy” the hypothesis of theorem (5.4): let $\g :P\lra (0,\infty)$ be a strictly plurisubharmonic exhaustive function and let $\L_\g$ be the line bundle $\L_Z$ equipped with the metric $$\begin{aligned}
\ltil_\g = \ltil_Z' \cdot e^{-\k \r^\m}\cdot e^{-\g}.\end{aligned}$$ We have $$\begin{aligned}
c_1 (\L_\g) -\text{Ric}(\ft_Z)
& = & (r+2) c_1(\L,\htil) + c_1 (\text{det}(E),g) \\
& & -\p^* \text{Ric}(\f) + dd^c (\k \r^\m) + dd^c \g \\
& \ge & \ft + dd^c \g\end{aligned}$$ for $\k \ge \k_o$ so that (5.14) hold. Hence for any $\g$ growing faster than $\chi$ we will have $$\begin{aligned}
\tag{5.18}
c_1 (\L_\g) - \text{Ric}(\ft_Z) \ \ge \ \ft_Z.\end{aligned}$$
Let now $Q$ be a polycylinder containing $x$ and with $\overline{Q} \subset P$. From (5.17) and the fact that $u$ is square integrable with respect to the metrics $\ltil \cdot
e^{-\k \r^\m}$ on $\L$ and $\ft$ on $Z$ we see that $u$ is square integrable on $\p^{-1}(Q \cap X)$ relative to the metrics $\ltil_Z' \cdot e^{-\k \r^\m}$ on $\L$ and $\ft_Z$ on the manifold.
Choosing a sequence of polycylinders $\{ Q\}$ which exhaust $P$ we prove that we can choose $\g$ growing so fast that $u$ be square integrable on $Z$ relative to the metric $\ltil_\g$ on $\L$ and $\ft_Z$ on $Z$. In other words $u$ belonds to L$^{0,q}(Z,\L_\g)$. The hypotheses of theorem (5.4) are fulfilled: the metric on $Z$ is complete and Kähler and $\L_\g$ satisfies (5.18). We conclude that there is a smooth $v\in \text{L}^{0,q-1}(Z,\L_\g)$ such that $\deb v=u$.
Finally, with $Q$ as above, we see that $v$ is square integrable on $\p^{-1}(Q\cap X)$ relative to the metrics $\ltil \cdot e^{-\k \r^\m}$ on $\L$ and $\ft$ on the manifold. This forces $v\in \A_\m^{0,q-1} (Q,E^*)$. The proofs of (i) and of the theorem are finished.
\
[**(5.19) Corollary:**]{} *Let $E$ satisfy the conditions from theorem (5.12). Let $\F$ be a coherent algebraic sheaf on $\overline{X}$ which is flat at all points at infinity $x\in \overline{X} \setminus X$. Then, for $q\ge 1$,* $$\begin{aligned}
H^q (\overline{X}, \F \tensor_{\O_{\overline{X}}} \O_\m (E^*))=0.\end{aligned}$$
*Proof:* By Hilbert’s syzygy theorem $\F$ has a finite resolution $$\begin{aligned}
0\lra \O_{\overline{X}}^{a_n} \stackrel{\a_n}{\lra} \O_{\overline{X}}^{a_{n-1}}
\stackrel{\a_{n-1}}{\lra} \ldots
\lra \O_{\overline{X}}^{a_1} \stackrel{\a_1}{\lra} \F \lra 0.\end{aligned}$$ We claim that tensoring the above with $\O_\m (E^*)$ we obtain a resolution for the sheaf $\F \tensor_{\O_{\overline{X}}} \O_\m (E^*)$. Indeed, the complex $$\begin{aligned}
\tag{5.20}
0\lra \O_{\overline{X}}^{a_n}\tensor \O_\m (E^*)
\lra \O_{\overline{X}}^{a_{n-1}}
\tensor \O_\m(E^*) \lra \ldots \lra \F \tensor \O_\m(E^*) \lra 0\end{aligned}$$ is exact at every point $x\in X$ by the mere fact that $\O_\m (E^*)_x \simeq
\O^r_{\overline{X},x}$ is a free, hence flat, $\O_{\overline{X},x}$-module. When $x\in \overline{X} \setminus X$ is a point at infinity the hypothesis that $\F_x$ be flat and standard arguments in homological algebra ensure that ${\mathcal Ker}(\a_i)_x$ is flat for all $1\le i \le n$. Assembling the exact sequences $$\begin{aligned}
0= \text{Tor}_1^{\O_{\overline{X},x}}
({\mathcal Ker}(\a_{i-1})_x ,\ \O_\m (E^*)_x) \lra
{\mathcal Ker}(\a_i)_x \tensor \O_\m (E^*)_x \lra \\
\O^{a_i}_{\overline{X},x} \tensor \O_\m (E^*)_x
\lra {\mathcal Ker}(\a_{i-1})_x \tensor \O_\m (E^*)_x \lra 0\end{aligned}$$ we conclude that (5.20) is exact. By the previous theorem we have the vanishment of cohomology $$\begin{aligned}
H^q (\Xb, \O_{\Xb}^{a_i} \tensor \O_\m(E^*))=0,\quad q\ge 1.\end{aligned}$$ The corollary now follows from standard long exact sequences in cohomology induced by (5.20).
\
[**(5.21) Corollary:**]{} *Assume that $E$ satisfies the conditions from theorem (5.12). Then $E^*$ is spanned at every point by global sections of order $\m$. More precisely, given $x_1, \ldots, x_N \in X$ and $e_1 \in E^*_{x_1},
\ldots, e_N \in E^*_{x_N}$ there is a global section $\s$ of $E^*$ of order $\m$ such that $\s (x_i)=e_i$ for $1 \le i \le N$.*
\
*Proof:* Let $\I \subset \O_{\overline{X}}$ be the ideal sheaf of $\{ x_1, \ldots, x_N \}$. It is a coherent algebraic sheaf on $\overline{X}$ which is flat at all $x \in \overline{X} \setminus X$. In fact, it is locally free there. By (5.19) we have $$\begin{aligned}
\tag{5.22}
H^1 (\overline{X}, \I \tensor_{\O_{\overline{X}}} \O_\m (E^*)) = 0.\end{aligned}$$ Tensoring the exact sequence $$\begin{aligned}
0 \lra \I \lra \O_{\overline{X}} \lra \O_{\{ x_1, \ldots, x_N \} } \lra 0\end{aligned}$$ with $\O_\m (E^*)$ we obtain a sequence $$\begin{aligned}
0 \lra \I \tensor \O_\m(E^*) \lra \O_\m(E^*) \stackrel{\a}{\lra}
\O(E^*)_{\{ x_1, \ldots, x_N \} }
\lra 0\end{aligned}$$ which is exact by the same discussion as in the proof of (5.19). Its long exact sequence in cohomology together with (5.22) show that $\a$ is surjective at the level of global sections. This finishes the proof because $\O(E^*)_{\{ x_1, \ldots, x_N \} }$ is a skyscraper sheaf with stalks $E^*_{x_i}$ at $x_i$ and zero outside $\{ x_1,\ldots,x_N\}$.
\
[**(5.23) Proposition:**]{} *Assume that $E$ satisfies the conditions from theorem (5.12). Then there are finitely many global sections $\s_1, \ldots, \s_N$ of $E^*$ of order $\m$ which span $E^*$ at every point. Moreover, these sections can be chosen in such a manner that for each $x\in X$ their differentials at $x$, written relatively to a trivialization of $E$ at $x$, span $E_x^* \tensor$ T$_x^* X$.*
\
*Proof:* We repeat here the arguments from §11 in [@griffiths-cornalba]. By (5.21) there exist linearly independent global sections $\s_1, \ldots, \s_r$ of $E^*$ of order $\m$. In this proof $r$ is the rank of $E$. Then $\s_1 \wedge \ldots
\wedge \s_r$ is a nontrivial section of det$(E^*)$. Its zero-set $Z$ is a proper analytic subset of $X$. We choose points $\{ x_i \}_{i \ge 1}$ on the relative interiors of each component of $Z$. By (11.6) in [@griffiths-cornalba] we can find a global section $\s_{r+1}$ of $E^*$ of order $\m$ such that $\s_{r+1}(x_i)
\notin$ span$\{ \s_1 (x_i), \ldots, \s_{r} (x_i) \}$ for all $i$.
Let us make this explicit: given a diverging sequence of points $\{ x_i\}_{i\ge 1}$ in $X$ and subspaces $F_i \subsetneqq E^*_{x_i}$, there exists a global section $\eta$ of $E^*$ of order $\m$ such that $\eta (x_i) \notin F_i$ for all $i$. To see this choose $e_i \in E^*_{x_i} \setminus F_i$ and global sections $\eta_i$ of $E^*$ satisfying: $$\begin{aligned}
\eta_i (x_j)= 0\ \text{for}\ j<i,\quad \eta_i (x_i) = e_i,\quad
\int_{\P(E)} |\tilde{\eta}_i |^2 \cdot e^{-\k \r^\m} \Phit
\ < \ \infty.\end{aligned}$$ Such sections exist by (5.21). The constant $\k$ can be assumed to be the same for all $i$ because the Vanishing Theorem and its corollary (5.21) hold with fixed large $\k$. We skip the details. Inductively on $i$ choose nonzero constants $c_i$ such that $$\begin{aligned}
\sum_{j\le i} c_j \eta_j (x_i) \notin F_i, \quad
\int_{\P(E)} |c_i \tilde{\eta}_i |^2 \cdot
e^{-\k \r^\m} \Phit \ < \ 2^{-i}, \quad
|c_i \eta_i | < 2^{-i}\ \ \text{on} \ K_i.\end{aligned}$$ Here $K_i,\ i\geq 1$, is an exhaustion of $X$ by compact subsets. The norm $|\cdot |$ is taken relative to $\ltil$. Clearly, $$\begin{aligned}
\eta =\sum_{i\ge 1} c_i \eta_i\end{aligned}$$ is well defined and satisfies our requirements.
Repeating this procedure we construct sections $\s_{r+1},\ldots,
\s_{2r}$ of $E^*$ of order $\m$ such that $\s_1, \ldots, \s_{2r}$ span $E^*$ at all points $x_i$. Thus $\s_1, \ldots, \s_{2r}$ span $E^*$ at all points outside an analytic subset included in $Z$ and of strictly smaller dimension. Repeating this procedure we can make $Z$ to be empty proving the first part of the proposition.
What we have proven is that for all $x\in X$ the images of $\s_1, \ldots,
\s_N$ inside the vector space $\O(E^*)_x \tensor \O_x / \mm_x$ span.
We turn now to the second part of the proposition. We notice that there is no canonical map $E^*\lra E^* \tensor $ T$^* X$. However, saying that $\s_1,\ldots, \s_N$ span $E^*$ at $x$ and $d \s_1, \ldots, d \s_N$ span $E^*_x \tensor$ T$_x^* X$ is equivalent to saying that the images of $\s_1, \ldots, \s_N$ inside $\O(E^*)_x \tensor \O_x / \mm_x^2$ span. Indeed, we have a non-cannonical decomposition (depending on the coordinate system) $$\begin{aligned}
\O_x / \mm_x^2 = \C \oplus \mm_x / \mm_x^2 = \C \oplus \text{T}_x^* X.\end{aligned}$$ To prove the second assertion we need only repeat the arguments from above with $E^*_{x_i}$ replaced by $\O(E^*)_{x_i} \tensor \O_{x_i} / \mm_{x_i}^2$.\
\
An Immersion
============
The sections $\s_0,\ldots \s_N$ from (5.23) are associated in a canonical fashion to sections $\st_0,\ldots,\st_N$ of $\L$. The latter span $\L$ at every point, cf. (6.1). Therefore they induce a holomorphic map $$\begin{aligned}
f:\P(E)\lra \P^N,\quad f(y)=(\st_0(y);\ldots;\st_N(y)).\end{aligned}$$ In the next section we will show that $f$ is of finite order. Our goal in the present section is to show that the sections from (5.23) can be chosen in such a manner that $f$ be an immersion. A holomorphic map $f$ is said to be an *immersion* if it is one-to-one and its Jacobian has maximal rank at every point. Equivalently, $f$ is an immersion if its image is a complex submanifold and $f$ is a biholomorphism onto its image. We note that the image of $f$ does not have to be closed and, in fact, in our situation, it will not be closed.
We begin by noting at (6.1) that already $f$ is a local immersion. However, to get the injectivity of $f$, the arguments from (5.23) are not sufficient. We will, instead, use arguments from the theory of Stein spaces, more precisely, the classical way of proving that a Stein manifold of dimension $n$ can be embedded into $\C^{2n+1}$. Our reference will be 5.3 in [@hormander]. Hörmander’s lemmas can be applied, quite generally, to a line bundle whose sections separate points and provide holomorphic coordinates at each point. We have checked at (5.23) that sections of finite order of $\L$ satisfy these two properties. One then uses the Baire category theorem to show that the set of sections giving a “good” map $f$ is dense in a certain topology. We will not be able to apply directly the category theorem because the topology on the space of sections of finite order of $\L$ is not complete. However, Baire’s method of proof still works, cf. (6.5).
\
[**(6.1) Remark:**]{} The sections $\st_0, \ldots, \st_N$ span $\L$ at every point, i.e. for every $y\in \P(E)$ some $\st_i (y)$ is nonzero. Indeed, if $y= (x,[v]), \ v\in E_x \setminus \{ 0\}$, then $<\st_i, v> =
<\s_i (x), v>$. But $\s_0 (x), \ldots, \s_N (x)$ span $E_x$, hence some $<\st_i, v>$ must be nonzero.
Thus, as already said above, there is an induced holomorphic map $$\begin{aligned}
f= (\st_0, \ldots, \st_N): \P(E) \lra \P^N.\end{aligned}$$ We claim that $f$ is a local immersion, i.e. for all $y\in \P(E)$ the Jacobian of $f$ at $y$ has maximal rank $n+r$. The accepted terminology is also “$f$ is regular at $y$”.
\
[**(6.2) Remark:**]{} Let $s_0, \ldots, s_m$ be global sections of $\L$ inducing a map $f$ from $Y$ to $\P^m$. Then $f$ is regular at a point $y\in Y$ if and only if the images of $s_0, \ldots, s_m$ under the canonical map $\L_y \lra \L_y \tensor \O_y / \mm_y^2$ generate the latter as a vector space.\
Let us denote by $\Si$ the space of global sections $\s \in \G (Y,\L)$ of order $\m$ in the L$^2$-sense: $$\begin{aligned}
||\s ||^2 := \int_{\P(E)} |\s |^2_{\ltil} \cdot e^{-\k \r^\m} \Phit \ < \
\infty.\end{aligned}$$ We choose $\k$ and $\m$ so large that the Vanishing Theorem (5.12) hold and its corollary (5.21) apply to $\Si$. In addition, we choose $\k$ so large that the sections $\st_0, \ldots, \st_N$ from (5.23) belong to $\Si$. We will fix these $\k$ and $\m$ for the remainder of this section. For the proofs of the following two lemmas we refer to [@hormander] (5.3).
\
[**(6.3) Lemma:**]{} *Let $K\subset Y$ be a compact subset. Then we can find an integer $m$ and sections $s_0, \ldots, s_m \in \Si$ inducing a map $f:Y\lra \P^m$ which regular and one-to-one on $K$.*
\
[**(6.4) Lemma:**]{} *Let $K\subset Y$ be a compact subset. Assume that some global sections $s_0, \ldots, s_{m+1}$ of $\L$ induce a map from $Y$ to $\P^{m+1}$ which is regular and one-to-one on $K$. Then, if $m\ge 2(n+r)+1$, we can find $(a_0, \ldots, a_m)\in \C^{m+1}$ arbitrarily close to the origin such that $s_0 - a_0 s_{m+1}, \ldots, s_m - a_m s_{m+1}$ induce a map $f:Y \lra \P^m$ which is regular and one-to-one on $K$. In fact, this is true for all $a\in \C^{m+1}$ outside a set of measure zero.*\
We are nearing our goal. The last step is to put a topology on $\Si$ and to show that the set of sections giving a regular one-to-one map from $Y$ to projective space is dense in $\Si^m$ equipped with the product topology. Let $\{ K_p \}_{p\ge 1}$ be an exhaustion of $Y$ by compact subsets. We equip $\Si$ with the topology of convergence on compact subsets: a sequence $\{ \s_q \}_{q\ge 1}$ converges to $\s$ if for each compact subset $K$ of $Y$ we have $$\begin{aligned}
\lim_{q\ra \infty} |\s_q - \s |_K = 0,\end{aligned}$$ where $$\begin{aligned}
|\s_q - \s |_K =\ \text{sup}\{ |\s_q (y) - \s (y)|_{\ltil},\ y\in K \}.\end{aligned}$$ This topology is given by a metric invariant under translations: $$\begin{aligned}
d(\s', \s'') = \sum_{p\ge 1} \frac{1}{2^p} \cdot
\frac{|\s' - \s'' |_{K_p}}{1+|\s' - \s''|_{K_p}}.\end{aligned}$$ It is not clear whether $(\Si, d)$ is a complete metric space; it is not clear that all Cauchy sequences in $\Si$ converge. However, Cauchy sequences which are bounded in $||\cdot ||$ are convergent. Indeed, assume that $\{ \s_q \}_{q\ge 1}$ is Cauchy relative to the distance $d$ and $||\s_q || \le M$ for some $M>0$ and all $q \ge 1$. There is a global section $\s$ of $\L$ such that $\{ \s_q \}_{q\ge 1}$ converges pointwise in $|\cdot |_{\ltil}$ to $\s$, uniformly on compact sets. We have $$\begin{aligned}
\int_{K_p} |\s_q |^2_{\ltil} \cdot e^{-\k \r^\m} \Phit \ < \ ||\s_q ||^2
\ \le \ M^2.\end{aligned}$$ Taking limit as $q\ra \infty$ we get $$\begin{aligned}
\int_{K_p} |\s |^2_{\ltil} \cdot e^{-\k \r^\m} \Phit \ \le \ M^2.\end{aligned}$$ Taking limit as $p\ra \infty$ we obtain $||\s || \le M$ forcing $\s \in \Si$. Thus $\{ \s_q \}_{q\ge 1}$ converges to $\s$ in $(\Si, d)$.
\
[**(6.5) Theorem:**]{} *Let $m\ge 2(n+r)+1$ be a given integer. Then the set of (m+1)-tuples $(\s_0, \ldots, \s_m) \in \Si^{m+1}$ which induce a regular injective map $f: Y\lra \P^m$ is dense in $\Si^{m+1}$.*
\
*Proof:* Let $G_p$ denote the set of (m+1)-tuples $(\s_0,
\ldots, \s_m) \in \Si^{m+1}$ which induce a map from $Y$ to $\P^m$ that is regular and one-to-one on $K_p$. Clearly $G_p$ is open. We claim that $G_p$ is dense. To see this choose an arbitrary (m+1)-tuple $\s = (\s_0, \ldots ,\s_m) \in \Si^{m+1}$. Lemma (6.3) provides us with sections $s_0, \ldots, s_N \in \Si$ which induce a map from $Y$ to $\P^N$ that is regular and one-to-one on $K_p$. Then $(\s_0, \ldots,
\s_m, s_0, \ldots, s_N)$ induce a map from $Y$ to $\P^{m+N+1}$ that is regular and one-to-one on $K_p$. Applying repeatedly lemma (6.4) to this (m+N+1)-tuple we deduce that we can find $a_{ij},\ 0\le i \le m,
\ 0\le j\le N$ arbitrarily close to the origin such that $\s_i' := \s_i - \sum_{j=0}^N a_{ij} s_j,\ \ 0\le i \le m$, induce a map from $Y$ to $\P^m$ that is regular and one-to-one on $K_p$. Thus $\s'$ belongs to $G_p$ and it can be made arbitrarily close to $\s$ relative to the distance $d$. We conclude that $G_p$ is dense.
Denoting $||\s || =$ max$\{ ||\s_j ||,\ 0\le j \le m \}$ we also notice that $\s'$ can be made arbitrarily close to $\s$ relative to $||\cdot ||$.
We claim that $\cap_{p\ge 1} G_p$ is dense in $\Si^{m+1}$. Had we known that $(\Si,d)$ is complete this would have been guaranteed by Baire’s category theorem. Nevertheless, in our situation, the arguments from the proof of Baire’s theorem can be carried out:
Choose arbitrary $\s \in \Si^{m+1},\ \varepsilon >0$. Choose $\s_1 \in G_1$ such that $d(\s, \s_1 ) < \varepsilon$ and $||\s - \s_1 || < 1$. Choose $\varepsilon_1 \in
(0, \varepsilon / 2)$ such that the closed ball $B [\s_1, 2\varepsilon_1 ]$ is contained in $G_1$. Inductively choose $\s_p \in G_1 \cap \ldots \cap G_p$ and $\varepsilon_{p+1} \in
(0, \varepsilon_p / 2)$ such that $d(\s_p, \s_{p+1}) <
\varepsilon_p,\ \ ||\s_{p+1}-\s_p||
< 2^{-p}$ and $B[\s_p, 2\varepsilon_p] \subset G_1 \cap
\ldots \cap G_p$. Then $\{ \s_p \}_{p\ge 1}$ is a Cauchy sequence in $\Si^{m+1}$. Also $||\s_p || \le ||\s || +2$ for all $p$. By the discussion preceding the theorem $\{ \s_p \}_{p\ge 1}$ converges to some $\s'
\in \Si^{m+1}$. For $q>p$ we have $d(\s_p, \s_q)\ < \ \varepsilon_p + \ldots +
\varepsilon_{q-1} \ < \ 2\varepsilon_p$. Taking limit as $q\ra \infty$ we obtain $d(\s_p, \s') \le 2\varepsilon_p$. By construction the closed ball of center $\s_p$ and radius $2\varepsilon_p$ is contained in $G_1 \cap \ldots \cap G_p$. Thus $\s' \in G_1 \cap \ldots \cap G_p$. Since $p$ is arbitrary we conclude that $\s' \in \cap_{p\ge 1} G_p$. We have $d(\s, \s') \le 2\varepsilon$. Since $\varepsilon > 0$ is arbitrary we conclude that $\cap_{p\ge 1} G_p$ is dense in $\Si^{m+1}$. Q.e.d.\
Nevanlinna Theory on $\P(E)$
============================
Let $X$ be a special affine variety of dimension $n$ and $E$ a holomorphic vector bundle of rank $r+1$ equipped with a Finsler metric of order $\l$ as defined at (5.5). Our goal is to provide means for measuring growth of holomorphic maps from $\P(E)$ to projective space or growth of analytic subsets of $\P(E)$. We will develop a theory along the same lines as in section 2.
For brevity we will write $Y=\P(E)$. We also write $\r, \f,\psi$ etc. instead of $\r \circ \p, \p^* \f, \p^* \psi$. Given $r\in (0,\infty)$ we put $$\begin{aligned}
Y[r]= \{ (x,[v])\in Y,\ \r'(x)\le r\},\quad
Y<r>= \{ (x,[v])\in Y,\ \r'(x)=r\}.\end{aligned}$$ By Sard’s theorem the sets $Y<r>$ are smooth for all $r$ outside a set of measure zero. In the sequel, each time we integrate over $Y<r>$, it will be tacitly assumed that the latter is smooth.
We will consider $\P(E)$ equipped with the Kähler metric induced by $\ft$. Our first observation is that the volume of the balls $Y[r]$ grows at most polynomially in $r$. Indeed, $$\begin{aligned}
\text{vol}(Y[r]) = \int_{Y[r]} \ft^{n+r}
& = & \int_{Y[r]} {n+r \choose r} (\ft^\H)^n
\wedge (\ft^\V)^r \\
& \le & \int_{Y[r]} {n+r \choose r}(2\k \r^{\l +c})^n \f^n
\wedge (\ft^\V)^r \quad \quad
\text{by (5.8).}\end{aligned}$$ But $\f^n$ is a volume form on $X$. Thus we can apply Fubini’s theorem: the integral from above can be computed by first integrating vertically along the fibers of $\P(E)$ and then horizontally along $X$. We get $$\begin{aligned}
\text{vol}(Y[r]) & \le & \int_{x\in X[r]} {n+r \choose n}
(2\k \r^{\l +c})^n \f^n
\int_{\P(E_x)} (\ft^\V_{|\P(E_x)})^r \\
& = & \int_{x\in X[r]} {n+r \choose n}
(2\k \r^{\l +c})^n \f^n \ \le \ r^a\end{aligned}$$ for some positive constant $a$. The above expression has polynomial growth because of (1.5) and of the fact that $\f$ has finite volume.
Given a holomorphic map $f:Y \lra \P^N$ we define its *characteristic function* $$\begin{aligned}
T_f(r,s)= \int_s^r \frac{dt}{t}
\int_{Y[t]} f^* \o \wedge \psi^{n-1}\wedge \ft^r,\end{aligned}$$ and, for $1\le k\le n$, the *higher characteristic functions* $$\begin{aligned}
T^{(k)}_f(r,s)= \int_s^r \frac{dt}{t}
\int_{Y[t]} f^* \o^k \wedge \psi^{n-k}\wedge \ft^r.\end{aligned}$$ Here $r>s>0$ are real numbers and $\o$ is the Fubini-Study form on $\P^N$. Given an analytic subset $Z\subset Y$ of pure dimension $k\ge r$ we define its *counting function* $$\begin{aligned}
N_Z(r,s)= \int_s^r \frac{dt}{t} \int_{Z[t]} \psi^{k-r}\wedge \ft^r,\end{aligned}$$ where $Z[t]=Z\cap Y[t]$. If the image of $Z$ under the projection $\pp \circ \p :Y\lra \C^n$ does not contain the origin then $N_Z (r,0)$ is well defined and we write $N_Z (r)=N_Z (r,0)$. Given a global section $\s$ of $\O_{\P^N}(1)$ we define the *proximity function* of $f$ to the zero-set of $\s$ $$\begin{aligned}
m_\s(r)= \int_{Y<r>} \text{log}\frac{1}{|\s \circ f|} d^c \t'
\wedge \psi^{n-1} \wedge \ft^r.\end{aligned}$$ Here $|\cdot |$ is taken relative to the canonical metric of $\O_{\P^N}(1)$. Notice that $d^c \t' \wedge \psi^{n-1} \wedge \ft^r$ is a volume form on $Y<r>$. By choosing $\s$ to have norm less than 1 at all points we can arrange that $m_\s (r)$ be non-negative.
The First Main Theorem and the Crofton Formula hold also in this context with virtually the same proofs as in the classical case:
\
[**(7.1) Theorem:**]{} *Let $f:\P(E)\lra \P^N$ be a holomorphic map. Let $H\subset \P^N$ be a hyperplane defined as the zero-set of a global section $\s$ of $\O_{\P^N}(1)$ and $Z=f^*H$. Assume that the image of $f$ is not contained in $H$. Then, for $r>s>0$, we have* $$\begin{aligned}
N_Z(r,s)+m_\s(r)-m_\s(s)=T_f(r,s).\end{aligned}$$
\
[**(7.2) Corollary:**]{} *Fix $s>0$. Then, for $r>s$, and under the hypotheses of the previous theorem, we have* $$\begin{aligned}
N_Z(r,s)\le T_f(r,s) +O(1).\end{aligned}$$ *Here $O(1)$ is a constant that may depend on $s$.*
\
[**(7.3) Crofton Formula:**]{} *Let $f:\P(E)\lra \P^N$ be a holomorphic map which is non-degenerate in the sense that the preimage of a plane $P$ of codimension $k$ in $\P^N$ is an analytic subset of pure codimension $k$ in $\P(E)$. Then, for $r>s>0$, we have* $$\begin{aligned}
T_f^{(k)} (r,s)= \int_{P\in G(N,k)} N_{f^* P} (r,s) d\m (P).\end{aligned}$$ *Here $G(N,k)$ is the Grassmannian of planes of codimension $k$ in $\P^N$ while $\m$ is the measure defined before (2.3).*
\
[**(7.4) Definition:**]{} Let $\l$ be a non-negative real number. Let $f: \P(E) \lra \P^N$ be a holomorphic map. We say that *$f$ has order $\l$* if there is $\k \ge 0$ such that for some $s$ and all $r>s$ we have $$\begin{aligned}
T_f (r,s) \le \k r^\l.\end{aligned}$$ Likewise, we say that an analytic subset $Z\subset \P(E)$ of pure dimension $k\ge r$ *has order $\l$* if $$\begin{aligned}
N_Z (r,s) \le \k r^\l.\end{aligned}$$
[**(7.5) Remark:**]{} Assume that $f$ has finite order and is linearly non-degenerate. Then (7.2) tells us that the preimage under $f$ of any hyperplane $H\subset \P(E)$ is a divisor of finite order. Conversely, assume that the pull-backs $f^{*}(H)$ have finite order in a uniform manner, i.e. there are $r_o, \k, \l \ge 0$ such that for all $r \ge r_o$ and $s < r$ we have $N_{f^{*}(H)}(r,s)
\le \k r^\l$. Then, by (7.3), $f$ has finite order, too.\
Before proceeding further let us notice that for an analytic subset $Z\subset \P(E)$ of pure dimension $k \ge r$ we can define a counting function $$\begin{aligned}
\hat{N}_Z (r,s) = \int_s^r \frac{dt}{t} \int_{Z[t]} \f^{k-r} \wedge
\ft^r\end{aligned}$$ by replacing $\psi$ with $\f$. One can see that $\hat{N}_Z$ and $N_Z$ have roughly the same growth, cf. [@maican]. What we mean is that there are constants $a,b \ge 1$ independent of $Z$ such that $$\begin{aligned}
N_Z (r) \le r^a \hat{N}_Z (r^b),\qquad \hat{N}_Z (r) \le r^a N_Z (r^b).\end{aligned}$$ In particular, both counting functions have polynomial growth at the same time. The reader may wonder why we didn’t use $\hat{N}_Z$ to measure growth in the first place. The problem with this is that it doesn’t satisfy the First Main Theorem.
\
[**(7.6) Remark:**]{} The theory we build in this section encompasses the theory we built in section 2. What we claim is the following: let $Z\subset X$ be a pure k-dimensional analytic subset and $\Zt = \p^{-1}Z$. Then $Z$ and $\Zt$ have roughly the same growth. In particular, both have finite order at the same time. Indeed, by Wirtinger’s theorem the restriction of $\f^k$ to $Z$ is a volume form on $Z$. Thus we can apply Fubini’s theorem in the definition of $\hat{N}_{\Zt}$: the integral of $\f^k \wedge
\ft^r$ on $\Zt$ can be computed by first integrating vertically along the fibers of $\Zt$ and then horizontally along $Z$. We get $\hat{N}_{\Zt} (r,s)= \hat{N}_Z (r,s)$. From §2 we know that $\hat{N}_Z$ has roughly the same growth at $N_Z$. From the comments preceding this remark we know that $\hat{N}_{\Zt}$ has roughly the same growth at $N_{\Zt}$. This finishes the proof of the claim.
\
[**(7.7) Proposition:**]{} *Assume that $E$ is equipped with a Finsler metric of order $\l$. Let $\s$ be a global section of $E^*$ of order $\l$ and $Z\subset \P(E)$ the zero-set of $\st$. Then there is $\m \ge \l$ depending only on $\l$ and $E$ such that $Z$ has order $\m$.*
\
*Proof:* Integrating the Poincaré-Lelong formula $$\begin{aligned}
[Z] = c_1 (\L,\ltil) + dd^c \text{log}|\st |^2_{\ltil}\end{aligned}$$ and taking logarithmic average we get $$\begin{aligned}
N_Z (r,s) & = & \int_s^r \frac{dt}{t}
\int_{Z[t]} \psi^{n-1} \wedge \ft^r \\
& = & \int_s^r \frac{dt}{t} \int_{Y[t]}
\big[ c_1 (\L,\ltil)+dd^c \text{log}|\st |^2 \big]
\wedge \psi^{n-1} \wedge \ft^r \\
& = & \int_s^r \frac{dt}{t} \int_{Y[t]}
c_1 (\L,\htil) \wedge \psi^{n-1} \wedge \ft^r \\
& & -\int_s^r \frac{dt}{t} \int_{Y[t]}
c_1 (\p^* \text{det}(E),\text{det}(G_{i\jb}))
\wedge \psi^{n-1} \wedge \ft^r \\
& & +\int_s^r \frac{dt}{t} \int_{Y[t]}
c_1 (\text{det}(E),g) \wedge \psi^{n-1} \wedge \ft^r \\
& & +\int_s^r \frac{dt}{t} \int_{Y[t]}
dd^c \text{log}|\st |^2 \wedge \psi^{n-1} \wedge \ft^r.\end{aligned}$$ Let us denote by (i), (ii), (iii) and (iv) the integrals from the right-hand-side above. First we notice that (i) has polynomial growth: by (5.7) and (1.6) $$\begin{aligned}
(i) & \le & \int_s^r \frac{dt}{t} \int_{Y[t]}
\ft \wedge \psi^{n-1} \wedge \ft^r \\
& \le & \int_s^r \frac{dt}{t} \int_{Y[t]} \r^{c(n-1)} \f^{n-1}
\wedge \ft^{r+1} \\
& \le & \int_s^r \frac{dt}{t} \int_{Y[t]} \r^{c(n-1)} \ft^{n+r} \\
& \le & r^{c(n-1)c'} \text{vol}(Y[r]) \ \le \ r^{a+c(n-1)c'}.\end{aligned}$$ Let $\m$ be the largest between $\l$ and the order of $g$. By hypothesis $|c_1 (\text{det}(E),g)| \le \r^\m \f$ yielding $$\begin{aligned}
(iii) & \le & \int_s^r \frac{dt}{t} \int_{Y[t]}
\r^\m \f \wedge \psi^{n-1} \wedge \ft^r \\
& \le & \int_s^r \frac{dt}{t} \int_{Y[t]} \r^{\m +c(n-1)} \ft^{n+r} \\
& \le & r^{a+\m c'+ c(n-1)c'}.\end{aligned}$$ By type considerations $\psi^{n-1} \wedge (\ft^\H)^2 =0$ forcing $$\begin{aligned}
\psi^{n-1} \wedge \ft^r & = & \psi^{n-1} \wedge (\ft^\H + \ft^\V)^r \\
& = & \psi^{n-1} \wedge (\ft^\V)^r + r\ \psi^{n-1} \wedge \ft^\H
\wedge (\ft^\V)^{r-1},\\
c_1 (G_{i\jb})\wedge \psi^{n-1} \wedge \ft^r
& = & c_1 (G_{i\jb})^\H \wedge \psi^{n-1} \wedge (\ft^\V)^r \\
& & + r\ c_1 (G_{i\jb})^\V \wedge \psi^{n-1} \wedge \ft^\H
\wedge (\ft^\V)^{r-1}.\end{aligned}$$ By (5.9), (5.10) and (5.8) we obtain $$\begin{aligned}
(ii) & \le & \int_s^r \frac{dt}{t} \int_{Y[t]}
\big[ \k \r^\l \f \wedge \psi^{n-1}\wedge (\ft^\V)^r
+ \k \r^{\l} (\ft)^\V \wedge \psi^{n-1} \wedge \ft^\H
\wedge (\ft^\V)^{r-1} \big] \\
& \le & 2\k r^{a+ \l c' + c(n-1)c'}.\end{aligned}$$
Performing the standard “integration twice” procedure from Nevanlinna theory we obtain $$\begin{aligned}
(iv) & = & \int_{Y<r>} \text{log}|\st | \cdot d^c \t' \wedge
\psi^{n-1} \wedge \ft^r \\
& & - \int_{Y<s>} \text{log}|\st | \cdot d^c \t' \wedge
\psi^{n-1} \wedge \ft^r.\end{aligned}$$ We fix $s$. By using the concavity of the logarithmic function we obtain the estimate $$\begin{aligned}
\tag{7.8}
\int_{Y<r>} \text{log}|\st | \cdot d^c \t' \wedge \psi^{n-1}
\wedge \ft^r \ \le \ \text{vol}(r) \cdot \text{log}
\Big\{ \frac{1}{\text{vol(r)}} \int_{Y<r>} |\st | \cdot d^c \t' \wedge
\psi^{n-1} \wedge \ft^r \Big\}\end{aligned}$$ where $$\begin{aligned}
\text{vol}(r) = \int_{Y<r>} d^c \t' \wedge \psi^{n-1} \wedge \ft^r.\end{aligned}$$ But $d^c \t' \wedge \psi^{n-1}$ is a volume form on $X<r>$ away from the branching set of the projection $\pp : X\lra \C^n$. This set has measure zero so we ignore it. Therefore, we can apply Fubini’s theorem on $Y<r>$. We get $$\begin{aligned}
\tag{7.9}
\text{vol}(r) = \int_{X<r>} d^c \t' \wedge \psi^{n-1}
= \int_{\C^n <r>} d^c \text{log} ||z||^2 \wedge
(dd^c \text{log} ||z||^2)^{n-1} = 1.\end{aligned}$$ It remains to estimate the integral $$\begin{aligned}
v(r) = \int_{Y<r>} |\st | \cdot d^c \t' \wedge \psi^{n-1} \wedge \ft^r.\end{aligned}$$ We will first estimate the integral $$\begin{aligned}
w(r)=\int_1^r \frac{v(t)}{t} dt.\end{aligned}$$ We have $$\begin{aligned}
w(r) & = & \int_{Y[1,r]} |\st |
\cdot d\t' \wedge d^c \t' \wedge \psi^{n-1} \wedge \ft^r\\
& & \text{(by Fubini's theorem)} \\
& \le & \int_{Y[1,r]}
|\st |\cdot \r^c \f \wedge (\r^c \f)^{n-1} \wedge \ft^r\\
& & (\text{by (1.6) and (1.7)}) \\
& \le & r^{c n c'} \int_{Y[1,r]} |\st | \cdot \f^n \wedge \ft^r \\
& & \text{(here $c'$ is such that $\t \le c' \t'$, cf. (1.5))} \\
& \le & r^{c n c'} \int_{Y[1,r]} |\st | \cdot \Phit \\
& \le & r^{c n c'}
\Big( \int_{Y[r]} |\st |^2 \cdot e^{-\k \r^\l} \cdot \Phit \Big)^{1/2}
\cdot \Big( \int_{Y[r]} e^{\k \r^\l} \cdot \Phit \Big)^{1/2} \\
& & (\text{by H\"older's Inequality}) \\
& \le & r^{c n c'} C^{1/2} e^{\k r^\l /2} \text{vol}(Y[r])^{1/2}\\
& \le & r^{c n c'} C^{1/2} e^{\k r^\l /2} r^{a/2}.\end{aligned}$$ Here $\k$ is so large that $\s$ be square integrable with respect to the metric $\ltil \cdot e^{-\k' \r^\l}$ on $\L$ and $\ft$ on $\P(E)$. Next we notice that $w(r)$ is an increasing function of class $\mathcal{C}^1$ with $w'(r)= v(r)/r$. Without loss of generality we may assume that $\lim_{r\ra \infty}w(r)= \infty$. We claim that for all $r\ge 1$ outside a set of finite Lebesgue measure we have $$\begin{aligned}
w'(r) \ \le \ w(r)^2.\end{aligned}$$ Indeed, let $B$ be the set of “bad” $r$ for which $w'(r)>w(r)^2$. Let $r_o$ be such that $w(r_o)=1$. For $r\ge r_o$ we have $$\begin{aligned}
\text{measure}(B\cap [1,w(r)]) < \int_1^{w(r)}
\frac{w'(t)}{w(t)^2} dt = \int_{r_o}^r \frac{dt}{t^2} < \frac{1}{r_o}.\end{aligned}$$ By letting $r$ converge to infinity we get that the measure of $B$ is at most $1/r_o$ which proves the claim.
We conclude that for all $r>0$ outside a set of finite Lebesgue measure we have $$\begin{aligned}
v(r) \le r^{1+ 2cnc'+a} C \cdot e^{\k r^\l}.\end{aligned}$$
Finally, taking logarithm and using (7.8) and (7.9) we arrive at the estimate $$\begin{aligned}
(iv) \le O(\text{log}(r)) + \k r^\l.\end{aligned}$$ We deduce the proposition holds with $\m$ depending on the order of $g$, on $\l$ and some constants but with $r$ outside a set of Lebesgue measure $1/r_o$. But $T_f(r)$ is increasing hence $T_f(r) = O((r+1/r_o)^\m)=
O(r^{\m + \ep})$ with arbitrary positive $\ep$. Q.e.d.
\
[**(7.10) Theorem:**]{} *Let $X$ be a special affine variety of dimension $n$. Let $E$ be a holomorphic vector bundle of rank $r+1$ on $X$. Assume that $E$ is equipped with a Finsler metric of order $\l$. Then there is $\m \ge \l$ such that for any integer $N\ge 2(n+r)+1$ there is a holomorphic immersion* $$\begin{aligned}
f: \P(E) \lra \P^N\end{aligned}$$ *of order $\m$ satisfying* $$\begin{aligned}
f^*\O_{\P^N}(1)\isom \L.\end{aligned}$$
*Proof:* Let $(\s_0,\ldots, \s_N)$ be as in (6.5). Let $\m$ be as in (7.7). For any linear combination $\s = a_0 \s_0 + \ldots +
a_N \s_N,\quad |a_i |\le 1$, the zero-set of $\st$ has order $\m$. In fact, going through the estimates of (7.7) one sees that these zero-sets have order $\m$ in a uniform manner. The theorem now follows from remark (7.5).\
\
List of Notations {#list-of-notations .unnumbered}
=================
--------------------------- ---------------------------------------------------------------------------
$X$ special affine variety of dimension $n$
$\O_{\fo}$ sheaf of holomorphic functions on $X$ of finite order, cf. (2.6)
$\tau,\ \rho=e^{\tau /2}$ strictly plurisubharmonic exhaustive functions on $X$, cf. (1.3)
$\t',\ \r'=e^{\t' /2}$ plurisubharmonic exhaustive functions on $X$
$d^c$ the twisted derivative $\frac{1}{4\pi \i}(\de - \deb)$
$dd^c$ the complex Hessian (also known as Levi form) $\frac{1}{2\pi \i}\de \deb$
$\f = dd^c \t$ Kähler form on $X$
$\Phi$ the volume-form $\f^n$ of $\f$
$\psi = dd^c \tau'$ semipositive-definite form on $X$ satisfying $\psi^n=0$
$\mathfrak{p}$ generic projection from $X$ to $\mathbb{C}^n$
$E$ holomorphic vector bundle of rank $r+1$ on $X$
$E^*$ the dual of $E$
$\O_{\fo}(E^*)$ sheaf of holomorphic sections of finite order of $E^*$, cf. (5.11)
$\mathbb{P}(E)$ projectivization of $E$
$E_o$ complement of the zero-section in $E$
$\V$ the vertical tangent bundle inside T$E_o$
$\H$ the horizontal tangent bundle inside T$E_o$
$p$ projection from $E$ onto the base $X$
$\pi$ projection from $\mathbb{P}(E)$ onto the base $X$
$\mathcal{L}$ the hyperplane line bundle $\mathcal{O}_{\mathbb{P}(E)}(1)$
$\st$ section of $\L$ corresponding to section $\s$ of $E^*$, cf. (4.7)
$g$ metric of finite order on det$(E)$
$h$ Hermitian or Finsler metric on $E$
$\tilde{h}$ Hermitian metric on $\mathcal{L}$ induced by $h$
$(G_{i\jb})$ Hermitian metric on $\V$ induced by $h$
$\ltil$ the Hermitian metric $\htil \cdot \text{det}(G_{i\jb})^{-1}
\cdot \p^* g$ on $\L$
$\ft$ Kähler form on $\P(E)$, cf. (5.7)
$\Phit$ the volume-form $\ft^{n+r}$ of $\ft$
$\ft^\V, \ft^\H$ the vertical, resp. the horizontal part of $\ft$,
cf. paragraph after (3.23)
--------------------------- ---------------------------------------------------------------------------
[30]{}
Marco Abate and Giorgio Patrizio. *Fisler Metrics- A Global Approach*, Lecture Notes in Mathematics [**1591**]{}, Springer Verlag, Berlin, 1994.
Aldo Andreotti and Eduardo Vesentini, *Carleman Estimates for the Laplace-Beltrami Equation on Complex Manifolds*, Publ. Math. I.H.E.S. [**25**]{} (1965), 81-130.
Jianguo Cao and Pit-Mann Wong, *Geometry of Projectivized Vector Bundles*, Journal of Math. of Kyoto Univ. [**43**]{}, no. 2, (2003), 369-410.
James Carlson and Phillip Griffiths, *The Order Function for Entire Holomorphic Mappings*, Value Distribution Theory, Part A, R. O. Kujala, A. L. Vitter (eds.), Marcel Dekker Inc., New York, 1974.
Maurizio Cornalba and Phillip Griffiths, *Analytic Cycles and Vector Bundles on Non-compact Algebraic Varieties*, Invent. Math. [**28**]{} (1975), 1-106.
Maurizio Cornalba and Bernard Shiffman, *A Counterexample to the Transcendental Bezôut Problem*, Ann. of Math. [**96**]{} (1972), 402-406.
Phillip Griffiths, *Function Theory of Finite Order on Algebraic Varieties*, Journal of Diff. Geom. [**6**]{} (1972), 285-306 and [**7**]{} (1972), 45-66.
Lars Hörmander, *L$^2$-estimates and Existence Theorems for the $\deb$-operator*, Acta. Math. [**113**]{} (1965), 89-52.
Lars Hörmander, *An Introduction to Complex Analysis in Several Variables*, North-Holland Publ. Comp., Amsterdam-London, 1973.
Kunihiko Kodaira and James Morrow, *Complex Manifolds*, Holt, Rinehart and Winston, Inc., 1971.
Pierre Lelong and Lawrence Gruman, *Entire Functions of Several Complex Variables*, Springer Verlag, Berlin, 1986.
Mario Maican, *Vector Bundles of Finite Order on Affine Manifolds*, Thesis at the University of Notre Dame, July 2005.
James Mulflur, Albert Vitter, and Pit-Mann Wong, *Holomorphic Functions of Finite Order on Affine Varieties*, Duke Math. Journal [**48**]{}, no. 2, (1981), 389-399.
Boris Shabat, *Distribution of Values of Holomorphic Mappings*, Translations of Mathematical Monographs [**61**]{}, American Mathematical Society, Providence, RI, 1985.
B. Shiffman, A. Sommese, *Vanishing Theorems on Complex Manifolds*, Progress in Mathematics [**56**]{}, Birkhäuser Boston, Inc., 1985.
Henri Skoda, *Solution a la croissance du second problème de Cousin dans $\C^n$*, Ann. Inst. Fourier, Grenoble [**21**]{}, no. 1, (1971), 11-23.
Henri Skoda, *Sous-ensembles analytiques d’ordre fini ou infini dans $\C^n$*, Bull. Soc. Math. France [**100**]{} (1972), 353-408.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Two, replica symmetry breaking specific, quantities of the Ising spin glass — the breakpoint $x_1$ of the order parameter function and the Almeida-Thouless line — are calculated in six dimensions (the upper critical dimension of the replicated field theory used), and also below and above it. The results comfirm that replica symmetry breaking does exist below $d=6$, and also the tendency of its escalation for decreasing dimension continues. As a new feature, $x_1$ has a nonzero and universal value for $d<6$ at criticality. Near six dimensions we have $x_{1c}=3\,(6-d)+O[(6-d)^2]$. A method to expand a generic theory with replica equivalence around the replica symmetric one is also demonstrated.'
author:
- 'G. Parisi'
- 'T. Temesvári'
title: Replica symmetry breaking in and around six dimensions
---
Introduction
============
Frustration in disordered systems gives rise to a complex equilibrium state with a nontrivial breaking of ergodicity (see [@MePaVi] for a review and important reprints of the field). In the mean field version of the Ising spin glass [@SK], the decomposition of the Gibbs state into ultrametrically organized pure states is (mathematically) encoded in the replica symmetry broken (RSB) solution of the replicated system [@MePaVi]. This solution has characteristics — such as the order parameter [*function*]{} $q(x)$, and the spin glass transition in nonzero external magnetic field along the so called Almeida-Thouless (AT) line — which fully distinguish it from the much simpler replica symmetric (RS) case. This RS solution is unstable in the mean field glassy phase [@AT].
From the physical point of view, RSB implies the presence of violations of nontrivial fluctuation-dissipation relations at off-equilibrium (during aging), while the off-equilibrium fluctuation-dissipation relations would be trivial in the RS case: in particular no aging of the response function is expected then, in variance with the experimental evidence in three dimensions at zero magnetic field. It is a very important task to determine the dimensional regime where the low temperature phase with aging response function survives. Evidently, there is no glassy phase in the one-dimensional system, whereas there is an ample numerical evidence against any transition in the two-dimensional case too. Generally speaking, we expect that the transitions disappear at the corresponding lower critical dimensions, i.e at $d_{SG}^{0}$ in zero magnetic field, and at $d_{SG}^{h}$ in the presence of a magnetic field. We cannot say a priori if these two lower critical dimensions are the same: in the case of an Ising ferromagnet with a random magnetic field, for instance, it is well known that $d_{IF}^{0}=1$, whereas $d_{IF}^{h}=0$. The situation in spin glasses is quite unclear: the different structure of the low momentum singularities in zero and nonzero magnetic field [@Gaussian_propagators] suggest that $d_{SG}^{0}<d_{SG}^{h}$, while the arguments based on domain wall energies give $d_{SG}^{0}=d_{SG}^{h}=2.5$ [@Franz_Parisi_Virasoro]. The existence of a low temperature phase with aging response function should be ultimately decided by investigating the structure of infrared divergences in the perturbative expansion, and by the analysis of nonperturbative contributions. This task goes by far beyond the goals of the present paper. We aim to study in details the properties of the low temperature phase near the critical temperature, and around the upper critical dimension (i.e. six) where the critical exponents at zero magnetic field become nontrivial. Our study also aims to correct some recent claims on the nonexistence of a RSB phase below six dimensions that are due to an incorrect analysis of the consequences of some renormalization group equations [@Moore_Bray_2011].
The mean field Ising spin glass, at least when studied with the replica trick, can be considered as the infinite-dimensional limit of the replica field theory representing the $d$-dimensional short ranged model defined on a hypercubic lattice [@AT2008]. The study of this replica field theory for decreasing dimensionalities seems to be a good strategy for reaching a full understanding of the three-dimensional Ising spin glass.
This project has had by now a long history whose first period was summarized in Ref. [@beyond]. It turns out from these studies that the RS glassy phase is notoriously unstable even down to $d\lesssim 6$, with a persistently escalating RSB phase (see, for instance, Fig. 1 of Ref. [@nucl]). A scaling picture was proposed in [@scaling_and_infrared] for helping to understand one-loop calculations in the (zero external magnetic field) RSB phase. Some of the results of this reference are reproduced and/or revised in the present paper, especially the behaviour of the breakpoint $x_1$ of $q(x)$ around six dimensions. The AT line was first found in Ref. [@GrMoBr83] for the range $6<d<8$, whereas it was followed up from mean field ($d=\infty$) to $d\lesssim 6$ (and also for nonzero replica number $n$) in [@AT2008].
Nevertheless, the RS spin glass phase has remained an alternative due to the so called droplet model [@FiHu86; @FiHu88; @BrMo86]. This theory predicts a unique Gibbs state (apart from spin inversion) for $T<T_c$ — that is why the replicated theory is RS — which is massless, and the glassy phase is unstable for any infinitesimal magnetic field, i.e. there is no AT line. A schematic picture of the two scenarios on the temperature-magnetic field plane is presented in Fig. \[RSB\_vs\_droplet\]. The phase boundary lies along the temperature axis in the droplet case, a zero-temperature fixed point governing its behaviour; the analogous attractive — and also zero-temperature — fixed point for the RSB scenario is shifted to a nonzero external field $h_c$. The other end of the phase boundary is, in both cases, the zero-field critical fixed point at $T_c$. Since the symmetry of the transition line — namely, an RS state with nonzero order parameter $q$, which is massless in the so called replicon sector, while massive in the longitudinal one — is the same (notwithstanding the fact that the AT line proceeds in nonzero magnetic field), the two renormalization group (RG) pictures can be studied in a common field theory. This is the generic replica symmetric field theory elaborated in Refs. [@rscikk; @nucl]. The vicinity of the (hypothetical) zero temperature fixed point can be studied in this field theory by assuming a hard (practically infinite) longitudinal mass, thus projecting the theory into the replicon sector. This was done decades ago by Bray and Roberts [@BrRo], who found a stable Gaussian fixed point for $d>6$, whereas it was impossible to find any physically relevant and stable fixed point for $d<6$. This was later interpreted [@BrMo86; @Moore_Bray_2011] as a sign that the AT line disappears below six dimensions, and the droplet scenario takes over. This is, however, a faulty argument, since — as we have explained above — the RG equations (those for instance of Ref. [@BrRo]) are not specific to the low temperature behaviour of the AT line. An effort to understand the crossover from the zero-field critical fixed point to the zero-temperature one was made in Ref. [@Iveta], where the whole set of RG equations was derived in a first order perturbative renormalization. (The Bray-Roberts equations are naturally included there.) The runaway flows found were discussed in details in [@AT2008], and it was argued in this reference that the RG scheme used could not be expected to detect a zero-temperature fixed point in epsilon expansion. But again, the lack of a fixed point with infinite longitudinal mass [*in the RG equations valid around the critical point*]{} is not specific to spin glasses, and this property cannot distinguish between the two rival spin glass theories.
![Schematic phase diagrams for a $d$-dimensional Ising spin glass in the temperature-magnetic field plane. There is an RSB glassy phase in (a) bordered by the AT line. On the other hand, the glassy phase is RS in (b), and lies in the zero-field subspace. Both the AT line and the zero-field glassy phase are represented by the same generic replica symmetric field theory with massive longitudinal and massless replicon modes. []{data-label="RSB_vs_droplet"}](figure1_2011_paper.eps)
In a recent paper [@Moore_Bray_2011], Moore and Bray suggest a proof that RSB disappears when six dimensions is approached form above. They take the $d\to 6^+$ limit of known first order results, using RG arguments, for $x_1$ (the breakpoint of the order parameter function) and the AT line, and find both going to zero. We reproduce their results in a more complete RG scheme in Sec. \[d>6\], and show what is the fundamental flaw in their argument. At this point, the reader is advised to jump to Fig. \[scaling\_variable\](b) in Sec. \[discussion\] where $x_1$ is plotted against dimension along with the so called scaling variable, which is effectively the relative error of the approximation. The breakpoint $x_1$ is monotonically increasing for decreasing dimension as long as the scaling variable is small. This is the range where the approximation is valid! However at around $d\approx 6.1$, the scaling variable starts to steeply increase (and actually goes to 1 for $d\to 6$), simultanously $x_1$ suddenly changes its behaviour, and falls to zero: this is the effect (and a similar scenario for the AT line) that has been found in [@Moore_Bray_2011], but it must be clear that these results fall outside the range of validity of the approximate RG equations. As a matter of fact, $x_1$ can be calculated directly in $d=6$ (Sec. \[d=6\]), its value is shown as the horizontal line in Fig. \[scaling\_variable\](b): it is visibly an extrapolation of the curve from the range where the approximation is good. (In fact, it is an old wisdom of the RG theories that the upper critical dimension requires special care.) There is only one case where the arguments of Ref. [@Moore_Bray_2011] are correct \[and interestingly enough, this is admitted there below Eq. (18) of that reference\], namely just at criticality. But that yields only the trivial results for the $d=6$ system: $x_1$ is zero for $T=T_c$, and the AT line starts at the origin, i.e. at $T=T_c$ and $h^2=0$, and does not say anything about the disappearence of RSB.[^1]
The outline of the paper is as follows: Section \[d>6\] is devoted to the study of the dimensional regime $6<d<8$, although the perturbative results of subsection \[pert\] are extensively used in later sections too. In Sec. \[d=6\], the renormalization group ideas are specifically applied to the $d=6$ case, simply following the lines explained in classical RG textbooks (see, for instance, [@PfeutyToulouse]). The breakpoint $x_1$ and the AT line are calculated at the upper critical dimension, both displaying logarithmic temperature corrections. A method for expanding a general (except that replica equivalence is assumed) RSB theory around the RS one is presented in Sec. \[RSB\_vs\_RS\], and applied to the ultrametric case. By this method, quantities of the RSB theory, like $x_1$, can be expressed in terms of vertices of the RS theory. In the next section, Sec. \[d<6\], we return to our original program, and study the case $d<6$: generic RG arguments are presented, and the calculation of $x_1$ and the AT line in $\epsilon$-expansion is performed. A new feature emerges below six dimensions, namely $x_1$ becomes nonzero and universal at criticality. In the last section, Sec. \[discussion\], special examples, both for $x_1$ and the AT line, are used to conclude that RSB escalates both in the regime above and below six dimensions.
Formulation of the spin glass problem for $6<d<8$ {#d>6}
=================================================
The simplest replicated field theory corresponding to the Ising spin glass in zero external magnetic field and below $d=8$ has two bare parameters defining the model: $\tau$ (measuring the distance from criticality and $w$ (the only bare cubic coupling compatible with the symmetrical — paramagnetic — state). Its Lagrangian is $$\label{simple_L}
\mathcal{L}=
\frac{1}{2}\sum_{\mathbf p}
\bigg(\frac{1}{2} p^2+\bar m\bigg)\sum_{\alpha\beta}
\phi^{\alpha\beta}_{\mathbf p}\phi^{\alpha\beta}_{-\mathbf p}
-\frac{1}{6N^{1/2}}\,\,w\,\sideset{}{'}\sum_{\mathbf {p_1p_2p_3}}
\sum_{\alpha\beta\gamma}\phi^{\alpha\beta}_{\mathbf p_1}
\phi^{\beta\gamma}_{\mathbf p_2}\phi^{\gamma\alpha}_{\mathbf p_3}$$ where the bare mass $\bar m=\bar m_c-\tau$, and the critical mass has been presented in the literature several times in leading order of the loop expansion: $$\bar m_c=\frac{1}{2}(n-2)w^2\frac{1}{N}\sum_{\mathbf p}\frac{1}{p^4}\quad.$$ In this $n(n-1)/2$ component field theory the fluctuating fields are symmetric in the replica indices with zero diagonals: $\phi^{\alpha\beta}_{\mathbf p}=\phi^{\beta\alpha}_{\mathbf p}$ and $\phi^{\alpha\alpha}_{\mathbf p}=0$, $\alpha$,$\beta=1,\dots,n$. \[Momentum conservation is indicated by the primed summation. The number $N$ of the Ising spins becomes infinite in the thermodynamic limit, rendering summations to integrals over the continuum of momenta in the diagrams of the perturbative expansion. A momentum cutoff $\Lambda$ is always understood to block ultraviolet divergences, although it can be (and will be) absorbed into the definition of different quantities.\] The replica number $n$ goes to zero in the spin glass limit.
Perturbative results {#pert}
--------------------
We are now going to recollect several results for the replica symmetric (RS) spin glass phase — see Refs. [@rscikk; @Iveta; @nucl; @free_energy_fluctuations] — which are needed for the following discussion. Due to the severe technical difficulties, only one-loop calculations have been accomplished ($\epsilon\equiv 6-d$ and $n=0$).
- RS order parameter $q$, i.e. the equation of state: $$\label{q1}
\frac{wq}{\tau}=1-2w^2\,\tau^{|\epsilon|/2}\,\,
\frac{1}{N}\sum_{\mathbf p}^{\frac{\Lambda}{\sqrt{\tau}}}
\frac{p^2-2}{p^4(p^2+2)^2} +\frac{1}{2}w\,\tau^{-2}\,h^2\quad.$$ (The last term with the external magnetic field $h$ has been included here for later reference. At the moment, it is to be considered as zero.) We can use $wq=\tau$ in the one-loop diagrams, and after rescaling the momentum as $p\rightarrow p/\sqrt \tau$, two different propagators remain: the replicon ($p^{-2}$) and the longitudinal \[$(p^2+2)^{-1}$\] ones. To make the formulae for the one-loop vertices more transparent, it is useful to introduce a common notation $I_{\dots}$ for the occuring integrals, as is illustrated below: $$I_{RRLL}\equiv \frac{1}{N}\sum_{\mathbf p}^{\frac{\Lambda}{\sqrt{\tau}}}
\frac{1}{p^4(p^2+2)^2}=\int^{\frac{\Lambda}{\sqrt{\tau}}}\frac{d^dp}{(2\pi)^d}
\frac{1}{p^4(p^2+2)^2}=K_d\int^{\frac{\Lambda}{\sqrt{\tau}}}
\frac{dp\,p^{-1+d}}{p^4(p^2+2)^2}\quad.$$
- The replicon mass: $$\label{GammaR1}
\Gamma_R=2m_1=-2\tau+2wq+4w^2\,\tau^{1+|\epsilon|/2}\,(4I_{RLL}-3I_{RRL})\quad.$$
- The basic cubic vertex of the $\text{Tr}\,\phi^3$ operator: $$\label{w11}
w_1=w+2w^3\,\tau^{|\epsilon|/2}\,(-8I_{RRL}+7I_{RRR}-14I_{RRLL}-8I_{RLLL})\quad.$$
- The quartic vertex of ${\phi^{\alpha\beta}}^4$: $$\label{u21}
u_2=24w^4\,\tau^{-1+|\epsilon|/2}\,I_{RRLL}\quad.$$
In fact, this last result is new. Details of the somewhat lengthy calculation of the replicon-type quartic vertices will be published later.
Simple two-parameter renormalization group
------------------------------------------
An extensive renormalization group (RG) study of the generic RS glassy phase was published in Ref. [@Iveta]. When close to the Gaussian fixed point[^2], i.e. $w\ll 1$ and $\tau\ll 1$, and only infinitesimally breaking the high-temperature (paramagnetic) symmetry of the system, we have the following simple two-parameter RG flow-equations: $$\label{RGflow}
\begin{aligned}
\dot{w^2} &=-|\epsilon|w^2-2w^4\,,\\[2pt]
\dot{\tau} &=\left(2-\frac{10}{3}w^2\right)\tau\, .
\end{aligned}$$ Physical quantities take simple scaling forms when, instead of $w$ and $\tau$, they are expressed in terms of the nonlinear scaling fields $\tilde w$ and $r$ defined by: $$\label{nonlinear}
\begin{aligned}
\dot{\tilde{w}^2\!\!} &=-|\epsilon|{\tilde w}^2\,,\\[2pt]
\dot{r} &=2r\, .
\end{aligned}$$ A straightforward calculation provides: $$\label{solution}
\begin{aligned}
%\tilde{w}^2 &=w^2\left(1+2\frac{w^2}{}
w^2&=\tilde{w}^2\left(1-2\frac{\tilde w^2}{|\epsilon|}\right)^{-1}\,,\\[2pt]
\tau &=r\left(1-2\frac{\tilde w^2}{|\epsilon|}\right)^{-5/3}\, .
\end{aligned}$$ We are now going to compute the quantities $q$, $\Gamma_R$, $w_1$ and $u_2$ by the RG in terms of $\tilde{w}^2$ and $r$. In this way, we can get more general results when approaching dimension six from above as compared with the perturbative computation: now we may have $|\epsilon|\ll w^2\ll 1$, although the scaling variable $\tilde{w}^2r^{|\epsilon|/2}$ must be small: $$\tilde{w}^2r^{|\epsilon|/2}\ll |\epsilon|,\quad\text{even when}\quad
|\epsilon|\ll w^2\,.$$
- The renormalization flow equation for $q$ is $$\label{qRG}
\dot{q}=\left(2+\frac{|\epsilon|}{2}+\frac{\eta_L}{2}\right)q$$ with $\eta_L=\eta_R=-\frac{2}{3}w^2$ in this approximation. It can be solved by using Eqs. (\[nonlinear\]) and (\[solution\]): $$\label{q2}
q=r^{1+\frac{|\epsilon|}{4}}\,\hat q\big(\tilde{w}^2r^{|\epsilon|/2}\big)
\,\left(1-2\frac{\tilde w^2}{|\epsilon|}\right)^{-1/6}\,,$$ and a comparison with (\[q1\]) makes it possible — after some manipulations — to get the leading terms of the scaling function: $$\label{q3}
\hat q(x)=\frac{1}{\sqrt{x}}(1+Cx+\dots),\quad \text{with the constant}\quad
C=2^{1+\frac{|\epsilon|}{2}}\,\Gamma\big(1+\frac{|\epsilon|}{2}\big)
\Gamma\big(1-\frac{|\epsilon|}{2}\big)\left(\frac{1}{|\epsilon|}+1\right)\, .$$
- The replicon mass evolves under renormalization as $$\label{GammaR_flow}
\dot{\Gamma}_R=(2-\eta_R)\,\Gamma_R=\left(2+\frac{2}{3}w^2\right)\Gamma_R\, ,$$ with the solution $$\label{GammaR2}
\Gamma_R=r\,\hat{\Gamma}_R\big(\tilde{w}^2r^{|\epsilon|/2}\big)
\,\left(1-2\frac{\tilde w^2}{|\epsilon|}\right)^{1/3}\,.$$ Substituting $q$ in Eq. (\[GammaR1\]) by $\tau$ from (\[q1\]) provides: $$\label{GammaR3}
\Gamma_R=-16w^2\,\tau^{1+\frac{|\epsilon|}{2}}\,I_{RRLL}+w\tau^{-1}h^2\,.$$ Keeping in mind that (\[GammaR3\]) is valid for $\tilde w^2\approx w^2
\ll |\epsilon|$ and $h^2$ is zero at the moment, it is straightforward to derive the scaling function in Eq. (\[GammaR2\]): $$\label{GammaR4}
\hat{\Gamma}_R(x)=C'x+\dots,\quad \text{with}\quad C'=
-2^{2+\frac{|\epsilon|}{2}}\,\Gamma\big(1+\frac{|\epsilon|}{2}\big)
\Gamma\big(1-\frac{|\epsilon|}{2}\big)\,.$$
- As for $w_1$, we have $$\label{w1RG}
\dot{w_1}=\left(-\frac{|\epsilon|}{2}-\frac{3}{2}\eta_R\right)w_1
=\left(-\frac{|\epsilon|}{2}+w^2\right)w_1$$ and $$\label{w12}
w_1=r^{-\frac{|\epsilon|}{4}}\,\hat w_1\big(\tilde{w}^2r^{|\epsilon|/2}\big)
\,\left(1-2\frac{\tilde w^2}{|\epsilon|}\right)^{1/2}\,.$$ Comparing (\[w11\]) and (\[w12\]) yields $$\label{w13}
\hat w_1(x)=\sqrt{x}\,(1+C''x+\dots),\quad \text{with}\quad C''=
2^{\frac{|\epsilon|}{2}}\,\Gamma\big(1+\frac{|\epsilon|}{2}\big)
\Gamma\big(1-\frac{|\epsilon|}{2}\big)\,
\left(\frac{16}{|\epsilon|}-9-|\epsilon|\right)\, .$$
- Finally, from the flow $$\label{u2RG}
\dot{u_2}=\big(-2-|\epsilon|-2\eta_R\big)\,u_2=
\left(-2-|\epsilon|+\frac{4}{3}w^2\right)u_2$$ follows the scaling form of the most important quartic vertex: $$\label{u22}
u_2=r^{-1-\frac{|\epsilon|}{2}}\,\hat u_2\big(\tilde{w}^2r^{|\epsilon|/2}\big)
\,\left(1-2\frac{\tilde w^2}{|\epsilon|}\right)^{2/3}\,.$$ From (\[u21\]) and (\[u22\]) results \[see also (\[GammaR4\])\] $$\label{u23}
\hat u_2(x)=-\frac{3}{2}C'x^2+\dots\,.$$
The calculation of $x_1$ and the Almeida–Thouless line
------------------------------------------------------
The leading contribution to the breakpoint of the order parameter function $q(x)$ is derived in Sec. (\[RSB\_vs\_RS\]), and has the simple form \[see (\[x1\_basic\]) and the more general considerations in that section about getting $x_1$ on the basis of the generic RS field theory\]: $$x_1=\frac{u_2}{w_1}\,q\,.$$ Inserting (\[q2\]), (\[w12\]) and (\[u22\]), the scaling equation of $x_1$ follows: $$x_1=\hat x_1\big(\tilde{w}^2r^{|\epsilon|/2}\big),\quad \text{with}
\quad \hat x_1(\dots)=\frac{\hat u_2(\dots)}{\hat w_1(\dots)}
\,\hat q(\dots)\,.$$ By the help of Eqs. (\[q3\]), (\[w13\]), (\[u23\]) and (\[GammaR4\]), we can conclude $$\label{x1_d>6}
x_1=6\times 2^{\frac{|\epsilon|}{2}}\,\Gamma\big(1+\frac{|\epsilon|}{2}\big)
\Gamma\big(1-\frac{|\epsilon|}{2}\big)\,
\,\tilde{w}^2r^{|\epsilon|/2}+\dots\,.$$ Inverting (\[solution\]), $x_1$ can be expressed by the original bare coupling $w$: $$\label{x1_Moore}
x_1\sim \frac{w^2}{1+2\frac{w^2}{|\epsilon|}}\,r^{|\epsilon|/2}\,.$$ This equation agrees with Eq. (21) of Ref. [@Moore_Bray_2011].[^3] The range of applicability of the above equation: $$\label{range}
w^2,r\ll 1,\quad 0<|\epsilon|<2\quad\text{and (most importantly)}\quad
\tilde{w}^2r^{|\epsilon|/2}\ll |\epsilon|\,.$$ If we fix the system’s bare coupling $w$ and approach six dimensions, then $\tilde w^2\to |\epsilon|/2$ and $x_1\sim |\epsilon|\,r^{|\epsilon|/2}$. This behaviour was interpreted by the authors of Ref. [@Moore_Bray_2011] as the sign of the end of RSB at six dimensions: a vanishing $x_1$ is consistent with RS. But, as Eq. (\[range\]) clearly shows, in this limit $r$ must go to zero,[^4] i.e. the breakpoint disappears at the critical surface in six dimensions — a property valid also for $d>6$ (but, as we will see later, not for $d<6$).[^5] In the next section we will show that below the critical surface $x_1>0$ and has a logarithmic temperature dependence at exactly six dimensions. We now turn to the problem of the Almeida–Thouless line. The introduction of a magnetic field $h^2$ involves a new nonlinear scaling field $\tilde{h^2\!}$ with $$\dot{\tilde{h^2\!}}=\left(4+\frac{|\epsilon|}{2}\right)\,\tilde{h^2\!}\,.$$ Eq. (\[GammaR2\]) remains valid, but the scaling function $\hat \Gamma_R$ has now two arguments: $x=\tilde{w}^2r^{|\epsilon|/2}$ and $y=\tilde{h^2\!}\,\,r^{-2-|\epsilon|/4}$. Realizing that the replicon mass starts at one-loop order, the bare parameters in (\[GammaR3\]) can be replaced by their corresponding nonlinear scaling fields, making it possible to read off the scaling function: $$\hat \Gamma_R(x,y)=C'x+\sqrt{x}\,y\,;$$ see also (\[GammaR4\]). The vanishing replicon mass defines the AT line, i.e. $y=-C'\,\sqrt{x}$ providing $$\label{AT}
\tilde{h^2\!}=-C'\,\tilde w \,r^{2+\frac{|\epsilon|}{2}}\,.$$ The connection between $h^2$ and $\tilde{h^2\!}$ may be found from the flow equation $$\label{h^2_flow}
\dot{h^2}=\left(4+\frac{|\epsilon|}{2}-\frac{\eta_L}{2}\right)\,h^2=
\left(4+\frac{|\epsilon|}{2}+\frac{1}{3}w^2\right)\,h^2\,,$$ with the solution \[see also (\[nonlinear\]) and (\[solution\])\]: $$\label{h_solution}
h^2=\tilde{h^2\!}\,\left(1-2\frac{\tilde w^2}{|\epsilon|}\right)^{1/6}\,.$$ It is useful to display the AT line (\[AT\]) in the original bare parameters by Eqs. (\[solution\]) and (\[h\_solution\]): $$\label{AT_d>6}
h^2=-C'\, \frac{w}{\left(1+2\frac{w^2}{|\epsilon|}\right)
^{4+\frac{5}{6}|\epsilon|}}
\,\tau ^{2+\frac{|\epsilon|}{2}}\,.$$ This equation is identical with Eq. (15) of Ref.[@Moore_Bray_2011], and the $|\epsilon|^4$ factor, arising when $|\epsilon|\to 0$ while fixing $w$, led those authors to conclude that the AT line disappears in six dimensions. But, again, Eq. (\[range\]) and the discussion below it shows that this limit provides results only on the critical surface ($\tau$ and $r$ zero), and it informs us only about the trivial fact that the AT line starts at the origin of the $\tau,h^2$ plain.
At the upper critical dimension: $d=6$ {#d=6}
======================================
As can be seen from the previous section, knowledge about the six dimensional system cannot be gained from the RG results in the $d\gtrapprox 6$ case. The fundamental reason for that is the impossibility to linearize the RG flow equations at exactly an upper critical dimension. Therefore, the scaling field $\tilde w$ is not defined for $d=6$, and we keep $w$ (although $r$ and $\tilde{h^2\!}$ are still meaningful). The RG flow (\[RGflow\]) is now: $$\label{RGflow_d=6}
\begin{aligned}
\dot{w^2} &=-2w^4\,,\\[2pt]
\dot{\tau} &=\left(2-\frac{10}{3}w^2\right)\tau\, .
\end{aligned}$$ The connection between $\tau$ and $r$ becomes \[instead of (\[solution\])\]: $$\label{tau_vs_r}
\tau=r\,w^{\frac{10}{3}}\,,$$ and the scaling variable with zero scaling dimension is now (instead of $\tilde{w}^2\,r^{|\epsilon|/2}$): $$\frac{w^2}{1-w^2\ln r}\,,$$ which can be easily checked by Eq. (\[RGflow\_d=6\]) and the nonlinear scaling field property $\dot r=2r$.
The calculation of $x_1$
------------------------
The renormalization group flow equations for the three relevant physical quantities $q$, $w_1$ and $u_2$ are as follows: $$\begin{gathered}
\dot q= \left(2-\frac{1}{3}w^2\right)\,q\,,\\[4pt]
\dot{w_1}= w^2\,w_1\,,\\[4pt]
\dot{u_2}=\left(-2+\frac{4}{3}w^2\right)\,u_2\,. \end{gathered}$$ They all have the same form, and their solutions are easily found in scaling form.
- The RS order parameter: $$\label{q_d=6}
\begin{gathered}
q=w^{\frac{1}{3}}\, r\,\hat q\left(\frac{w^2}{1-w^2\ln r}\right)\,,\\[4pt]
\hat q(x)=x\,\left[1+(2+\ln 2)\,x+\frac{5}{3}\,x\ln x+\dots\right]\,.
\end{gathered}$$ The scaling function $\hat q(x)$ has been obtained by evaluating (\[q1\]) in $d=6$ (in zero magnetic field at the moment) and using the connection between $\tau$ and $r$ in (\[tau\_vs\_r\]).
- The cubic vertex $w_1$ in six dimensions: $$\label{w1_d=6}
\begin{gathered}
w_1=w^{-1}\,\hat w_1\left(\frac{w^2}{1-w^2\ln r}\right)\,,\\[4pt]
\hat w_1=x\,\left[1+\Big(-\frac{39}{2}+8\ln 2-7\ln n\Big)\,x
+\frac{5}{3}\,x\ln x+\dots\right]\,.
\end{gathered}$$ Eqs. (\[w11\]) and (\[tau\_vs\_r\]) has been used to get the scaling function. One important remark is appropriate here. The term with the logarithm of the replica number, $\ln n$, comes from $I_{RRR}$ in (\[w11\]), and is a prominent example of the severe infrared divergences caused by the replicon propagator. Similar contributions enter in higher order vertices, such as $I_{RRRR}$ in the quartic vertex belonging to the operator $\text{Tr}\,\phi^4$. This is a clear indication — beside the instability of the replicon mode — that the replica symmetric theory is ill-defined in the spin glass limit. In fact, these infrared divergent terms can be resummed when we build up the RSB theory on the basis of the RS one, as explained in Sec. \[RSB\_vs\_RS\]. What is gained in this resummation, after setting $n$ to zero, is the small mass regime of the RSB solution which effectively acts as an infrared cutoff. It must be stressed that without this resummation, the theory is infrared divergent in any arbitrarily high dimension.
- As for the quartic vertex $u_2$, its scaling form and the leading term of the scaling function are \[see (\[u21\]) and (\[tau\_vs\_r\])\]: $$\label{u2_d=6}
\begin{gathered}
u_2=w^{-\frac{4}{3}}\,r^{-1}\,\hat u_2\left(\frac{w^2}{1-w^2\ln r}\right)\,,\\[4pt]
\hat u_2(x)=6x+\dots\,.
\end{gathered}$$
By Eqs. (\[q\_d=6\]), (\[w1\_d=6\]) and (\[u2\_d=6\]) $x_1$ turns out to be a function of the scaling variable, as it must be: $$x_1=\frac{u_2}{w_1}\,q=\hat x_1\left(\frac{w^2}{1-w^2\ln r}\right)
\quad\text{with}\quad \hat x_1(\dots)=\frac{\hat u_2(\dots)}{\hat w_1(\dots)}
\,\hat q(\dots)\,.$$ The leading order of the scaling function is simply $\hat x_1(x)=6x+\dots$, providing one of our basic results $$\label{x1_d=6}
x_1=6\,\left(\frac{w^2}{1-w^2\ln r}\right)+\dots\,;\quad w,r\ll 1\quad \text{and}
\quad r=\tau\,w^{-\frac{10}{3}},\quad d=6\,.$$ It is clear from the above equation that $x_1$ is zero at criticality ($r=\tau=0$), and for fixed $w$ the approach to zero is logarithmic: $$x_1=6\,|\ln r|^{-1}+\dots\,;\quad r,\tau\to 0\quad \text{and}\quad w\quad\text{fixed},\quad d=6\,.$$
Almeida–Thouless line in six dimensions
---------------------------------------
The flow equation for the replicon mass is unchanged as compared with the $d>6$ case, and is given by Eq. (\[GammaR\_flow\]). The nonlinear scaling field corresponding to the external magnetic field satisfies $\dot{\tilde{h^2\!}}=4\,\tilde{h^2\!}$, therefore the second variable with zero scaling dimension is $\tilde{h^2\!}/r^2$. Straightforward considerations lead us to $$\label{GammaR_d=6}
\Gamma_R=w^{-\frac{2}{3}}\, r\,\hat \Gamma_R\left(\frac{w^2}{1-w^2\ln r}\,\,,\,
\frac{\tilde{h^2\!}}{r^2}\right)\,.$$ The evolution of the “bare” magnetic field, i.e. $\dot{h^2}=\left(4+\frac{1}{3}w^2\right)\,h^2$ \[see (\[h\^2\_flow\])\] and (\[RGflow\_d=6\]) yield $$\label{h_vs_h}
h^2=\tilde{h^2\!}\,\,w^{-\frac{1}{3}}\,.$$ Evaluating Eq. (\[GammaR3\]) at $d=6$, and replacing the bare parameters $\tau$ and $h^2$ by $r$ and $\tilde{h^2\!}$ according to (\[tau\_vs\_r\]) and (\[h\_vs\_h\]), respectively, makes it possible to read off the scaling function in leading order: $$\hat \Gamma_R(x,y)=\frac{1}{x}\,(-4x^4+y+\dots)\,.$$ From its zero, the AT line is obtained as follows: $$\label{AT_d=6}
\tilde{h^2\!}=4\,r^2\,\left(\frac{w^2}{1-w^2\ln r}\right)^4
+\dots\,;\quad w,r\ll 1\quad \text{and}
\quad r=\tau\,w^{-\frac{10}{3}},\quad \tilde{h^2\!}=h^2\,w^{\frac{1}{3}};\quad d=6\,.$$ For a given cubic coupling $w$, the magnetic field vs. temperature relationship for the boundary of the RS phase when approaching the critical point becomes: $$\tilde{h^2\!}=4\,r^2\,|\ln r|^{-4}+\dots\,;
\quad r,\tau\to 0\quad \text{and}\quad w\quad\text{fixed},\quad d=6\,.$$
Formulation of replica symmetry breaking on the basis of the generic replica symmetric theory {#RSB_vs_RS}
=============================================================================================
The considerations in this section are quite general and, therefore, the paramagnetic system (i.e. an RS system with zero order parameter) must be represented — instead of the simple case of (\[simple\_L\]) which is sufficient around $d=6$ — by a model which includes all the invariants compatible with its higher symmetry [@droplet]. The replicated field theory is now defined by the Lagrangian $\mathcal L$ of the symmetrical (high-temperature and zero-field) theory: $$\begin{gathered}
\label{L}
\mathcal{L}=
\frac{1}{2}\sum_{\mathbf p}
\bigg(\frac{1}{2} p^2+\bar m_1\bigg)\sum_{\alpha\beta}
\phi^{\alpha\beta}_{\mathbf p}\phi^{\alpha\beta}_{-\mathbf p}
-\frac{1}{6N^{1/2}}\,\sideset{}{'}\sum_{\mathbf {p_1p_2p_3}}
\bar w_1\sum_{\alpha\beta\gamma}\phi^{\alpha\beta}_{\mathbf p_1}
\phi^{\beta\gamma}_{\mathbf p_2}\phi^{\gamma\alpha}_{\mathbf p_3}
-\frac{1}{24N}\,\sideset{}{'}\sum_{\mathbf {p_1p_2p_3p_4}}\\[2pt]
\bigg(\bar u_1\!\!\sum_{\alpha\beta\gamma\delta}\phi^{\alpha\beta}_{\mathbf p_1}
\phi^{\beta\gamma}_{\mathbf p_2}\phi^{\gamma\delta}_{\mathbf p_3}
\phi^{\delta\alpha}_{\mathbf p_4}+\bar u_2\!\sum_{\alpha\beta}
\phi^{\alpha\beta}_{\mathbf p_1}\phi^{\alpha\beta}_{\mathbf p_2}
\phi^{\alpha\beta}_{\mathbf p_3}\phi^{\alpha\beta}_{\mathbf p_4}
+\bar u_3\!\sum_{\alpha\beta\gamma}\phi^{\alpha\gamma}_{\mathbf p_1}
\phi^{\alpha\gamma}_{\mathbf p_2}\phi^{\beta\gamma}_{\mathbf p_3}
\phi^{\beta\gamma}_{\mathbf p_4}+
\bar u_4\!\!\sum_{\alpha\beta\gamma\delta}
\phi^{\alpha\beta}_{\mathbf p_1}\phi^{\alpha\beta}_{\mathbf p_2}
\phi^{\gamma\delta}_{\mathbf p_3}\phi^{\gamma\delta}_{\mathbf p_4}
\bigg)\\[2pt]
-\frac{1}{120N^{3/2}}\,\sideset{}{'}\sum_{\mathbf {p_1p_2p_3p_4p_5}}
\bigg(\bar v_1\!\!\sum_{\alpha\beta\gamma\delta\mu}\phi^{\alpha\beta}_{\mathbf p_1}
\phi^{\beta\gamma}_{\mathbf p_2}\phi^{\gamma\delta}_{\mathbf p_3}
\phi^{\delta\mu}_{\mathbf p_4}\phi^{\mu\alpha}_{\mathbf p_5}+
\bar v_2\!\sum_{\alpha\beta\gamma}
\phi^{\alpha\beta}_{\mathbf p_1}\phi^{\alpha\beta}_{\mathbf p_2}
\phi^{\alpha\beta}_{\mathbf p_3}\phi^{\alpha\gamma}_{\mathbf p_4}
\phi^{\beta\gamma}_{\mathbf p_5}+\\[2pt]
\bar v_3\!\sum_{\alpha\beta\gamma\delta}\phi^{\alpha\beta}_{\mathbf p_1}
\phi^{\beta\gamma}_{\mathbf p_2}\phi^{\gamma\alpha}_{\mathbf p_3}
\phi^{\gamma\delta}_{\mathbf p_4}\phi^{\gamma\delta}_{\mathbf p_5}+
\bar v_4\!\sum_{\alpha\beta\gamma\mu\nu}\phi^{\alpha\beta}_{\mathbf p_1}
\phi^{\beta\gamma}_{\mathbf p_2}\phi^{\gamma\alpha}_{\mathbf p_3}
\phi^{\mu\nu}_{\mathbf p_4}\phi^{\mu\nu}_{\mathbf p_5}
\bigg)+\dots\end{gathered}$$ where the bare mass $\bar m_1\equiv\bar m=\bar m_c-\tau$, with $\tau$ measuring the distance from criticality, has been also used in (\[simple\_L\]), and $\bar w_1\equiv w$ (momentum conservation is indicated by the primed summations). The fifth order invariants with the $\bar v$ bare couplings were also included here. In what follows, we use the same notation for an [*exact*]{} vertex (e.g. $u_2$) and its corresponding bare coupling ($\bar u_2$), the bar indicating always a bare quantity.
As explained in details in Appendix D of Ref. [@nucl], the generic Legendre-transformed free energy can be expanded around the RS spin glass state with order parameter $q$; see (D.5) of this reference: $$\begin{gathered}
\label{F}
\frac{1}{N}\mathcal F(q_{\alpha\beta})=\\[4pt]
\frac{1}{N} \mathcal F(q)+
\frac{1}{2}\left[ m_1\sum_{\alpha\beta}(q_{\alpha\beta}-q)^2+m_2\sum_{\alpha\beta\gamma}
(q_{\alpha\gamma}-q)(q_{\beta\gamma}-q)+m_3\sum_{\alpha\beta\gamma\delta}
(q_{\alpha\beta}-q)(q_{\gamma\delta}-q)\right] \\[4pt]
-\frac{1}{6}\left[w_1\sum_{\alpha\beta\gamma}(q_{\alpha\beta}-q)(q_{\beta\gamma}-q)
(q_{\gamma\alpha}-q)+w_2\sum_{\alpha\beta}(q_{\alpha\beta}-q)^3+
w_3\sum_{\alpha\beta\gamma}(q_{\alpha\beta}-q)^2(q_{\alpha\gamma}-q)+\dots
\right]\\[4pt]
-\frac{1}{24}\left[u_1\sum_{\alpha\beta\gamma\delta}(q_{\alpha\beta}-q)(q_{\beta\gamma}-q)
(q_{\gamma\delta}-q)(q_{\delta\alpha}-q)
+u_2\sum_{\alpha\beta}(q_{\alpha\beta}-q)^4
%+u_3\sum_{\alpha\beta\gamma}(q_{\alpha\gamma}-q)^2(q_{\beta\gamma}-q)^2
+\dots\right]\\[4pt]
-\frac{1}{120}\left[v_1\sum_{\alpha\beta\gamma\delta\mu}(q_{\alpha\beta}-q)(q_{\beta\gamma}-q)
(q_{\gamma\delta}-q)(q_{\delta\mu}-q)(q_{\mu\alpha}-q)
%+v_2\sum_{\alpha\beta\gamma}(q_{\alpha\beta}-q)^3(q_{\alpha\gamma}-q)
%(q_{\beta\gamma}-q)
+\dots\right]+\dots\quad.\end{gathered}$$ In zero external field $\mathcal F(q_{\alpha\beta})$ has the same symmetry as $\mathcal L$ of Eq. (\[L\]) — which is higher than that of a generic RS system —, even when $T<T_c$, and using this symmetry, a set of equations can be found between the exact vertices of the generic RS theory (see Refs. [@droplet; @nucl]). The most effective way to get the required vertex relationships is demanding that invariants incompatible with the symmetrical theory, e.g. $\sum_{\alpha\beta}q_{\alpha\beta}^3$, must finally disappear from (\[F\]). In this manner, all the vertices of the lower symmetry: $m_2$, $m_3$; $w_2$,…$w_8$; $u_5$,…$u_{23}$; …etc., (see Appendix A of Ref. [@nucl] for the classification of cubic and quartic vertices) and, as a bonus, $m_1$ can be expressed in terms of $w_1$, $u_1$, $u_2$, $u_3$, $u_4$, $v_1$, $v_2$, $v_3$, $v_4$, and higher order symmetrical vertices.[^6] We than have $$\begin{aligned}
m_1&=\frac{1}{2}nw_1\,q+\frac{1}{6}(n^2u_1-2u_2)\,q^2+\frac{1}{24}
n(n^2v_1-2v_2)\,q^3+\dots,\\[3pt]
m_2&=-w_1\,q-\frac{1}{3}(nu_1+u_3)\,q^2+\frac{1}{60}[5n(3n^2-5n+1)v_1+2v_2-4nv_3]
\,q^3+\dots,\\[3pt]
m_3&=-\frac{1}{6}(u_1+2u_4)\,q^2-\frac{1}{60}[5(5n-4)v_1+2v_3+6nv_4]\,q^3+\dots,%\\[3pt]\end{aligned}$$ and furthermore $$\begin{aligned}
w_2&=u_2\,q+\frac{1}{20}nv_2\,q^2+\dots,\\[3pt]
w_3&=u_3\,q+\frac{1}{10}(3v_2+nv_3)\,q^2+\dots,\\[3pt]
w_4&=u_4\,q+\frac{1}{20}(v_3+3nv_4)\,q^2+\dots,\\[3pt]
w_5&=u_1\,q+\frac{1}{20}(5nv_1+4v_3)\,q^2+\dots,\\[3pt]
w_6&=\frac{1}{10}v_3\,q^2+\dots,\\[3pt]
w_7&=\frac{1}{40}(10v_1+12v_4)\,q^2+\dots,\\[3pt]
w_8&=O(q^3).\end{aligned}$$ Of the quartic vertices, only those are listed below which are required up to the order of the present calculation: $$\begin{aligned}
u_5&=\frac{3}{5}v_2\,q+\dots,& u_6&=\frac{2}{5}v_3\,q+\dots,& u_7&=\frac{2}{5}v_4\,q+\dots,\\[3pt]
u_8&=\frac{2}{5}v_2\,q+\dots,& u_{10}&=\frac{1}{5}v_3\,q+\dots,&
u_{11}&=\frac{2}{5}v_3\,q+\dots,\\[3pt]
u_{14}&=\frac{3}{5}v_4\,q+\dots,& u_{16}&=v_1\,q+\dots.&&\end{aligned}$$ By exploiting these expressions, the free energy functional in Eq. (\[F\]) can now be written (omitting an additive term depending only on $q$): $$\begin{gathered}
\frac{1}{N}\mathcal F(q_{\alpha\beta})=\frac{1}{4}Mq\,\sum_{\alpha\beta}q_{\alpha\beta}^2-
\frac{1}{6}W\,\sum_{\alpha\beta\gamma}q_{\alpha\beta}q_{\beta\gamma}q_{\gamma\alpha}
-\frac{1}{24}\Big[(u_1+v_1q+\dots)\,\sum_{\alpha\beta\gamma\delta}
q_{\alpha\beta}q_{\beta\gamma}q_{\gamma\delta}q_{\delta\alpha}\\
+\big(u_2+\frac{2}{5}v_2q+\dots\big)\,\sum_{\alpha\beta}q_{\alpha\beta}^4+
\big(u_3+\frac{3}{5}v_3q+\dots\big)\,\sum_{\alpha\beta\gamma}q_{\alpha\gamma}^2q_{\beta\gamma}^2
+\big(u_4+\frac{3}{5}v_4q+\dots\big)\,\Big(\sum_{\alpha\beta}q_{\alpha\beta}^2\Big)^2\Big]\\
-\frac{1}{120}\Big[(v_1+\dots)\,\sum_{\alpha\beta\gamma\delta\mu}q_{\alpha\beta}q_{\beta\gamma}
q_{\gamma\delta}q_{\delta\mu}q_{\mu\alpha}+(v_2+\dots)\,\sum_{\alpha\beta\gamma}q_{\alpha\beta}^3
q_{\alpha\gamma}q_{\beta\gamma}+(v_3+\dots)\,\sum_{\alpha\beta\gamma\delta}q_{\alpha\beta}
q_{\beta\gamma}q_{\gamma\alpha}q_{\gamma\delta}^2\\
+(v_4+\dots)\,\Big(\sum_{\alpha\beta\gamma}
q_{\alpha\beta}q_{\beta\gamma}q_{\gamma\alpha}\Big)\Big(\sum_{\alpha\beta}q_{\alpha\beta}^2\Big)
\Big]-\dots\quad,\end{gathered}$$ with the following notations for $M$ and $W$: $$\begin{aligned}
M&\equiv (n-2)w_1\\[3pt]
&\!\!\!+\frac{1}{3}[(n^2-3)u_1+u_2+(n-1)\tilde u_3]\,q+
\frac{1}{60}[5(n^3-4)v_1-2(2n-8)v_2+2(n-1)(n+4)\tilde v_3]\,q^2+\dots,\\[3pt]
W&\equiv w_1+u_1\,q+\frac{1}{20}[10v_1-3v_2-(n-1)\tilde v_3]\,q^2+\dots\quad.\end{aligned}$$ The tilded vertices $\tilde u_3\equiv u_3+nu_4$ and $\tilde v_3\equiv v_3+nv_4$ were introduced here; in fact, only these combinations will enter the equation of state.
Stationarity of the free energy functional provides the equation of state: $$\begin{gathered}
\label{basic}
0=Mq\,q_{\alpha\beta}-W\,(q^2)_{\alpha\beta}-\frac{1}{3}\big[(u_1+v_1q+\dots)\,(q^3)_{\alpha\beta}+
(u_2+\frac{2}{5}v_2q+\dots)\,q_{\alpha\beta}^3\\[4pt]
+(\tilde u_3+\frac{3}{5}\tilde v_3q+\dots)\,
(q^2)_{\alpha\alpha}\,q_{\alpha\beta}\big]%\\
-\frac{1}{60}\Big\{5(v_1+\dots)\,(q^4)_{\alpha\beta}+(v_2+\dots)\,\big[3(q^2)_{\alpha\beta}\,q_{\alpha\beta}^2+
\sum_{\gamma}(q_{\alpha\gamma}^3q_{\beta\gamma}+q_{\beta\gamma}^3q_{\alpha\gamma})\big]\\[2pt]+
(\tilde v_3+\dots)\,\big[2(q^3)_{\alpha\alpha}\,q_{\alpha\beta}+3(q^2)_{\alpha\alpha}\,
(q^2)_{\alpha\beta}\big]\Big\}-\dots\quad.\end{gathered}$$ Only replica equivalence was used in the derivation of this equation — $(q^2)_{\alpha\alpha}$, for instance, is independent of the replica number —, otherwise it is quite general: it provides an RSB solution in terms of the RS order parameter $q$ (which measures the distance from criticality now), and of the exact RS vertices. It can equally be used in any regime where some kind of perturbation theory is valid.
We now turn to the case of infinite step, ultrametrically organized RSB. The small parameter making possible a perturbative treatment is $x_1$, the breakpoint of the order parameter function: it is proportional to $q$ in the SK model and for the field theory above 8 dimensions, to $q^{d/2-3}$ between 6 and 8 dimensions, whereas it is of order $\epsilon$ below 6 dimensions. $q(x)$, the order parameter function, has the form:[^7] $$\label{q(x)}
q(x)=q_1\,[r+x_1^2\,\delta\bar q(r)],\qquad \text{with}\qquad r\equiv x/x_1\qquad\text{and}\qquad
\delta\bar q(1)=0.$$
The contributions of the various vertices to Eq. (\[basic\]) are listed below. The definition of the bilinear expression $\{\dots;\dots\}$ used extensively in that list is as follows: $$\{f(r);g(r)\}\equiv f(r)g(1)+f(1)g(r)-f(r)\int_r^1du g(u) -g(r)\int_r^1du f(u)-rf(r)g(r)
-\int_0^rdu f(u)g(u).$$
- $w_1$: $$2(q_1-q)\,r-x_1q_1\,(r-\frac{1}{3}r^3)+2x_1^2(q_1-q)\,\delta\bar q(r)-2x_1^3q_1\,\{r;\delta\bar q(r)\}+O(x_1^4),$$
- $u_1$: $$-(q_1-q)^2\,r+x_1(q_1-q)q_1\,(r-\frac{1}{3}r^3)-\frac{1}{3}x_1^2q_1^2\,(\frac{3}{4}r-\frac{1}{2}r^3+\frac{3}{20}r^5)
+O(x_1^4),$$
- $u_2$: $$\frac{1}{3}q^2\,r-\frac{1}{3}q_1^2\,r^3+\frac{1}{3}x_1^2q^2\,\delta\bar q(r)-x_1^2q_1^2\,r^2\delta\bar q(r)
+O(x_1^4),$$
- $\tilde u_3$: $$\frac{1}{3}(q_1^2-q^2)\,r-\frac{2}{9}x_1q_1^2\,r+\frac{1}{3}x_1^2(q_1^2-q^2)\,\delta\bar q(r)
-\frac{2}{3}x_1^3q_1^2\,r\{r;\delta\bar q(r)\}_{r=1}-\frac{2}{9}x_1^3q_1^2\,\delta\bar q(r)+O(x_1^4),$$
- $v_1$: $$\begin{gathered}
\frac{1}{3}(q_1-q)^3\,r-\frac{1}{2}x_1(q_1-q)^2q_1\,(r-\frac{1}{3}r^3)+\frac{1}{3}x_1^2(q_1-q)
q_1^2\,(\frac{3}{4}r-\frac{1}{2}r^3+\frac{3}{20}r^5)\\[5pt]
-\frac{1}{12}x_1^3q_1^3\,(\frac{1}{2}r-\frac{1}{2}r^3
+\frac{3}{10}r^5-\frac{1}{14}r^7)+O(x_1^4),\end{gathered}$$
- $v_2$: $$\begin{gathered}
-\frac{1}{5}(q_1-q)q^2\,r+\frac{2}{15}(q_1-q)q_1^2\,r^3
-\frac{1}{30}x_1q_1^3\,(\frac{3}{4}r+\frac{1}{2}r^3-\frac{9}{20}r^5)
+\frac{3}{20}x_1q^2q_1\,(r-\frac{1}{3}r^3)\\[5pt]
-\frac{1}{20}x_1q_1^3\,r^2(r-\frac{1}{3}r^3)+\frac{1}{10}(q_1-q)^2q\,r+\frac{1}{30}(q_1-q)^3\,r
-\frac{1}{5}x_1^2(q_1-q)q_1^2\,(1-2r^2)\delta\bar q(r)\\[5pt]
-\frac{1}{10}x_1^3q_1^3\,r(r-\frac{1}{3}r^3)\delta\bar q(r)+\frac{3}{10}x_1^3q_1^3\,\{r;\delta\bar q(r)\}
-\frac{1}{10}x_1^3q_1^3\,r^2\{r;\delta\bar q(r)\}-\frac{1}{30}x_1^3q_1^3\,\{r^3;\delta\bar q(r)\}\\[5pt]
-\frac{1}{10}x_1^3q_1^3\,\{r;r^2\delta\bar q(r)\}+O(x_1^4),\end{gathered}$$
- $\tilde v_3$: $$\begin{gathered}
\Big[-\frac{3}{10}(q_1-q)^2q_1+\frac{2}{15}x_1(q_1-q)q_1^2-\frac{1}{75}x_1^2q_1^3
+\frac{2}{15}(q_1-q)^3\Big]\,r\\[5pt]
+\frac{1}{20}x_1q_1\Big[2(q_1-q)q_1-\frac{2}{3}x_1q_1^2-(q_1-q)^2\Big]\,(r-\frac{1}{3}r^3)+O(x_1^4).\end{gathered}$$
Inserting the above expressions into Eq. (\[basic\]) and demanding that the coefficients of $r$ and $r^3$ disappear, $x_1$ can be read off with some effort. It is best to give $x_1$ as the zero, $f(x_1)=0$, of the following function: $$\begin{gathered}
f(x)\equiv\\
\Big[-\big(\frac{u_2}{w_1}q\big)+\frac{1}{2}\big(\frac{y_2}{w_1}q^3\big)+\dots\Big]
+\Big[1-\frac{13}{60}\big(\frac{v_2}{w_1}q^2\big)+\dots\Big]\,x+\Big[-\frac{1}{3}+\frac{1}{6}
\big(\frac{u_1}{w_1}q\big)+\dots\Big]\,x^2+\Big[-\frac{1}{9}+\dots\Big]\,x^3+\dots\,.%O(x^4).\end{gathered}$$ The leading contribution is the well-known formula $$\label{x1_basic}
x_1=\frac{u_2}{w_1}q$$ which is used extensively throughout this paper. As a byproduct, the shift of $q_1$ from the RS order parameter is given by $$\label{q_1}
q_1-q=\frac{1}{3}\,x_1\,q\big(1+\frac{2}{3}x_1+\dots\big)\quad.$$ \[To preserve consistency, a sixth order contribution $-\frac{1}{6!}\,y_2\,\sum_{\alpha\beta}(q_{\alpha\beta}-q)^6$ should have been included in the free energy expansion (\[F\]), as it enters the constant of $f(x)$ at the third order, i.e. at the highest order studied here.\]
Below six dimensions {#d<6}
====================
The renormalization group: fixed point and nonlinear scaling fields
-------------------------------------------------------------------
In $d=6-\epsilon$ the Gaussian fixed point becomes unstable, and the zero field spin glass transition is governed by the non-trivial one. Here we collect and present the available results for this fixed point (in the results for the fixed point below, a generic $n$ is kept, although $n=0$ is taken in the further parts of the section) : $$\bar {w}_1^{*2}\equiv w^{*2}=\frac{1}{2-n}\,\epsilon\,,\quad
\bar{m}_1^*\equiv \bar{m}^*=-\frac{2-n}{4}\,{w^*}^2\,;\quad \text{see
Refs.\ \cite{HaLuCh76} and \cite{Iveta}}\,.$$ Although they will not be used in this paper, the fixed point values of the quartic couplings (which — according to our knowledge — have not been published before) are also listed here: $$\bar{u}_1^*=\frac{3}{2}n\,w^{*4},\quad\bar{u}_2^*=12\,w^{*4},
\quad\bar{u}_3^*=-24\,w^{*4}, \quad\bar{u}_4^*=\frac{9}{2}\,w^{*4}\,.$$ The renormalization flow equations for the bare couplings of the generic RS theory were displayed in Ref. [@Iveta]. Using these equations, a new set of parameters $g_i$ — the so called nonlinear scaling fields introduced by Wegner [@Wegner] — can be defined with the following properties:[^8]
- $g_i\equiv 0$ at the fixed point for all $i$.
- An infinitesimally small $g_i$, with all the others kept zero, gives an eigenvector belonging to the eigenvalue $\lambda_i$ of the linearized renormalization group equations around the fixed point.
- They satisfy [*exactly*]{} the equations $\dot g_i=\lambda_i\,g_i$.
The RG flow of an observable $y$ — the order parameter or an irreducible vertex, for instance — can be written in terms of the $g_i$’s as follows: $$\label{yRG}
\dot y=\Big(k+\sum_i k_i\,g_i+\sum_{ij} k_{ij}\, g_ig_j+\dots\Big)\,\,y\quad .$$ The solution of this equation, i.e. $y$ in terms of the scaling fields is easily found: $$\label{y}
y(g_1,g_2,\dots)=g_1^{k/\lambda_1}\,
\hat y\Big(g_2\,g_1^{-\lambda_2/\lambda_1},\dots,g_i\,g_1^{-\lambda_i/\lambda_1},\dots\Big)
\times \exp\Big(\sum_i \frac{k_i}{\lambda_i}\,g_i
+\sum_{ij} \frac{k_{ij}}{\lambda_i+\lambda_j}\, g_ig_j+\dots\Big)\,,$$ the scaling function $\hat y(\dots)$ can be determined by perturbative methods.
In our two-parameter system defined by $\tau$ and $w$ the two nonzero scaling fields[^9] $r\equiv g_1$ and $\tilde g\equiv g_2$ (the notations are chosen to keep connection with previous sections) can be found by starting with the RG equations (\[RGflow\])[^10] and taking the temperature-like relevant eigenvalue $\lambda_r$ and the irrelevant one, $\lambda_{\tilde g}$, from Ref. [@Iveta] as $$\label{lambda}
\lambda_r\equiv \frac{1}{\nu}=2-\frac{5}{3}\,\epsilon+\dots\,,\qquad\qquad
\lambda_{\tilde g}=-\epsilon+\dots\,.$$ The bare parameters are then straightforwardly expressed by the scaling fields as $$\label{r_g}
\begin{aligned}
%\tau&=r\,(1-2\frac{\tilde g}{\epsilon})^{-5/3}\\[3pt]
w^2&=w^{*2}+\frac{\tilde g}{1-2\frac{\tilde g}{\epsilon}}=
\frac{\epsilon/2}{1-2\frac{\tilde g}{\epsilon}}\\[5pt]%\,\quad.
\tau&=r\,\Big(1-2\frac{\tilde g}{\epsilon}\Big)^{-5/3}\,.
\end{aligned}$$
$x_1$ below the upper critical dimension
----------------------------------------
For the calculation of $x_1$ to first order in $\epsilon$, the RG study of $q$, $w_1$ and $u_2$ is required. The truncated (one-loop) renormalization group equations (\[qRG\]), (\[w1RG\]) and (\[u2RG\]) — see also footnote \[note\] — can be used whenever $w^2\ll 1$ and $\tau \ll 1$. We can solve these truncated equations in a similar way as (\[y\]) was derived from the generic equation (\[yRG\]). The scaling exponents and the relations between bare and scaling parameters are taken from Eqs. (\[lambda\]) and (\[r\_g\]), respectively. The scaling functions, which are always denoted by the “hat” symbol, cannot be determined by the RG equations alone, but the perturbative results of Eqs. (\[q1\]), (\[w11\]) and (\[u21\]) make it possible to get them to first order in $\epsilon$. \[The bare values must be replaced by the scaling fields using (\[r\_g\]), and take into account again footnote \[note\].\] In the following, the results for $q$, $w_1$ and $u_2$ are listed in itemized form. The $k$ and $k_2\equiv k_{\tilde g}$ quantities defined in (\[yRG\]) are also presented for completeness.
- $$\begin{gathered}
q=r^{1+\frac{\epsilon}{2}}\,\,\hat q\big(\tilde g\,r^{\frac{\epsilon}{2}}\big)
\times\Big(1-2\,\frac{\tilde g}{\epsilon}\Big)^{-1/6}\,,\label{q_scaling}\\[4pt]
\intertext{with the scaling function}
w^*\hat q(x)=\Big[1+\big(\frac{1}{2}\ln 2+1\big)\,\epsilon+\dots\Big]+
2\,\Big[1+\big(\ln 2+\frac{17}{6}\big)\,\epsilon+\dots\Big]\,\,
\Big(\frac{x}{\epsilon}\Big)+O\Big[\Big(\frac{x}{\epsilon}\Big)^2\Big]\,;
\label{q_hat}\\[10pt]
k=2-\frac{\epsilon}{2}+\frac{1}{2}\,\eta^*_L=2-\frac{2}{3}\,\epsilon+O(\epsilon^2)
\qquad\text{and}\qquad k_{\tilde g}=-\frac{1}{3}+O(\epsilon)\,.
\label{q_k}%\notag\end{gathered}$$
- $$\begin{gathered}
w_1=r^{\frac{\epsilon}{2}}\,\,\hat w_1\big(\tilde g\,r^{\frac{\epsilon}{2}}\big)
\times\Big(1-2\,\frac{\tilde g}{\epsilon}\Big)^{1/2}\,,
\label{w1_scaling}\\[4pt]
\intertext{with the scaling function}
\frac{\hat w_1(x)}{w^*}=\Big[1+\big(4\ln 2-\frac{39}{4}-\frac{7}{2}\ln n\big)\,\epsilon+\dots\Big]+
2\,\Big[1+\big(8\ln 2-\frac{56}{3}-7\ln n\big)\,\epsilon+\dots\Big]\,\,
\Big(\frac{x}{\epsilon}\Big)\notag
\\[10pt]+O\Big[\Big(\frac{x}{\epsilon}\Big)^2\Big]\,;\qquad\qquad\qquad\qquad
\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\,\,
\label{w1_hat}\\[10pt]
k=\frac{\epsilon}{2}-\frac{3}{2}\,\eta^*_R=\epsilon+O(\epsilon^2)
\qquad\text{and}\qquad k_{\tilde g}=1+O(\epsilon)\,.
\notag\end{gathered}$$
- $$\begin{gathered}
u_2=r^{-1}\,\,\hat u_2\big(\tilde g\,r^{\frac{\epsilon}{2}}\big)
\times\Big(1-2\,\frac{\tilde g}{\epsilon}\Big)^{2/3}\,,
\label{u2_scaling}\\[4pt]
\intertext{with the scaling function}
\frac{\hat u_2(x)}{w^{*4}}=6\,\Big[1+O(\epsilon)\Big]+12\,\Big[1+O(\epsilon)\Big]
\,\,\Big(\frac{x}{\epsilon}\Big)+O\Big[\Big(\frac{x}{\epsilon}\Big)^2\Big]\,;
\label{u2_hat}\\[10pt]
k=-2+\epsilon-2\,\eta^*_R=-2+\frac{5}{3}\,\epsilon+O(\epsilon^2)
\qquad\text{and}\qquad k_{\tilde g}=\frac{4}{3}+O(\epsilon)\,.
\notag\end{gathered}$$
The following remarks are appropriate here: Firstly, according to Eqs. (\[y\]), (\[lambda\]) and (\[q\_k\]) the temperature exponent for the order parameter $q$ is [*exactly*]{} $k/\lambda_r=(2-\epsilon/2+\eta^*/2)\,\nu
\equiv \beta=1+\epsilon/2+O(\epsilon^2)$; see (\[q\_scaling\]). The temperature exponents in (\[q\_scaling\]), (\[w1\_scaling\]) and (\[u2\_scaling\]) are correct up to $\epsilon$ order. Secondly, the discussion below Eq. (\[w1\_d=6\]) concerning the fully replicon, infrared divergent contribution to $w_1$ is equally valid for the $\ln n$ terms in (\[w1\_hat\]). Thirdly, the $O(\epsilon)$ corrections in the scaling functions $\hat q$ and $\hat w_1$ are unnecessary for the leading order calculation of $x_1$; they are displayed here to demonstrate the general form of the $\epsilon$ expansion. The corresponding corrections for $\hat u_2$ are not even available, see (\[u2\_hat\]), as they would require a two-loop level calculation.
The leading contribution in the $\epsilon$ expansion for $x_1$ follows from substituting $q$, $w_1$ and $u_2$ from Eqs. (\[q\_scaling\])-(\[q\_hat\]), (\[w1\_scaling\])-(\[w1\_hat\]) and (\[u2\_scaling\])-(\[u2\_hat\]), respectively, into the basic formula in Eq. (\[x1\_basic\]). A remarkably simple formula reflecting the invariance of $x_1$ under renormalization can be concluded: $$\label{x1_d<6}
\begin{gathered}
x_1=6\,w^{*2}\,\hat x_1\big(\tilde g\,r^{\frac{\epsilon}{2}}\big)
=3\,\epsilon \,\hat x_1\big(\tilde g\,r^{\frac{\epsilon}{2}}\big)\,,\quad
\text{ with the scaling function}\\[6pt]
\hat x_1(x)=\Big[1+O(\epsilon)\Big]+2\,\Big[1+O(\epsilon)\Big]
\,\,\Big(\frac{x}{\epsilon}\Big)+O\Big[\Big(\frac{x}{\epsilon}\Big)^2\Big]\,.
\end{gathered}$$
Almeida-Thouless line for $d<6$ {#AT_below}
-------------------------------
The external magnetic field $h^2$ evolves under renormalization according to Eq. (\[h\^2\_flow\]), with $|\epsilon|$ replaced by $-\epsilon$. The corresponding nonlinear scaling field $g_3\equiv\tilde{h^2}$ has now the relevant eigenvalue $$\lambda_{\tilde{h^2}}=4-\frac{\epsilon}{2}-\frac{\eta^*}{2}\equiv \frac{\delta\,\beta}{\nu}\,.$$ The flow equation for the replicon mass — Eq. (\[GammaR\_flow\]) — does not contain explicitly the magnetic field, therefore it enters the solution only through the invariant $\tilde{h^2}\,r^{-\delta\,\beta}$; see Eqs. (\[yRG\]), (\[y\]) and (\[lambda\]). According to the generic scheme (\[y\]), we have $$\Gamma_R=r^{(2-\eta^*)\,\nu}\,\,\hat\Gamma_R\big(\tilde g\,r^{-\lambda_{\tilde g}\,\nu},
\tilde{h^2}\,r^{-\delta\,\beta},\dots\big)\times\exp \Big(\frac{2}{3}\,\frac{\tilde g}{\lambda_{\tilde g}}
+\dots\Big)\,.$$ The exponential part can again be calculated in the truncated, one-loop approximation, in the usual way, providing (note that $\lambda_{\tilde g}=-\epsilon+\dots$) $$\Big (1-2\,\frac{\tilde g}{\epsilon}\Big)
^\frac{1}{3}\,,$$ whereas a comparison with the perturbative result (\[GammaR3\]) — after substituting the bare parameters by their corresponding nonlinear scaling fields \[see Eq. (\[r\_g\]) and also $$\label{h}
h^2=\tilde{h^2}\,\Big (1-2\,\frac{\tilde g}{\epsilon}\Big)
^{1/6}$$ which follows from (\[h\^2\_flow\])\] — gives the scaling function: $$\label{hat_GammaR}
\hat\Gamma_R(x,y)=w^{*2}\,\Big\{[-4+O(\epsilon)]+[-24+O(\epsilon)]\,\Big(\frac{x}{\epsilon}\Big)
+[1+O(\epsilon)]\,\Big(\frac{y}{w^*}\Big)+[-2+O(\epsilon)]\,\Big(\frac{x}{\epsilon}\Big)
\,\Big(\frac{y}{w^*}\Big)+\dots
\Big\}\,.$$ The zero of the scaling function gives the AT-line: $$\label{AT_d<6}
\tilde{h^2}=4\,w^*r^{\delta\,\beta}=4\,w^*r^{2+\dots}\,,\qquad \frac{\tilde g}{\epsilon}\,
r^{\epsilon/2}\ll 1\,\quad\text{and}\,\quad 0<\epsilon\ll 1\,.$$ For the fixed point system, $w=w^*$ implies $\tilde g=0$ and $\tilde{h^2}=h^2$, $r=\tau$. The result in (\[AT\_d<6\]) is then identical with Eq. (18) of Ref. [@AT2008].
Discussion of the results and conclusions {#discussion}
=========================================
For a thorough analysis of the $d$-dependence of $x_1$ while crossing the upper critical dimension, we recollect here the one-loop truncated results from previous sections; see Eqs. (\[x1\_d>6\]), (\[x1\_d=6\]) and (\[x1\_d<6\]). The goodness of these approximations depends on the smallness of the scaling variable, which is defined and expressed in terms of the bare parameters $w^2$ and $\tau$ as follows: $$\label{cases}
\mathord{\text{scaling variable}}=
\begin{cases}
\frac{2}{|\epsilon|}\,\tilde w^2\,r^{|\epsilon|/2}=
\frac{2}{|\epsilon|}\,w^2\,\tau^{|\epsilon|/2}\,
\big(1+\frac{2}{|\epsilon|}w^2\big)^{-1-\frac{5}{6}|\epsilon|} & \quad\text{$d>6$,
see (\ref{solution}),}\\[10pt]
w^2\,\big(1-w^2\ln r\big)^{-1}=w^2\,\big(1+\frac{5}{3}w^2\ln w^2-w^2\ln \tau\big)^{-1}
& \quad\text{$d=6$, see (\ref{tau_vs_r}),}\\[10pt]
\frac{2}{\epsilon}\,\tilde g\,r^{\epsilon/2}=\tau^{w^{*2}}\,
\Big(\frac{w^{*2}}{w^2}\Big)^{\frac{5}{3}{w^{*2}}}\,\Big(1-\frac{w^{*2}}{w^2}\Big),
\quad w^{*2}=\frac{\epsilon}{2}
& \quad\text{$d<6$, see (\ref{r_g}).}
\end{cases}$$ This scaling variable is displayed — for a chosen pair of bare values $w^2=0.005$ and $\tau=0.0001$, both much smaller than one, as it should be in this approximation — as a function of dimension $d$ below \[Fig. (\[scaling\_variable\]a)\] and above \[Fig. (\[scaling\_variable\]b)\] 6, where it takes $\approx 0.005$. $x_1$ is also shown in this figure, with the awkward behaviour of approaching zero from both sides of the upper critical dimension six, while $x_1\approx 0.03$ in $d=6$. It is clear, however, from the figure that our approximation breaks down when approaching $d=6$ from either side, as the scaling variable goes to unity in that limit. As a matter of fact, it must be stipulated that the scaling variable be at least as good as in $d=6$, i.e. $\lesssim 0.005$. Therefore, the range of applicability of our approximation (for the given $w$ and $\tau$) is constrained to $d\approx 5.99$ and $d\gtrsim 6.4$, respectively. (Note that the chosen $w$ is just the fixed point when $d=5.99$.) Representative values of $x_1$ in these ranges, together with the six-dimensional case, are presented in Tab. \[table\]. It can be concluded from this example that $x_1$ keeps on being monotonically increasing when lowering dimensions through 6. Nevertheless, a discontinuity of $x_1(w^2,\tau)$ at $d=6$ cannot be excluded. An extrapolation of the data from the range $d\gtrsim 6.4$, using an exponential and/or a power law fit, provides $x_1(0.005,0.0001)\approx 0.026-0.028$, a value somewhat lower than the six-dimensional one, $0.030$, when considering the scaling variable as a measure of the relative error (it is $\approx 0.005$ in six dimensions, see Table \[table\]). A similar extrapolation from the $d<6$ side, however, does not exist.
![The scaling variable (left vertical axis) measures the goodness of the approximation. (a): $d<6$ and (b): $d>6$. The dependence of $x_1$ is also shown in both regimes, together with its $d=6$ value (horizontal lines). $w^2=0.005$ and $\tau=0.0001$ are fixed in this figure. The approximation breaks down when approaching $d=6$ from both sides.[]{data-label="scaling_variable"}](d_below_6.eps "fig:")![The scaling variable (left vertical axis) measures the goodness of the approximation. (a): $d<6$ and (b): $d>6$. The dependence of $x_1$ is also shown in both regimes, together with its $d=6$ value (horizontal lines). $w^2=0.005$ and $\tau=0.0001$ are fixed in this figure. The approximation breaks down when approaching $d=6$ from both sides.[]{data-label="scaling_variable"}](d_above_6.eps "fig:")
(a)(b)
$d$ $\quad x_1\times 10^2\quad$ scaling variable
--------- ----------------------------- ------------------
6.8 0.1287 0.000308
6.6 0.2648 0.001026
6.4 0.5649 0.003834
6 2.9943 0.004991
5.99005 2.9993 0.004776
5.99 3.0000 0
5.98995 3.0006 -0.004774
: $x_1$ around six dimensions shows monotonically increasing behaviour with decreasing dimensionality. The smallness of the scaling variable verifies the approximation. The bare parameters $w^2=0.005$ and $\tau=0.0001$ are the same as in Fig. \[scaling\_variable\].[]{data-label="table"}
Below six dimensions $x_1$ has only a slight temperature dependence, and it becomes nonzero and universal at criticality: $$x_1=[3\,\epsilon+O(\epsilon^2)]+C\,\tau^{\frac{\epsilon}{2}+\dots}+\dots\,,\qquad d=6-\epsilon\,;$$ $C$ is a nonuniversal, i.e. $w$-dependent, amplitude. The typical behaviour for both below and above six dimensions is displayed in Fig. \[x1\_vs\_tau\], the value of the cubic coupling is kept $w^2=0.005$. The tendency of an increasing $x_1$ while lowering the dimension is again obvious. The vertical scale was magnified in the left subfigure (a) to show the qualitative difference between the 6- and 5.99-dimensional curves.
![$x_1$ as a function of the reduced temperature $\tau$; $w^2=0.005$. (a): $d\le 6$ and (b): $d\ge 6$. $x_1$ is zero at criticality, i.e. for $\tau=0$, when $d\ge 6$. On the contrary, it is nonzero for $d<6$ and has the universal value $x_1=3\,\epsilon+O(\epsilon^2)$ at $T_c$. []{data-label="x1_vs_tau"}](x1_vs_tau_3.eps "fig:") ![$x_1$ as a function of the reduced temperature $\tau$; $w^2=0.005$. (a): $d\le 6$ and (b): $d\ge 6$. $x_1$ is zero at criticality, i.e. for $\tau=0$, when $d\ge 6$. On the contrary, it is nonzero for $d<6$ and has the universal value $x_1=3\,\epsilon+O(\epsilon^2)$ at $T_c$. []{data-label="x1_vs_tau"}](x1_vs_tau_1.eps "fig:")
(a)(b)
The critical field along the AT-line, for a given pair of bare parameters $w^2$ and $\tau$, can be analysed using results from previous sections. See Eqs. (\[GammaR4\]), (\[AT\_d>6\]) for $d>6$, and (\[AT\_d=6\]) for $d=6$. Below six dimensions, if we wish to move somewhat away from the fixed point, the zero of the expanded equation (\[hat\_GammaR\]) must be found, providing \[instead of (\[AT\_d<6\])\]: $$\tilde{h^2}=4\,w^*r^{\delta\,\beta}\,\big(1+8\,\frac{\tilde g}{\epsilon}\,r^{\epsilon/2}\big)\,
,\qquad \frac{\tilde g}{\epsilon}\,
r^{\epsilon/2}\ll 1\,\quad\text{and}\,\quad 0<\epsilon\ll 1\,.$$ Eqs. (\[r\_g\]), (\[h\]) and (\[cases\]), together with $\delta\beta=2+\frac{3}{2}\epsilon$, give us the critical field as $$w\,h^2=4\,w^2\,\tau^{2+3\,w^{*2}}\,\Big(\frac{w^{*2}}{w^2}\Big)^{4+5\,w^{*2}}\,
\bigg[1+4\,\tau^{w^{*2}}\,
\Big(\frac{w^{*2}}{w^2}\Big)^{\frac{5}{3}{w^{*2}}}\,\Big(1-\frac{w^{*2}}{w^2}\Big)\bigg],
\quad w^{*2}=\frac{\epsilon}{2}\,.$$
The critical field where RSB sets in as a function of temperature (i.e. the AT line) — or more precisely $wh^2$ as a function of $\tau$ — is shown in Fig. \[AT\_figure\] for three different dimensions at fixed cubic coupling $w^2=0.005$. The curve for $d=5.99$ (note that the system is at exactly the fixed point then) is significantly below the six-dimensional one. It is easy to see that this follows directly from the exponent inequality $\delta\beta-2=\frac{3}{2}\epsilon+\dots>0$. To see clearly the behaviour of the critical field above and below six dimensions for decreasing $d$, it is tabulated in Table \[AT\_table\] for the system with $w^2=0.005$ and $\tau=0.0001$. The last three rows of this table show that the kind of monotonicity found above six dimensions is restored below it, i.e. the critical field increases with decreasing dimensions. It must be remarked, however, that around the last dimension value $d=5.98995$, the error[^11] of our approximation starts to be the same order of magnitude as the variation of the critical field itself. The range where this one-loop perturbative method is applicable below the upper critical dimension is certainly very narrow.
![Almeida-Thouless line ($wh^2$ versus $\tau$) of the field theoretic model with $w^2=0.005$ for three different dimensions.[]{data-label="AT_figure"}](AT.eps)
$d$ $\quad wh^2\times 10^{10}\quad$ scaling variable
--------- --------------------------------- ------------------
6.8 0.0827 0.000308
6.6 0.1680 0.001026
6.4 0.3497 0.003834
6 1.9849 0.004991
5.99005 1.7410 0.004776
5.99 1.7419 0
5.98995 1.7421 -0.004774
: Critical field values around six dimensions for $w^2=0.005$ and $\tau=0.0001$. Below the critical field replica symmetry is broken. The scaling variable’s values are, of course, the same as in Table \[table\].[]{data-label="AT_table"}
As a conclusion, we can confidently claim that RSB survives below six dimensions in the cubic replica field theory representing the Ising spin glass. We focused on two quantities which are strongly related to RSB: the breakpoint of the order parameter function $x_1$ and the Almeida-Thouless line. A combination of the perturbative one-loop method with a simple two-parameter renormalization group (which is correct near the critical fixed point) provided reliable results in all the three ranges of dimensionalities, i.e. for $d$ larger, equal, and less than six. The calculations above and below six dimensions are rather different, due to the Gaussian versus nontrivial fixed point governing critical behaviour in the two cases. The applied perturbative method makes it impossible to approach closely the upper critical dimension: the range of dimensions where the approximation is correct for a given system (i.e. for given $w$ and $\tau$) is very narrow and close to 6 when $d<6$, whereas it is $d\gtrsim 6.2$ when $d>6$ (and the farther we are from $d=6$, the better the approximation). The six-dimensional case needs special care along the way systems at their upper critical dimension are commonly studied [@PfeutyToulouse]. The logarithmic temperature dependences obtained are quite similar to those in ordinary systems at their upper critical dimension.
Above six dimensions both $x_1$ and the critical field are monotonically increasing for decreasing $d$, and this tendency persists for $d<6$. There seems to be, however, a discontinuity of the critical field at $d=6^-$: the AT line for $d\lessapprox 6$ is significantly below the six-dimensional one, see Fig. \[AT\_figure\] and Table \[AT\_table\]. Nevertheless, we can notice that the trend of increasing dominance of RSB for decreasing space dimensions persists even below six dimensions.
As a final remark, we recall that for $d<6$, $x_1$ gains the qualitatively new feature of being nonzero (and universal!) at criticality. This might suggest a kind of first order transition. That this is not the case can be clearly seen by displaying the order parameter function using Eqs. (\[q(x)\]) and (\[q\_1\]): $$q(x)=q_1\,\hat q(x/x_1)\,,\qquad \text{where} \qquad q_1\sim q\sim \tau^{\beta}\,.$$ An elaboration of the equation of state along the lines of Sec. \[RSB\_vs\_RS\] for $d<6$ (which is out of the scope of the present paper, and is left for a future publication), proves that, next to the spin glass transition, $\hat q$ is a function independent of temperature,[^12] and thus nontrivial even at criticality. The prefactor $q_1$, however, disappears at $T_c$ ensuring continuity of the order parameter through the spin glass transition.
We are extremely grateful to Imre Kondor for his thorough review of the paper prior to publication, and also for his useful suggestions.
[22]{} natexlab\#1[\#1]{}bibnamefont \#1[\#1]{}bibfnamefont \#1[\#1]{}citenamefont \#1[\#1]{}url \#1[`#1`]{}urlprefix\[2\][\#2]{} \[2\]\[\][[\#2](#2)]{}
, , , **, vol. of ** (, , ).
, ****, ().
, ****, ().
, ****, ().
, , , ****, ().
, ****, (), .
, ****, (), .
, , , ** (, ), vol. of **, p. , .
, ****, (), .
, , , ****, ().
, , , ****, ().
, ****, ().
, ****, ().
, in **, edited by (, ), vol. of **, .
, , , ****, (), .
, ****, ().
, , , ****, (), .
, ** (, , ).
, ****, (), .
, ****, (), .
, , , ****, ().
, ****, ().
[^1]: Somewhat surprisingly, [@Moore_Bray_2011] neglects discussing and even citing Ref. [@AT2008], where the AT line is followed up from mean field to $d\lesssim 6$. Subsection \[AT\_below\] reconsiders and comfirms the existence of an AT line below six dimensions.
[^2]: From now on, we redefine the parameters by suitably absorbing the geometrical factor $K_d$ and $\Lambda$: $\tau/\Lambda^2\rightarrow
\tau$, $w^2K_d\Lambda^{|\epsilon|}\rightarrow w^2$ and $h^2K_d^{-1/2}
\Lambda^{-4-|\epsilon|/2}\rightarrow h^2$.
[^3]: $|r(0)|=|r|$ in that paper is what we call $\tau$ here, whereas $w(0)=w$ agrees with our notation for the bare cubic coupling.
[^4]: That point has been noticed in Ref. [@Moore_Bray_2011], but was completely misinterpreted. We will return to this problem in Sec. \[discussion\]; see the first row of Eq. (\[cases\]) showing the impossibility of the limit $|\epsilon|\to 0$ in this approximation.
[^5]: The multiplicative factor $|\epsilon|$ in $x_1$ has its origin in the termination of the definition of the nonlinear scaling field $\tilde w$ in $d=6$. This is a feature of the RS renormalization group, and is not related to the problem of replica symmetry breaking.
[^6]: A vertex is called symmetrical if it is nonzero in the zero order parameter RS system.
[^7]: We hope that the [*ratio*]{} $r$ of $x$ to $x_1$ introduced here cannot be confused with the temperature-like scaling field of previous sections.
[^8]: The summary presented in this paragraph about the use of nonlinear scaling fields is quite general, not limited to the nontrivial fixed point of the RS replica field theory.
[^9]: In references [@droplet; @AT2008] an alternative scheme was used with a second relevant scaling field entering after appropriately redefining the field theory for getting rid of “tadpole” diagrams. The irreducible vertices are the same in the two schemes.
[^10]: \[note\]But be careful to replace $-|\epsilon|$ with $\epsilon$.
[^11]: The relative error can be estimated as being proportional to the square of the scaling variable.
[^12]: This has been suggested in Ref. [@beyond], see Eq. (155) of it.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The thermoelectric power factor of hierarchically nanostructured materials is investigated using the non-equilibrium Green’s function method for quantum transport, including interactions of electrons with acoustic and optical phonons. We describe hierarchical nanostructuring by superlattice-like potential barriers/wells, combined with quantum dot barriers/wells nanoinclusions as well as voids in the intermediate region. We show that these structures can be designed in a way that the power factor is not only largely immune to the presence of the nanostructure features, but under certain conditions benefits can be achieved as well. Interestingly, we show that these design approaches are linked to the energy relaxation of the current flow and whether charge carrier scattering is limited by elastic or inelastic processes. In particular, when nanostructures form potential barriers, the power factor can be substantially enhanced under elastic scattering conditions, irrespective of nanostructuring density and potential barrier heights. When inelastic scattering processes dominate, however, the power factor is inevitably degraded. In the case in which nanostructures form potential wells, despite a slight decrease, the power factor is quite resilient under either elastic or inelastic scattering processes. These nanostructuring design approaches could help open the path to the optimization of new generation nanostructured thermoelectric materials by not only targeting reductions in thermal conductivity, but simultaneous improvements in the power factor as well.'
author:
- Vassilios Vargiamidis
- Neophytos Neophytou
title: |
Hierarchical nanostructuring approaches for thermoelectric materials\
with high power factors
---
Introduction
============
Thermoelectric (TE) materials convert heat from temperature gradients into electricity and vice versa. Their performance is quantified by the dimensionless figure of merit $ZT = \sigma S^2 T / (\kappa_{\text{e}} + \kappa_{\ell})$ where $\sigma$ is the electrical conductivity, $S$ is the Seebeck coefficient, $T$ is the operating temperature, $\kappa_e$ is the electronic thermal conductivity, and $\kappa_\ell$ is the lattice thermal conductivity. The product $\sigma S^2$ is known as the power factor ($PF$). Traditional TE materials, which are mostly semiconductor doped alloys of Sb and Bi$_2$Te$_3$ at room temperature, and PbTe or SiGe at higher temperatures, reach $ZT \approx 1$. Over the last several years, however, numerous other materials have been explored, such as transition-metal dichalcogenides (TMDC) [@GDing16; @HHuang16; @WHuang14], skutterudites [@YTang15ncom; @YTang15nmat], phonon-glass-electron crystal structures [@Beek15nmat], half-Heuslers [@Stern16ntech; @CFu15ncom], oxides [@YYin17], etc. A large number of these materials demonstrate $ZT$ above $1$, primarily due to the reduction of their thermal conductivity [@Beretta19].
In order to achieve even further reductions in thermal conductivity, the majority of these materials are explored in the context of nanostructuring. Many approaches towards this route are common practice, such as superlattice-like geometries [@Mizuno15], alloying [@XWWang08], heavy doping [@Ikeda10], nanoporous materials [@Verdier16; @RYang05; @JHLee08], nanograining [@Neo13ntech; @Bennett16], nanoinclusions (NIs) [@Biswas12; @Gayner16; @Zou15; @Hopkins11; @Vineis10; @Popescu09], etc. Nanoinclusions, in particular, cause scattering of short wavelength phonons with mean-free-paths in the order of nanometers, which otherwise contribute significantly in the thermal conductivity of common TE materials, such as in PbTe [@Bo11]. This technique is applied to a broad range of materials, including Bi$_2$Te$_3$ [@Fan11; @Keshavarz14], PbTe [@Heremans05; @Biswas12; @Hsu04], SiGe [@Ahmad16; @XWWang08; @Zhu09], MnSi [@Saleemi15], and SnTe [@Tan14], to name a few.
Furthermore, undoubtedly, one of the most successful approaches to reduce thermal conductivity is hierarchical nanostructuring, where distortion features are placed within a matrix material at the mesoscale (grains/boundaries), microscale (NIs) and nanoscale (atomic defects). These scatter phonons of various wavelengths and reduce phonon transport across the entire spectrum. Indeed, by nanostructuring PbTe in a hierarchical manner, record high value of $ZT=2.2$ was achieved due to drastic reductions in $\kappa_\ell$ [@Biswas12]. More recent works have achieved even higher $ZT$ up to $2.5$ at $923$ K [@GTan16]. For such a success, the thermal conductivities in these materials reached values well below the amorphous limit which is $1-2$ W/mK at room temperature [@Kearney18], and thus, cannot be further reduced easily. Therefore, further benefits to $ZT$ can only be achieved from the enhancement of the $PF$. In the majority of cases, however, nanostucturing degrades the electrical conductivity and the $PF$ as well. In addition, the adverse interdependence of the electrical conductivity and Seebeck coefficient does not allow flexibility in $PF$ improvements.
The importance of retaining high $PF$s has recently become more appreciated in nanostructured TE materials. Reference [@Zhao14], in particular, emphasizes the importance of matrix/inclusion band alignment to retain the original conductivity of the material and to avoid degradation in the $PF$. Large $PF$ improvements were also demonstrated in highly doped nanocrystalline Si [@Neo13ntech], especially in the presence of nanovoids [@Lorenzi14]. While the impact of nanostructuring on the thermal conductivity can be more clearly understood as a general increase in phonon scattering, the same cannot be assumed for the $PF$. Due to the complexity in geometry, theoretical works to date (by us and others) focus on one type of nanostructured feature at a time, i.e. only superlattices (SLs) [@Thes16JEM], only nanocrystalline boundaries [@Kearney18], only NIs [@Fan11; @Ahmad16; @Liu12; @Foster17; @Peng14; @Zhou11], etc. In several cases, the conclusions vary substantially, from reports of large $PF$ benefits to only moderate or none; a consequence of the difficulty in accurately simulating and optimizing the complexity of geometries coupled with the complexities of the nanoscale transport physics in the presence of various types of nanostructured features. Indeed, the complexity of the electronic transport, combining semiclassical effects, quantum effects (i.e. quantization, tunneling, interferences, resonances), ballistic and diffusive regimes, as well as the geometry details with multiple features and feature types, makes accurate modelling a difficult task. Thus, it is imperative to shed more light and establish a high level of understanding of the $PF$ behavior in the presence of more than one nanoscale feature type, both qualitatively and quantitatively, if $ZT$ is to be maximized.
In this work we use the non-equilibrium Green’s Function (NEGF) method to calculate the electron transport properties of two-dimensional (2D) nanostructures when superlattice-type boundaries and NIs are present simultaneously (see Fig. 1). NEGF provides a unified, geometry flexible, fully quantum mechanical simulation approach well-suited for this problem. We present a systematic investigation of how such complex geometries affect the $PF$ in the cases where the nanofeatures impose potential barriers, or potential wells for charge carriers. We explore the influence of the heights of those barriers, as well as their number density. We then identify the design approaches that will allow for $PF$ immunity in hierarchically nanostructured materials, and in some cases, even significant improvements.
The paper is organized as follows. In Sec. II we describe our NEGF approach including our calibration procedure and indicate the geometries we study. In Sec. III we present and discuss our results, in Sec. IV we present a discussion on optimal nanostructuring, and finally, in Sec. V we conclude.
![\[fig1\] (Colour online) Geometries of the hierarchical nanostructured materials that we consider in this paper. (a) The pristine channel. (b) Channel with SL-type barriers. (c) Channel with SL barriers and quantum-dot potential barrier NIs. (d) Channel with SL-type wells. (e) Channel with SL-type wells and quantum-dot potential well NIs. ](Figure_1.png "fig:"){width="7cm" height="15.2cm"}\
Approach
========
We employ a 2D quantum transport simulator based on the NEGF formalism including electron-phonon (e-ph) scattering in the self-consistent Born approximation. We include both scattering of electrons with acoustic phonons (elastic scattering) and with optical phonons (inelastic scattering). The formalism and the details of the specific 2D simulator, which we developed in order to capture phonon scattering, as well as its convergence details, are described in several works of ours and of others [@Koswatta07; @Datta05; @Anantram08; @Foster17; @Var17pssa]. However, in order to be able to better discuss certain characteristics of e-ph scattering in the context of this work, we include here a brief description of the model with notation adopted to our system.
In the NEGF method a system/channel, described by a Hamiltonian $H$, is connected to two contacts (left and right), which are represented by self-energy functions $\Sigma_{\text{L}}$ and $\Sigma_{\text{R}}$. The Hamiltonian is constructed using a 2D effective mass (single-orbital) tight-binding grid uniformly spaced in the $x$- and $y$-directions, resulting in a banded matrix. The 2D channels we consider within the effective mass approximation, have a uniform $m^\ast = m_0$ in the entire channel, where $m_0$ is the rest mass of the electron. Thus, we do not consider a specific material, rather our study aims in providing first-order qualitative guidance into the design of high power factor nanostructured materials. These self-energies represent the influence of the semi-infinite Left and Right leads on the channel, respectively. Note that $\Sigma_{\text{L}}$ and $\Sigma_{\text{R}}$ are energy dependent, and non-Hermitian. They are formed using the first/last channel layers from which electrons are injected into the channel (and thus, they have the same size as these layers), and are calculated using the Sancho-Rubio iterative scheme [@Sancho_Rubio_85]. They are added to the first/last layer elements of the channel Hamiltonian. The e-ph scattering process in the device enters the NEGF formalism through the self-energy function $\Sigma_{\text{S}}$. One can view the scattering process as just another contact described by $\Sigma_{\text{S}}$, similar to the actual contacts described by $\Sigma_{\text{L}}$ and $\Sigma_{\text{R}}$, however, $\Sigma_{\text{S}}$ is added to all diagonal elements of the Hamiltonian.
The retarded Green’s function for the device is given by [@Datta05] $$G(E) = \left[ \left( E + i \eta^+ \right) I - H - \Sigma(E) \right]^{-1} ,
\label{eq1}$$ where $\eta^+$ is an infinitesimal positive number which pushes the poles of $G$ to the lower half plane in complex energy, $I$ is the identity matrix, and $\Sigma(E)$ is the sum of the self-energies $$\Sigma(E) = \Sigma_{\text{L}}(E) + \Sigma_{\text{R}}(E) + \Sigma_{\text{S}}(E) .
\label{eq2}$$ It proves useful and convenient to define the in-scattering self-energies due to contacts as: $$\Sigma_{\text{L, R}}^{\text{in}} (E) = - 2 \text{Im} \left[ \Sigma_{\text{L, R}} (E) \right] f_{\text{L, R}} (E) ,
\label{eq3}$$ where $\text{Im}[...]$ is the imaginary part and $f_{\text{L, R}}$ is the Fermi distribution for the left and right leads. Similarly, the out-scattering self-energies are defined as: $$\Sigma_{\text{L, R}}^{\text{out}} (E) = - 2 \text{Im} \left[ \Sigma_{\text{L, R}} (E) \right] \left[ 1 - f_{\text{L, R}} (E) \right] .
\label{eq4}$$ With $\Sigma_{\text{L, R}}^{\text{in/out}} (E)$ one can express the electron and hole correlation functions as: $$G^{\text{n}} (E) = G(E) \Sigma_{\text{L, R}}^{\text{in}} (E) G^\dagger (E) ,
\label{eq5}$$ $$G^{\text{p}} (E) = G(E) \Sigma_{\text{L, R}}^{\text{out}} (E) G^\dagger (E) .
\label{eq6}$$
Assuming that the system consists of 2D grid/lattice points with uniform spacing $a$, and making the nearest neighbour tight-binding approximation, the current density between grid points $j$ and $j+1$ is given by $$\begin{aligned}
\nonumber J_{j, j+1} = \frac{i e}{\hbar}
\\* &&\hspace*{-1in} \times (2) \int_{-\infty}^{\infty} \frac{d E}{2 \pi} \left[ H_{j+1,j} G_{j, j+1}^{\text{n}} (E) - H_{j,j+1} G_{j+1, j}^{\text{n}} (E) \right] ,
\label{eq7}$$ where $H_{j+1, j} = H_{j, j+1}^\dagger$ are the hopping matrix elements of the Hamiltonian, and $(2)$ is for the two spin directions.
A second source for in-scattering and out-scattering of electrons from an occupied state is the e-ph interaction. The self-energy at point $j$ and energy $E$ has two terms corresponding to scattering from $(j, E+\hbar \omega)$ and $(j, E-\hbar \omega)$. Within the Born approximation the in-scattering self-energy into a fully empty state is [@Mahan87] $$\Sigma_{\text{S}}^{\text{in}} (E) = D_{0} [ n_{\text{B}} G^{\text{n}} (E-\hbar \omega) + (n_{\text{B}} + 1) G^{\text{n}} ( E + \hbar \omega) ] .
\label{eq8}$$ where $D_0$ represents the e-ph scattering strength at grid point $j$, $n_{\text{B}}$ is the Bose-Einstein distribution function for phonons of energy $\hbar \omega$, and $G^{n} (E-\hbar \omega)$ is the electron density at $E-\hbar \omega$. The first and second terms in Eq. (\[eq8\]) represent in-scattering of electrons from $E - \hbar \omega$ (phonon absorption) and $E + \hbar \omega$ (phonon emission) to $E$, respectively. The out-scattering self energy, $\Sigma_{\text{S}}^{\text{out}} (E)$, from a fully filled state at energy $E$ is given by [@Mahan87] $$\Sigma_{\text{S}}^{\text{out}} (E) = D_{0} [ (n_{\text{B}} + 1) G^{\text{p}} (E-\hbar \omega) + n_{\text{B}} G^{\text{p}} ( E + \hbar \omega) ] ,
\label{eq9}$$ where $G^{\text{p}} (E-\hbar \omega) $ and $G^{\text{p}} (E+\hbar \omega)$ are the densities of unoccupied states at $E-\hbar \omega$ and $E+\hbar \omega$. The first and second terms in Eq. (\[eq9\]) represent out-scattering of electrons from $E$ to $E-\hbar \omega$ (phonon emission) and $E+\hbar \omega$ (phonon absorption), respectively. In the case of acoustic phonons, $\hbar \omega \rightarrow 0$, and so in Eqs. (\[eq8\]) and (\[eq9\]) we let $D_0 n_{\text{B}} \rightarrow D_{\text{AP}}$ (making use of the commonly employed equipartition approximation – see details in Appendix A) while in the case of optical phonons $D_0 \rightarrow D_{\text{OP}}$, which are taken to be constant throughout the channel.
The strength of the phonon scattering is adjusted such that the mean-free-path of electrons is $\lambda = 15$ nm. The way we calibrate this, is that we initially simulate a channel with length $L = 15$ nm in the ballistic regime, and then we increase the electron - acoustic phonon scattering strength $D_{\text{AP}}$ in the NEGF formalism, until the channel conductance drops to 50% of its ballistic value ($D_{\text{AP}}=0.0026$ eV$^2$). This effectively fixes a mean-free-path of $15$ nm for the channel (under acoustic phonon scattering conditions alone), a value that is comparable to that of common semiconductors such as silicon [@Neo11PRB]. The nanostructured geometries that we consider are shown in Fig. 1. The channels have length $L=100$ nm and width $W=15$ nm. Thus, with a mean-free-path of $\lambda = 15$ nm, the channel we consider is long enough for the transport to be diffusive. Note that the computational cost of NEGF simulations scales with the third power in channel width, and thus we only consider narrow channels of widths as given above. However, using wider structures should not affect our final conclusions. Throughout the paper we assume room temperature $T = 300$ K.
In the case that we consider electron-optical phonon scattering only, we use an optical phonon energy of $\hbar\omega = 0.06$ eV (a value similar to that of silicon longitudinal optical phonons, see for example [@Jacoboni83]), and simply use the same value of the scattering strength in the simulations, i.e., $D_0 = D_{\text{OP}}=D_{\text{AP}}=0.0026$ eV$^2$. This gives an energy relaxation length of $\lambda_{\text{E}} = 13$ nm, which, in the well region of length $60$ nm, dictates semi-relaxation of the electron energy, after they pass over the SL barriers. We consider this semi-relaxation of energy because it has been shown that it provides optimal conditions for the $PF$ of SL materials [@Neo13ntech; @Thes15JAP; @Kim09]. Note, however, that in this way the conductance of the channel with only optical phonon scattering will be in general larger compared to that with only acoustic phonon scattering. This is due to the fact that electron scattering with optical phonons has weaker phonon absorption rates, as a consequence of lower than unity phonon occupation number. Thus, we cannot compare the two cases quantitatively. In the case where we consider both acoustic and optical phonon scattering, we simply divide the scattering strengths $D_{\text{AP}}$ and $D_{\text{OP}}$ by a factor of two, which (as we show later) gives similar conductances for their pristine channels as in the acoustic phonon scattering case.
In order to calculate the power factor $G S^2$, where $G$ is the channel conductance and $S$ the Seebeck coefficient, we use the fact that the Seebeck coefficient is the average energy of the current flow [@Kim11] (see Appendix B for derivation details) $$S = \frac{1}{q T L} \int_{0}^{L} \langle E(x) - E_F \rangle dx ,
\label{eq13}$$ where $q$ is the carrier charge ($q=-\vert e \vert$ for electrons and $q=\vert e \vert$ for holes) and $\langle E ( x ) \rangle$ is the energy of the current flow along the transport direction, defined as: $$\langle E ( x ) \rangle = \frac{\int I_{ch} (E, x) E dE}{\int I_{ch} ( E, x) dE} ,
\label{eq14}$$ where $I_{ch}(E,x)$ is the energy and position resolved current. This is the quantity we integrate in order to get in Eq. (\[eq7\]) the total current as $J = \int I_{ch}(E) dE$. Note that the current is constant along the channel at each cross section, however, its energy is not constant, i.e. the charge carriers can gain or lose energy as they propagate. This happens in the presence of inelastic scattering (optical phonons). Thus, the energy of the current is position dependent and it is its scaled integral in the channel which provides the overall value of the Seebeck coefficient. In experimental settings, one extracts the Seebeck coefficient from the open circuit voltage upon the application of a thermal gradient along the channel, as $S = \Delta V / \Delta T$, which equivalently can also be computed by $S = I_{ch (\Delta V = 0)} / G \Delta T$. In [@Kim11], it was validated that the two methods of extracting the Seebeck coefficient are equivalent, which makes it easier in time consuming simulations (as the ones we undertake) to only run the $\Delta V \neq 0$ case and still be able to extract the Seebeck coefficient by integrating the energy of the current flow over the length of the channel, and that is how we extract the Seebeck coefficient in this work.
Note that from NEGF we obtain the conductance $G$, rather than the conductivity $\sigma$, because NEGF simulates a 2D channel with specific dimensions, thus, the units of the two quantities are also different. This could be converted to conductivity using the channel dimensions, however since we do not have a specific thickness associated with our 2D simulation ($W=15$ nm, $L=100$ nm only), we use the conductance $G$ from here on. Also note that in all our results below we refer to $G$ as the conductance, not to be confused with the Green’s function in Eqs. (\[eq1\])-(\[eq9\]), for which it is also customary to use $G$.
Figure 1(a) shows the pristine channel that we begin with. The conduction band reference is set at $E_{\text{C}} = 0$ eV. The Fermi level is placed also at $E_{\text{F}} = 0$ eV as this provides the highest $PF$ [@Kim09; @Thes15JAP; @Foster17]. Thus, we begin with an optimal channel as our basis, and we then proceed with inserting nanostructured features. The first nanostructured feature is SL type boundaries that form potential barriers for electrons as shown in Fig. 1(b). We consider thickness of $L_{\text{SL}} = 5$ nm and arbitrarily choose barrier heights of $V_{\text{SL}} = 0.05$ eV $(\approx 2 k_B T)$. We then add NIs with diameter $d=3$ nm each in between those boundaries as indicated in Fig. 1(c). In these structures, the Fermi level is placed at $E_{\text{F}} = V_{\text{SL}} = 0.05$ eV, so that the carriers are allowed easily to flow over the SL barriers. The NIs are modeled as potential barriers of cylindrical shape in rectangular arrangements within the matrix material, and their number density and barrier heights are varied, as discussed in the text below. The choice of the SL and NI sizes are such as to minimize the influence of quantum tunneling, which becomes strong and detrimental for the $PF$ for feature sizes below $3$ nm [@Thes15JAP]. We also consider nanostructured geometries where the potential barriers are replaced with potential wells (Figs. 1(d)-(e)). Finally, we also consider the situation in which the NIs are replaced by voids (not shown). The NEGF approach is ideal for such geometries as they can be described precisely when the Hamiltonian of the system is constructed.
![\[fig2\] (Colour online) Average energy of the current flow $\langle E ( x ) \rangle$ as defined in Eq. (\[eq14\]) along the channel length with (a) SL barriers, (b) SL barriers and NI barriers, and (c) SL wells and NI wells. The black lines (solid and dashed) represent the potential barriers. The yellow-green colormap shows the energy resolved current $I_{ch}(E,x)$, with yellow corresponding to regions of large flow in the case of OP scattering. The blue lines represent the position of the Fermi level $E_F$. In (a) the red-solid line shows $\langle E ( x ) \rangle$ in the presence of AP only (i.e., elastic scattering), in which case charge carriers have constant energy above the barriers/wells. The magenta-dashed-dotted line shows $\langle E ( x ) \rangle$ in the presence of OP only (i.e., inelastic scattering), in which case carriers relax their energy in the regions between the barriers and absorb phonons to gain energy and overpass the barriers. The black-dashed line shows $\langle E ( x ) \rangle$ in the presence of AOP. In (b) the red-solid and magenta-dashed-dotted lines are the same as in (a) for the SL structure. The green-dashed and black-dotted lines show $\langle E ( x ) \rangle$ in the presence of AP and OP, respectively, for the SL+NIs structure. In (c) the red-solid line shows $\langle E ( x ) \rangle$ in the presence of AP, the magenta-dashed-dotted for OP, and the black-dashed for AOP. ](Figure_2.png "fig:"){width="8.6cm" height="14.2cm"}\
It turns out that a lot about the thermoelectric transport can be understood by looking at the energy of the current flow, $\langle E(x) \rangle$, along the transport direction, as defined in Eq. (\[eq14\]). This states, as expected, that the higher in energy the current flows with respect to the Fermi level, the higher the Seebeck coefficient is. It also provides some indication about the electrical conductivity, i.e. the higher the energy of the current flow, the more electrons with higher velocities are utilized (assuming no complex bandstructure effects cause velocity reductions), and the higher the conductivity could be. As we will show below, it proves to be a very useful feature in understanding TE transport particularly in the nanostructured materials we consider, and below we describe how we use it to interpret our simulation results.
Figure 2 shows $\langle E ( x ) \rangle$ for different structure cases and scattering conditions. Figure 2(a) considers a channel with two SL-type barriers inserted (black SL barrier lines). The colormap in all subfigures shows the energy and spatial regions where the current flows (yellow) in the case where both acoustic and optical phonon (AOP) scattering are taken into account, i.e., elastic and inelastic scattering, respectively. In this situation, $\langle E ( x ) \rangle$ is indicated by the dashed-black line. Clearly, electrons absorb optical phonons, overpass the potential barriers and then they relax into the wells by emitting optical phonons. The solid-red line shows $\langle E ( x ) \rangle$ under purely acoustic phonon (AP) scattering-limited conditions. In this case scattering is elastic, and the energy of the current is constant along the channel. The dashed-dotted-magenta line shows the optical phonon (OP) scattering-limited transport case, where the energy of the current flow is now slightly lower and the degree of energy relaxation slightly larger compared to the AOP case. Note that we do not compare the two cases on equal basis as the scattering rates for the OP case and AOP case are different, i.e. the OP scattering strength is halved in the AOP case, so relaxation is weaker.
Figure 2(b) shows $\langle E ( x ) \rangle$ in the case where NIs are placed in between the SL barriers (dashed line barriers). The horizontal dashed-green line shows again that under acoustic (elastic) scattering conditions $\langle E ( x ) \rangle$ remains constant along the channel length, and is almost identical to the SL case (solid-red line). In the case of optical (inelastic) scattering conditions (dotted-black line), $\langle E ( x ) \rangle$ is relaxing in the region between the SLs, as expected. However, the presence of NIs reduces the number of available states that carriers can fall into after emitting phonons of energy $\hbar \omega$. In addition, since NIs disturb the low energy electrons, the average energy of the current $\langle E ( x ) \rangle$ slightly increases compared to the SL alone case [@Foster17]. Consequently, the degree of energy relaxation in the presence of NIs (dotted-black line) is smaller than that in the absence of NIs (dashed-dotted magenta line-repeated here from Fig. 2(a)), signaling a higher Seebeck coefficient. In Fig. 2(c), we show the corresponding case where the SL regions and the NIs form potential wells, and thus, less obstruction of transport is expected. Indeed, in this case $\langle E ( x ) \rangle$ is uniform throughout the channel, not only in the case of elastic scattering, but also in the case of inelastic scattering, as the wells are too narrow for the electrons to relax into.
![\[fig3\] (Colour online) Row-wise, summary of conductance $G$, Seebeck coefficient $S$, and power factor $PF$, for key structure examples as shown in Fig. 1. The blue bar indicates the pristine channel (Fig. 1a), the red bars the SL barrier channels (Fig. 1b), the magenta bars the SL+NI barrier channel (Fig. 1c), the light green bar the SL well channel (Fig. 1d), and the green bar the SL+NI well structure (Fig. 1e). Column-wise, the three different groups are results for AP scattering only, OP scattering only, and AOP scattering. Notice that the $PF$ in the SL case is higher than that of the pristine channel by $23\%$ under elastic AP scattering conditions only. This is retained when NIs are inserted. ](Figure_3.png "fig:"){width="8.4cm" height="14cm"}\
{width="\textwidth"}
Results and discussion
======================
In the basis of the above observations, we will explain the $PF$ behavior in these nanostructures. It turns out that the dominance of elastic or inelastic scattering processes has a significant effect in the $PF$ [@Kim11; @Thes16JEM], since relaxation lowers $\langle E ( x ) \rangle$ and degrades the Seebeck coefficient. Therefore, for each one of the five geometries shown in Fig. 1, we consider AP scattering alone, OP scattering alone, and finally both AOP scattering combined. We begin by presenting in Figs. 3(a)-(c) some key simulation results for those basic structures with respect to the electronic conductivity, Seebeck coefficient, and $PF$, respectively. In the case of the SL barrier structures, we align the Fermi level with $V_{\text{SL}}$ at $0.05$ eV for optimal performance, whereas in the other cases with the band edge $E_{\text{C}}$. The different bars correspond to the geometries of Fig. 1 as follows: i) blue bars-pristine structure, ii) red bars-SL barrier structure, iii) magenta bars-SL barriers plus NIs structure, iv) light green bars-SL wells structure, and v) green bars-SL wells with NIs structure. The three different column groups show the corresponding values for AP scattering, OP scattering, and AOP scattering, respectively.
**Elastic scattering improves the optimal PF:** We first focus on AP (elastic) scattering conditions, where the results for $G$, $S$, and $PF$ are shown in the first column group of Fig. 3. We compare how each quantity changes with respect to the pristine material (blue bars) for each nanostructured geometry. Interestingly, by raising $E_{\text{F}}$ high in the bands, the conductance $G$ (first column, first row, red bar) is increased by $55\%$, despite the introduction of the SL barriers (we discuss the reasons behind this behavior below). The Seebeck coefficient naturally drops (red bar in second row, first column), but overall the $PF$ is *increased* in the SL structure by $23\%$ compared to the pristine structure. This indicates that the energy filtering, provided by potential barriers that cut lower parts of the Fermi distribution, is more effective at degenerate conditions as long as energy does not relax, as also pointed out in earlier studies [@Kim09; @Neo13ntech; @Thes15JAP]. However, the reason the $PF$ improves originates from the significant increase of the conductivity, rather than from the Seebeck coefficient which is actually degraded. The introduction of NIs in the region between the SL barriers (magenta bars), has a small degrading effect in the electrical conductance of the channel, as the NIs introduce additional scattering. However, this reduction is not strong, and the $PF$ is retained at values higher than those of the pristine material by $19\%$. Thus, we demonstrate here that it is indeed possible to achieve significant $PF$ improvements (rather than reductions) in a hierarchically nanostructured material. This $PF$ improvement combined with the expected very low thermal conductivity, can lead to high $ZT$.
It is important, thus, to clarify the reason behind the increase in $G$, which is responsible for this $PF$ improvement. The physical origin for this behavior lies in the fact that the charge carriers propagate on average at higher energies, which allow higher group velocities, and thus higher mobility. This is evident in Fig. 4(a), which shows the transmission versus energy of the pristine material (solid-blue line) and of the SL material (dashed-red line), extracted at every $x$-point as $Tr(x) = \left( h / e^2 \right) \left( I(x) /(f_1 - f_2) \right)$. Note that $I(x)$ is constant in the case of AP which makes the $Tr(x)$ constant, but not in the case of OP and AOP. In the semiclassical Boltzmann transport formalism, the transmission is related to the transport distribution function (TD) $\Xi ( E )$ via $Tr_n (E) = \left( W / L \right) 2\pi\hbar \Xi_n (E)$, where $\Xi_n(E) = g_n(E) \upsilon_n^2 (E) \tau_n (E)$ is the TD function per subband $n$ [@NeoNanoLett; @Neo11PRB; @Mahan96PNAS; @Scheidemantel03]. In the usual case where $\tau_n (E) \propto 1/g_n(E)$, then $\Xi_n(E) \propto \upsilon_n^2 (E) \propto 2 E / m^\star$, which is linear in energy as we also observe within NEGF in Fig. 4(a). In the case of the SL, the transmission opens up for energies above the SL barrier at $V_{\text{SL}} = 0.05$ eV, but when this happens, the slope is *larger* compared to that of the pristine material. This indicates that carriers, after passing over the SL barriers, relax on the higher velocity states of the intermediate region, and propagate with larger group velocities. Figure 4(b) shows the transmission scaled over the derivative of the Fermi function $Tr \left( \partial f / \partial E \right)$, which captures the part of the transmission that actually contributes to transport. Clearly, the higher peak in the case of the SL structure (dashed-red line), indicates larger conductivity. Thus, the deeper the well between the SL barriers, the higher the mobility and conductivity of the material. **Inelastic scattering degrades the power factor:** What actually causes reduction in the $PF$, is the presence of energy relaxation, which is a result of inelastic scattering, in our case OP scattering. In the second column of Fig. 3, we show how the TE coefficients change in the presence of inelastic scattering alone. Because the phonon absorption process is weaker in the OP case as a consequence of the lower than unity phonon occupation number, the conductance of the pristine channel case (blue bars) is larger compared to that of the AP only case. Thus, we do not intent to quantitatively compare the two cases anyway. Furthermore, it is clearly seen that both the conductance $G$ and the Seebeck coefficient $S$ are reduced in the SL structure, $G$ even more in the SL plus NIs structure (compare the blue bar to the red and magenta bars in the second column group of Fig. 3). This reduction in both $G$ and $S$ leads to a large reduction in the $PF$ by $29\%$. As shown in Fig. 2 above, the energy relaxation process of charge carriers in the region between the SL barriers causes electrons to propagate at lower velocity states, which leads to reduction of the conductivity and of the Seebeck coefficient. This is again shown clearly in Figs. 4(c) and 4(d), where we plot the transmission $Tr$ and $Tr \left( \partial f / \partial E \right)$ for the pristine (blue lines) and the SL structure (red-dashed lines) when only OP scattering is considered. The peak in the pristine case at low energies is a consequence of not having phonon emission processes for carriers with energies smaller than the OP energies considered here $\hbar \omega = 0.06$ eV. In the case of the SL structure, however, where the $E_{\text{F}}$ is raised at the $V_{\text{SL}}$ level, emission is actually possible. In this case the scattering rate into lower energy states is larger, leading to reduction of the electronic conductivity compared to the pristine channel, despite the higher carrier energies and velocities. (Note that in the presence of inelastic scattering, the current flow, although constant along the channel, varies in energy. Thus, the transmission versus energy function is also spatially varying. In this case we still have transmission for energies below $V_{\text{SL}}$ because we extract the transmission at a point in the middle of the channel, where relaxation allows current flow at lower energies).
In the case where we introduce NIs in the region between the SL barriers (Fig. 3, column 2, magenta lines), the conductance suffers even more. However, the $PF$ is slightly increased compared to the SL case (red bar), but is significantly reduced by $27\%$ compared to the pristine case (blue bar). We mention here that our simulations (not shown here), indicate that the degrading effect of energy relaxation can be prevented when OP emission is suppressed. This can be achieved by utilizing lower energies for transport compared to $\hbar \omega$ (lower Fermi level and consequently lower $V_{\text{SL}}$) or materials with large $\hbar \omega$, such that there is not enough energy range for emission to happen.
In a realistic scenario, however, the scattering is dominated by both elastic and inelastic processes. In the third column of Fig. 3 we show $G$, $S$, and $PF$, respectively, in the case in which AOP scattering is taken into account. Note that here the strengths of both scattering mechanisms are reduced to half of their initial values in order to have similar conductance numbers as in the AP case (for the pristine structure, blue bars) for a more reasonable comparison. Since the scattering rates and relaxation rates change, we cannot map quantitatively the results of this column to the previous two, but we treat it as a separate case and we only draw qualitative conclusions. In this scenario, the conductance in the case of SL barriers only (red bar) naturally increases in comparison to that of the pristine channel (blue bar) as the electrons propagate at higher velocity states (see Fig. 4(a)). Even if OP emission takes place, in this case the reduction in $G$ is not significant enough to lower it below that of the pristine channel. The Seebeck coefficient on the other hand, is reduced largely in the case of SL barriers in comparison to the pristine channel, which is similar to the case in which only OP scattering is considered (middle column). The origin of this similarity is that when both types of scattering processes are considered in this context, the energy relaxation is still determined by the OP scattering (note that as before we raise the $E_{\text{F}}$ in the SL channel). We also note in the inset of Fig. 3(c) that the average energy of the current flow measured from the Fermi level (i.e. Eq. (\[eq14\])), in the case of the pristine channel, $\langle E_{\text{pristine}} \rangle - E_F^{\text{pristine}} = 0.053$ eV (left), while in the case of a channel with SL barriers, $\langle E_{\text{SL}} \rangle - E_{\text{F}}^{\text{SL}} = 0.042$ eV (right), which is a factor of $\approx 0.79 \times$ smaller than that of the pristine case. Thus, similarly, the Seebeck coefficient shown in the bar chart drops from $S_{\text{pristine}} = 1.77 \times 10^{-4}$ V/K to $S_{\text{SL}} = 1.39 \times 10^{-4}$ V/K, i.e., by a factor of $\approx 0.78 \times$. Due to the large reduction in the Seebeck coefficient, the $PF$ is also degraded.
**Potential wells reduce $PF$ only slightly:** We now investigate the structures where the SL layers and the NIs introduce *potential wells* for transport electrons (green-colored bars in Fig. 3). In the AP case, either of the two features reduce the conductance slightly, increase the Seebeck coefficient again very slightly, and thus the $PF$ suffers only slightly by $1.1\%$ and $2.6\%$, respectively, compared to the pristine material. As expected, potential wells cause some obstruction to transport due to reflections at the interfaces of the SL and NI boundaries, but this is not enough to cause significant reduction of the $PF$. The degradation of the $PF$ in the case in which inelastic scattering alone is taken into account is $11\%$ and $19\%$ for the SL wells structure and for the SL wells plus NIs structure, respectively (second column, third row of Fig. 3, compare blue vs green bars). This reduction is significantly less than that in the case of SL barriers and SL barriers plus NIs. This is due to the fact that, since the wells formed are thin in our case ($5$ nm), the carrier energy cannot relax easily in there, which therefore leads to a weak degrading influence.
![\[fig5\] (Colour online) Transmission $Tr$ versus electron energy $E$ for the channel with SL barriers and nanoinclusions (NIs) in the elastic scattering regime (AP scattering only) for: (a) increasing NI barrier height $V_{\text{N}}$, and (b) increasing number of NIs. In (b) the height of the NIs is set to $V_{\text{N}} = 0.1$ eV. The insets show schematics of the channels considered. ](Figure_5.png "fig:"){width="8.6cm" height="10.4cm"}\
{width="19.2cm" height="12.5cm"}\
To summarize, therefore, in materials in which transport is dominated by elastic scattering, or if the inelastic scattering energy relaxation length is much larger that the characteristic geometrical features of the channel, it is beneficial to utilize nanostructures that form potential barriers, while setting high Fermi levels at the level of the SL barriers. In that case, benefits to the $PF$ by $>20\%$ can be achieved. In the case where the dominant scattering mechanisms are inelastic, then nanostructuring using potential wells is more beneficial. Although in this case improvements cannot be achieved, at least the reduction to the $PF$ is minimal. Later on we analyze these two cases in more detail.
**Robustness to $V_{\text{N}}$ and NI number density:** In the first case, where NIs form barriers, and under elastic scattering conditions, where benefits are observed, it is important to note that these benefits seem to be robust to the barrier heights of the NIs and their number density. Figure 5, for example, shows the transmission versus energy in the case where only AP scattering is considered for increasing values of $V_{\text{N}}$ (Fig. 5(a)) and for increasing number densities (Fig. 5(b)), as also illustrated in the insets. In Fig. 5a, by changing the barrier height from zero to $V_{\text{N}} = 0.05$ eV and then to $0.1$ eV, the transmission changes only slightly. The same is observed in Fig. 5(b) when the number of NIs changes from zero to $4$ and then to $10$. Minimal changes to the transmission are observed, indicating that the performance will be robust to such variations. It is known from a previous simulation work that the NI number density does not have a significant influence on the $PF$ when the Fermi level is raised in the bands and $V_{\text{NI}} \leq E_{\text{F}}$, because the moderate decrease they cause in $G$ is compensated by an increase in $S$ [@Foster17]. Experimental observations, where NIs embedded within matrix TE materials also point to this direction [@Fan11; @Peng14; @Zou14]. Here, in the hierarchical architecture, even better, the NI density does not affect the transmission, which means it affects neither $G$, nor $S$, nor the $PF$. The density, however, can drastically affect the thermal conductivity by increasing phonon scattering as shown in several works [@Biswas16; @Biswas17; @Shenghong18; @Honarvar18; @Yang18PRB], which will benefit the overall $ZT$ figure of merit. We can attribute this difference in behavior to simple scattering theory, i.e. from Matthiessen’s rule the scattering mechanism with the smaller mean-free-path will have the largest effect. For example in Si the mean-free-path for electrons is of the order of few nanometers, but for phonons the dominant mean-free-path is $\sim 135$ nm - $300$ nm [@Sellan10; @Dettori15], which largely increases the influence of closely packed NIs on phonons, rather than electrons.
**Comprehensive analysis:** In Fig. 6 we present a comprehensive analysis for the TE coefficients in the SL plus NIs barrier case, as functions of the NI barrier height $V_{\text{N}}$ and for increasing NI number density. In a similar manner to Fig. 3, row-wise we show the TE coefficients $G$, $S$, and $PF$, while column-wise we show results for simulations that consider only AP scattering, only OP scattering, and AOP scattering, respectively. We consider the initial structure as the one which contains the SL barriers of height $V_{\text{B}} = 0.05$ eV, and we plot data versus the heights of the NI barriers $V_{\text{N}}$. In each sub-figure we show results for three structures, containing $4$, $6$ and $10$ NIs in the regions between the SL barriers (as shown in the inset of Fig. 6(b)). In all cases we observe that, as the NI barrier height increases, the conductance is reduced, however not strongly. The Seebeck coefficient demonstrates only a small increase, as the NIs tend to push $\langle E ( x ) \rangle$ slightly upward as shown in Fig. 2(b). Therefore, in the case of AP scattering in SL channels, the $PF$ exhibits a slight degradation of the order of $10\%$ when NIs are introduced (Fig. 6(c)). However, even at the high NI density and high $V_{\text{N}}$, the $PF$ is higher than that of the pristine material we began with (horizontal dashed-green line in Fig. 6(c)). A $\sim 10\%$ reduction in the $PF$ is observed in the case in which only OP scattering is considered (middle column, Fig. 6(f)), but in this case the $PF$ is already $\sim 30\%$ below the pristine material $PF$ value (not shown). In the third column of Fig. 6, where we consider the influence of AOP scattering, we observe again a $\sim4\%$ decrease in the $PF$ in comparison to the pristine case, an intermediate percentage value between the two extreme cases (although due to the scattering rates chosen, we cannot quantitatively compare this case to the previous two directly). In addition, as the number density of NIs increases, the conductance and $PF$ drop slightly, however their effect is not significant, even at high NI densities.
**The effect of voids:** The far right points connected by the black-dotted lines in the sub-figures of Fig. 6, indicate the corresponding results in the case where the NIs are replaced with voids. For simulation purposes, we increase $V_{\text{N}}$ in those geometries to very large numbers, effectively leading to vanishing wave function in those regions, which resembles a void structure. We notice that voids cause significant degradation in the conductance and in the $PF$; namely, there is $30\%$ - $50\%$ reduction from the SL reference depending on the NI number density (it turns out that in this case the density has a stronger effect). The Seebeck coefficient also seems to be reduced in the presence of voids, and interestingly it can be reduced to values below the Seebeck value of the SL channel without NIs or voids that we began with (left-most data points in Figs. 6(b), (e), and (h). The reasons behind this counter-intuitive simultaneous reduction in both conductance and Seebeck coefficient will be discussed later on. It is important to note, however, that voids degrade the thermal conductivity drastically, compared to NIs [@Dunham16; @Taborda16]. Thus, despite the $\sim 50\%$ reduction in the $PF$, a large increase in the $ZT$ figure of merit is expected to be achieved in the void structures.
Note also that although in the presence of voids the conductance is degraded, the degradation is smaller when the Fermi level is placed at degenerate conditions, i.e., when $E_{\text{F}}$ is placed at the level of the SL barriers, in which case the carriers have higher velocities and are affected somewhat less. In Fig. 7(a) we show the transmission of the pristine channel and of the SL channel with/without voids, plotted versus the carrier energy. The blue/red lines correspond to the absence/presence of voids in a pristine channel (solid lines) and the channel with SL barriers (dashed lines), where the conduction opens after $V_{\text{SL}}$. The transmission in the SL case with or without voids (dashed lines), has a large slope after the energy crosses $V_{\text{SL}}$ (around the SL Fermi level), larger than the slope of the pristine channel (at $E=0$ eV, around the pristine channel Fermi level). In the pristine channel, upon the introduction of voids (red solid line), the transmission slope decreases significantly, starting and remaining close to zero for several meV above $E_{\text{F}}=0$ eV. Notice on the other hand, that in the SL case upon the introduction of voids (red-dashed line), the transmission is not degraded around $E = V_{\text{SL}}$ where the Fermi level is placed. As also seen earlier for the transmission of SL barriers in Fig. 5, this means that the conductance of the pore structures suffers less if we operate at degenerate conditions.
![\[fig7\] (Colour online) Comparison of the effect of voids on the transmission of the pristine channel (with $E_{\text{F}}$ at $0$ eV) and the SL channel, in which case the latter operates at highly degenerate conditions ($E_{\text{F}} = 0.05$ eV). (a) Transmission versus carrier energy of a pristine channel in the presence (red solid line) and in the absence (blue solid line) of voids. The dashed lines show the respective transmissions for a channel with SL barriers. (b) Transmission ratios $Tr_{\text{voids}} / Tr_{\text{prist}}$ and $Tr_{\text{SL+voids}} / Tr_{\text{SL}}$ plotted versus carrier energy. ](Figure_7.png "fig:"){width="11cm" height="11.5cm"}\
![\[fig8\] (Colour online) Transmission $Tr$ versus electron energy $E$ for the channel with SL wells and for the channel with SL wells plus NI wells between them for (a) elastic-AP scattering (b) inelastic-OP scattering, and (c) elastic/inelastic – AOP scattering regimes combined. The solid blue line shows results for the pristine channel, the green dashed line corresponds to the SL wells channel, and the green dashed-dotted line corresponds to the SL wells+NI channel, as indicated in the inset of (a). ](Figure_8.png "fig:"){width="8.5cm" height="15cm"}\
This is reflected more clearly in Fig. 7(b) where we show the ratio of the transmission of the pristine channel with voids to that of the pristine channel, $Tr_{\text{voids}} / Tr_{\text{prist}}$ (solid line) and the ratio of the transmission of the SL channels with voids to that of the SL channels, $Tr_{\text{SL+voids}} / Tr_{\text{SL}}$ (dashed line) plotted versus energy. In the SL channels the ratio of the transmissions starts from unity at $E = V_{\text{SL}}$, indicating the weak influence of the voids. In the pristine channel, on the other hand, the ratio begins at zero at $E=0$ eV, indicating that in this case the effect of voids is detrimental. This is quite important, indicating that the electronic conductance in highly disordered structures, which can slow down phonons significantly, can be less affected if they are operated at degenerate conditions, which will help the $PF$.
Making use of the observations in Fig. 7, we discuss briefly now the reason for which the values of the Seebeck coefficient of the SL channel with voids are smaller than those of the SL channel without voids (see Figs. 6(b), (e), and (h)). As can be observed in Fig. 7(a), for energies close to $V_{\text{SL}}$ (and also $E_\text{F}$) the SL+voids channel (red-dashed line) has similar transmission to that of the SL channel (blue-dashed line). However, at higher energies the transmission of the SL+voids channel grows with energy at a slower rate and it merges with the lower transmission of the pristine+voids channel. As a consequence the Seebeck coefficient, which is proportional to the slope of the transmission (in a similar manner that is proportional to the derivative of the density of states at the Fermi level), drops to lower values compared to those of the SL channel, i.e., the right-most points in Figs. 6(b), (e), and (h) are lower than the left-most points which correspond to $V_\text{N}=0$.
**Transport in the case of potential wells:** Finally, we examine the transport behavior of the structures in which the SL barriers and the NIs form potential wells. As indicated above in Fig. 3, the degradation in the conductance and in the $PF$ is very small, of the order of $\sim 1\%$. In Figs. 8(a)-(c) we show the transmission function versus energy in the cases where only AP scattering is present, only OP scattering is present, and where both elastic and inelastic scatterings are present, respectively. In each sub-figure we show three cases: i) pristine channel, ii) SL structure, and iii) SL structure plus NIs, all forming potential wells for electrons. These geometries are depicted in the insets of Fig. 8(a). Under elastic scattering conditions (Fig. 8(a)), in all three cases the transmission functions are almost identical, indicating that the nanostructuring does not obscure electronic transport. This is expected as the carriers flow at higher energies compared to the well energy levels. Some quantum reflections, however, are always present [@Thes15JAP; @Foster17], and thus some minor signatures are evident in the transmission features. In the case of OP scattering conditions in Fig. 8(b), a large peak is observed in the pristine case at low energies due to the lack of OP emission (blue line). At energies higher than the phonon energies $\hbar \omega = 0.06$ eV, a sharp drop is encountered, a result of the fact that the electrons have enough energy now to emit a phonon and lose energy all the way to the band edge. In the case of SL structures (green-dashed line) and the SL plus NIs geometries (green-dashed-dotted line), the initial peak is slightly suppressed, since now electrons have narrow regions (of the size of the SL and NI wells, $\sim 3$ nm - $5$ nm) to emit phonons and move to lower energies. However, this process is weak and the transmission is not significantly changed, which also shows why the conductance and $PF$ do not degrade noticeably compared to the pristine structure. In the case of Fig. 8(c), where AOP scattering is considered, the transmission function among the three geometries differs again slightly, the degree being intermediate between the completely elastic and completely inelastic behavior. This directly reflects the fact that the $PF$ does not change significantly from the pristine case when the wells are introduced. We note here, that our model considers nanostructuring as causing simple shifts in the band edges. In reality, however, the effective masses of the NIs will vary, strain fields will build around them, charging effects and interface resistances will appear, and the phonon-scattering details will also change. These will most probably add an additional reduction in the transmission which needs to be examined more carefully taking into account material specific parameters. However, our results demonstrate that to first order we should not expect large conductivity reductions from potential wells, as also is the case observed in experiments [@Zhao14].
Discussion on optimal nanostructuring for high power factors
============================================================
In our simulations we have chosen the simplest possible system that can account for energy filtering through barriers and energy relaxation through wells (i.e., only 2 barriers) without strong $PF$ degradation We did not intent to optimize the power factor of the hierarchical structure, it could be that higher power factors can be obtained for different geometries, but that would require significantly more work. Our purpose was to show that hierarchical architectures could provide benefits for the power factor, using geometrical features that we believe are close to optimal. Thus, here we elaborate on the reasons some of our choices are justified, which could also serve as useful guidelines to experimentalists in the design of high power factor advanced nanostructured thermoelectric materials in the presence of energy filtering:
i\) *$E_\text{F}$ and $V_\text{B}$ choice:* Degenerate conditions (with $E_F$ into the bands) are beneficial to employ high velocity electrons, and in that case the barrier heights $V_\text{B}$ and the Fermi level $E_\text{F}$ need to be positioned at similar levels. Energy filtering from potential barriers increases the Seebeck coefficient, however, not so easily the power factor unless these conditions are satisfied.
ii\) *Distance between filtering barriers:* This needs to be large enough because the more closely spaced the barriers are, the larger the resistance and the lower the conductivity, but on the other hand short enough, such that carriers do have space to relax their energy completely (this reduces the Seebeck coefficient and conductivity). Therefore, the distance is determined by the energy relaxation mean-free-path $\lambda_\text{E}$ (which is typically larger than the momentum relaxation mean-free-path). In our systems we have chosen an energy relaxation mean-free-path of $15.5$nm, and thus, a distance between the barriers of $50$nm, which is almost 3.5 times larger and allows semi-relaxation of the carrier energy. Indeed, for similar simulation scattering parameters, in Ref. [@Thes16JAP] we showed that the distance between barriers is optimized at around $50$nm.
iii\) *Spacing between the nanoinclusions (NIs) within*
*the barriers:* From simple scattering mean-free-path considerations, it is common to assume that the NIs introduce a mean-free-path similar to the distance between them, and this needs to be combined with the momentum relaxation mean-free-path (not the energy relaxation-mean-free-path only), through Matthiessen’s rule. In our simulations, the distance between the NIs for the high density case is $d_{NI} = 6$nm, which is quite smaller compared to the mean-free-path of electron-phonon scattering. In principle, NIs degrade the power factor from its optimal value, and should be avoided if we only consider power factor improvements. However, they bring significant degradation in the thermal conductivity, for example in semiconductor materials for which the phonon mean-free path is $10$s-$100$s of nanometers. The important observation in this work, however, is that the degrading influence of NIs on the power factor of *hierarchical architectures* is suppressed (at a larger degree compared to material cases that do not include the SL barriers). This is because the SL barriers (in combination with elevated Fermi levels) utilize charge carriers of higher energies, which are less susceptible to scattering from NIs. The scattering rate of high energy/large wavevector carriers by potential barriers is weaker, especially when the NI barrier height $V_{\text{NI}}$ is lower compared to the carrier energies (or negative in the case of wells). Thus, the recommendation for practical design of such hierarchical geometries is that the degrading effects on the power factor will be suppressed even if the NIs are placed at distances smaller compared to the mean-free-path of charge carriers.
With regards to improving the performance of thermoelectric materials, we need to mention here that although the electronic conductivity can be designed to be immune to the presence of NIs to a large degree, this is not the case for the thermal conductivity. A large number of literature reports indicate that NIs indeed cause significant degradation in the thermal conductivity [@Dettori15; @Hahn_14; @Neo_14JEM]. The combination of these two effects could decouple the electrical with the thermal conductivities and improve the $ZT$ figure of merit. There are two reasons why NIs affect phonons more than electrons: i) The distance between NIs can be thought of as the mean-free-path for scattering on the NIs. From simple Matthiessen’s rule scattering rate combination, the carrier with the longer mean-free-path will experience the larger relative reduction in its conductivity from a given NI geometry. Therefore, the thermal conductivity, carried by phonons with dominant mean-free-paths in the $10$s-$100$s of nanometers (in common semiconductors like Si), will experience a stronger reduction compared to the electronic conductivity, where electrons have mean-free-paths of a few to $10$s of nanometers. ii) Scattering of electrons on NIs is caused by the potential barriers that the NIs form. The electrons that contribute to conductivity are located energetically in a narrow window around the Fermi level, which can be shifted at high energies, where carriers are less obscured (especially if the barrier height $V_{\text{NI}}$ is small or negative). Phonon scattering on the other hand, does not offer this degree of freedom in the design of thermal conductivity. At room temperature for example, most phonons in the spectrum contribute to transport, and since phonons are lattice vibrations, they are affected by lattice interruptions. Therefore, although different nanostructuring can affect phonons with different mean-free-paths differently, all phonons are affected by NIs and the SL barriers.
Summary and Conclusions
=======================
In this work we investigated the influence of hierarchical nanostructuring on the thermoelectric coefficients of nanomaterials, which are the primer candidates for achieving ultra-low thermal conductivities and high thermoelectric $ZT$ figures of merit. Using the fully quantum mechanical NEGF transport formalism, we studied systematically two-dimensional materials with embedded SL-type barriers/wells combined with quantum dot-like NIs and voids. We found novel effects and presented design strategies for such materials, and stated the conditions under which the $PF$ is not only immune to nanostructuring, but it can also be improved. In summary, we showed that: 1) Nanostructuring using superlattice-like potential barriers and nanoinclusions can have up to $20\%$ $PF$ improvements even at very high nanoinclusion densities, as long as the Fermi level is placed well into the bands and charge carrier relaxation is avoided; 2) Nanostructuring using potential wells causes only minor reduction in the $PF$, even at very high nanostructuring densities. Thus, designing the nanostructured geometry of such materials should take into account the energy resolved mean-free-path of carriers, as well as their energy relaxation length caused by inelastic processes, in this case the optical phonon energies and the electron-optical phonon interaction strength. Such insight is currently not being explored in hierarchical nanostructured materials, where current strategies only focus on thermal conductivity reduction. It can, however, offer significant benefits to the thermoelectric figure of merit by simultaneously improving, or at least not degrading the power factor as well.
[**Acknowledgments**]{} This work has received funding from the European Research Council (ERC) under the European Union’s Horizon 2020 Research and Innovation Programme (Grant Agreement No. 678763). We thank Dr Mischa Thesberg for helping with the construction of the NEGF simulator. We also thank Samuel Foster and Dhritiman Chakraborty for helpful discussions.
The acoustic deformation potential scattering rates
===================================================
For elastic acoustic deformation potential (ADP) scattering, where $\hbar \omega \rightarrow 0$, we use the commonly employed equipartition approximation. This results in the acoustic scattering rates to become proportional to the density of final states at the energy of the electronic state under consideration with the proportionality constant determined by the acoustic phonon deformation potential $D_{\text{A}}$, the temperature, and other material parameters. This process is described in detail in Ref. [@Lundstrom], whereas the connection of the constant $D_{\text{AP}}$ used within NEGF to the actual deformation potential $D_{\text{A}}$ is presented in Ref. [@Koswatta07].
The ADP scattering rate is determined by [@Lundstrom]: $$\frac{1}{\tau} = \frac{m^\ast D_{\text{A}}^2}{4 \pi \hbar \rho v_{\text{s}} p} \int_{\beta_{min}}^{\beta_{max}}
\left ( n_{\text{B}} + \frac{1}{2} \mp \frac{1}{2} \right ) \beta^2 d \beta ,
\label{eqA1}$$ where $m^\ast$ is the effective mass of the material, $D_{\text{A}}$ is the deformation potential, $\rho$ is the mass density, $v_{\text{s}}$ is the sound velocity, and $p$ is the carrier momentum. $n_{\text{B}}$ is the number of phonons, determined by the Bose-Einstein distribution, and the integration is performed over all phonon wavevectors that participate in ADP scattering. Since the number of acoustic phonons at room temperature is large so that $n_{\text{B}} \simeq n_{\text{B}}+1$, and because $k_{\text{B}} T \gg \hbar \omega$, we can use the equipartition approximation $n_{\text{B}} \approx k_{\text{B}} T / \hbar \omega$. Taking also into account that the phonon dispersion is linear for acoustic phonons $(\omega = \upsilon_{\text{s}}\beta)$ the above equation can be simplified as $$\frac{1}{\tau} = \frac{m^\ast D_{\text{A}}^2 k_{\text{B}} T}{2 \pi \hbar^2 \rho v_{\text{s}}^2 p} \int_{\beta_{min}}^{\beta_{max}}
\beta d \beta ,
\label{eqA2}$$ where a factor of $2$ has been inserted due to both emission and absorption processes. In order to ensure momentum and energy conservation, $\beta_{min} = 0$ and $\hbar \beta_{max} = 2 p$, and therefore we get $$\frac{1}{\tau} = \frac{m^\ast D_{\text{A}}^2 k_{\text{B}} T}{\pi \hbar^2 \rho v_{\text{s}}^2} \frac{k}{\hbar} ,
\label{eqA3}$$ where $k = p / \hbar$ is the electron wave vector. Using $k = \left ( 2 m^\ast E \right )^{1/2} / \hbar$, and the density of states, $g(E)$, in 3D: $$g(E) = \frac{1}{2 \pi^2} \left ( \frac{2 m^\ast}{\hbar^2} \right )^{3/2} E^{1/2} ,
\label{eqA4}$$ we can finally write the rate equation as: $$\frac{1}{\tau} = \frac{\pi D_{\text{A}}^2 k_{\text{B}} T}{\hbar \rho v_{\text{s}}^2} g(E) .
\label{eqA5}$$ Thus, the ADP scattering rate can be approximated with a constant times the density-of-states, which is standard practice.
Seebeck coefficient as average energy of the current flow
=========================================================
In order to extract the Seebeck coefficient $S$ we proceed as follows. In the Boltzmann transport formalism, the Seebeck coefficient is given by: $$S = \frac{k_{\text{B}}}{q \sigma} \int_{-\infty}^{\infty} dE \left( - \frac{\partial f}{\partial E} \right) \Xi (E) \left( \frac{E - E_{F}}{k_B T} \right) ,
\label{eqB1}$$ where $\Xi(E)$ is the transport distribution function. In terms of the energy and position resolved current $I_{ch}(E,x)$ the Seebeck coefficient can be expressed as $$S^{\prime}(x) = \frac{1}{q I_{ch}} \int_{-\infty}^{\infty} dE I_{ch}(E,x) \left( \frac{E - E_F}{T} \right) ,
\label{eqB2}$$ which can be rewritten as $$\begin{aligned}
\nonumber S^{\prime}(x) = \frac{1}{q T} \left( \frac{1}{I_{ch}} \int_{-\infty}^{\infty} I_{ch}(E,x) E dE - E_F \right)
\\* &&\hspace*{-2.49in} = \frac{1}{qT} \langle E(x) - E_F \rangle ,
\label{eqB3}$$ where the integral in the last equation is the definition of the average energy of the current. The same derivation can be expressed in terms of the transmission by using the definition: $Tr_n(E) = (W/L) 2 \pi\hbar\Xi_n(E)$. The total Seebeck coefficient of the channel is then given as $$S = \frac{1}{L} \int_{0}^{L} S^{\prime} ( x ) dx = \frac{1}{q T L} \int_{0}^{L} \langle E(x) - E_F \rangle dx ,
\label{eqB4}$$ where $q$ is the carrier charge ($q=-\vert e \vert$ for electrons and $q=\vert e \vert$ for holes) and $\langle E ( x ) \rangle$ is the energy of the current flow along the transport direction as given in Eq. (\[eq14\]).
[9]{}
G. Ding, G. Y. Gao, Z. Huang, W. Zhang and K. Yao, Nanotechnology **27**, 375703 (2016).
H. Huang, Y. Cui, Q. Li, C. Dun, W. Zhou, W. Huang, L. Chen, C. A. Hewitt, and D. L. Carroll, Nano Energy **26**, 172 (2016).
W. Huang, X. Luo, C. K. Gan, S. Y. Quek, and G. Liang, Phys. Chem. Chem. Phys. **16**, 10866 (2014).
Y. Tang, R. Hanus, S.-W. Chen, and G. J. Snyder, Nat. Commun. **6**, 7584 (2015).
Y. Tang, Z. M. Gibbs, L. A. Agapito, G. Li, H.-S. Kim, M. B. Nardelli, S. Curtarolo, and G. J. Snyder, Nat. Mater. **14**, 1223 (2015).
M. Beekman, D. T. Morelli, and G. S. Nolas, Nat. Mater. **14**, 1182 (2015).
R. Stern, B. Dongre, and G. K. H. Madsen, Nanotechnology **27**, 334002 (2016).
C. Fu, S. Bai, Y. Liu, Y. Tang, L. Chen, X. Zhao, and T. Zhu, Nat. Commun. **6**, 8144 (2015).
Y. Yin, B. Tudu, and A. Tiwari, Vacuum **146**, 356 (2017).
D. Beretta, N. Neophytou, J. M. Hodges, M. G. Kanatzidis, D. Narducci, M. M.- Gonzalez, M. Beekman, B. Balke, G. Cerrettii, W. Tremel *et al.,* Mater. Sci. Eng. R Rep. (to be published).
H. Mizuno, S. Mossa, and J.-L. Barrat, Sci.Rep. **5**, 14116 (2015).
X. W. Wang, H. Lee, Y. C. Lan, G. H. Zhu, G. Joshi, D. Z. Wang, J. Yang, A. J. Muto, M. Y. Tang, J. Klatsky, S. Song, M. S. Dresselhaus, G. Chen, and Z. F. Ren, Appl. Phys. Lett. **93**, 193121 (2008).
H. Ikeda and F. Salleh, Appl. Phys. Lett. **96**, 012106 (2010).
M. Verdier, K. Termentzidis, and D. Lacroix, J. Appl. Phys. **119**, 175104 (2016).
R. Yang, G. Chen, and M. S. Dresselhaus, Phys. Rev. B **72**, 125418 (2005).
J.-H. Lee, G. A. Galli, and J. C. Grossman, Nano Lett. **8**, 3750 (2008).
N. Neophytou, X. Zianni, H. Kosina, S. Frabboni, B. Lorenzi, and D. Narducci, Nanotechnology **24**, 205402 (2013).
N. S. Bennett, D. Byrne, A. Cowley, and N. Neophytou, Appl. Phys. Lett. **109**, 173905 (2016).
K. Biswas, J. He, I. D. Blum, C.-I. Wu, T. P. Hogan, D. N. Seidman, V. P. Dravid, and M. G. Kanatzidis, Nature **489**, 414 (2012).
C. Gayner and K. K. Kar, Prog. Mater. Sci. **83**, 330 (2016).
T. Zou, X. Qin, Y. Zhang, X. Li, Z. Zeng, D. Li, J. Zhang, H. Xin, W. Xie, and A. Weidenkaff, Sci. Rep. **5**, 17803 (2015).
P. E. Hopkins, C. M. Reinke, M. F. Su, R. H. Olsson, E. A. Shaner, Z. C. Leseman, J. R. Serrano, L. M. Phinney, and I. El-Kady, Nano Lett. **11**, 107 (2011).
C. J. Vineis, A. Shakouri, A. Majumdar, and M. G. Kanatzidis, Adv. Mater. **22**, 3970 (2010).
A. Popescu, L. Woods, J. Martin, and G. Nolas, Phys. Rev. B **79**, 205302 (2009).
B. Qiu, H. Bao, G. Zhang, Y. Wu, and X. Ruan, Comput. Mater. Sci. **53**, 278 (2012).
S. Fan, J. Zhao, Q. Yan, J. Ma, and H. H. Hng, J. Electron. Mater. **40**, 1018 (2011).
M. K. Keshavarz, D. Vasilevskiy, R. A. Masut, and S. Turenne, Mater. Charact. **95**, 44 (2014).
J. P. Heremans, C. M. Thrush, and D. T. Morelli, J. Appl. Phys. **98**, 063703 (2005).
K. F. Hsu, S. Loo, F. Guo, W. Chen, J. S. Dyck, C. Uher, T. P. Hogan, E. K. Polychroniadis, and M. G. Kanatzidis, Science **303**, 818 (2004).
S. Ahmad, A. Singh, A. Bohra, R. Basu, S. Bhattacharya, R. Bhatt, K. N. Meshram, M. Roy, S. K. Sarkar, Y. Hayakawa, A. K. Debnath, D. K. Aswal, and S. K. Gupta, Nano Energy **27**, 282 (2016).
G. H. Zhu, H. Lee, Y. C. Lan, X. W. Wang, G. Joshi, D. Z. Wang, J. Yang, D. Vashaee, H. Guilbert, A. Pillitteri, M. S. Dresselhaus, G. Chen, and Z. F. Ren, Phys. Rev. Lett. **102**, 19 (2009).
M. Saleemi, A. Famengo, S. Fiameni, S. Boldrini, S. Battiston, M. Johnsson, M. Muhammed, and M. S. Toprak, J. Alloy. Comp. **619**, 31 (2015).
G. Tan, L. D. Zhao, F. Shi, J. W. Doak, S. H. Lo, H. Sun, C. Wolverton, V. P. Dravid, C. Uher, and M. G. Kanatzidis, J. Am. Chem. Soc. **136**, 7006 (2014).
A. Banik, B. Vishal, S. Perumal, R. Datta, and K. Biswas, Energy Environ. Sci. **9**, 2011 (2016).
G. Tan, F. Shi, S. Hao, Li-D. Zhao, H. Chi, X. Zhang, C. Uher, C. Wolverton, V. P. Dravid, and M. G. Kanatzidis, Nat. Commun. **7**, 12167 (2016).
B. T. Kearney, B. Jugdersuren, D. R. Queen, T. H. Metcalf, J. C. Culbertson, P. A. Desario, R. M. Stroud, W. Nemeth, Q. Wang, and X. Liu, J. Phys. Condens. Matter **30**, 085301 (2018).
L.-D. Zhao, V. P. Dravid, and M. G. Kanatzidis, Energy Environ. Sci. **7**, 251 (2014)
B. Lorenzi, D. Narducci, R. Tonini, S. Frabboni, G. C. Gazzadi, G. Ottaviani, N. Neophytou, and X. Zianni, J. Electron. Mater. **43**, 3812 (2014).
M. Thesberg, M. Pourfath, N. Neophytou, and H. Kosina, J. Electron. Mater. **45**, 1584 (2016).
M. Liu and X. Y. Qin, Appl. Phys. Lett. **101**, 132103 (2012).
S. Foster, M. Thesberg, and N. Neophytou, Phys. Rev. B **96**, 195425 (2017).
J. Peng, L. Fu, Q. Liu, M. Liu, J. Yang, D. Hitchcock, M. Zhou, and J. He, J. Mater. Chem. A **2**, 73 (2014).
J. Zhou and R. Yang, J. Appl. Phys. **110**, 084317 (2011).
S. O. Koswatta, S. Hasan, M. S. Lundstrom, M. P. Anantram, and D. E. Nikonov, IEEE Trans. Electron Devices **54**, 2339 (2007).
S. Datta, Quantum Transport: Atom to Transistor, Cambridge, UK: Cambridge Univ. Press, 2005.
M. P. Anantram, M. Lundstrom, and D. Nikonov, Proc. IEEE **96**, 1511 (2008).
V. Vargiamidis, S. Foster, and N. Neophytou, Phys. Status Solidi A **215**, 1700997 (2018).
M. P. L. Sancho, J. M. L. Sancho, and J. Rubio, J. Phys. F: Met. Phys. **15**, 851 (1985); **14**, 1205 (1984).
G. D. Mahan, Phys. Rep. **145**, 251 (1987).
N. Neophytou and H. Kosina, Phys. Rev. B **83**, 245305 (2011).
C. Jacoboni and L. Reggiani, Rev. Mod. Phys. **55**, 645 (1983).
M. Thesberg, M. Pourfath, H. Kosina, and N. Neophytou, J. Appl. Phys. **118**, 224301 (2015).
R. Kim, S. Datta, and M. S. Lundstrom, J. Appl. Phys. **105**, 034506 (2009).
R. Kim and M. Lundstrom, J. Appl. Phys. **110**, 034511 (2011).
N. Neophytou and H. Kosina, Nano Lett. **10**, 4913 (2010).
G. D. Mahan and J. O. Sofo, Proc. Natl. Acad. Sci. USA **93**, 7436 (1996).
T. J. Scheidemantel, C. A.-Draxl, T. Thonhauser, J. V. Badding, and J. O. Sofo, Phys. Rev. B **68**, 125210 (2003).
T. H. Zou, X. Y. Qin, D. Li, G. L. Sun, Y. C. Dou, Q. Q. Wang, B. J. Ren, J. Zhang, H. X. Xin, and Y. Y. Li, Appl. Phys. Lett. **104**, 013904 (2014).
S. Perumal, P. Bellare, U. S. Shenoy, U. V. Waghmare, and K. Biswas, Chem. Mater. **29**, 10426 (2017).
S. Ju, T. Shiga, L. Feng, and J. Shiomi, Phys. Rev. B **97**, 184305 (2018).
H. Honarvar and M. I. Hussein, Phys. Rev B **97**, 195413 (2018).
X. Yang and W. Li, Phys. Rev. Materials **2**, 015401 (2018).
D. P. Sellan, E. S. Landry, J. E. Turney, A. J. H. McGaughey, and C. H. Amon, Phys. Rev. B. **81**, 214305 (2010).
R. Dettori, C. Melis, X. Cartoixà, R. Rurali, and L. Colombo, Phys. Rev. B **91**, 054305 (2015).
M. T. Dunham, B. Lorenzi, S. C. Andrews, A. Sood, M. Asheghi, D. Narducci, and K. E. Goodson, Appl. Phys. Lett. **109**, 253104 (2016).
J. A. P.-Taborda, M. M.-Rojo, J. Maiz, N. Neophytou, and M. M. González, Nature Sci. Rep. **6**, 32778 (2016).
M. Thesberg, H. Kosina, and N. Neophytou, J. Appl. Phys. **120**, 234302 (2016).
K. R. Hahn, C. Melis, and L. Colombo, Eur. Phys. J. B 87:50 (2014).
N. Neophytou, X. Zianni, H. Kosina, S. Frabboni, B. Lorenzi, and D. Narducci, J. Elec. Mater. **43**, 1896 (2014).
M. Lundstrom, Fundamentals of Carrier Transport, Cambridge, UK: Cambridge Univ. Press, 2000.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In this expository and survey paper, along one of main lines of bounding the ratio of two gamma functions, we look back and analyse some inequalities, the complete monotonicity of several functions involving ratios of two gamma or $q$-gamma functions, the logarithmically complete monotonicity of a function involving the ratio of two gamma functions, some new bounds for the ratio of two gamma functions and divided differences of polygamma functions, and related monotonicity results.'
address: 'Research Institute of Mathematical Inequality Theory, Henan Polytechnic University, Jiaozuo City, Henan Province, 454010, China'
author:
- Feng Qi
date:
- 'Completed on 21 October 2008 at VU’s Student Village in Melbourne'
-
title: 'Bounds for the ratio of two gamma functions—From Gautschi’s and Kershaw’s inequalities to completely monotonic functions'
---
[^1]
[^2]
Introduction
============
For the sake of proceeding smoothly, we briefly introduce some necessary concepts and notation.
The gamma and $q$-gamma functions
---------------------------------
It is well-known that the classical Euler gamma function may be defined by $$\label{egamma}
\Gamma(x)=\int^\infty_0t^{x-1} e^{-t}\operatorname{d\mspace{-2mu}}t$$ for $x>0$. The logarithmic derivative of $\Gamma(x)$, denoted by $\psi(x)=\frac{\Gamma'(x)}{\Gamma(x)}$, is called the psi or digamma function, and $\psi^{(k)}(x)$ for $k\in \mathbb{N}$ are called the polygamma functions. It is common knowledge that special functions $\Gamma(x)$, $\psi(x)$ and $\psi^{(k)}(x)$ for $k\in\mathbb{N}$ are fundamental and important and have much extensive applications in mathematical sciences.
The $q$-analogues of $\Gamma$ and $\psi$ are defined [@andrews pp. 493–496] for $x>0$ by $$\begin{gathered}
\label{q-gamma-dfn}
\Gamma_q(x)=(1-q)^{1-x}\prod_{i=0}^\infty\frac{1-q^{i+1}}{1-q^{i+x}},\quad 0<q<1,\\
\label{q-gamma-dfn-q>1}
\Gamma_q(x)=(q-1)^{1-x}q^{\binom{x}2}\prod_{i=0}^\infty\frac{1-q^{-(i+1)}}{1-q^{-(i+x)}}, \quad q>1,\end{gathered}$$ and $$\begin{aligned}
\label{q-gamma-1.4}
\psi_q(x)=\frac{\Gamma_q'(x)}{\Gamma_q(x)}&=-\ln(1-q)+\ln q \sum_{k=0}^\infty\frac{q^{k+x}}{1-q^{k+x}}\\
&=-\ln(1-q)-\int_0^\infty\frac{e^{-xt}}{1-e^{-t}}\operatorname{d\mspace{-2mu}}\gamma_q(t) \label{q-gamma-1.5}\end{aligned}$$ for $0<q<1$, where $\operatorname{d\mspace{-2mu}}\gamma_q(t)$ is a discrete measure with positive masses $-\ln q$ at the positive points $-k\ln q$ for $k\in\mathbb{N}$, more accurately, $$\gamma_q(t)=
\begin{cases}
-\ln q\sum\limits_{k=1}^\infty\delta(t+k\ln q),&0<q<1,\\ t,&q=1.
\end{cases}$$ See [@Ismail-Muldoon-119 p. 311].
The $q$-gamma function $\Gamma_q(z)$ has the following basic properties: $$\lim_{q\to1^+}\Gamma_q(z)=\lim_{q\to1^-}\Gamma_q(z)=\Gamma(z)\quad \text{and}\quad \Gamma_q(x)=q^{\binom{x-1}2}\Gamma_{1/q}(x).$$
The generalized logarithmic mean
--------------------------------
The generalized logarithmic mean $L_p(a,b)$ of order $p\in\mathbb{R}$ for positive numbers $a$ and $b$ with $a\ne b$ may be defined [@bullenmean p. 385] by $$L_p(a,b)=
\begin{cases}
\left[\dfrac{b^{p+1}-a^{p+1}}{(p+1)(b-a)}\right]^{1/p},&p\ne-1,0;\\[1em]
\dfrac{b-a}{\ln b-\ln a},&p=-1;\\[1em]
\dfrac1e\left(\dfrac{b^b}{a^a}\right)^{1/(b-a)},&p=0.
\end{cases}$$ It is well-known that $$\begin{gathered}
L_{-2}(a,b) =\sqrt{ab}\,=G(a,b),\quad L_{-1}(a,b)=L(a,b),\\
L_0(a,b)=I(a,b)\quad \text{and} \quad L_1(a,b)=\frac{a+b}2=A(a,b)\end{gathered}$$ are called respectively the geometric mean, the logarithmic mean, the identric or exponential mean, and the arithmetic mean. It is also known [@bullenmean pp. 386–387, Theorem 3] that the generalized logarithmic mean $L_p(a,b)$ of order $p$ is increasing in $p$ for $a\ne b$. Therefore, inequalities $$\label{mean-ineq}
G(a,b)<L(a,b)<I(a,b)<A(a,b)$$ are valid for $a>0$ and $b>0$ with $a\ne b$. See also [@abstract-jipam; @abstract-rgmia; @qi1]. Moreover, the generalized logarithmic mean $L_p(a,b)$ is a special case of $E(r,s;x,y)$ defined by , that is, $L_p(a,b)=E(1,p+1;a,b)$.
Logarithmically completely monotonic functions
----------------------------------------------
A function $f$ is said to be completely monotonic on an interval $I$ if $f$ has derivatives of all orders on $I$ and $$\label{cmf-dfn-ineq}
(-1)^{n}f^{(n)}(x)\ge0$$ for $x \in I$ and $n \ge0$.
\[p.161-widder\] A necessary and sufficient condition that $f(x)$ should be completely monotonic for $0<x<\infty$ is that $$f(x)=\int_0^\infty e^{-xt}\operatorname{d\mspace{-2mu}}\alpha(t),$$ where $\alpha(t)$ is nondecreasing and the integral converges for $0<x<\infty$.
\[p.83-bochner\] If $f(x)$ is completely monotonic on $I$, $g(x)\in I$, and $g'(x)$ is completely monotonic on $(0,\infty)$, then $f(g(x))$ is completely monotonic on $(0,\infty)$.
A positive function $f(x)$ is said to be logarithmically completely monotonic on an interval $I\subseteq\mathbb{R}$ if it has derivatives of all orders on $I$ and its logarithm $\ln f(x)$ satisfies $$(-1)^k[\ln f(x)]^{(k)}\ge0$$ for $k\in\mathbb{N}$ on $I$.
The notion “logarithmically completely monotonic function” was first put forward in [@Atanassov] without an explicit definition. This terminology was explicitly recovered in [@minus-one] whose revised and expanded version was formally published as [@minus-one.tex-rev].
It has been proved once and again in [@CBerg; @clark-ismail-nonlinear; @clark-ismail-rgmia; @compmon2; @absolute-mon.tex; @minus-one; @minus-one.tex-rev; @schur-complete] that a logarithmically completely monotonic function on an interval $I$ must also be completely monotonic on $I$. C. Berg points out in [@CBerg] that these functions are the same as those studied by Horn [@horn] under the name infinitely divisible completely monotonic functions. For more information, please refer to [@CBerg; @e-gam-rat-comp-mon; @auscm-rgmia] and related references therein.
Outline of this paper
---------------------
In this expository and survey paper, along one of main lines of bounding the ratio of two gamma functions, we look back and analyse Gautschi’s double inequality and Kershaw’s second double inequality, the complete monotonicity of several functions involving ratios of two gamma or $q$-gamma functions by Alzer, Bustoz-Ismail, Elezović-Giordano-Pečarić and Ismail-Muldoon, the logarithmically complete monotonicity of a function involving the ratio of two gamma functions, some new bounds for the ratio of two gamma functions and the divided differences of polygamma functions, and related monotonicity results by Batir, Elezović-Pečarić, Qi and others.
Gautschi’s and Kershaw’s double inequalities
============================================
In this section, we begin with the papers [@gaut; @kershaw] to introduce a kind of inequalities for bounding the ratio of two gamma functions.
Gautschi’s double inequalities
------------------------------
The first result of the paper [@gaut] was the double inequality $$\label{gaut-3-ineq}
\frac{(x^p+2)^{1/p}-x}2<e^{x^p}\int_x^\infty e^{-t^p}\operatorname{d\mspace{-2mu}}t\le c_p\biggl[\biggl(x^p+\frac1{c_p}\biggr)^{1/p}-x\biggr]$$ for $x\ge0$ and $p>1$, where $$c_p=\biggl[\Gamma\biggl(1+\frac1p\biggr)\biggr]^{p/(p-1)}$$ or $c_p=1$. By an easy transformation, the inequality was written in terms of the complementary gamma function $$\Gamma(a,x)=\int_x^\infty e^{-t}t^{a-1}\operatorname{d\mspace{-2mu}}t$$ as $$\label{gaut-4-ineq}
\frac{p[(x+2)^{1/p}-x^{1/p}]}2<e^x\Gamma\biggl(\frac1p,x\biggr)\le pc_p\biggl[\biggl(x+\frac1{c_p}\biggr)^{1/p}-x^{1/p}\biggr]$$ for $x\ge0$ and $p>1$. In particular, if letting $p\to\infty$, the double inequality $$\frac12\ln\biggl(1+\frac2x\biggr)\le e^xE_1(x)\le\ln\biggl(1+\frac1x\biggr)$$ for the exponential integral $E_1(x)=\Gamma(0,x)$ for $x>0$ was derived from , in which the bounds exhibit the logarithmic singularity of $E_1(x)$ at $x=0$. As a direct consequence of the inequality for $p=\frac1s$, $x=0$ and $c_p=1$, the following simple inequality for the gamma function was deduced: $$\label{gaut-none-ineq}
2^{s-1}\le\Gamma(1+s)\le1,\quad 0\le s\le 1.$$
The second result of the paper [@gaut] was a sharper and more general inequality $$\label{gaut-6-ineq}
e^{(s-1)\psi(n+1)}\le\frac{\Gamma(n+s)}{\Gamma(n+1)}\le n^{s-1}$$ for $0\le s\le1$ and $n\in\mathbb{N}$ than . It was obtained by proving that the function $$f(s)=\frac1{1-s}\ln\frac{\Gamma(n+s)}{\Gamma(n+1)}$$ is monotonically decreasing for $0\le s<1$ and that $$\lim_{s\to1^-}f(s)=-\lim_{s\to1^-}\psi(n+s)=-\psi(n+1).$$
For more information on refining the inequality , please refer to [@incom-gamma-L-N; @qi-senlin-mia; @Qi-Mei-99-gamma] and related references therein.
The left-hand side inequality in can be rearranged as $$\label{gaut-ineq-1}
\frac{\Gamma(n+s)}{\Gamma(n+1)}\exp((1-s)\psi(n+1))\ge1$$ or $$\label{gaut-ineq-2-exp}
\biggl[\frac{\Gamma(n+s)}{\Gamma(n+1)}\biggr]^{1/(s-1)}e^{-\psi(n+1)}\le1$$ for $n\in\mathbb{N}$ and $0\le s\le 1$. Since the limit $$\lim_{n\to\infty}\biggl\{\biggl[\frac{\Gamma(n+s)}{\Gamma(n+1)}\biggr]^{1/(s-1)} e^{-\psi(n+1)}\biggr\}=1$$ can be verified by using Stirling’s formula in [@abram p. 257, 6.1.38]: For $x>0$, there exists $0<\theta<1$ such that $$\label{Stirling-formula}
\Gamma(x+1)=\sqrt{2\pi}\,x^{x+1/2}\exp\biggl(-x+\frac{\theta}{12x}\biggr),$$ it is natural to guess that the function $$\label{gaut-funct}
\biggl[\frac{\Gamma(x+s)}{\Gamma(x+1)}\biggr]^{1/(s-1)}e^{-\psi(x+1)}$$ for $0\le s<1$ is possibly increasing with respect to $x$ on $(-s,\infty)$.
For information on the study of the right-hand side inequality in , please refer to [@bounds-two-gammas.tex; @Wendel-Gautschi-type-ineq.tex; @Wendel2Elezovic.tex] and a great amount of related references therein.
Kershaw’s second double inequality and its proof {#kershaw-sec}
------------------------------------------------
In 1983, over twenty years later after the paper [@gaut], among other things, D. Kershaw was motivated by the left-hand side inequality in [@gaut] and presented in [@kershaw] the following double inequality for $0<s<1$ and $x>0$: $$\begin{gathered}
\label{gki2}
\exp\big[(1-s)\psi\big(x+\sqrt{s}\,\big)\big] <\frac{\Gamma(x+1)}{\Gamma(x+s)}
<\exp\biggl[(1-s)\psi\biggl(x+\frac{s+1}2\biggr)\biggr].\end{gathered}$$ It is called in the literature Kershaw’s second double inequality.
Define the function $f_\alpha$ by $$\label{kershaw-f-dfn}
f_\alpha(x)=\frac{\Gamma(x+1)}{\Gamma(x+s)}\exp((s-1)\psi(x+\alpha))$$ for $x>0$ and $0<s<1$, where the parameter $\alpha$ is to be determined.
It is not difficult to show, with the aid of Stirling’s formula, that $$\label{kershaw-2.3}
\lim_{x\to\infty}f_\alpha(x)=1.$$
Now let $$\label{kershaw-F-dfn}
F(x)=\frac{f_\alpha(x)}{f_\alpha(x+1)}=\frac{x+s}{x+1}\exp\frac{1-s}{x+\alpha}.$$ Then $$\frac{F'(x)}{F(x)}=(1-s)\frac{(\alpha^2-s)+(2\alpha-s-1)x}{(x+1)(x+s)(x+\alpha)^2}.$$ It is easy to show that
1. if $\alpha=s^{1/2}$, then $F'(x)<0$ for $x>0$;
2. if $\alpha=\frac{s+1}2$, then $F'(x)>0$ for $x>0$.
Consequently if $\alpha=s^{1/2}$ then $F$ strictly decreases, and since $F(x)\to1$ as $x\to\infty$ it follows that $F(x)>1$ for $x>0$. But, from , this implies that $f_\alpha(x)>f_\alpha(x+1)$ for $x>0$, and so $f_\alpha(x)>f_\alpha(x+n)$. Take the limit as $n\to\infty$ to give the result that $f_\alpha(x)>1$, which can be rewritten as the left-hand side inequality in . The corresponding upper bound can be verified by a similar argument when $\alpha=\frac{s+1}2$, the only difference being that in this case $f_\alpha$ strictly increases to unity.
The idea contained in the above stated proof of was also utilized by other mathematicians. For detailed information, please refer to related contents and references in [@bounds-two-gammas.tex].
The inequality can be rearranged as $$\label{gki2-rew-1}
\frac{\Gamma(x+s)}{\Gamma(x+1)}\exp\big[(1-s)\psi\big(x+\sqrt{s}\,\big)\big] <1 <\frac{\Gamma(x+s)}{\Gamma(x+1)}\exp\biggl[(1-s)\psi\biggl(x+\frac{s+1}2\biggr)\biggr]$$ or $$\begin{gathered}
\label{gki2-rew-2}
\biggl[\frac{\Gamma(x+s)}{\Gamma(x+1)}\biggr]^{1/(s-1)}\exp\big[-\psi\big(x+\sqrt{s}\,\big)\big] >1\\
>\biggl[\frac{\Gamma(x+s)}{\Gamma(x+1)}\biggr]^{1/(s-1)}\exp\biggl[-\psi\biggl(x+\frac{s+1}2\biggr)\biggr]. \end{gathered}$$ By Stirling’s formula , we can prove that $$\lim_{x\to\infty}\biggl\{\biggl[\frac{\Gamma(x+s)}{\Gamma(x+1)}\biggr]^{1/(s-1)} \exp\big[-\psi\big(x+\sqrt{s}\,\big)\big]\biggr\}=1$$ and $$\lim_{x\to\infty}\biggl\{\biggl[\frac{\Gamma(x+s)}{\Gamma(x+1)}\biggr]^{1/(s-1)} \exp\biggl[-\psi\biggl(x+\frac{s+1}2\biggr)\biggr]\biggr\}=1.$$ These clues make us to conjecture that the functions in the very ends of inequalities and are perhaps monotonic with respect to $x$ on $(0,\infty)$.
Several complete monotonicity results
=====================================
The complete monotonicity of the functions in the very ends of inequalities were first demonstrated in [@Bustoz-and-Ismail], and then several related functions were also proved in [@Alzer1; @egp; @laj-7.pdf] to be (logarithmically) completely monotonic.
Bustoz-Ismail’s complete monotonicity results {#Bustoz-Ismail-sec}
---------------------------------------------
In 1986, motivated by the double inequality and other related inequalities, J. Bustoz and M. E. H. Ismail revealed in [@Bustoz-and-Ismail Theorem 7 and Theorem 8] that
1. the function $$\label{bustol-ismail-AM}
\frac{\Gamma(x+s)}{\Gamma(x+1)}\exp\biggl[(1-s)\psi\biggl(x+\frac{s+1}2\biggr)\biggr]$$ for $0\le s\le1$ is completely monotonic on $(0,\infty)$; When $0<s<1$, the function satisfies $(-1)^nf^{(n)}(x)>0$ for $x>0$;
2. the function $$\label{bustol-ismail-AMM}
\frac{\Gamma(x+1)}{\Gamma(x+s)}\exp\bigl[(s-1)\psi\bigl(x+s^{1/2}\bigr)\bigr]$$ for $0<s<1$ is strictly decreasing on $(0,\infty)$.
The proof of the complete monotonicity of the function in [@Bustoz-and-Ismail Theorem 7] relies on the inequality $$\label{lemma3.1-ism}
(y+a)^{-n}-(y+b)^{-n}>(b-a)n\biggl(y+\frac{a+b}2\biggr)^{-n-1},\quad n>0$$ for $y>0$ and $0<a<b$, the series representation $$\label{series-repr}
\psi(x)=-\gamma-\frac1x+\sum_{n=1}^\infty\biggl(\frac1n-\frac1{x+n}\biggr)$$ in [@er p. 15], and the above Theorem \[p.83-bochner\] applied to $f(x)=e^{-x}$.
The inequality verified in [@Bustoz-and-Ismail Lemma 3.1] can be rewritten as $$\label{lemma3.1-ism-rew}
\biggl[\frac1{-n}\cdot\frac{(y+a)^{-n}-(y+b)^{-n}}{(y+a)-(y+b)}\biggr]^{1/[(-n)-1]}
<\frac{(y+a)+(y+b)}2,\quad n>0$$ for $y>0$ and $0<a<b$, which is equivalent to $$\label{E-E(1,2)}
E(1,-n;y+a,y+b)<E(1,2;y+a,y+b),$$ where $E(r,s;x,y)$ stands for extended mean values and is defined for two positive numbers $x$ and $y$ and two real numbers $r$ and $s$ by $$\begin{aligned}\label{emv-dfn}
E(r,s;x,y)&=\biggl(\frac{r}{s}\cdot\frac{y^s-x^s}
{ y^r-x^r}\biggr)^{{1/(s-r)}}, & rs(r-s)(x-y)&\ne 0; \\
E(r,0;x,y)&=\biggl(\frac{1}{r}\cdot\frac{y^r-x^r}
{\ln y-\ln x}\biggr)^{{1/r}}, & r(x-y)&\ne 0; \\
E(r,r;x,y)&=\frac1{e^{1/r}}\biggl(\frac{x^{x^r}}{y^{y^r}}\biggr)^{ {1/(x^r-y^r)}},& r(x-y)&\ne 0; \\
E(0,0;x,y)&=\sqrt{xy}, & x&\ne y; \\
E(r,s;x,x)&=x, & x&=y.
\end{aligned}$$ Actually, the inequality is an immediate consequence of monotonicity of $E(r,s;x,y)$, see [@ls2]. For more information, please refer to and related references therein.
The proof of the decreasing monotonicity of the function just used the formula and and the above Theorem \[p.83-bochner\] applied to $f(x)=e^{-x}$.
Indeed, J. Bustoz and M. E. H. Ismail had proved in [@Bustoz-and-Ismail Theorem 7] that the function is logarithmically completely monotonic on $(0,\infty)$ for $0\le s\le1$. However, because the inequality strictly holds for a completely monotonic function $f$ on $(a,\infty)$ unless $f(x)$ is constant (see [@Dubourdieu p. 98], [@e-gam-rat-comp-mon p. 82] and [@haerc1]), distinguishing between the cases $0\le s\le1$ and $0<s<1$ is not necessary.
Alzer’s and related complete monotonicity results {#alzer-comp-sec}
-------------------------------------------------
Stimulated by the complete monotonicity obtained in [@Bustoz-and-Ismail], including those mentioned above, H. Alzer obtained in [@Alzer1 Theorem 1] that the function $$\label{alzer-func}
\frac{\Gamma(x+s)}{\Gamma(x+1)}\cdot\frac{(x+1)^{x+1/2}}{(x+s)^{x+s-1/2}}
\exp\biggl[s-1+\frac{\psi'(x+1+\alpha)-\psi'(x+s+\alpha)}{12}\biggr]$$ for $\alpha>0$ and $s\in(0,1)$ is completely monotonic on $(0,\infty)$ if and only if $\alpha\ge\frac12$, so is the reciprocal of for $\alpha\ge0$ and $s\in(0,1)$ if and only if $\alpha=0$.
As consequences of the monotonicity of the function , the following inequalities are deduced in [@Alzer1 Corollary 1 and Corollary 2]:
1. The inequalities $$\label{alzer-fun-ineq}
\begin{gathered}
\exp\biggl[s-1+\frac{\psi'(x+1+\beta) -\psi'(x+s+\beta)}{12}\biggr] \le\frac{(x+s)^{x+s-1/2}}{(x+1)^{x+1/2}}\cdot\frac{\Gamma(x+1)}{\Gamma(x+s)}\\
\le \exp\biggl[s-1+\frac{\psi'(x+1+\alpha) -\psi'(x+s+\alpha)}{12}\biggr],\quad \alpha>\beta\ge0
\end{gathered}$$ are valid for all $s\in(0,1)$ and $x\in(0,\infty)$ if and only if $\beta=0$ and $\alpha\ge\frac12$.
2. If $$a_n=\frac32\biggl\{1+\ln\biggl[\frac{2[\Gamma((n+1)/2)]^2} {[\Gamma(n/2)]^2}\cdot{n^{n-1}}{(n+1)^n}\biggr]\biggr\},$$ then $$\label{sum-alzer-ineq}
a_n<(-1)^{n+1}\Biggl[\frac{\pi^2}{12}-\sum_{k=1}^n(-1)^{k+1}\frac1{k^2}\Biggr]<a_{n+1},\quad n\in\mathbb{N}.$$
The inequality follows from the formula $$\frac14\biggl[\psi'\biggl(\frac{n}2+1\biggr)-\psi'\biggl(\frac{n+1}2\biggr)\biggr] =\sum_{k=1}^\infty\frac{(-1)^k}{(n+k)^2} =(-1)^{n}\Biggl[\sum_{k=1}^n\frac{(-1)^{k+1}}{k^2}-\frac{\pi^2}2\Biggr]$$ and the inequality applied to $s=\frac12$, $\alpha=\frac12$ and $\beta=0$.
The proof of the complete monotonicity of the function in [@Alzer1] is based on Theorem \[p.83-bochner\] applied to $f(x)=e^{-x}$, the formulas $$\frac1x=\int_0^\infty e^{-xt}\operatorname{d\mspace{-2mu}}t,\quad \ln\frac{y}x=\int_0^\infty\frac{e^{-xt}-e^{-yt}}t\operatorname{d\mspace{-2mu}}t$$ and $$\psi(x)=-\gamma+\int_0^\infty\frac{e^{-t}-e^{-xt}}{1-e^{-t}}\operatorname{d\mspace{-2mu}}t$$ for $x,y>0$, and discussing the positivity of the functions $$\frac{12-t^2e^{-\alpha t}}{12(1-e^{-t})}-\frac12-\frac1t\quad\text{and}\quad
\frac12+\frac1t-\frac{12-t^2}{12(1-e^{-t})}$$ for $x\in(0,\infty)$ and $\alpha\ge\frac12$. Therefore, H. Alzer essentially gave in [@Alzer1 Theorem 1] necessary and sufficient conditions for the function to be logarithmically completely monotonic on $(0,\infty)$.
In [@laj-7.pdf Theorem 3], a slight extension of [@Alzer1 Theorem 1] was presented: The function $$\label{li-ext-fun}
\frac{\Gamma(x+s)}{\Gamma(x+t)}\cdot\frac{(x+t)^{x+t-1/2}}{(x+s)^{x+s-1/2}}
\exp\biggl[s-t+\frac{\psi'(x+t+\alpha) -\psi'(x+s+\alpha)}{12}\biggr]$$ for $0<s<t$ and $x\in(0,\infty)$ is logarithmically completely monotonic if and only if $\alpha\ge\frac12$, so is the reciprocal of if and only if $\alpha=0$.
The decreasing monotonicity of and its reciprocal imply that the double inequality $$\begin{gathered}
\label{li-ext-fun-ineq}
\exp\biggl[t-s+\frac{\psi'(x+s+\beta) -\psi'(x+t+\beta)}{12}\biggr] \le\frac{(x+t)^{x+t-1/2}}{(x+s)^{x+s-1/2}}\cdot\frac{\Gamma(x+s)}{\Gamma(x+t)}\\
\le \exp\biggl[t-s+\frac{\psi'(x+s+\alpha) -\psi'(x+t+\alpha)}{12}\biggr]\end{gathered}$$ for $\alpha>\beta\ge0$ are valid for $0<s<t$ and $x\in(0,\infty)$ if and only if $\beta=0$ and $\alpha\ge\frac12$.
It is obvious that the inequality is a slight extension of the double inequality obtained in [@Alzer1 Corollary 2].
In [@Ismail-Muldoon-119 Theorem 3.4], the following complete monotonicity were established: Let $0<q<1$ and $$\label{g{alpha,q}(x)}
g_{\alpha,q}(x)=(1-q)^x(1-q^x)^{1/2}\Gamma_q(x) \exp\biggl[\frac{F(q^x)}{\ln q}-\frac{\psi_q'(x+\alpha)}{12}\biggr],$$ where $$F(x)=\sum_{n=1}^\infty\frac{x^n}{n^2}=-\int_0^x\frac{\ln(1-t)}{t}\operatorname{d\mspace{-2mu}}t.$$ Then $[\ln g_{\alpha,q}(x)]'$ is completely monotonic on $(0,\infty)$ for $\alpha\ge\frac12$, $-[\ln g_{\alpha,q}(x)]'$ is completely monotonic on $(0,\infty)$ for $\alpha\le0$, and neither is completely monotonic on $(0,\infty)$ for $0<\alpha<\frac12$.
As a consequence of [@Ismail-Muldoon-119 Theorem 3.4], the following result was deduced in [@Ismail-Muldoon-119 Corollary 3.5]: Let $0<q<1$, $0<s<1$ and $$\begin{split}\label{f-alpha-q}
f_{\alpha,q}(x)&=\frac{g_\alpha(x+s)}{g_\alpha(x+1)}\\
&=\frac{(1-q)^{s-1}(1-q^{x+s})^{1/2}\Gamma_q(x+s)}{(1-q^{x+1})^{1/2}\Gamma_q(x+1)}\\ &\quad\times\exp\biggl[\frac{F(q^{x+s})-F(q^{x+1})}{\ln q}+\frac{\psi_q'(x+1+\alpha)-\psi_q'(x+s+\alpha)}{12}\biggr].
\end{split}$$ Then $[\ln f_{\alpha,q}(x)]'$ is completely monotonic on $(0,\infty)$ for $\alpha\ge\frac12$, $-[\ln f_{\alpha,q}(x)]'$ is complete monotonic on $(0,\infty)$ for $\alpha\le0$, and neither is completely monotonic on $(0,\infty)$ for $0<\alpha<\frac12$.
Taking the limit $q\to1^-$ in yields [@Ismail-Muldoon-119 Corollary 3.6], a recovery of [@Alzer1 Theorem 1] mentioned above.
It is clear that [@laj-7.pdf Theorem 3] can be derived by taking the limit $$\lim_{q\to1^-}\frac{g_\alpha(x+s)}{g_\alpha(x+t)}$$ for $0<s<t$, where $g_\alpha(x)$ is defined by .
Ismail-Muldoon’s complete monotonicity results
----------------------------------------------
Inspired by inequalities and , Ismail and Muldoon proved in [@Ismail-Muldoon-119 Theorem 3.2] the following conclusions: For $0<a<b$ and $0<q<1$, let $$\label{Gamma-q(x+a)}
h(x)=\ln\biggl\{\frac{\Gamma_q(x+a)}{\Gamma_q(x+b)}\exp[(b-a)\psi_q(x+c)]\biggr\}.$$ If $c\ge\frac{a+b}2$, then $-h'(x)$ is completely monotonic on $(-a,\infty)$; If $c\le a$, then $h'(x)$ is completely monotonic on $(-c,\infty)$; Neither $h'(x)$ or $-h'(x)$ is completely monotonic for $a<c<\frac{a+b}2$. Consequently, the following inequality was deduced in [@Ismail-Muldoon-119 Theorem 3.3]: If $0<q<1$, the inequality $$\label{q(x+1)}
\frac{\Gamma_q(x+1)}{\Gamma_q(x+s)} <\exp\biggl[(1-s)\psi_q\biggl(x+\frac{s+1}2\biggr)\biggr],\quad 0<s<1$$ holds for $x>-s$.
Influenced by , H. Alzer posed in the final of the paper [@Alzer-Math-Nachr-2001 p. 13] the following open problem: For real numbers $0<q\ne1$ and $s\in(0,1)$, determine the best possible values $a(q,s)$ and $b(q,s)$ such that the inequalities $$\exp[(1-s)\psi_q(x+a(q,s))]<\frac{\Gamma_q(x+1)}{\Gamma_q(x+s)} <\exp[(1-s)\psi_q(x+b(q,s))]$$ hold for all $x>0$.
Since the paper [@Ismail-Muldoon-119] was published in a conference proceedings, it is not easy to acquire it, so the completely monotonic properties of the function $h(x)$, obtained in [@Ismail-Muldoon-119 Theorem 3.2], were neglected in most circumstances.
Elezović-Giordano-Pečarić’s inequality and monotonicity results
---------------------------------------------------------------
Inspired by the double inequality , the following problem was posed in [@egp p. 247]: What are the best constants $\alpha$ and $\beta$ such that the double inequality $$\psi(x+\alpha)\le\frac1{t-s}\int_s^t\psi(u)\operatorname{d\mspace{-2mu}}u\le\psi(x+\beta)$$ holds for $x>-\min\{s,t,\alpha,\beta\}$?
An answer to the above problem was procured in [@egp Theorem 4]: The double inequality $$\label{second-egp-thm4}
\psi\biggl(x+\psi^{-1}\biggl(\frac1{t-s}\int_s^t\psi(u)\operatorname{d\mspace{-2mu}}u\biggr)\biggr)
<\frac1{t-s}\int_s^t\psi(x+u)\operatorname{d\mspace{-2mu}}u<\psi\biggl(x+\frac{s+t}2\biggr)$$ is valid for every $x\ge0$ and positive numbers $s$ and $t$.
Moreover, the function $$\label{gamma-arithmetic-funct}
\psi\biggl(x+\frac{s+t}2\biggr)-\frac1{t-s}\ln\frac{\Gamma(x+t)}{\Gamma(x+s)}$$ for $s,t>0$ and $r=\min\{s,t\}$ was proved in [@egp Theorem 5] to be completely monotonic on $(-r,\infty)$.
It is clear that [@egp Theorem 5] stated above extends or generalizes the complete monotonicity of the function .
By the way, the complete monotonicity in [@egp Theorem 5] was extended and iterated in [@notes-best.tex-mia Proposition 5] and [@notes-best.tex-rgmia Proposition 5] as follows: The function $$\label{cmf-lcmf}
\bigg[\dfrac{\Gamma(x+t)}{\Gamma(x+s)}\bigg]^{1/(s-t)} \exp\biggl[\psi\biggl(x+\frac{s+t}2\biggr)\biggr]$$ is logarithmically completely monotonic with respect to $x$ on $(-\alpha,\infty)$, where $s$ and $t$ are real numbers and $\alpha=\min\{s,t\}$.
Along the same line as proving the inequality in [@egp], the inequality was generalized in [@Chen-Ai-Jun-rgmia-07 Theorem 2] as $$\begin{gathered}
\label{chen-ai-jun-rgmia-07-ineq}
(-1)^n\psi^{(n)}\biggl(x+\bigl(\psi^{(n)}\bigr)^{-1}\biggl(\frac1{t-s}\int_s^t\psi^{(n)}(u)\operatorname{d\mspace{-2mu}}u\biggr)\biggr)<\\ \frac{(-1)^n\bigl[\psi^{(n-1)}(x+t)-\psi^{(n-1)}(x+s)\bigr]}{t-s} <(-1)^n\psi^{(n)}\biggl(x+\frac{s+t}2\biggr)\end{gathered}$$ for $x>0$, $n\ge0$, and $s,t>0$, where $\bigl(\psi^{(n)}\bigr)^{-1}$ denotes the inverse function of $\psi^{(n)}$.
Since the inverse functions of the psi and polygamma functions are involved, it is much difficult to calculate the lower bounds in and .
In [@kershaw-anal.appl], by the method used in [@kershaw], it was proved that the double inequality $$\label{kershaw-singapore-ineq}
\psi\bigl(x+\sqrt{st}\,\bigr)<\frac{\ln\Gamma(x+t)-\ln\Gamma(x+s)}{t-s}<\psi\biggl(x+\frac{s+t}2\biggr)$$ holds for $s,t>0$. It s clear that the upper bound in is a recovery of and an immediate consequence of the complete monotonicity of the function .
Two logarithmically complete monotonicity results
=================================================
Suggested by the double inequality , it is natural to put forward the following problem: What are the best constants $\delta_1(s,t)$ and $\delta_2(s,t)$ such that $$\label{gki2-gen}
\exp[\psi(x+\delta_1(s,t))] \le\biggl[\frac{\Gamma(x+t)}{\Gamma(x+s)}\biggr]^{1/(t-s)} \le\exp[\psi(x+\delta_2(s,t))]$$ is valid for $x>-\min\{s,t,\delta_1(s,t),\delta_2(s,t)\}$? where $s$ and $t$ are real numbers.
It is clear that the inequality can also be rewritten as $$\label{gki2-gen-rew-1}
\biggl[\frac{\Gamma(x+t)}{\Gamma(x+s)}\biggr]^{1/(s-t)}\exp[\psi(x+\delta_1)] \le1 \le\biggl[\frac{\Gamma(x+t)}{\Gamma(x+s)}\biggr]^{1/(s-t)} \exp[\psi(x+\delta_2)]$$ which suggests some monotonic properties of the function $$\label{gamma-delta-ratio}
\biggl[\frac{\Gamma(x+t)}{\Gamma(x+s)}\biggr]^{1/(t-s)}\exp[-\psi(x+\delta(s,t))],$$ since the limit of the function as $x\to\infty$ is $1$ by using .
This problem was considered in [@ratio-gamma-polynomial.tex-jcam; @ratio-gamma-polynomial.tex-rgmia; @gamma-batir.tex-jcam; @gamma-batir.tex-rgmia] along two different approaches and the following results of different forms were established.
\[gamma-ratio-multply\] Let $a$, $b$, $c$ be real numbers and $\rho=\min\{a,b,c\}$. Define $$\label{gamma-multply}
F_{a,b;c}(x)=
\begin{cases}
\biggl[\dfrac{\Gamma(x+b)}{\Gamma(x+a)}\biggr]^{1/(a-b)}\exp[\psi(x+c)], &a\ne
b\\
\exp[\psi(x+c)-\psi(x+a)],&a=b\ne c
\end{cases}$$ for $x\in(-\rho,\infty)$. Furthermore, let $\theta(t)$ be an implicit function defined by equation $$\label{implicit}
e^t-t =e^{\theta(t)}-\theta(t)$$ on $(-\infty,\infty)$. Then $\theta(t)$ is decreasing and $t\theta(t)<0$ for $\theta(t)\ne t$, and
1. $F_{a,b;c}(x)$ is logarithmically completely monotonic on $(-\rho,\infty)$ if $$\label{d1}
\begin{split}
(a,b;c)&\in \{c\ge a,c\ge b\}\cup\{c\ge a,0\ge c-b\ge\theta(c-a)\}\\*
&\quad\cup\{c\le a,c-b\ge\theta(c-a)\}\setminus\{a=b=c\};
\end{split}$$
2. $[F_{a,b;c}(x)]^{-1}$ is logarithmically completely monotonic on $(-\rho,\infty)$ if $$\label{d2}
\begin{split}
(a,b;c)&\in \{c\le a,c\le b\}\cup\{c\ge a,c-b\le\theta(c-a)\}\\*
&\quad\cup\{c\le a,0\le c-b\le\theta(c-a)\}\setminus\{a=b=c\}.
\end{split}$$
\[nu-log-mon\] For real numbers $s$ and $t$ with $s\ne t$ and $\theta(s,t)$ a constant depending on $s$ and $t$, define $$\label{nudef}
\nu_{s,t}(x)=\frac1{\exp\big[\psi\bigl(x+\theta(s,t)\bigr)\big]}
\biggl[\frac{\Gamma(x+t)}{\Gamma(x+s)}\biggr]^{1/(t-s)}.$$
1. The function $\nu_{s,t}(x)$ is logarithmically completely monotonic on the interval $(-\theta(s,t),\infty)$ if and only if $\theta(s,t)\le\min\{s,t\}$;
2. The function $[\nu_{s,t}(x)]^{-1}$ is logarithmically completely monotonic on the interval $(-\min\{s,t\},\infty)$ if and only if $\theta(s,t)\ge\frac{s+t}2$.
In [@ratio-gamma-polynomial.tex-jcam; @ratio-gamma-polynomial.tex-rgmia], it was deduced by standard argument that $$\begin{gathered}
(-1)^i[\ln F_{a,b;c}(x)]^{(i)}
=\int_0^\infty\biggl[\frac{e^{(c-a)u}-e^{(c-b)u}}{u(b-a)}-1\biggr]
\frac{u^ie^{-(x+c)u}}{1-e^{-u}}\operatorname{d\mspace{-2mu}}u\\
=\int_0^\infty\biggl[\frac{[e^{(c-a)u}-(c-a)u]-[e^{(c-b)u}-(c-b)u]} {[(c-a)-(c-b)]u}\biggr] \frac{u^ie^{-(x+c)u}}{1-e^{-u}}\operatorname{d\mspace{-2mu}}u\end{gathered}$$ for $i\in\mathbb{N}$ and $a\ne b$. Therefore, the sufficient conditions in [@ratio-gamma-polynomial.tex-jcam Theorem 1] and [@ratio-gamma-polynomial.tex-rgmia Theorem 1] are stated in terms of the implicit function $\theta(t)$ defined by .
In [@gamma-batir.tex-jcam; @gamma-batir.tex-rgmia], the logarithmic derivative of $\nu_{s,t}(x)$ was rearranged as $$\ln\nu_{s,t}(x)=\int_0^\infty\frac{e^{-[x+\theta(s,t)]u}}{1-e^{-u}} \Bigl\{1-e^{u[\theta(s,t)+\ln p_{s,t}(u)]}\Bigr\}\operatorname{d\mspace{-2mu}}u,$$ where $$p_{s,t}(u)=\biggl(\frac1{t-s}\int_s^t e^{-uv}\operatorname{d\mspace{-2mu}}v\biggr)^{1/u}.$$ Since the function $p_{s,t}(u)$ is increasing on $[0,\infty)$ with $$\lim_{u\to0}p_{s,t}(u)=e^{-(s+t)/2}\quad \text{and}\quad \lim_{u\to\infty}p_{s,t}(u)=e^{-\min\{s,t\}},$$ the necessary and sufficient conditions in [@gamma-batir.tex-jcam Theorem 1] and [@gamma-batir.tex-rgmia Theorem 1] may be derived immediately by considering Theorem \[p.161-widder\].
However, the necessary conditions in [@gamma-batir.tex-jcam Theorem 1] and [@gamma-batir.tex-rgmia Theorem 1] were proved by establishing the following inequalities involving the polygamma functions and their inverse functions in [@gamma-batir.tex-jcam Proposition 1] and [@gamma-batir.tex-rgmia Proposition 1]:
1. If $m>n\ge0$ are two integers, then $$\label{m>n>0}
\left(\psi^{(m)}\right)^{-1}\left(\frac1{t-s} \int_s^t\psi^{(m)}(v)\operatorname{d\mspace{-2mu}}v\right) \le\left(\psi^{(n)}\right)^{-1}\left(\frac1{t-s}
\int_s^t\psi^{(n)}(v)\operatorname{d\mspace{-2mu}}v\right),$$ where $\left(\psi^{(k)}\right)^{-1}$ stands for the inverse function of $\psi^{(k)}$ for $k\ge0$;
2. The inequality $$\label{log-mean-ineq}
\psi^{(i)}(L(s,t))\le\frac1{t-s}\int^t_s\psi^{(i)}(u)\operatorname{d\mspace{-2mu}}u$$ is valid for $i$ being positive odd number or zero and reversed for $i$ being positive even number;
3. The function $$\label{increas-conc}
\left(\psi^{(\ell)}\right)^{-1}\left(\frac1{t-s} \int_s^t\psi^{(\ell)}(x+v)\operatorname{d\mspace{-2mu}}v\right)-x$$ for $\ell\ge0$ is increasing and concave in $x>-\min\{s,t\}$ and has a sharp upper bound $\frac{s+t}{2}$.
Note that if taking $m=1$, $n=0$, $i=0$ and $\ell=0$ in , and , then [@f-mean Lemma 1] and [@f-mean Theorem 6] may be derived straightforwardly.
New bounds and monotonicity results
===================================
Elezović-Pečarić’s lower bound
------------------------------
The inequality for $i=0$, that is, [@f-mean Lemma 1], may be rewritten as $$\frac{\ln\Gamma(t)-\ln\Gamma(s)}{t-s}\ge\psi(L(s,t))$$ or $$\label{Elezovic-Pecaric-ineq-lower}
\biggl[\frac{\Gamma(t)}{\Gamma(s)}\biggr]^{1/(t-s)}\ge e^{\psi(L(s,t))}$$ for positive numbers $s$ and $t$.
From the left-hand side inequality in , it is easy to see that the inequality refines the traditionally lower bound $e^{\psi(G(s,t))}$.
In [@gamma-fun-ineq-batir Theorem 2.4], the following incorrect double inequality was obtained: $$\label{kershaw-batir}
e^{(x-y)\psi(L(x+1,y+1)-1)} \le\frac{\Gamma(x)}{\Gamma(y)} \le e^{(x-y)\psi(A(x,y))},$$ where $x$ and $y$ are positive real numbers.
Allasia-Giordano-Pečarić’s inequalities
---------------------------------------
In Section 4 of [@Allasia-Gior-Pecaric-MIA-02], as straightforward consequences of Hadamard type inequalities obtained in [@agpit], the following double inequalities for bounding $\ln\frac{\Gamma(y)}{\Gamma(x)}$ were listed: For $y>x>0$, $n\in\mathbb{N}$ and $h=\frac{y-x}n$, we have $$\begin{gathered}
\frac{h}2[\psi(x)+\psi(y)]+h\sum_{k=1}^{n-1}\psi(x+kh)<\ln\frac{\Gamma(y)}{\Gamma(x)} <h\sum_{k=0}^{n-1}\psi\biggl(x+\biggl(k+\frac12\biggr)h\biggr),\\
0<h\sum_{k=0}^{n-1}\psi\biggl(x+\biggl(k+\frac12\biggr)h\biggr) -\ln\frac{\Gamma(y)}{\Gamma(x)}
<\ln\frac{\Gamma(y)}{\Gamma(x)} -\frac{h}2[\psi(x)+\psi(y)]-h\sum_{k=1}^{n-1}\psi(x+kh),\\
\frac{h}2[\psi(x)+\psi(y)]+h\sum_{k=1}^{n-1}\psi(x+kh) -\sum_{i=1}^{m-1}\frac{B_{2i}h^{2i}}{(2i)!}\bigl[\psi^{(2i-1)}(y)-\psi^{(2i-1)}(x)\bigr]< \\
\ln\frac{\Gamma(y)}{\Gamma(x)}
<h\sum_{k=0}^{n-1}\psi\biggl(x+\biggl(k+\frac12\biggr)h\biggr) -\sum_{i=1}^{m-1}\frac{B_{2i}h^{2i}}{(2i)!}\bigl[\psi^{(2i-1)}(y)-\psi^{(2i-1)}(x)\bigr],\\
0<h\sum_{k=0}^{n-1}\psi\biggl(x+\biggl(k+\frac12\biggr)h\biggr) -\sum_{i=1}^{m-1}\frac{B_{2i}(1/2)h^{2i}}{(2i)!}\bigl[\psi^{(2i-1)}(y)-\psi^{(2i-1)}(x)\bigr]
-\ln\frac{\Gamma(y)}{\Gamma(x)}\\
<\ln\frac{\Gamma(y)}{\Gamma(x)}-\frac{h}2[\psi(x)+\psi(y)]-h\sum_{k=1}^{n-1}\psi(x+kh)
+\sum_{i=1}^{m-1}\frac{B_{2i}h^{2i}}{(2i)!}\bigl[\psi^{(2i-1)}(y)-\psi^{(2i-1)}(x)\bigr],\\
h\sum_{k=0}^{n-1}\psi\biggl(x+\biggl(k+\frac12\biggr)h\biggr) -\sum_{i=1}^{m-2}\frac{B_{2i}(1/2)h^{2i}}{(2i)!}\bigl[\psi^{(2i-1)}(y)-\psi^{(2i-1)}(x)\bigr]
<\ln\frac{\Gamma(y)}{\Gamma(x)}\\
<h\sum_{k=0}^{n-1}\psi\biggl(x+\biggl(k+\frac12\biggr)h\biggr) -\sum_{i=1}^{m-1}\frac{B_{2i}(1/2)h^{2i}}{(2i)!}\bigl[\psi^{(2i-1)}(y)-\psi^{(2i-1)}(x)\bigr],\\
\frac{h}2[\psi(x)+\psi(y)]+h\sum_{k=1}^{n-1}\psi(x+kh) -\sum_{i=1}^{m-1}\frac{B_{2i}h^{2i}}{(2i)!}\bigl[\psi^{(2i-1)}(y)-\psi^{(2i-1)}(x)\bigr]
<\ln\frac{\Gamma(y)}{\Gamma(x)}\\
<\frac{h}2[\psi(x)+\psi(y)]+h\sum_{k=1}^{n-1}\psi(x+kh) -\sum_{i=1}^{m-2}\frac{B_{2i}h^{2i}}{(2i)!}\bigl[\psi^{(2i-1)}(y)-\psi^{(2i-1)}(x)\bigr],\end{gathered}$$ where $m$ is an odd and positive integer, $$B_k\biggl(\frac12\biggr)=\biggl(\frac1{2^{k-1}}-1\biggr)B_k,\quad k\ge0$$ and $B_i$ for $i\ge0$ are Bernoulli numbers defined by $$\frac{t}{e^t-1}=\sum_{i=0}^\infty B_i\frac{t^i}{i!} =1-\frac{x}2+\sum_{j=1}^\infty B_{2j}\frac{x^{2j}}{(2j)!}, \quad\vert x\vert<2\pi.$$ If replacing $m$ by an even and positive integer, then the last four double inequalities are reversed.
Batir’s double inequality for polygamma functions
-------------------------------------------------
It is clear that the double inequality can be rearranged as $$\label{kershaw-rearr}
\psi\bigl(x+\sqrt{s}\,\bigr)<\frac{\ln\Gamma(x+1)-\ln\Gamma(x+s)}{1-s} <\psi\biggl(x+\frac{s+1}2\biggr)$$ for $0<s<1$ and $x>1$. The middle term in can be regarded as a divided difference of the function $\ln\Gamma(t)$ on $(x+s,x+1)$. Stimulated by this, N. Batir extended and generalized in [@batir-jmaa-06-05-065 Theorem 2.7] the double inequality as $$\label{batir-psi-ineq}
-\bigl\vert \psi^{(n+1)}(L_{-(n+2)}(x,y))\bigr\vert <\frac{\bigl\vert \psi^{(n)}(x)\bigr\vert -\bigl\vert \psi^{(n)}(y)\bigr\vert }{x-y}<-\bigl\vert \psi^{(n+1)}(A(x,y))\bigr\vert$$ where $x,y$ are positive numbers and $n\in\mathbb{N}$.
Chen’s double inequality in terms of polygamma functions
--------------------------------------------------------
In [@chen-mean-GK-ineq Theorem 2], by virtue of the composite Simpson rule $$\int_a^bf(t)\operatorname{d\mspace{-2mu}}t=\frac{b-a}6\biggl[f(a)+4f\biggl(\frac{a+b}2\biggr)+f(b)\biggr] -\frac{(b-a)^5}{2880}f^{(4)}(\xi),\quad \xi\in(a,b)$$ in [@Hammerlin-Hoffmann-bbok-91] and the formula $$\frac1{y-x}\int_x^yf(t)\operatorname{d\mspace{-2mu}}t=\sum_{k=0}^\infty\frac1{(2k+1)!} \biggl(\frac{y-x}2\biggr)^{2k}f^{(2k)}\biggl(\frac{x+y}2\biggr)$$ in [@Neuman-Sandor-05-Aus], the following double inequalities and series representations were trivially shown: For $n\in\mathbb{N}$ and positive numbers $x$ and $y$ with $x\ne y$, $$\begin{gathered}
\frac13A(\psi(x),\psi(y))+\frac23\psi(A(x,y))-\frac{(y-x)^4}{2880}\psi^{(4)}(\max\{x,y\}) <\frac{\ln\Gamma(y)-\ln\Gamma(x)}{y-x} \\
<\frac13A(\psi(x),\psi(y))+\frac23\psi(A(x,y))-\frac{(y-x)^4}{2880}\psi^{(4)}(\min\{x,y\}),\\
(-1)^{n-1}\biggl[\frac{A\bigl(\psi^{(n)}(x),\psi^{(n)}(y)\bigr)}3 +\frac{2\psi^{(n)}(A(x,y))}3-\frac{(y-x)^4\psi^{(n+4)}(\min\{x,y\})}{2880}\biggr]\\
<\frac{(-1)^{n-1}\bigl[\psi^{(n-1)}(y)-\psi^{(n-1)}(x)\bigr]}{y-x} \\
<(-1)^{n-1}\biggl[\frac{A\bigl(\psi^{(n)}(x),\psi^{(n)}(y)\bigr)}3 +\frac{2\psi^{(n)}(A(x,y))}3-\frac{(y-x)^4\psi^{(n+4)}(\max\{x,y\})}{2880}\biggr],\\
\frac{\ln\Gamma(y)-\ln\Gamma(x)}{y-x}=\sum_{k=0}^\infty\frac1{(2k+1)!} \biggl(\frac{y-x}2\biggr)^{2k}\psi^{(2k)}\biggl(\frac{x+y}2\biggr),\\
\frac{\psi^{(n-1)}(y)-\psi^{(n-1)}(x)}{y-x}=\sum_{k=0}^\infty\frac1{(2k+1)!} \biggl(\frac{y-x}2\biggr)^{2k}\psi^{(2k+n)}\biggl(\frac{x+y}2\biggr).\end{gathered}$$
Recent monotonicity results by Qi and his coauthors
---------------------------------------------------
Motivated by the left-hand side inequality in , although it is not correct, several refinements and generalizations about inequalities and were established by Qi and his coauthors in recent years.
###
In [@gamma-psi-batir.tex-jcam Theorem 1] and [@gamma-psi-batir.tex-rgmia Theorem 1], by virtue of the method used in [@gamma-fun-ineq-batir Theorem 2.4] and the inequality for $i=0$, the inequality and the right-hand side inequality in were recovered.
###
In [@gamma-psi-batir.tex-jcam Theorem 2] and [@gamma-psi-batir.tex-rgmia Theorem 2], the decreasing monotonicity of the function and the right-hand side inequality in were extended and generalized to the logarithmically complete monotonicity, and the inequality was generalized to a decreasing monotonicity.
\[log-complete-fcn\] For $s,t\in\mathbb{R}$ with $s\ne t$, the function $$\label{f_s,t}
\biggl[\frac{\Gamma(x+s)}{\Gamma(x+t)}\biggr]^{1/(s-t)}
\frac1{e^{\psi(L(s,t;x))}}$$ is decreasing and $$\label{g_s,t}
\biggl[\frac{\Gamma(x+s)}{\Gamma(x+t)}\biggr]^{1/(t-s)} e^{\psi(A(s,t;x))}$$ is logarithmically completely monotonic on $(-\min\{s,t\},\infty)$, where $$L(s,t;x)=L(x+s,x+t)\quad \text{and}\quad A(s,t;x)=A(x+s,x+t).$$
###
In [@new-upper-kershaw.tex; @new-upper-kershaw-JCAM.tex], the upper bounds in , , , and related inequalities in [@ratio-gamma-polynomial.tex-jcam; @ratio-gamma-polynomial.tex-rgmia; @gamma-batir.tex-jcam; @gamma-batir.tex-rgmia] were refined and extended as follows.
\[identric-kershaw-thm\] The inequalities $$\label{identric-kershaw-ineq-equiv}
\biggl[\frac{\Gamma(a)}{\Gamma(b)}\biggr]^{1/(a-b)}\le e^{\psi(I(a,b))}$$ and $$\label{batir-psi-ineq-ref-equiv}
\frac{(-1)^{n}\bigl[\psi^{(n-1)}(a) -\psi^{(n-1)}(b)\bigr]}{a-b} \le(-1)^n\psi^{(n)}(I(a,b))$$ for $a>0$ and $b>0$ hold true.
The basic tools to prove and are an inequality in [@cargo] and and a complete monotonicity in [@subadditive-qi-3.tex] respectively. They may be recited as follows:
1. If $g$ is strictly monotonic, $f$ is strictly increasing, and $f\circ g^{-1}$ is convex (or concave, respectively) on an interval $I$, then $$\label{carton-ineq}
g^{-1}\left(\frac1{t-s}\int_s^tg(u)\operatorname{d\mspace{-2mu}}u\right) \le
f^{-1}\left(\frac1{t-s}\int_s^tf(u)\operatorname{d\mspace{-2mu}}u\right)$$ holds (or reverses, respectively) for $s,t\in I$. See also [@bullenmean p. 274, Lemma 2] and [@f-mean p. 190, Theorem A].
2. The function $$\label{psi-abs-minus-cm-1}
x\bigl|\psi^{(i+1)}(x)\bigr|-\alpha\bigl|\psi^{(i)}(x)\bigr|,\quad i\in\mathbb{N}$$ is completely monotonic on $(0,\infty)$ if and only if $0\le\alpha\le i$. See also [@subadditive-qi-guo.tex].
By the so-called G-A convex approach, the inequality was recovered in [@Zhang-Morden]: For $b>a>0$, $$[b-L(a,b)]\psi(b)+[L(a,b)-a]\psi(a)<\ln\frac{\Gamma(b)}{\Gamma(a)}<(b-a)\psi(I(a,b)).$$ See also [MR2413632](http://www.ams.org/mathscinet-getitem?mr=2413632). Moreover, by the so-called geometrically convex method, the following double inequality was shown in [@Xiao-Ming-JIPAM Theorem 1.2]: For positive numbers $x$ and $y$, $$\frac{x^x}{y^y}\biggl(\frac{x}y\biggr)^{y[\psi(y)-\ln y]}e^{y-x}\le \frac{\Gamma(x)}{\Gamma(y)} \le\frac{x^x}{y^y}\biggl(\frac{x}y\biggr)^{x[\psi(x)-\ln x]}e^{y-x}.$$
###
In [@subadditive-qi-guo.tex; @subadditive-qi-3.tex], the function $$\label{psi-abs-minus-cm-2}
\alpha\bigl|\psi^{(i)}(x)\bigr|-x\bigl|\psi^{(i+1)}(x)\bigr|$$ was proved to be completely monotonic on $(0,\infty)$ if and only if $\alpha\ge i+1$. Utilizing the inequality and the completely monotonic properties of the functions and yields the following double inequality.
\[new-upper-2-thm-1\] For real numbers $s>0$ and $t>0$ with $s\ne t$ and an integer $i\ge0$, the inequality $$\label{new-upper-main-ineq}
(-1)^i\psi^{(i)}(L_p(s,t))\le \frac{(-1)^i}{t-s}\int_s^t\psi^{(i)}(u)\operatorname{d\mspace{-2mu}}u \le(-1)^i\psi^{(i)}(L_q(s,t))$$ holds if $p\le-i-1$ and $q\ge-i$.
The double inequality recovers, extends and refines inequalities , , and .
A natural question is whether the above sufficient conditions $p\le-i-1$ and $q\ge-i$ are also necessary for the inequality to be valid.
###
As generalizations of the inequalities , , the decreasing monotonicity of the function , and the left-hand side inequality in , the following monotonic properties were presented.
\[new-upper-2-thm-3\] If $i\ge0$ is an integer, $s,t\in\mathbb{R}$ with $s\ne t$, and $x>-\min\{s,t\}$, then the function $$\label{psi-minus-mon}
(-1)^i\left[\psi^{(i)}(L_p(s,t;x)) -\frac{1}{t-s}\int_s^t\psi^{(i)}(x+u)\operatorname{d\mspace{-2mu}}u\right]$$ is increasing with respect to $x$ for either $p\le-(i+2)$ or $p=-(i+1)$ and decreasing with respect to $x$ for $p\ge1$, where $L_p(s,t;x)=L_p(x+s,x+t)$.
It is not difficult to see that the ideal monotonic results of the function should be as follows.
\[new-upper-2-thm-3-conj\] Let $i\ge0$ be an integer, $s,t\in\mathbb{R}$ with $s\ne t$, and $x>-\min\{s,t\}$. Then the function is increasing with respect to $x$ if and only if $p\le-(i+1)$ and decreasing with respect to $x$ if and only if $p\ge-i$.
Corresponding to Conjecture \[new-upper-2-thm-3-conj\], the complete monotonicity of the function and its negative may also be discussed.
Acknowledgements {#acknowledgements .unnumbered}
----------------
This article was ever reported on 16 February 2009 as a talk in the seminar held at the RGMIA, School of Engineering and Science, Victoria University, Australia, while the author was visiting the RGMIA between March 2008 and February 2009 by the grant from the China Scholarship Council. The author expresses thanks to Professors Pietro Cerone and Server S. Dragomir and other local colleagues at Victoria University for their invitation and hospitality throughout this period.
[99]{}
M. Abramowitz and I. A. Stegun (Eds), *Handbook of Mathematical Functions with Formulas, Graphs, and Mathematical Tables*, National Bureau of Standards, Applied Mathematics Series **55**, 9th printing, Washington, 1970.
G. Allasia, C. Giordano, and J. Pečarić, *Hadamard-type inequalities for $(2r)$-convex functions with applications*, Atti Accad. Sci. Torino Cl. Sci. Fis Mat Natur. **133** (1999), 1–14.
G. Allasia, C. Giordano and J. Pečarić, *Inequalities for the gamma function relating to asymptotic expasions*, Math. Inequal. Appl. **5** (2002), no. 3, 543–555.
H. Alzer, *Sharp bounds for the ratio of $q$-gamma functions*, Math. Nachr. **222** (2001), no. 1, 5–14.
H. Alzer, *Some gamma function inequalities*, Math. Comp. **60** (1993), no. 201, 337–346.
G. E. Andrews, R. A. Askey, and R. Roy, *Special Functions*, Cambridge University Press, Cambridge, 1999.
R. D. Atanassov and U. V. Tsoukrovski, *Some properties of a class of logarithmically completely monotonic functions*, C. R. Acad. Bulgare Sci. **41** (1988), no. 2, 21–23.
N. Batir, *On some properties of digamma and polygamma functions*, J. Math. Anal. Appl. **328** (2007), no. 1, 452–465; Available online at <http://dx.doi.org/10.1016/j.jmaa.2006.05.065>.
N. Batir, *Some gamma function inequalities*, RGMIA Res. Rep. Coll. **9** (2006), no. 3, Art. 5; Available online at <http://www.staff.vu.edu.au/rgmia/v9n3.asp>.
C. Berg, *Integral representation of some functions related to the gamma function*, Mediterr. J. Math. **1** (2004), no. 4, 433–439.
S. Bochner, *Harmonic Analysis and the Theory of Probability*, California Monographs in Mathematical Sciences, University of California Press, Berkeley and Los Angeles, 1960.
J. Bustoz and M. E. H. Ismail, *On gamma function inequalities*, Math. Comp. **47** (1986), 659–667.
P. S. Bullen, *Handbook of Means and Their Inequalities*, Mathematics and its Applications, Volume 560, Kluwer Academic Publishers, Dordrecht/Boston/London, 2003.
G. T. Cargo, *Comparable means and generalized convexity*, J. Math. Anal. Appl. **12** (1965), 387–392.
Ch.-P. Chen, *On some inequalities for means and the second Gautschi-Kershaw’s inequality*, RGMIA Res. Rep. Coll. **11** (2008), Suppl., Art. 6; Available online at <http://www.staff.vu.edu.au/rgmia/v11(E).asp>.
Ch.-P. Chen and A.-J. Li, *Monotonicity results of integral mean and application to extension of the second Gautschi-Kershaw’s inequality*, RGMIA Res. Rep. Coll. **10** (2007), no. 4, Art. 2; Available online at <http://www.staff.vu.edu.au/rgmia/v10n4.asp>.
Ch.-P. Chen and F. Qi, *An alternative proof of monotonicity for the extended mean values*, Austral. J. Math. Anal. Appl. **1** (2004), no. 2, Art. 11; Available online at <http://ajmaa.org/cgi-bin/paper.pl?string=v1n2/V1I2P11.tex>.
J. Dubourdieu, *Sur un théorème de M. S. Bernstein relatif à la transformation de Laplace-Stieltjes*, Compositio Math. **7** (1939-40), 96–111.
N. Elezović, C. Giordano and J. Pečarić, *The best bounds in Gautschi’s inequality*, Math. Inequal. Appl. **3** (2000), 239–252.
N. Elezović and J. Pečarić, *Differential and integral $f$-means and applications to digamma function*, Math. Inequal. Appl. **3** (2000), no. 2, 189–196.
A. Erdélyi, W. Magnus, F. Oberhettinger and F. G. Tricomi (Editors), *Higher Transcendental Functions*, Vol. 1, McGraw-Hill, New York, 1953.
W. Gautschi, *Some elementary inequalities relating to the gamma and incomplete gamma function*, J. Math. Phys. **38** (1959/60), 77–81.
B.-N. Guo and F. Qi, *A simple proof of logarithmic convexity of extended mean values*, Numer. Algorithms (2009), in press; Available online at <http://dx.doi.org/10.1007/s11075-008-9259-7>.
G. Hämmerlin and K.-H. Hoffmann, *Numerical Mathematics*, Translated from the German by Larry Schumaker. Undergraduate Texts in Mathematics. Readings in Mathematics. Springer, New York, 1991.
R. A. Horn, *On infinitely divisible matrices, kernels and functions*, Z. Wahrscheinlichkeitstheorie und Verw. Geb **8** (1967), 219–230.
M. E. H. Ismail and M. E. Muldoon, *Inequalities and monotonicity properties for gamma and $q$-gamma functions*, in: R.V.M. Zahar (Ed.), Approximation and Computation: A Festschrift in Honour of Walter Gautschi, ISNM, Vol. **119**, BirkhRauser, Basel, 1994, 309–323.
D. Kershaw, *Some extensions of W. Gautschi’s inequalities for the gamma function*, Math. Comp. **41** (1983), 607–611.
D. Kershaw, *Upper and lower bounds for a ratio involving the gamma function*, Anal. Appl. (Singap.) **3** (2005), no. 3, 293–295.
A. Laforgia and P. Natalini, *Supplements to known monotonicity results and inequalities for the gamma and incomplete gamma functions*, J. Inequal. Appl. **2006** (2006), Article ID 48727, 1–8.
E. B. Leach and M. C. Sholander, *Extended mean values [II]{}*, J. Math. Anal. Appl. **92** (1983), 207–223.
A.-J. Li, W.-Zh. Zhao and Ch.-P. Chen, *Logarithmically complete monotonicity properties for the ratio of gamma function*, Adv. Stud. Contemp. Math. (Kyungshang) **13** (2006), no. 2, 183–191.
E. Neuman and J. Sándor, *On the Ky Fan inequality and related inequalities*, II, Bull. Aust. Math. Soc. **72** (2005), no. 1, 87–107.
J. Pečarić, F. Qi, V. Šimić and S.-L. Xu, *Refinements and extensions of an inequality, III*, J. Math. Anal. Appl. **227** (1998), no. 2, 439–448.
F. Qi, *A class of logarithmically completely monotonic functions and application to the best bounds in the second Gautschi-Kershaw’s inequality*, J. Comput. Appl. Math. **224** (2009), no. 2, 538–543; Available online at <http://dx.doi.org/10.1016/j.cam.2008.05.030>.
F. Qi, *A class of logarithmically completely monotonic functions and application to the best bounds in the second Gautschi-Kershaw’s inequality*, RGMIA Res. Rep. Coll. **9** (2006), no. 4, Art. 11; Available online at <http://www.staff.vu.edu.au/rgmia/v9n4.asp>.
F. Qi, *A new lower bound in the second Kershaw’s double inequality*, J. Comput. Appl. Math. **214** (2008), no. 2, 610–616; Available online at <http://dx.doi.org/10.1016/j.cam.2007.03.016>.
F. Qi, *A new lower bound in the second Kershaw’s double inequality*, RGMIA Res. Rep. Coll. **10** (2007), no. 1, Art. 9; Available online at <http://www.staff.vu.edu.au/rgmia/v10n1.asp>.
F. Qi, *A note on Schur-convexity of extended mean values*, RGMIA Res. Rep. Coll. [**4**]{} (2001), no. 4, Art. 4, 529–533; Available online at <http://www.staff.vu.edu.au/rgmia/v4n4.asp>.
F. Qi, *A note on Schur-convexity of extended mean values*, Rocky Mountain J. Math. **35** (2005), no. 5, 1787–1793.
F. Qi, *Bounds for the ratio of two gamma functions*, RGMIA Res. Rep. Coll. **11** (2008), no. 3, Art. 1; Available online at <http://www.staff.vu.edu.au/rgmia/v11n3.asp>.
F. Qi, *Bounds for the ratio of two gamma functions—From Wendel’s and related inequalities to logarithmically completely monotonic functions*, Available online at <http://arxiv.org/abs/0904.1048>. F. Qi, *Bounds for the ratio of two gamma functions—From Wendel’s limit to Elezović-Giordano-Pečarić’s theorem*, Available online at <http://arxiv.org/abs/0902.2514>.
F. Qi, *Certain logarithmically $N$-alternating monotonic functions involving gamma and $q$-gamma functions*, Nonlinear Funct. Anal. Appl. **12** (2007), no. 4, 675–685.
F. Qi, *Certain logarithmically $N$-alternating monotonic functions involving gamma and $q$-gamma functions*, RGMIA Res. Rep. Coll. **8** (2005), no. 3, Art. 5, 413–422; Available online at <http://www.staff.vu.edu.au/rgmia/v9n3.asp>.
F. Qi, *Generalized abstracted mean values*, J. Inequal. Pure Appl. Math. **1** (2000), no. 1, Art. 4; Available online at <http://jipam.vu.edu.au/article.php?sid=97>.
F. Qi, *Generalized abstracted mean values*, RGMIA Res. Rep. Coll. **2** (1999), no. 5, Art. 4, 633–642; Available online at <http://www.staff.vu.edu.au/rgmia/v2n5.asp>.
F. Qi, *Generalized weighted mean values with two parameters*, R. Soc. Lond. Proc. Ser. A Math. Phys. Eng. Sci. **454** (1998), no. 1978, 2723–2732.
F. Qi, *Logarithmic convexity of extended mean values*, Proc. Amer. Math. Soc. **130** (2002), no. 6, 1787–1796.
F. Qi, *Refinements, extensions and generalizations of the second Kershaw’s double inequality*, RGMIA Res. Rep. Coll. **10** (2007), no. 2, Art. 8; Available online at <http://www.staff.vu.edu.au/rgmia/v10n2.asp>.
F. Qi, *The extended mean values: Definition, properties, monotonicities, comparison, convexities, generalizations, and applications*, Cubo Mat. Educ. **5** (2003), no. 3, 63–90.
F. Qi, *The extended mean values: Definition, properties, monotonicities, comparison, convexities, generalizations, and applications*, RGMIA Res. Rep. Coll. **5** (2002), no. 1, Art. 5, 57–80; Available online at <http://www.staff.vu.edu.au/rgmia/v5n1.asp>.
F. Qi and Ch.-P. Chen, *A complete monotonicity property of the gamma function*, J. Math. Anal. Appl. **296** (2004), no. 2, 603–607.
F. Qi and B.-N. Guo, *A class of logarithmically completely monotonic functions and the best bounds in the second Kershaw’s double inequality*, J. Comput. Appl. Math. **212** (2008), no. 2, 444–456; Available online at <http://dx.doi.org/10.1016/j.cam.2006.12.022>.
F. Qi and B.-N. Guo, *A class of logarithmically completely monotonic functions and the best bounds in the second Kershaw’s double inequality*, RGMIA Res. Rep. Coll. **10** (2007), no. 2, Art. 5; Available online at <http://www.staff.vu.edu.au/rgmia/v10n2.asp>.
F. Qi and B.-N. Guo, *A property of logarithmically absolutely monotonic functions and the logarithmically complete monotonicity of a power-exponential function*, Available online at <http://arxiv.org/abs/0903.5038>.
F. Qi and B.-N. Guo, *Complete monotonicities of functions involving the gamma and digamma functions*, RGMIA Res. Rep. Coll. **7** (2004), no. 1, Art. 8, 63–72; Available online at <http://www.staff.vu.edu.au/rgmia/v7n1.asp>.
F. Qi and B.-N. Guo, *Some logarithmically completely monotonic functions related to the gamma function*, submitted.
F. Qi and B.-N. Guo, *The function $(b^x-a^x)/x$: Logarithmic convexity and applications to extended mean values*, Available online at <http://arxiv.org/abs/0903.1203>.
F. Qi, B.-N. Guo and Ch.-P. Chen, *Some completely monotonic functions involving the gamma and polygamma functions*, J. Aust. Math. Soc. **80** (2006), 81–88.
F. Qi, B.-N. Guo and Ch.-P. Chen, *Some completely monotonic functions involving the gamma and polygamma functions*, RGMIA Res. Rep. Coll. **7** (2004), no. 1, Art. 5, 31–36; Available online at <http://www.staff.vu.edu.au/rgmia/v7n1.asp>.
F. Qi, B.-N. Guo and Ch.-P. Chen, *The best bounds in Gautschi-Kershaw inequalities*, Math. Inequal. Appl. **9** (2006), no. 3, 427–436.
F. Qi, B.-N. Guo and Ch.-P. Chen, *The best bounds in Gautschi-Kershaw inequalities*, RGMIA Res. Rep. Coll. **8** (2005), no. 2, Art. 17; Available online at <http://www.staff.vu.edu.au/rgmia/v8n2.asp>.
F. Qi and S. Guo, *Inequalities for the incomplete gamma and related functions*, Math. Inequal. Appl. **2** (1999), no. 1, 47–53.
F. Qi and S. Guo, *New upper bounds in the second Kershaw’s double inequality and its generalizations*, RGMIA Res. Rep. Coll. **10** (2007), no. 2, Art. 1; Available online at <http://www.staff.vu.edu.au/rgmia/v10n2.asp>.
F. Qi, S. Guo and Sh.-X. Chen, *A new upper bound in the second Kershaw’s double inequality and its generalizations*, J. Comput. Appl. Math. **220** (2008), no. 1-2, 111–118; Available online at <http://dx.doi.org/10.1016/j.cam.2007.07.037>.
F. Qi, S. Guo and B.-N. Guo, *Complete monotonicity of some functions involving polygamma functions*, submitted.
F. Qi, S. Guo and B.-N. Guo, *Note on a class of completely monotonic functions involving the polygamma functions*, RGMIA Res. Rep. Coll. **10** (2007), no. 1, Art. 5; Available online at <http://www.staff.vu.edu.au/rgmia/v10n1.asp>.
F. Qi, X.-A. Li and Sh.-X. Chen, *Refinements, extensions and generalizations of the second Kershaw’s double inequality*, Math. Inequal. Appl. **11** (2008), no. 3, 457–465.
F. Qi, W. Li and B.-N. Guo, *Generalizations of a theorem of I. Schur*, RGMIA Res. Rep. Coll. **9** (2006), no. 3, Art. 15; Available online at <http://www.staff.vu.edu.au/rgmia/v9n3.asp>.
F. Qi and Q.-M. Luo, *A simple proof of monotonicity for extended mean values*, J. Math. Anal. Appl. **224** (1998), no. 2, 356–359.
F. Qi and Q.-M. Luo, *Refinements and extensions of an inequality*, Mathematics and Informatics Quarterly **9** (1999), no. 1, 23–25.
F. Qi and J.-Q. Mei, *Some inequalities of the incomplete gamma and related functions*, Z. Anal. Anwendungen **18** (1999), no. 3, 793–799.
F. Qi and S.-L. Xu, *Refinements and extensions of an inequality, II*, J. Math. Anal. Appl. **211** (1997), 616–620.
F. Qi and S.-L. Xu, *The function $(b^x-a^x)/x$: Inequalities and properties*, Proc. Amer. Math. Soc. **126** (1998), no. 11, 3355–3359.
F. Qi, S.-L. Xu and L. Debnath, *A new proof of monotonicity for extended mean values*, Intern. J. Math. Math. Sci. **22** (1999), no. 2, 415–420.
H. van Haeringen, *Completely monotonic and related functions*, J. Math. Anal. Appl. **204** (1996), no. 2, 389–408.
D. V. Widder, *The Laplace Transform*, Princeton University Press, Princeton, 1941.
S.-L. Zhang, Ch.-P. Chen and F. Qi, *Another proof of monotonicity for the extended mean values*, Tamkang J. Math. **37** (2006), no. 3, 207–209.
X.-M. Zhang and Y.-M. Chu, *An inequality involving the Gamma function and the psi function*, Int. J. Mod. Math. **3** (2008), no. 1, 67–73.
X.-M. Zhang, T.-Q. Xu, and L.-B. Situ, *Geometric convexity of a function involving gamma function and applications to inequality theory*, J. Inequal. Pure Appl. Math. **8** (2007), no. 1, Art. 17; Available online at <http://jipam.vu.edu.au/article.php?sid=830>.
[^1]: The author was partially supported by the China Scholarship Council
[^2]: This paper was typeset using -LaTeX
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The CRESST experiment monitors $300{\, \mathrm{g}}$ ${\mathrm{CaWO_4}}$ crystals as targets for particle interactions in an ultra low background environment. In this paper, we analyze the background spectra that are recorded by three detectors over many weeks of data taking. Understanding these spectra is mandatory if one wants to further reduce the background level, and allows us to cross-check the calibration of the detectors. We identify a variety of sources, such as intrinsic contaminations due to primordial radioisotopes and cosmogenic activation of the target material. In particular, we detect a $3.6{\, \mathrm{keV}}$ X-ray line from the decay of ${\mathrm{{}^{41}Ca}}$ with an activity of $(26\pm4){\, \mathrm{\mu Bq}}$, corresponding to a ratio ${\mathrm{{}^{41}Ca}}/{\mathrm{{}^{40}Ca}}=(2.2\pm0.3)\times10^{-16}$.'
address:
- 'Max-Planck-Institut für Physik, Föhringer Ring 6, D-80805 München, Germany'
- 'Physik-Department E15, Technische Universität München, D-85747 Garching, Germany'
- 'Department of Physics, University of Oxford, Oxford OX1 3RH, United Kingdom'
- 'Eberhard-Karls-Universität Tübingen, D-72076 Tübingen, Germany'
- 'INFN, Laboratori Nazionali del Gran Sasso, I-67010 Assergi, Italy'
- 'on leave from: Departamento de Fisica, Universidade de Coimbra, P3004 516 Coimbra, Portugal'
- Deceased
author:
- 'R. F. Lang'
- 'G. Angloher'
- 'M. Bauer'
- 'I. Bavykina'
- 'A. Bento'
- 'A. Brown'
- 'C. Bucci'
- 'C. Ciemniak'
- 'C. Coppi'
- 'G. Deuter'
- 'F. von Feilitzsch'
- 'D. Hauff'
- 'S. Henry'
- 'P. Huff'
- 'J. Imber'
- 'S. Ingleby'
- 'C. Isaila'
- 'J. Jochum'
- 'M. Kiefer'
- 'M. Kimmerle'
- 'H. Kraus'
- 'J.-C. Lanfranchi'
- 'B. Majorovits'
- 'M. Malek'
- 'R. McGowan'
- 'V. B. Mikhailik'
- 'E. Pantic'
- 'F. Petricca'
- 'S. Pfister'
- 'W. Potzel'
- 'F. Pröbst'
- 'S. Roth'
- 'K. Rottler'
- 'C. Sailer'
- 'K. Schäffner'
- 'J. Schmaler'
- 'S. Scholl'
- 'W. Seidel'
- 'L. Stodolsky'
- 'A. J. B. Tolhurst'
- 'I. Usherov'
- 'W. Westphal'
title: Electron and Gamma Background in CRESST Detectors
---
CRESST, ${\mathrm{CaWO_4}}$, Low Background, Calcium-41, Dark Matter 29.40.Vj ,95.35.+d
Introduction
============
The Cryogenic Rare Event Search with Superconducting Thermometers CRESST [@angloher2005; @angloher2009] aims to detect rare nuclear recoils from elastic scattering of Dark Matter particles [@jungman1996]. Since less than one such recoil is expected per kilogram of target mass and week of exposure, efficient shielding of the target against ambient radioactivity is mandatory.
To this end, the CRESST experiment is located in the Laboratori Nazionali del Gran Sasso under an average of $1400{\, \mathrm{m}}$ of rock overburden, where the cosmic muon flux is reduced by about six orders of magnitude with respect to the surface. Additionally, the target is surrounded (from the outside to the inside) by $45{\, \mathrm{cm}}$ of polyethylene to reduce the neutron flux, a radon box that is constantly flushed with nitrogen gas to reduce the concentration of radioactive radon and its daughters inside the shielding, as well as $20{\, \mathrm{cm}}$ of lead and $14{\, \mathrm{cm}}$ of pure copper against the external gamma background [@angloher2009]. The understanding of the remaining radioactive background present in the experiment demonstrates its excellent performance and is a necessary step to improve the sensitivity of CRESST to rare processes.
Setup
=====
The target consists of individual $300{\, \mathrm{g}}$ ${\mathrm{CaWO_4}}$ crystals that are operated as calorimeters [@angloher2005]. Up to 33 such detectors can be arranged in a compact layout [@angloher2009]. To enhance the temperature signal following an interaction, the crystals are cooled to $\sim 15{\, \mathrm{mK}}$ where heat capacities are low. The high frequency phonons that are created in interactions in the crystal are collected by a thermometer, which is a thin tungsten film that is evaporated directly onto the crystals. This film becomes superconducting at these temperatures and is stabilized in its transition to the superconducting state. A small temperature rise results in an increase of resistance that is measured with a SQUID-based readout circuit [@seidel1990].
A resistive heater structure is evaporated onto each thermometer. This structure is used to provide a constant heating bias as well as for periodic injection of additional heat pulses. Each such injected heat pulse briefly drives the thermometer out of its transition. The amplitude response of the thermometer to the injected heat is controlled to stay at a constant value with small adjustments to the heating bias by means of a feedback system. In this way the operating point of the thermometer is stabilized over several months to within the required precision [@angloher2005; @angloher2009]. This can also be seen from the sharpness of gamma lines discussed here.
Pulse Height Evaluation
=======================
Our thermometers yield rather slow pulses with rise times of about $1{\, \mathrm{ms}}$. These pulses are digitized with a time base of $40{\, \mathrm{\mu s}}$. $4096$ such samples comprise a record, of which the first $1024$ samples are taken from the time before the trigger to define the baseline of the pulse.
Template pulses are fitted to the record in order to evaluate the amplitude of the pulse as a measure of the deposited energy. These templates are an average of appropriate pulses, often obtained from a gamma calibration line. Since the SQUIDs only measure relative changes in current, the baseline needs to be defined individually for each recorded pulse. Hence, the free parameters in the fit are the level of the baseline, the onset of the pulse, and its amplitude.
The pulse samples the transition curve of the thermometer from its superconducting to the normal conducting state. For interactions with recoil energies below $\sim200{\, \mathrm{keV}}$, the resultant pulse generally samples only the linear region of the thermometer’s transition curve between the superconducting and normal conducting state. For interactions with larger energies, the pulse samples a region of the transition curve that is non-linear, distorting the resulting pulse shape. Therefore, in the template pulse fit, only those parts of the pulse are included that show a linear behavior. This linearizes the energy response of the fitted amplitudes over a wide energy range, even up to the MeV range [@cozzini2004].
Energy Calibration
==================
To reach the cryogenic temperatures, the detectors are enclosed by a fivefold thermal shield made from copper. Calibration sources are placed outside this shield, which has a total thickness of $12{\, \mathrm{mm}}$ of copper. This requires gamma energies $\gtrsim 100{\, \mathrm{keV}}$ in order to penetrate this barrier. Hence, a method is required to take the energy calibration to lower energies.
To this end, additional heater pulses corresponding to a number of fixed energies are injected into the heater structure every 30 seconds. The pulse height of the response pulse from the thermometer is also evaluated with a template fit. Non-linearities are due to the shape of the transition curve and due to the readout circuit. They manifest themselves in the amplitude of the response pulse, but, for given amplitude, not in its shape. Therefore, the response pulses give the correct energy assessment even if there are small differences in the pulse shape of heater response pulse and particle pulse.
The heater pulses are calibrated at relatively high energies, typically with the $122{\, \mathrm{keV}}$ or $136{\, \mathrm{keV}}$ line from a ${\mathrm{{}^{57}Co}}$ source. The amplitude response of the thermometer as a function of energy injected into the heater can then be fitted with a low-order polynomial function. Figure \[fig:cpeb\] depicts this procedure for the calibration of particle pulses in the complete energy range covered by the heater pulses. In the following we identify a variety of lines at the expected energies, thus validating our calibration procedure.
![Calibration at low energies with heater pulses of various discrete energies. Fitting a low order polynomial to the thermometer response (small points, including error bars) gives the transfer function (solid line) between injected voltage (scale on the left) and fitted amplitude (scale on the bottom). The $136{\, \mathrm{keV}}$ peak gives a calibration point (dotted line) for the injected energy (scale on the right). Then, for each particle pulse, the particle’s energy is found by evaluating the transfer function at the fitted amplitude (dashed line).[]{data-label="fig:cpeb"}](figure01.pdf){width="1\columnwidth"}
Calibration Spectrum {#sec:copperfluorescence}
====================
{width="100.00000%"}
Figure \[fig:CoSpectrumDaisyRun31g\] shows a typical spectrum from a ${\mathrm{{}^{57}Co}}$ calibration. The two strong peaks at $122{\, \mathrm{keV}}$ and $136{\, \mathrm{keV}}$ originate from gamma transitions in the ${\mathrm{{}^{57}Co}}$ source. They show pronounced shoulders towards lower energies, which originate from photons that are Compton scattered in the copper thermal shields and the detector support structures that surround the detectors.
A gamma can interact in the crystal via the photoelectric effect, ejecting an electron from its shell. The X-ray produced in the subsequent recombination may escape from the crystal. In this case, we measure the energy of the original gamma minus the energy of the X-ray (escape peak). If the escaping X-ray reaches a neighboring crystal, its energy will be measured there. In figure \[fig:CoSpectrumDaisyRun31g\] we observe a set of peaks in the energy range between $50{\, \mathrm{keV}}$ and $80{\, \mathrm{keV}}$, originating from the dominant $K_{{\mathrm{\alpha1}}}$, $K_{{\mathrm{\alpha2}}}$, $K_{{\mathrm{\beta1}}}$/$K_{{\mathrm{\beta3}}}$ and $K_{{\mathrm{\beta2}}}$ escape lines of tungsten X-rays. The origin of the various observed lines are labeled in the figure. In addition, a line is visible at $8{\, \mathrm{keV}}$ from copper fluorescence due to interactions in the copper surrounding the cryostat (see section \[sec:copper\]).
The calibration spectrum allows confirmation of our energy calibration method. This is shown in table \[tab:ecal\], where measured energies are compared to the literature values [@firestone1996] for a few lines. The values agree within the resolution of our detectors.
Background Spectra
==================
In the following, spectra from exposure to background radiation alone are analyzed, from one detector during the prototyping phase of the experiment [@angloher2005] (figures \[fig:daisystrontium\], \[fig:daisyrun28c\]) and two detectors during the commissioning phase [@angloher2009] (figures \[fig:verenarun30c\], \[fig:verenarun30m\] and \[fig:zorarun30c\]). All these crystals were grown by the General Physics Institute, Moscow, Russia, but in different batches.
Strontium and Yttrium
---------------------
Strontium belongs to the same chemical group as calcium and can thus easily be incorporated in our ${\mathrm{CaWO_4}}$ crystals in the Czochralski growth process. Indeed this is the case, as can be seen in figure \[fig:daisystrontium\] where we show the spectrum of one crystal up to $1{\, \mathrm{MeV}}$. We observe a continuous background which can be attributed to the beta decay of ${\mathrm{{}^{90}Sr}}$ (with an endpoint energy of $546{\, \mathrm{keV}}$) together with the subsequent beta decay of ${\mathrm{{}^{90}Y}}$ ($2282{\, \mathrm{keV}}$ endpoint energy).
![Background spectrum observed with one crystal ($m=306.8{\, \mathrm{g}}$) from an exposure to background radiation of $15.00{\, \mathrm{kg\,d}}$ early 2004 (detector <span style="font-variant:small-caps;">Daisy</span>, run 28). The black line is a fit above $80{\, \mathrm{keV}}$ of a parametrization of the ${\mathrm{{}^{90}Sr}}$/${\mathrm{{}^{90}Y}}$ beta spectrum to the data.[]{data-label="fig:daisystrontium"}](figure03.pdf){width="1\columnwidth"}
A fit above $80{\, \mathrm{keV}}$ to a parametrization of the expected beta spectrum from ${\mathrm{{}^{90}Sr}}$ in equilibrium with ${\mathrm{{}^{90}Y}}$ [@kossert2009] can be seen to describe the data in this energy range. From the fit we deduce an activity of $(4.37\pm0.15){\, \mathrm{mBq}}$ of ${\mathrm{{}^{90}Sr}}$ for this crystal. In addition to this common and continuous background, the three crystals show quite different features in the energy range below $100{\, \mathrm{keV}}$.
![Zoom on the energy region below $100{\, \mathrm{keV}}$ of the spectrum shown in figure \[fig:daisystrontium\].[]{data-label="fig:daisyrun28c"}](figure04.pdf){width="1\columnwidth"}
![Background spectrum of a second crystal ($m=305.5{\, \mathrm{g}}$) from an exposure of $21.20{\, \mathrm{kg\,d}}$ in 2007 (detector <span style="font-variant:small-caps;">Verena</span>, run 30).[]{data-label="fig:verenarun30c"}](figure05.pdf){width="1\columnwidth"}
![Background spectrum recorded in 2007 from a third detector ($m=307.0{\, \mathrm{g}}$) with an exposure of $22.81{\, \mathrm{kg\,d}}$ (detector <span style="font-variant:small-caps;">Zora</span>, run 30).[]{data-label="fig:zorarun30c"}](figure06.pdf){width="1\columnwidth"}
Lead
----
Isotopes from natural decay chains are present in the crystals and can be identified by their alpha decays [@cozzini2004]. A step in the natural decay of ${\mathrm{{}^{238}U}}$ is ${\mathrm{{}^{210}Pb}}$, which beta-decays into ${\mathrm{{}^{210}Bi}}$ with an endpoint energy of $63.5{\, \mathrm{keV}}$. In $84\percent$ of these decays a $46.5{\, \mathrm{keV}}$ gamma is emitted. If the decays happen in the vicinity of our crystals, we observe a line at $46.5{\, \mathrm{keV}}$ (see figure \[fig:daisyrun28c\]), which reflects an external activity of $(31\pm5){\, \mathrm{\mu Bq}}$. If, on the other hand, ${\mathrm{{}^{210}Pb}}$ is a contamination intrinsic to the crystal, our calorimetric measurement gives the energy of the gamma plus that of the emitted electron. Hence, we observe a beta spectrum starting at $46.5{\, \mathrm{keV}}$ that extends up to $63.5{\, \mathrm{keV}}$, as can prominently be seen in figure \[fig:verenarun30c\]. The activity of this decay due to sources both internal and external to the crystal is much higher, namely $(672\pm28){\, \mathrm{\mu Bq}}$. Figure \[fig:zorarun30c\] also shows this lead feature for the third crystal with an activity of $(241\pm25){\, \mathrm{\mu Bq}}$.
Actinium
--------
A similar case is that of ${\mathrm{{}^{227}Ac}}$, a step in the natural decay of ${\mathrm{{}^{235}U}}$. ${\mathrm{{}^{227}Ac}}$ beta-decays with an endpoint energy of $44.8{\, \mathrm{keV}}$ into ${\mathrm{{}^{227}Th}}$, where two excited levels may lead to the emission of $24.5{\, \mathrm{keV}}$ and $9.3{\, \mathrm{keV}}$ gammas. This results in two beta spectra, one starting at $9.3{\, \mathrm{keV}}$, and one at $24.5{\, \mathrm{keV}}$, as can prominently be seen in figure \[fig:zorarun30c\] but also in figure \[fig:verenarun30c\]. This contamination is not observable in the spectrum shown in figure \[fig:daisyrun28c\], which indicates that this detector contains less impurities from the decay of ${\mathrm{{}^{235}U}}$.
Activated Tungsten
------------------
If the raw material or the crystals themselves are exposed to cosmic radiation, the tungsten in the ${\mathrm{CaWO_4}}$ crystals can be activated. A possible channel is ${\mathrm{{}^{182}W}} (p,\alpha) {\mathrm{{}^{179}Ta}}$. ${\mathrm{{}^{179}Ta}}$ then decays via electron capture with a half-life of 1.8 years into ${\mathrm{{}^{179}Hf}}$. The energy signature of this decay in our crystals is the binding energy of the captured electron, mostly from the K-shell, which is $E_{{\mathrm{K,Hf}}}=65.4{\, \mathrm{keV}}$. The activation via ${\mathrm{{}^{183}W}} (p,t)$ results in ${\mathrm{{}^{181}W}}$. This decays with a half-life of 121 days via electron capture into an excited ${\mathrm{{}^{181}Ta}}$ nucleus, which de-excites with emission of a $6.2{\, \mathrm{keV}}$ gamma. Hence, the dominant energy signature of this decay is the binding energy of a tantalum K-shell electron, $E_{{\mathrm{K,Ta}}}=67.4{\, \mathrm{keV}}$, plus the energy of the gamma, $6.2{\, \mathrm{keV}}$, adding up to $73.7{\, \mathrm{keV}}$.
Lines at these energies can be seen in all spectra. For the crystal of which the spectrum is shown in figure \[fig:daisyrun28c\], a measurement with a similar exposure to background radiation, about three months earlier, is available [@majorovits2006]. This allows us to extract the ${\mathrm{{}^{179}Ta}}$ and ${\mathrm{{}^{181}W}}$ activities from two measurements. A comparison to the expected decay times is shown in figure \[fig:activated\], confirming the origin of these lines.
![Two measurements of the ${\mathrm{{}^{181}W}}$ and ${\mathrm{{}^{179}Ta}}$ activities in one crystal during two runs with similar exposure to background radiation. The expectations from decays with $T_{1/2}=121{\, \mathrm{d}}$ and $T_{1/2}=665{\, \mathrm{d}}$ are drawn as dashed lines and seen to be consistent with the measurement. Error bars are uncertainties of the fit to the spectral lines (each containing tens of counts) and the running time of the respective measurements (detector <span style="font-variant:small-caps;">Daisy</span>, run 27 and run 28).[]{data-label="fig:activated"}](figure07.pdf){width="1\columnwidth"}
Activated Calcium
-----------------
Capture of thermal neutrons can activate calcium present in the ${\mathrm{CaWO_4}}$ crystals. ${\mathrm{{}^{41}Ca}}$ has a half-life of $10^5$ years and decays via electron-capture to ${\mathrm{{}^{41}K}}$. The K-shell binding energy of potassium is only $3.61{\, \mathrm{keV}}$, but our detectors are capable of detecting such low energies. Figure \[fig:verenarun30m\] shows again the spectrum of figure \[fig:verenarun30c\] but for lower energies and with a finer binning. We observe a prominent line at $(3.71\pm0.02){\, \mathrm{keV}}$ (mean and error from the fit of a Gaussian). This is only $3\percent$ above the expected energy from ${\mathrm{{}^{41}Ca}}$, which we therefore attribute to this line. The width of this line is only $(130\pm19){\, \mathrm{eV}}$, consistent with the calibration. For illustration, figure \[fig:ca41event\] shows a typical pulse at this low energy, still clearly visible above baseline noise.
![Low-energy part of the spectrum shown in figure \[fig:verenarun30c\]. In addition to the lines from ${\mathrm{{}^{41}Ca}}$ and copper fluorescence there is a line at $\sim11.5{\, \mathrm{keV}}$ which so far could not be identified in a consistent way (detector <span style="font-variant:small-caps;">Verena</span>, run 30).[]{data-label="fig:verenarun30m"}](figure08.pdf){width="1\columnwidth"}
![A typical pulse with an energy of $3.7{\, \mathrm{keV}}$ (grey), and its 50-sample ($2{\, \mathrm{ms}}$) moving average (black).[]{data-label="fig:ca41event"}](figure09.pdf){width="1\columnwidth"}
From the observed activity of $(26\pm4){\, \mathrm{\mu Bq}}$ we can readily calculate that there are $(1.4\pm0.2)\times10^8$ ${\mathrm{{}^{41}Ca}}$ atoms present in the crystal, only a fraction of $(2.2\pm0.3)\times10^{-16}$ of all calcium atoms. This is an order of magnitude more accurate than other measurements of this fraction which are typically carried out by accelerator mass spectrometry (AMS) [@merchel2009; @fink1990; @henning1987]. Our present sensitivity is in the $10^{-17}$ range even with ${\mathrm{CaWO_4}}$ crystals that are not optimized for this purpose. This demonstrates the extreme sensitivity of our detectors for rare decays and opens new possibilities for radioactive dating using ${\mathrm{{}^{41}Ca}}$ [@raisbeck1979]. In fact, having an in situ calibration line at these low energies is an advantage for the direct Dark Matter search, since it independently confirms our energy calibration procedure in this otherwise inaccessible region.
Copper Fluorescence {#sec:copper}
-------------------
Another line, seen at $(8.17\pm0.04){\, \mathrm{keV}}$ in figure \[fig:verenarun30m\], can be attributed to copper fluorescence. This originates from gammas or electrons that eject a photoelectron from the copper of the cryostat that surrounds the detectors, resulting in a copper X-ray absorbed in our crystals. Our energy calibration puts the line only $2\percent$ higher than the expected $8.04{\, \mathrm{keV}}$ from the literature. This is consistent with the slight energy overestimate seen when identifying the ${\mathrm{{}^{41}Ca}}$ spectral line, again confirming its origin.
Lutetium
--------
All single crystal spectra above $100{\, \mathrm{keV}}$ are rather featureless. Yet, two crystals were operated simultaneously and in close vicinity during the commissioning of the setup. This allows us to search for coincident events, of which a spectrum is shown in figure \[fig:coincidentb\].
![Spectrum of coincident events. To increase the counting statistics, the spectra of the two individual detectors were added to produce this figure. Without the coincidence requirement, the added spectrum would have a count rate of ${\mathcal{O}}(500)$ counts per $2{\, \mathrm{keV}}$ bin (detectors <span style="font-variant:small-caps;">Verena</span> and <span style="font-variant:small-caps;">Zora</span>, run 30).[]{data-label="fig:coincidentb"}](figure10.pdf){width="1\columnwidth"}
At $59{\, \mathrm{keV}}$, tungsten ${\mathrm{K_{\alpha}}}$ escape events can be seen. In addition, two lines at $202{\, \mathrm{keV}}$ and $307{\, \mathrm{keV}}$ become very prominent, and we attribute those to a contamination with lutetium intrinsic to the crystals. ${\mathrm{{}^{176}Lu}}$ beta-decays to ${\mathrm{{}^{176}Hf}}$ with a total decay energy of $1191.7{\, \mathrm{keV}}$. Each decay produces a gamma cascade with gamma energies of $306.8{\, \mathrm{keV}}$, $201.8{\, \mathrm{keV}}$ and $88.3{\, \mathrm{keV}}$. In particular the higher energy gammas may escape from one crystal and hit the other. Hence, for a coincident $306.8{\, \mathrm{keV}}$ event observed in one detector, the other crystal will mostly absorb the other two gammas, plus the energy of the beta electron. This explains the structure seen in the scatter plot of figure \[fig:zoravsverenab\], showing the energies of coincident events in each detector. We observe a beta spectrum starting at $201.8{\, \mathrm{keV}}+88.3{\, \mathrm{keV}}=290.1{\, \mathrm{keV}}$ for most events at $306.8{\, \mathrm{keV}}$ in the other detector (lines marked (a) and (b) in figure \[fig:zoravsverenab\]). Figure \[fig:beta\] shows the energy spectrum of these events in one detector, tagged by an energy of $306.8{\, \mathrm{keV}}$ in the other one, in agreement with the expectation from the lutetium beta spectrum.
![Scatter plot of coincident events in two detectors operated during commissioning. The vertical line (a) going up from $290{\, \mathrm{keV}}$ is due to lutetium decaying in the second detector, in coincidence with the $307{\, \mathrm{keV}}$ gamma escaped to the first detector. The horizontal line (b) is due to the same process with detectors reversed, at $307{\, \mathrm{keV}}$ in the second detector and starting at $290{\, \mathrm{keV}}$ in the first one. At (c) we observe a hint of a horizontal line starting at $395{\, \mathrm{keV}}$ due to lutetium decaying in the first detector with the energy of an escaped $202{\, \mathrm{keV}}$ gamma in the second detector.[]{data-label="fig:zoravsverenab"}](figure11.pdf){width=".8\columnwidth"}
![Histogram of all coincident events in one detector that appear at $(307\pm3){\, \mathrm{keV}}$ in the respective other detector: The ${\mathrm{{}^{176}Lu}}$ beta spectrum starting at $290{\, \mathrm{keV}}$ is visible despite the low statistics.[]{data-label="fig:beta"}](figure12.pdf){width="1\columnwidth"}
Conclusions
===========
The excellent energy resolution of less than $1{\, \mathrm{keV}}$ over the full energy range used in the Dark Matter search [@angloher2005; @angloher2009] allows us to identify a variety of different background sources in the CRESST experiment. Since the CRESST experiment can discriminate nuclear recoil signal from electron recoil background based on the light output, the presence of these features does not compromise the search for Dark Matter, yet allows an audit of the performance of the experiment. The energy resolution ($1\sigma$) is $130{\, \mathrm{eV}}$ at $3.6{\, \mathrm{keV}}$, and the observed sharpness of the lines proves the high stability of the detectors during the background measurements running for several months. The observed line from activated calcium provides an independent in situ energy calibration even below the energy range used for the Dark Matter search, and in addition proves the high sensitivity of the experiment to rare decays.
Acknowledgments
===============
R. F. L. acknowledges useful discussions with G. Korschinek on ${\mathrm{{}^{41}Ca}}$. This work was partially supported by funds of the DFG (SFB 375 and Transregio 27 “Neutrinos and Beyond”), the Munich Cluster of Excellence (“Origin and Structure of the Universe"), the EU networks for Cryogenic Detectors (ERB-FMRXCT980167) and for Applied Cryogenic Detectors (HPRN-CT2002-00322), and the Maier-Leibnitz-Laboratorium (Garching). Support was provided by the Science and Technology Facilities Council.
[00]{}
G. Angloher et al. (The CRESST Collaboration), Astroparticle Physics **23** (2005) 325, [[`arXiv:astro-ph/0408006`](http://arxiv.org/abs/astro-ph/0408006)]{}.
G. Angloher et al. (The CRESST Collaboration), accepted for publication in Astroparticle Physics (2009), [[`arXiv:0809.1829`](http://arxiv.org/abs/0809.1829)]{}.
G. Jungman, M. Kamionkowski and K. Griest, Physics Reports **267** (1996) 195, [[`arXiv:hep-ph/9506380`](http://arxiv.org/abs/hep-ph/9506380)]{}.
W. Seidel et al., Physics Letters B **236** (1990) 483.
C. Cozzini et al., Physical Review C **70** (2004) 064606, [[`arXiv:nucl-ex/0408006`](http://arxiv.org/abs/nucl-ex/0408006)]{}.
R. B. Firestone et al., *Table of Isotopes (CD ROM Edition)*, 1st ed., John Wiley & Sons, New York, 1996.
K. Kossert, *private communication*, 2009.
B. Majorovits et al. (2006), in: The Identification of Dark Matter, Proceedings of the Sixth International Workshop, edited by G. Fanourakis, M. Axenides and J. Vergados, 192, World Scientific.
S. Merchel et al., Nuclear Instruments and Methods B (2009), to be published.
D. Fink et al., Nuclear Instruments and Methods B **47** (1990) 79.
W. Henning et al., Nuclear Instruments and Methods A **257** (1987) 60.
G. M. Raisbeck and F. Yiou, Nature **277** (1979) 42.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We develop a quasi-classical theory of high harmonic generation in semiconductors based on an interband current that has been transformed from Bloch to Wannier basis. The Wannier quasi-classical approach reveals a complete picture of the mechanisms shaping high harmonic generation, such that quantitative agreement with full quantum calculations is obtained. The intuitive picture revealed by quasi-classical wavepacket propagation will be helpful in the interpretation and design of high harmonic and attosecond experiments. Beyond that, the capacity to quantitatively model quantum dynamics with classical trajectories should prove useful for a wider spectrum of condensed matter research, including coherent control, transport theory, and strong field physics.'
author:
- 'A. M. Parks[1]{}'
- 'G. Ernotte[1]{}'
- 'A. Thorpe[1]{}'
- 'C. R. McDonald[1]{}'
- 'P. B. Corkum[1]{}'
- 'M. Taucer[1]{}'
- 'T. Brabec[1]{}'
bibliography:
- 'manuscript\_jmp.bib'
title: 'Wannier quasi-classical approach to high harmonic generation in semiconductors'
---
Introduction
============
High harmonic generation (HHG) in solids has been demonstrated in a wide range of materials [@Ghimire2011; @Zaks2012; @Schubert2014; @Hohenleutner2015; @Luu2015; @Garg2016; @Vampa2015_2; @Vampa2015_3; @Liu2016; @You2017; @Banks2017; @Worner2018; @Uzan2019]; it has caught attention as a source for ultrashort xuv radiation and as a tool to measure ultrafast dynamics and structural properties, such band structure [@Luu2015; @Vampa2015_2] and the Berry curvature [@Banks2017; @Worner2018]. HHG in solids is driven by interband and intraband currents. While the interband current is more dominant in wide-band materials, such as semiconductors [@Vampa2015_2], HHG in narrow-band dielectrics is driven more by the intraband current [@Luu2015]. This work focuses on interband HHG in wide-band materials.
Although some experimental features can be reasonably well reproduced by numerical models [@Tancogne2017; @Wu2017; @Jiang2018; @Li2019], a thorough understanding of all the components shaping harmonic spectra is still missing. This inhibits progress in optimizing HHG as a radiation source and in further developing HHG as a diagnostic tool.
The principal mechanism of interband HHG has been clarified by saddle point integration of the interband current derived in the Bloch basis [@Vampa2014]. Electron and hole are born at the same lattice site in real space by tunnel ionization and quiver in the laser field. When they recollide at some lattice site, a harmonic photon is emitted. Its energy is equal to the bandgap at the crystal momentum of the electron-hole pair at recollision. Despite its merits, the Bloch quasi-classical model falls short of accounting for the lattice structure; quantum mechanics allows recombination of electrons and holes at different lattice sites, as was clearly demonstrated in recent work [@You2017; @Osika2017; @Yue2020].
Here we develop a generalized quasi-classical approach that accounts for the lattice structure; this is achieved by transforming the interband current from Bloch to Wannier basis followed by saddle point integration. The basis change has a substantial effect. The resulting Wannier quasi-classical (WQC) model is found to be in quantitative agreement with quantum calculations. So far, quasi-classical $k$-space analysis has been used to qualitatively investigate strong field effects in gases and in the condensed matter phase; quantitative agreement has not been demonstrated yet. Whether quantitative agreement can be obtained in the Bloch basis remains to be seen, however the richer physics revealed by the WQC picture indicates that this might not be the case. The more refined WQC picture arises from the fact that the transition dipole moment enters the classical action in the exponent, and therewith the saddle point equations.
The quantitative agreement with full quantum calculations suggests that the physical picture for HHG in semiconductors revealed by the WQC analysis is complete. An electron and hole can ionize and recombine at different lattice sites with a probability determined by the tunneling exponent and Wannier dipole moments; birth and recombination sites are connected by classical trajectories; quantum effects are included by a quadratic expansion of the classical action about the classical trajectories. Beyond that, our WQC analysis allows unprecedented insight into the real-space aspects of tunnel ionization in solids; it gives access to the tunnel ionized wavefunction in real space and therewith, to the birth location of the electron hole pair.
More generally, our analysis opens an avenue for modeling quantum dynamics of wavepackets by propagating classical trajectories. This is potentially relevant for a wide spectrum of applications ranging from strong field physics to transport phenomena [@Datta1997; @Ludwig2019] and coherent control [@Stevens2005; @Muniz2014]. On a fundamental level, the WQC approach could open an alternative pathway to modeling noise and few electron-hole dynamics in solids; as propagation from initial to final Wannier wavepacket is done by classical trajectories, the space in between does not need to be resolved in contrast to a full quantum approach.
Theory
======
Two Band WQC Model
------------------
Our formalism is developed for a 3D, two-band model. We first summarize derivation of HHG in the Bloch basis [@Vampa2014]; it starts from the time-dependent Hamiltonian $H(t) = H_{0} +
{\bf x}\cdot{\bf F}(t)$; ${\bf F}(t)$ represents the laser field; $H_{0}$ is the unperturbed lattice Hamiltonian with Bloch eigenstates $\Phi_{{m},{\bf k}}({\bf x}) = 1/\sqrt{V}
u_{m,{\bf k}}({\bf x}) \exp(i{\bf k\cdot x})$ and with energies $E_{m}({\bf k})$ in band $m$ with crystal momentum $\bf k$; the band index $m=v,c$ refers to valence and conduction band, respectively; $u_{{m},{\bf k}}$ is the periodic part of the Bloch function, $\langle \Phi_{{m},{\bf k}} \vert \Phi_{{m},{\bf k}} \rangle = 1$, and $\langle u_{{m},{\bf k}}
\vert u_{{m},{\bf k}} \rangle = \upsilon$. Finally, $V= N \upsilon$ is the volume of the solid, with $N$ and $\upsilon$ the number and volume of primitive unit cells. Hartree atomic units are used, unless otherwise noted.
In the presence of the laser field the wavefunction becomes time-dependent. In the length gauge it is represented as $$\Psi({\bf x},t) = \sum_{m=v,c} \int_{\rm BZ} a_{m}({\bf k}, t) \Phi_{{m},{\bf k}}({\bf x}) \, d^3{\bf k} \text{,}
\label{ansatz}$$ where $a_{m}({\bf k},t)$ are the probability amplitudes and integration is over the full Brillouin zone (BZ). As initial conditions we choose an empty conduction band $a_{c}({\bf k},t=0)
= 0$, and a filled valence band, $a_{v}({\bf k},t=0) = 1/\sqrt{V_{BZ}}$, where $V_{BZ}$ is the Brillouin zone volume. The Ansatz (\[ansatz\]) is substituted into the time-dependent Schrödinger equation, and the interband polarization and current are found to be [@Vampa2014]
\[jerom\] $$\begin{aligned}
\mathbf{p}_{er}(t) & \!=\! -i\!\!\int_{\rm BZ} \!\!\!\!\! d\mathbf{k} \, \mathbf{d}(\mathbf{k}) \!\!\! \int_{-\infty}^{t} \!\!\!\!\!\!\! dt' \mathbf{F}(t')\! \cdot\!\mathbf{d}^*[\mathbf{k}(t',t)]
e^{-i S(\mathbf{k},t',t)}\!+\!\text{c.c.} \label{pert} \\
\tilde{\mathbf{j}}_{er}(\omega) & = \! i\omega \!\! \int_{-\infty}^{\infty} \!\!\! dt e^{-i\omega t}\mathbf{p}_{er}(t) \label{jerom1} \end{aligned}$$
with $S(\mathbf{k},t',t) = \int_{t'}^{t} \varepsilon(\mathbf{k}(t'',t) ) dt'' - i (t-t')/T_2$, $T_2$ the dephasing time, $\mathbf{k}(t',t) = \mathbf{k} + \mathbf{A}(t) -
\mathbf{A}(t')$ with $\mathbf{A}(t)$ the vector potential satisfying $\mathbf {F}=-\partial_t\mathbf{A}$, and $\varepsilon = E_c-E_{v}$. Here, we have used the relation [@Blount1962] $\langle\Phi_{m,\mathbf{k}}|\mathbf{x}|\Phi_{m',\mathbf{k}'}\rangle=\delta(\mathbf{k}-\mathbf{k}')[i\delta_{m,m'}\nabla_{\mathbf{k}}+\mathbf{d}_{mm'}(\mathbf{k})]$, with $\mathbf{d}_{mm'}(\mathbf{k})=i\langle u_{m,\mathbf{k}}|\nabla_\mathbf{k}|u_{m,\mathbf{k}}\rangle$ the transition dipole moment. For a two-band system, we denote $$\begin{aligned}
\mathbf{d}(\mathbf{k})= \mathbf{d}_{vc}(\mathbf{k})= i\langle u_{v,\mathbf{k}}|\nabla_\mathbf{k}|u_{c,\mathbf{k}}\rangle \text{,}
\label{kdipole}\end{aligned}$$ and we assume a centro-symmetric system for which the diagonal elements $\mathbf{d}_{mm}(\mathbf{k})$ can be set to zero [@Li2019_phaseinv].
In the following we will translate HHG, as described by the interband current of (\[jerom\]), from $k$-space to real space by using Wannier functions. The Bloch and Wannier basis functions are connected by a Fourier transform according to
\[ft\] $$\begin{aligned}
u_{m,\mathbf{k}}(\mathbf{x}) & = \sum_{j} w_{m}(\mathbf{x} - \mathbf{x}_j) e^{-i\mathbf{k}\cdot(\mathbf{x}-\mathbf{x}_j)} \label{wtob} \\
w_{m}(\mathbf{x} - \mathbf{x}_j) & = \frac{1}{\upsilon} \int_{\rm BZ} u_{m,\mathbf{k}}(\mathbf{x}) e^{i\mathbf{k}\cdot(\mathbf{x}-\mathbf{x}_j)}
d\mathbf{k} \text{.} \label{btow}\end{aligned}$$
Here, $w_{m}(\mathbf{x} - \mathbf{x}_j)$ is the Wannier function of band $m$ corresponding to the primitive unit cell at position $\mathbf{x}_j$. By virtue of (\[btow\]), the initial wavefunction, $$\begin{aligned}
\Psi({\bf x},0) = \int_{\rm BZ}d{\bf k} \Phi_{{v},{\bf k}}({\bf x}) a_{v}({\bf k},t=0) \, = w_{m}(\mathbf{x}) \text{,}
\label{initial}\end{aligned}$$ corresponds to the Wannier function at position $\mathbf{x}_j = 0$. HHG can start from any other site $\mathbf{x}_j$. The initial Wannier function can be shifted to $\mathbf{x}_j$ by setting $a_{v}({\bf k},t=0) = \exp(-i \mathbf{k}\cdot \mathbf{x}_j)$. As all lattice sites are identical, it is sufficient to investigate $\mathbf{x}_j = 0$.
In order to translate the interband current (\[jerom\]) into real space, the Bloch functions in the transition dipole moment (\[kdipole\]) are replaced by the Wannier functions with the help of relation (\[wtob\]). This leads to $$\begin{aligned}
& \mathbf{d}(\mathbf{k}) \! = \sum_{j,k} \int_{\upsilon} \!\! w_{v}^*(\mathbf{x}-\mathbf{x}_k) [\mathbf{x}-\mathbf{x}_j] w_{c}(\mathbf{x}-\mathbf{x}_j)
e^{i\mathbf{k}\cdot(\mathbf{x}_j-\mathbf{x}_k)} d\mathbf{x} \nonumber \\
~ & = \sum_{j,l} \int_{\upsilon} \!\! w_{v}^*(\mathbf{x}-(\mathbf{x}_j+\mathbf{x}_l)) [\mathbf{x}-\mathbf{x}_j] w_{c}(\mathbf{x}-\mathbf{x}_j)
e^{-i\mathbf{k}\cdot\mathbf{x}_l} d\mathbf{x} \nonumber \\
~ & = \sum_{l} e^{-i\mathbf{k}\cdot\mathbf{x}_l} \! \int_{V} \!\! w_{v}^*(\mathbf{x} - \mathbf{x}_l) \, \mathbf{x} \, w_{c}(\mathbf{x}) d\mathbf{x}
= \sum_{l} \mathbf{d}_l e^{-i\mathbf{k}\cdot\mathbf{x}_l} \text{,}
\label{dipow}\end{aligned}$$ where the second line was obtained by setting $\mathbf{x}_k = \mathbf{x}_j + \mathbf{x}_l$ and by replacing summation index $k$ with $l$ in the first line. Also, note that performing $\sum_j$ in the second line changes the integration volume from a unit cell to the whole crystal volume. The Wannier dipole moments are equivalent to the Fourier series expansion coefficients of the Bloch dipole moment $\mathbf{d}(\mathbf{k})$. Interpreted in real space, the Wannier dipole moment $\mathbf{d}_l$ describes a transition where an electron is born $l$ lattice cells away from the hole. Bloch and Wannier dipole moments are not unique; $\Phi_{{m},{\bf k}} \rightarrow \Phi_{{m},{\bf k}}
\exp[i \alpha(\mathbf{k}) ]$ is also an eigenfunction for any real function $\alpha$ that is periodic in $k$-space. Although the full equations, including the diagonal dipole elements $\mathbf{d}_{mm}$, are gauge invariant [@Blount1962; @Li2019_phaseinv], it is computationally advantageous to choose strongly confined Wannier basis functions [@Kohn1959; @Mostofi2014] in order to keep the number of relevant lattice sites small. In the 1D examples discussed further down we chose maximally localized Wannier basis functions [@Kohn1959] for which $\mathbf{d}_{mm} = 0$.
Inserting (\[dipow\]) into (\[jerom\]), the interband current follows as
\[jeromw\] $$\begin{aligned}
\tilde{\mathbf{j}}_{er}(\omega) & =\! \sum_{j,l} \left\{ \mathbf{d}_j [\mathbf{d}_l^*\cdot \mathbf{T}_{jl}(\omega)]\! -\! \mathbf{d}_j^\ast[\mathbf{d}_l\cdot
\mathbf{T}_{jl}^\ast(-\omega)]\right\} \nonumber \\
&=\! \sum_{j,l}\left[ \mathbf{P}_{jl}(\omega) - \mathbf{P}^\ast_{jl}(-\omega)\right] \text{,} \label{jeromw1} \\
\mathbf{T}_{jl}(\omega) & = \omega\!\! \int_{\rm BZ} \!\!\!\! d\mathbf{k} \!\! \int_{-\infty}^{\infty} \!\!\!\!\! dt \! \int_{-\infty}^{t}
\!\!\!\!\!\! dt' \mathbf{F}(t') e^{i \varphi(\mathbf{k},t',t,\mathbf{x}_l,\mathbf{x}_j) } \text{,} \label{jeromw2} \end{aligned}$$
Here $\varphi = -S(\mathbf{k},t',t) -\omega t + \mathbf{k}\cdot (\mathbf{x}_l - \mathbf{x}_j) + [\mathbf{A}(t)-\mathbf{A}(t')]\cdot \mathbf{x}_l$; $\mathbf{P}_{jl}(\omega)$ represents the probability amplitude that the harmonic $\omega$ is generated by an electron-hole pair that is born with a relative distance $|\mathbf{x}_l|$ between electron and hole and later recombines with relative distance $|\mathbf{x}_j|$, and the propagator $\mathbf{T}_{jl}$ describes the evolution between $\mathbf{d}_l^*$ and $\mathbf{d}_j$.
Saddle Point Integration
------------------------
{width="16cm"}
The integrals in (\[jeromw2\]) are solved by saddle point integration. The saddle point equations,
\[saddle\] $$\begin{aligned}
& \varepsilon[ \mathbf{k}(t',t) ] + \mathbf{F}(t')\cdot \mathbf{x}_l = 0 \label{stp} \text{,} \\
& \varepsilon(\mathbf{k}) - \mathbf{F}(t) \cdot[ \boldsymbol{\xi}(t',t) - \mathbf{x}_l ] = \varepsilon(\mathbf{k}) + \mathbf{F}(t)\cdot \mathbf{x}_j = \mp \omega \label{st} \text{,} \\
& \boldsymbol{\xi}(t',t) = \mathbf{x}_l - \mathbf{x}_j \label{sk} \text{,}\end{aligned}$$
result from $\partial \varphi / \partial \mu\!\!=\!\!0$ with $\mu = t', t, \mathbf{k}$, respectively. The field quiver motion between times $t'$ and $t$ is given by the distance $ \boldsymbol{\xi}(\mathbf{k},t',t) = \int_{t'}^t \mathbf{v}( \mathbf{k}(t'',t) ) \, dt''$, where $\mathbf{v}(\mathbf{k}) = \boldsymbol{\nabla}_{\mathbf{k}} \varepsilon$ is the band velocity. Note that the classical action depends on the difference between conduction and valence band. As a result, the above quantities represent the difference between electron and hole band velocity and excursion distance. Finally, the $\mp$ in (\[st\]) accounts for the complex conjugate term in (\[jerom1\]).
The set of equations (\[saddle\]) are solved for a linearly polarized laser field $\mathbf{F} = F \hat{\mathbf{x}}$; further $\mathbf{A} = A \hat{\mathbf{x}}$ and $k_x = k$. The solutions of the saddle point equations are denoted by $t' = t_b + i \delta$, $t = t_r$, $\mathbf{k}_{s}$. For $\delta \ll 1$, (\[stp\]) can be solved analytically; it determines the saddle point momentum $\mathbf{k}_s = (k_{s},k_{ys},k_{zs}) = (A(t_b) - A(t_r),0,0)$, as well as $$\begin{aligned}
\delta = \sqrt{ \frac{2 (E_g + F(t_b) x_l)}{\beta_{xx}(0) F^2(t_b)} } \text{,}
\label{delta}\end{aligned}$$ where we have approximated the bandgap as $$\begin{aligned}
\label{Equad}
\varepsilon(\mathbf{k})\approx E_g + \frac{1}{2}\sum_{i,j} k_i k_j\beta_{ij}(0)\end{aligned}$$ with $i,j = x,y,z$; $\beta_{ij}(\mathbf{k}) = \partial^2 \varepsilon / \partial k_i \partial k_j$ the inverse mass tensor; and $E_g$ the minimum bandgap. The positive sign in (\[delta\]) is chosen to obtain an exponentially decaying tunneling rate.
The two remaining saddle point equations (\[st\]) and (\[sk\]) determine $t_b$ and $t_r$. They have to be solved numerically for each possible birth site $\mathbf{x}_l$ and recombination site $\mathbf{x}_j$; for instance, by running through $t_b$ and finding all $t_r(t_b)$’s that fulfill (\[sk\]). From those, the pairs $[t_b, t_r](\omega)$ are selected that produce a given harmonic $\omega$ via (\[st\]). The physical implications of the saddle point equations are discussed at the end of this subsection.
Next, the integrand of (\[jeromw2\]) is evaluated at the saddle point, where the small imaginary birth time determines the tunneling exponent. Further, the phase $\varphi$ is expanded to second order, which gives the multivariate Gaussian integral $$\begin{aligned}
\int_{-\infty}^{\infty} d\mathbf{q} \exp((i/2)\mathbf{q}^T \mathcal{H} \mathbf{q}) = (2\pi)^{5/2} / \sqrt{-i \lvert \mathcal{H} \rvert} \text{,}
\label{saddle2}\end{aligned}$$ where $\mathbf{q} = (t',t,\mathbf{k})$, and $\mathcal{H}$ is the Hessian $\mathcal{H}_{ij} = \partial^2 \varphi / \partial_i \partial_j$ with $i,j \in \mathbf{q}$. The full expression for the determinant of the Hessian is provided in appendix \[app\_Hess\]. Putting everything together, we obtain the WQC propagator
\[WQCprop\] $$\begin{aligned}
\mathbf{T}_{jl} & = \!\!\!\!\!\! \sum_{[t_b,t_r](\omega,\mathbf{x}_l,\mathbf{x}_j) } \!\!\!\! \mathbf{g}(t_b+i\delta, t_r) \, e^{-t_x} e^{-i \chi(t_b,t_r) + i \pi/4} \text{,}
\label{tprop} \\
t_x & = \textrm{Im}[\varphi(t_b+i\delta)] \approx \frac{ \sqrt{2} [E_g + F(t_b) x_l]^{3/2}}{ [\beta_{xx}(0) F^2(t_b)]^{1/2} } \text{,} \label{tunnel} \\
\chi & = \int_{t_b}^{t_r} \varepsilon(A(t_b) - A(\tau)) d\tau + \omega t_r + \mathbf{k}_s\cdot \mathbf{x}_j \text{,} \label{phchi} \end{aligned}$$
where $\mathbf{g} = \omega \mathbf{F}(t_b+i\delta) (2 \pi)^{5/2} / \sqrt{\lvert \mathcal{H} \rvert} $ and to leading order the determinant from the Gaussian integral $\lvert \mathcal{H}
\rvert \approx v_x(\mathbf{k}_s) f(t_b+i\delta,t,\mathbf{k}_s)$ [@Uzan2019], see appendix \[app\_Hess\]. Further, it is convenient to split the phase in (\[phchi\]) into $\chi =
\chi_1 + \chi_2$, where $\chi_1 = \int_{t_b}^{t_r} \varepsilon[A(t_b) - A(\tau)] \, d\tau \,+\, \omega t_r$ contains the classical action and the harmonic frequency Fourier term. The second term is the Fourier term of the recombination dipole moment, $\chi_2 = \mathbf{k}_s \cdot\mathbf{x}_j$. The total probability amplitude $$\begin{aligned}
\mathbf{P}_{jl}= e^{i\pi/4}\sum_{[t_b,t_r](\omega,\mathbf{x}_l,\mathbf{x}_j) }&\left[ \mathbf{g}(t_b+i\delta)\mathbf{d}_l^\ast e^{-t_x}e^{-i\chi_1(t_b,t_r)}\right.\nonumber\\
&\left.\,\times\,\mathbf{d}_j e^{-i\chi_2(t_b,t_r)}\right]\end{aligned}$$ is governed by the prefactor $\mathbf{g}$, the ionization amplitude $\mathbf{d}_l^\ast e^{-t_x}$, the quantum mechanical phase factor $e^{-i\chi_1}$ acquired along the classical trajectory, and the recombination amplitude $\mathbf{d}_j e^{-i\chi_2}$. For each possible birth site $\mathbf{x}_l$ and recombination site $\mathbf{x}_j$ in the lattice, the summation runs over all birth and recombination times $t_r,t_b$ that satisfy the saddle point conditions for a particular harmonic frequency $\omega$.
The propagator (\[WQCprop\]) together with the saddle point equations (\[saddle\]) and the interband current ([\[jeromw1\]]{}) represent the WQC description of HHG in semiconductors. They reveal a complete and detailed picture of the physical mechanisms driving HHG in real and reciprocal space, summarized in figures \[fig1\](a) and (b), respectively. The empty circles in figure \[fig1\](a) represent the centers of the atomic unit cells $\mathbf{x}_l$, where $l = (l_x, l_y)$ in the 2D schematic. A Wannier basis function is located at each center. Initially, all Wannier sites of the valence band are filled. As all lattice sites are identical, it is sufficient to investigate $\mathbf{x}_l = 0$, see below (\[initial\]). Following the notation of our calculation we chose indices $l,j$ to represent birth and recombination sites, respectively. HHG proceeds in three steps.
**Step 1 - creation of electron-hole pair by ionization.** At birth time $t_b$, a valence band electron localized at lattice site $\mathbf{x}_0$ transitions to the conduction band, and is localized at lattice site $\mathbf{x}_0+\mathbf{x}_l$. The tunneling probability is determined by the tunneling exponent $t_x$ and by the Wannier dipole moment $\mathbf{d}_l^*$, see figure \[fig1\](a). The potential energy experienced by the created electron-hole dipole in the laser field makes the effective ionization potential $E_g + F(t_b) x_l$ birth site dependent, see (\[stp\]) and (\[tunnel\]). In reciprocal space the electron transitions from valence to conduction band at the $\Gamma$-point at time $t_b$, see figure \[fig1\](b). Step 1 is of quantum mechanical nature.
**Step 2 - electron-hole evolution in laser field.** The electron-hole pair quivers in the laser field. In real space it follows the classical trajectory $\boldsymbol{\xi}(t_b,t_r)$ in figure \[fig1\](a) until electron and hole revisit each other and are separated by $|\mathbf{x}_j|$ at time $t_r$, see (\[sk\]). The propagation step is dominantly classical; of quantum mechanical nature are the phase $\chi_1(t_b,t_r)$ picked up between birth and recombination time, and the quasiclassical factor $\mathbf{g}$ coming from the quadratic expansion of the classical action $S$ about the classical trajectory. The shaded green area about the classical trajectory in figure \[fig1\](a) indicates the quantum correction up to second order. In reciprocal space in figure \[fig1\](b) the electron-hole pair evolves from initial crystal momentum zero to saddle point crystal momentum $\mathbf{k}_s(t_b,t_r)$, defined below (\[saddle\]).
**Step 3 - recombination.** At time $t_r$ electron and hole recombine with probability amplitude $\mathbf{d}_j e^{-i \mathbf{k}_s(t_b,t_r)\cdot \mathbf{x}_j}$, see figure \[fig1\](a). The harmonic energy is given by the bandgap energy at $\mathbf{k}_s(t_b,t_r)$, see figure \[fig1\](b), plus the energy of the electron hole dipole in the field $F(t_r)$, see (\[st\]). Due to the second term, harmonics with energies somewhat larger than the maximum bandgap can be generated.
Results
=======
For the remainder of the paper, the WQC approach and its physical significance are explored within a 1D model system. In this case the interband current, WQC propagator, and probability amplitude reduce to scalars; namely $\tilde{j}_{er}$, $T_{jl}$, and $P_{jl}$. Specifically, we use a 1D delta function model potential, $V(x)=\Omega\sum_{n=-\infty}^\infty\delta[x-(n+1/2)a]$ with unit cell size $a$ and barrier penetration parameter $\Omega$. Details of the delta function model are given in appendix \[dltref\]. For the investigated parameters the bandgap is well approximated by the nearest neighbour dispersion $\varepsilon(k)=E_g + \Delta[1-\cos(ka)]$, where $E_g$ is the minimum bandgap and $2\Delta$ represents the bandwidth. We chose $a=7$ and considered two values $\Omega=0.5,1.5$ to model a weakly and tightly bound semiconductor, respectively. The corresponding bandgap parameters are $E_g=0.141,0.269$; $\Delta=0.269,0.17$. Finally, for all runs we use a dephasing time $T_2 = T_0/2$ so that only returns within a single cycle are relevant.
\[t\] ![\[fig2\] Harmonic yield $|h_n|^2$ versus harmonic order $n$; $a = 7$, $T_2 = T_0/2$ (a,b); empty blue circles with lines (exact) and filled blue circles (WQC) refer to $\Omega = 0.5$, $\omega_0 = 0.01425$ ($\lambda = 3.2 \mu$m), and $F_0 = 0.0025$ in (a) and $F_0 = 0.0015$ in (b); empty red squares connected by lines (exact) and filled red squares (WQC) refer to $\Omega = 1.5$, $\omega_0 = 0.0285$ ($\lambda = 1.6 \mu$m), and $F_0 = 0.008$ in (a) and $F_0 = 0.005$ in (b); lines are used to guide the eye. ](fig2.eps "fig:"){width="8.6cm"}
In figure [\[fig2\]]{} the exact (quantum mechanical) harmonic spectrum, as obtained from numerical integration of (\[jerom\]), is compared with the Wannier quasiclassical solution, (\[jeromw1\]), (\[saddle\]), (\[delta\]), and (\[tprop\]). For the exact approach we use $F(t) = F_0 \sin(\omega_0 t) \exp(-(t/\tau)^2)$ where $F_0$ is the maximum field strength, and the pulse duration, $\tau = 40 T_0$, is long enough to approach the continuous wave (cw) limit; $\omega_0$ is the laser center frequency and $T_0 = 2\pi / \omega_0$ denotes the optical cycle. We plot the harmonic intensity $\lvert h_n \rvert^2 = \int_{\omega_{-}}^{\omega_{+}}
d\omega \lvert \tilde{{j}}_{er}(\omega) \rvert^2$ integrated over the frequency interval $\omega_{\pm} = (n\pm 1/2) \omega_0$.
![\[fig3\] The contribution of the long and short classical trajectories to the probability amplitude $|P_{jl}|$ for harmonic order $n=15$ in a wide-band semiconductor; parameters $a=7$, $\Omega=0.5$, $\omega=0.01425$, and $F_0=0.0025$ corresponding to filled blue circles in figure \[fig2\](a). Figure (a) depicts the combinations of birth ($l$) and recombination ($j$) site indices for which each trajectory exists and contributes to $P_{jl}$; black regions indicate no solution. Figure (b) shows the contribution of the long trajectory to $|P_{jl}|$, while figure (c) shows the contribution from the short trajectory. Note that the values of the colorscale differ by two orders of magnitude in (b) and (c).](fig3.eps){width="8.6cm"}
\[!t\] ![\[fig4\] Probability amplitude $|P_{jl}|$ versus birth ($l$) and recombination ($j$) site indices for harmonic order $n=15$ in a narrow-band semiconductor; parameters $a=7$, $\Omega=1.5$, $\omega=0.0285$, and $F_0=0.008$ corresponding to filled red circles in figure \[fig2\](a). Here we plot the total probability amplitude $|P_{jl}|$, but note that the long trajectory is dominant; the individual contributions are similar to the behaviour depicted for the wide-band semiconductor in figure \[fig3\](b).](fig4.eps "fig:"){width="8.7cm"}
For the WQC calculation we assume the continuous wave limit, $F(t) = F_0 \sin(\omega_0 t)$, in order to facilitate interpretation of the results. Equation (\[tprop\]) has been derived for finite pulses employing the Fourier transform. For a transition to the cw limit, the Fourier transform has to be replaced by a Fourier series; as a result, $\omega \rightarrow n \omega_0$, pre-factor $g \rightarrow g / (2 \pi T_0)$, where the $ 1/(2\pi)$ comes from the 1D nature of our model. The harmonic yield becomes $\lvert h_n \rvert^2 = \lvert
\tilde{{j}}_{er}(n \omega_0) \rvert^2$ with $T_{jl}$ given by the WQC propagator (\[tprop\]).
In figure \[fig2\] the blue empty circles (exact) and blue filled circles (WQC) refer to results for the weakly bound model semiconductor, with $\Omega = 0.5$, $\omega_0 = 0.01425$. Red empty squares (exact) and red filled squares (WQC) refer to the tightly bound semiconductor, with $\Omega = 1.5$, $\omega_0 = 0.0285$. Plots with the same symbols in figures \[fig2\](a) and (b) correspond to the same values of $\Omega$ and $\omega_0$, but differ in $F_0$.
The WQC approach agrees well with the exact solution, with most data points being off by less than a factor 2. Even the first 1-2 cutoff harmonics are described fairly well, which demonstrates that they are of quasi-classical origin. The good agreement allows us to interpret semiconductor quantum dynamics such as ionization, electron/hole transport, and HHG in terms of classical trajectories. The quantum contributions to HHG are captured by the tunneling exponent $t_x$, by the pre-exponential factor $g$ in (\[tprop\]), and by the Wannier dipole moments in (\[jeromw\]).
A few points disagree by a larger factor of up to 6. In particular, figure \[fig2\](a) shows that the WQC result for harmonic $n=15$ exhibits larger discrepancy for the weakly bound semiconductor ($\Omega=0.5$) compared to the more tightly bound semiconductor ($\Omega=1.5$). The reason for this behaviour is identified in figure \[fig3\] and will be discussed later.
Numerical solution of the full saddle point equations reveals two distinct classical trajectories that contribute to the probability amplitude $P_{jl}$; one long trajectory and one short. Moreover, each solution exists for only certain combinations of birth ($l$) and recombination ($j$) lattice sites. Figure \[fig3\] shows the contributions arising from the different classical trajectories for the fifteenth harmonic $(n=15)$ with $\Omega = 0.5$, $F_0 = 0.0025$, corresponding to the filled blue circles in figure \[fig2\](a). Figure \[fig3\](a) depicts the regions in the $j$-$l$ plane where each trajectory contributes to $|P_{jl}(n=15)|$. No solution exists for the dark region in the top-right, and the probability amplitude here is zero. Figures \[fig3\](b) and (c) show the individual contributions to the probability amplitude from the long and short trajectories, respectively. The long trajectory is dominant, as the electron-hole pair is born close to the field peak, whereas the short trajectory is born closer to the nodal point. This outweighs the effect of the short dephasing time, which favors the short trajectory. As a result, the contribution of each data point to the WQC propagator is dominantly determined by a factor $\sim ge^{-t_x}$ of a single (long) trajectory. The full probability amplitude $|P_{jl}(n=15)|$ is essentially identical to figure \[fig3\](b).
In figure \[fig4\] the total probability amplitude for the fifteenth harmonic $\lvert P_{jl}(n=15) \rvert$ is plotted as a function of birth and recombination site indices $l,j$ for $\Omega = 1.5$, $F_0 = 0.008$, which corresponds to the filled red squares in figure \[fig2\](a). For this system the long trajectory is also dominant, and analysis of the individual contributions would reveal a picture qualitatively similar to figure \[fig3\].
In both figures \[fig3\] and \[fig4\], harmonic $n=15$ has been selected, as the WQC result for the weakly bound semiconductor exhibits a more pronounced difference, while it agrees well for the tightly bound semiconductor. For both systems, the maximum probability is shifted towards negative birth site indices; it is more likely for electron and hole to be born apart than at the same site. Tunnel ionization probability is determined by $e^{-t_x}$ and by birth dipole moment $d_l^*$. The tunnel exponent $t_x$ depends on the ionization potential $E_g + F(t_b)
x_l$, see (\[tprop\]). Thus, for positive field the electron-hole pair gains energy when born at increasingly negative distances which reduces $t_x$. When $-x_l = E_g/F(t_b)$, $t_x$ vanishes; in other words, the valence and conduction band levels separated by $-l$ sites align, and the electron hops from the valence to the conduction band site. The penalty to be paid is a rapidly dropping dipole moment $d_l$. As such, the birth site index at which ionization is maximum is determined by a tradeoff between tunnel exponent and Wannier dipole moment. The dipole elements for the parameters of figure \[fig3\](a) drop more slowly with increasing $\lvert l \rvert$ than for (b); see appendix \[dltref\]. Therefore, the site of highest ionization probability is shifted more strongly towards negative $l$. Recombination is most probable for $j=0$ in figures \[fig3\](a) and (b) which is consistent with previous findings [@Osika2017]. The drop in probability for increasing $j$ is due to $d_j$, which is why $\lvert P_{jl} \rvert$ extends to larger $j$ in figure \[fig3\](b).
The results in figures \[fig3\] and \[fig4\] are displayed for birth times in the positive field cycle $0 \le t_b \le T_0/2$; the negative half cycle would show the same picture, but mirrored about the $x-$ and $y-$axis ($j,l \rightarrow -j,-l$).
Recall that exact and quasiclassical results do not agree well for harmonic $n=15$ in figure \[fig2\](a) ($\Omega = 0.5$). The reason is found in figure \[fig3\](b); disagreement is due to the point $(j,l) = (4,-2)$ that exhibits unusually high probability. We find that at this point $k_s$ is approximately zero, and therewith $|\mathcal{H}|\approx 0$. Since $g\propto 1/\sqrt{|\mathcal{H}|}$, this leads to a large value of the prefactor $g$. This behaviour indicates that the quadratic saddle point expansion is no longer sufficient and the next higher order term(s) must be included. In contrast, agreement for harmonic $n=15$ in figure \[fig2\](a) for $\Omega = 1.5$ is good. This is consistent with the fact that in figure \[fig4\], $k_s \approx 0$ does not occur in areas of high probability.
Finally, the WQC method hinges on saddle point integration which works well when the exponent is rapidly oscillating. This is fulfilled for wide-band semiconductors with large bandwidth ($\Delta$) and in the long wavelength limit. When transitioning to smaller $\Delta$ (dielectrics) and shorter wavelengths, saddle point integration is expected to fail at some point. This will be subject to further research. Also, it is generally possible for transitions involving higher conduction bands to contribute to the harmonic spectrum, but this is beyond the scope of the two band model considered here.
Conclusion
==========
In summary, we have shown that the full quantum dynamics driving HHG in wide band materials, such as semiconductors, can be quantitatively explained in terms of quasi-classical trajectory propagation. The physical insight offered by trajectory analysis will prove useful for optimization and design of strong field and attosecond experiments and for the development of novel diagnostic applications of HHG, such as reconstruction of the dipole moment [@Zhao2019]. We believe that our approach presents a versatile tool for investigating open issues in strong field solid state physics, such as the role of noise and many-body effects in strong field processes. Beyond that, quantitatively accurate quasi-classical analysis should be of interest for a wider range of topics in material science.
P. B. Corkum acknowledges the support of AROSR grant number FA9550-16-0109. G. Ernotte was supported by the Vanier Canada Graduate Scholarship program.
Hessian {#app_Hess}
=======
Here we provide expressions for the determinant of the Hessian $\mathcal{H}_{ij}=\partial^2\varphi/\partial_i\partial_j$ appearing in (\[WQCprop\]). Evaluation of the second derivatives yields
$$\begin{aligned}
\lvert \mathcal{H} \rvert =
\left \lvert {\begin{array}{ccccc}
\mathbf{F}(t') \cdot\mathbf{v}(\mathbf{k}(t',t)) & - \mathbf{F}(t)\cdot \mathbf{v}(\mathbf{k}(t',t)) & v_x(\mathbf{k}(t',t)) & v_y(\mathbf{k}(t',t)) & v_z(\mathbf{k}(t',t)) \\
+ \dot{\mathbf{F}}(t') \cdot\mathbf{x}_l & & & & \\[5pt]
- \mathbf{F}(t)\cdot \mathbf{v}(\mathbf{k}(t',t)) & \mathbf{F}(t)\cdot \mathbf{v}(\mathbf{k}) - \dot{\mathbf{F}}(t)\cdot \mathbf{x}_j - & -v_x(\mathbf{k}) + & -v_y(\mathbf{k}) + & -v_z(\mathbf{k}) + \\
~& F_i(t) D_{ij}(t',t) F_j(t) & F_i(t) D_{ix}(t',t) & F_i(t) D_{iy}(t',t) & F_i(t) D_{iz}(t',t) \\[5pt]
v_x(\mathbf{k}(t',t)) & -v_x(\mathbf{k}) + F_i(t) D_{xi}(t',t) & -D_{xx}(t',t) & -D_{xy}(t',t) & -D_{xz}(t',t) \\[5pt]
v_y(\mathbf{k}(t',t)) & -v_y(\mathbf{k}) + F_i(t) D_{yi}(t',t) & -D_{yx}(t',t) & -D_{yy}(t',t) & -D_{yz}(t',t) \\[5pt]
v_z(\mathbf{k}(t',t)) & -v_z(\mathbf{k}) + F_i(t) D_{zi}(t',t) & -D_{zx}(t',t) & -D_{zy}(t',t) & -D_{zz}(t',t) \\[5pt]
\end{array} } \right \rvert_{(t'=t_b+i\delta, t=t_r, \mathbf{k} = \mathbf{k}_s)} \text{,}
\label{hessian1}\end{aligned}$$
Using linear dependence between column 2 and columns 3,4, and 5, see the supplement of [@Uzan2019], the determinant can be simplified to
$$\begin{aligned}
\lvert \mathcal{H} \rvert =
\left \lvert {\begin{array}{ccccc}
\mathbf{F}(t')\cdot \mathbf{v}(\mathbf{k}(t',t))+\dot{\mathbf{F}}(t') \cdot\mathbf{x}_l & 0 & v_x(\mathbf{k}(t',t)) & v_y(\mathbf{k}(t',t)) & v_z(\mathbf{k}(t',t)) \\[5pt]
- \mathbf{F}(t)\cdot \mathbf{v}(\mathbf{k}(t',t)) & -\dot{\mathbf{F}}(t)\cdot \mathbf{x}_j & -v_x(\mathbf{k})\ + & -v_y(\mathbf{k})\ + & -v_z(\mathbf{k})\ + \\
& & F_i(t) D_{ix}(t',t) & F_i(t) D_{iy}(t',t) & F_i(t) D_{iz}(t',t) \\[5pt]
v_x(\mathbf{k}(t',t)) & -v_x(\mathbf{k}) & -D_{xx}(t',t) & -D_{xy}(t',t) & -D_{xz}(t',t) \\[5pt]
v_y(\mathbf{k}(t',t)) & -v_y(\mathbf{k}) & -D_{yx}(t',t) & -D_{yy}(t',t) & -D_{yz}(t',t) \\[5pt]
v_z(\mathbf{k}(t',t)) & -v_z(\mathbf{k}) & -D_{zx}(t',t) & -D_{zy}(t',t) & -D_{zz}(t',t) \\[5pt]
\end{array} } \right \rvert_{(t'=t_b+i\delta, t=t_r, \mathbf{k} = \mathbf{k}_s)} \text{.}
\label{hessian2}\end{aligned}$$
Here, $i,j\!\!\in\!\!\{x,y,z\}$, summation is implied when indices $i$ or $j$ are repeated, $D_{ij} = \int_{t'}^t d\tau \beta_{ij}(\mathbf{k}(t'',t))$, $\beta_{ij} = \partial_{k_i}
v_j(\mathbf{k})$, and $\dot{\mathbf{F}}(t)=\partial_t\mathbf{F}(t)$. For completeness $\lvert \mathcal{H} \rvert$ is given for a general field $\mathbf{F}(t)$; for the case treated here, set $F_y = F_z = 0$. To leading order $\lvert \mathcal{H} \rvert = v_x(\mathbf{k}) \mathbf{f}(t',t,\mathbf{k}) + \dot{\mathbf{F}}(t) \cdot\mathbf{x}_l h(t',t,\mathbf{k})$, where $h, \mathbf{f}$ are minors of $\lvert \mathcal{H} \rvert$. For completeness, we have included time derivatives of the laser field which are however small in the long wavelength limit. As a result the leading order term is $\lvert \mathcal{H} \rvert = v_x(\mathbf{k}) \mathbf{f}(t',t,\mathbf{k})$.
Delta function potential {#dltref}
========================
![\[figB\] (a) Bloch dipole transition elements Im$[d^\ast(k)]$ versus $k$; (b) Wanner dipole transition elements $d_j$ versus $j$ which represents the difference in lattice sites at which electron and hole are born. 1D model parameters: $a=7$, $\Omega = 0.5$ (black), $\Omega = 1.5$ (red).](figB.eps){width="8.6cm"}
The WQC approach and its physical significance are explored by means of a 1D delta-function model potential, $V(x) = \Omega
\sum_{n=-\infty}^{\infty} \delta[x-(n+1/2)a]$ with unit cell size $a$ and barrier penetration parameter $\Omega$. For the investigated parameters the bandgap is well approximated by the nearest neighbor approximation, $\varepsilon = E_g + \Delta [1 - \cos(ka) ]$, where $E_g$ is the minimum bandgap and $2 \Delta$ represents the bandwidth.
The binding energy is determined by $2 E_m = K_m^2$, where $m=v,c$ and $K_m$ is determined by $$\begin{aligned}
\cos(ka) = \cos(K_m a) + \frac{\Omega}{K_m} \sin(K_m a) \text{.}
\label{senergy}\end{aligned}$$ The wavefunction is given by $$\begin{aligned}
\Phi_{{m},{k}}({x}) & = \sqrt{\frac{1}{a}} u_{m,k}(x) \exp(ikx) \label{swf} \\
u_{m,k}(x) & = A_m(k) \left[ e^{i(K_m-k)x} + r_m e^{-i(K_m+k)x} \right] \nonumber \\
A_m(k) & = 1 / \sqrt{1 + r_m^2 + 2 r_m \sin(K_m a) / (K_m a)} \nonumber \\
r_m(k) & = \frac{\sin[(K_m-k)a/2]}{\sin[(K_m + k)a/2]} \nonumber\end{aligned}$$ From the wavefunction the Bloch dipole moment is found to be $$\begin{aligned}
d_{cv}(k) & = d^*(k) = i \frac{2 A_c A_v}{E_c - E_v} \times \label{sdipole} \\
& \left\{ [ (K_v-k)r_c - (K_v+k)r_v ] \frac{\sin[ (K_v+K_c) a/2]}{(K_v+K_c) a} + \right. \nonumber \\
& \left. + [ (K_v-k) - (K_v+k)r_v r_c ] \frac{\sin[ (K_v-K_c) a/2]}{(K_v-K_c) a} \right\} \nonumber \text{.} \end{aligned}$$
We chose $a\!\!=\!\!7$ and $\Omega\!\!=\!\!0.5,1.5$ to model a weakly and red more tightly bound semiconductor, respectively. The corresponding bandgap parameters are $E_g\! =\! 0.141 , 0.269$; $\Delta \! =\! 0.269 , 0.17$. The Bloch dipole elements $d(k)$ and Wannier dipole elements $d_j$ are plotted in figure \[figB\]. As expected, $d_j$ drops faster for the more tightly bound model. Finally, we have chosen the coordinate center at the point of inversion symmetry which corresponds with choosing a maximally localized Wannier basis [@Kohn1959]. For this choice the diagonal (intraband) dipole moments are zero and the phase of the interband dipole moment is constant.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Assuming that neutrinos have non-vanishing magnetic moments, we discuss the possibility of pair production through annihilation of charged fermions in high-energy collisions. Adopting the Pauli interaction for photon-neutrino coupling, we calculate the neutrino pair production cross section in the photon channel and compare the result with the standard model in $ Z^{0} $ channel. we demonstrated that the enhancement of the production rate for Majorana neutrino pairs over the standard model rate can be possible at the center-of-mass energy of $10 - 100$ TeV for the Large Hadron Collider or the ultra-high-energy cosmic Ray when the transition magnetic moment is not smaller than $10^{-9} - 10^{-10} \mu_B$.'
author:
- 'Y. M.'
- 'H. K.'
- 'W.-G.'
- 'Y.'
title: Pair Production of Majorana Neutrinos by Annihilation of Charged Particles in High Energy Collision
---
INTRODUCTION
============
Atmospheric and solar neutrino observations [@ashie; @aharmim], as well as reactor experiments [@Araki], provide us strong evidence of the oscillation between different flavors of neutrinos, which can’t be possible if neutrinos are massless. Neutrinos are massive, electrically neutral fermions with spin 1/2. Although they are neutral fermions, it has been an interesting question how neutrinos can couple to photons. One of the possibilities is the non-vanishing magnetic moment of neutrinos, which can induce a spin-dependent coupling to photons.
Experimental bounds for the neutrino magnetic moments have been obtained from experiments for the solar neutrinos, accelerator neutrinos and reactor neutrinos [@wong]. The upper bound of the neutrino magnetic moment is found to be in the range of $10^{-10} - 10^{-7}$ $\mu_B$. Here, $\mu_B$ is the Bohr magneton. A model-dependent bound can also be obtained, for example, in big-bang nucleosynthesis and SN87a, to be $10^{-12} - 10^{-10}$ $\mu_B$ [@gg; @elmfors; @ag]. The theoretical bounds [@bell] were also discussed recently to get a rather wider range of upper bounds, $10^{-15} - 10^{-7}$ $\mu_B$. In the standard model, the neutrino magnetic moment induced by the one-loop effect [@fujikawa] is $
\mu_{\nu} = 3 \times
10^{-19} \left(\frac{m_{\nu}}{eV}\right) \mu_{B}$, which is much smaller than the above bounds.
In this work, we take the magnetic moment as a parameter, which might probe the physics beyond the standard model. It is interesting to note that the effect of the magnetic moments of neutrinos on the vacuum instability has recently been investigated in the presence of a strong external magnetic field [@yoon] to find out that with non-vanishing magnetic moment the vacuum instability appears beyond the critical field strength, $B_c = \frac{m_{\nu}}{\mu_{\nu}}$, against the pair production of neutrinos.
We consider a process of charged fermion-antifermion annihilation into neutrino pairs. The standard process is a pair production through the $Z^0$ channel. If neutrinos have non-vanishing magnetic moments, they can also be produced in the photon channel through the Pauli interaction [@pauli] on top of the standard process. Previously, the production of a massive neutrino through a magnetic interaction has been calculated and discussed for obtaining the experimental bounds [@barut; @schgal; @deshpande; @hklyoon].
In this work, we calculate the differential and the total cross sections for pair production of Majorana neutrinos with transition magnetic moments in the photon channel through the Pauli interaction. The Majorana neutrino is known to have only a transition magnetic moment, which implies that the lepton flavor number is not conserved: a pair produced via the transition magnetic moment consists of two different lepton flavors. We consider an extreme process with ultra-high energy, $E_{CM} > 10$ TeV , which is possible for the hadronic collision in the Large Hadron Collider (LHC) and $E^{GZK}_{CM} \sim 100$ TeV [@note] in the ultra-high-energy cosmic ray (UHECR). Because of the momentum-dependent coupling in the Pauli interaction, the total cross section in the photon channel becomes constant in the high-energy region, $E \gg m_i $, while the cross section in the $Z^0$ channel decreases as the energy scale increases. Hence, there is a critical energy beyond which the Pauli interaction dominates the SM process. We demonstrate that the critical energy can be in the energy region of the LHC or the UHECR provided the transition magnetic moment is not much smaller than $10^{-9} - 10^{-10}$ $\mu_B$. It is also found that the angular distribution of the differential cross section in the Pauli interaction has a maximum at $\theta =
\pi/2$ in the center-of-mass system.
In Section II, The basic feature of the Pauli interaction and the cross section in the center-of-mass system are discussed. In section III, the cross section for Majorana neutrino pairs is discussed in detail, and discussions are given in Section IV.
Pauli interaction
=================
In relativistic quantum theory, the standard picture is that the motion of a charged fermion is governed by the Dirac equation with an interaction with the external electromagnetic field, $A^{\mu}$, in a gauge-covariant way: $$\begin{aligned}
i\hbar \frac{\partial \psi}{\partial t}=\left[c \boldsymbol{ \alpha } \cdot \left(\boldsymbol{p}-\frac{e}{c}\boldsymbol{A}\right)
+ \beta m c^{2} +e \Phi \right] \psi.\end{aligned}$$ Pauli introduced a new form of the interaction with a magnetic moment [@pauli]: $$\begin{aligned}
\mathcal{L}^{Pauli}= l \frac{1}{2} \psi^{\dagger} \gamma^{\mu}\gamma^{\nu} \psi F_{\mu \nu},\end{aligned}$$ where the $F_{\mu \nu}$ are the external field strengths in natural units and $l$ has the dimension of a length. The corresponding Dirac-Pauli Lagrangian is given by $$\begin{aligned}
{\cal L} =
\bar{\psi}({p\!\!\!/}+\frac{\mu}{2}\sigma^{\mu\nu}F_{\mu\nu}-m)\psi,\label{pauli}\end{aligned}$$ where $\sigma^{\mu\nu}=\frac{i}{2}[\gamma^{\mu},\gamma^{\nu}]$ and $g_{\mu\nu}=(+,-,-,-)$. $\mu$ in the Pauli term measures the magnitude of the magnetic moment of the neutral fermion. This Lagrangian describes the interaction of the neutral fermion, but with a non-vanishing magnetic moment coupled with the external electromagnetic field through the Pauli interaction. The quantum mechanics with the Pauli interaction has been investigated for various types of electromagnetic fields [@lin]. The vacuum instability against the pair production rate of neutral fermions in linear magnetic fields through the Pauli interaction has been calculated [@yoon].
Neutrinos do not have electric charges, but they are found to have a non-vanishing mass. Also, it is natural to ask about the possibility of magnetic moments. However, it is not clear so far whether they have non-vanishing magnetic moments through which they can interact with photons directly. We assume, in this work, the case where the massive neutrinos have non-vanishing magnetic moments or transition magnetic moments. Then, we can consider a process in which the neutrino pairs can be produced through photon exchange, in addition to the weak process by $Z_0$ boson exchange.
In this work, we calculate the cross section of neutrino pair production through the Pauli interaction to investigate the observational effect of the magnetic moment of neutrinos. We first consider a process in which neutrino pairs are produced by the annihilation of a charged fermion, $q$, with charge $Q e$, which has minimal coupling to a photon, lepton, or parton in the hadron collision. In high-energy hadron collisions, the charged fermions can be considered to be those of partons in hadrons.
The process we are considering is the annihilation of charged fermions $q$ and $\bar{q}$ into a neutrino through the photon channel with the Pauli interaction. The interaction Lagrangian for neutrino pair production is $$\begin{aligned}
\mathcal{L}_{int} = \frac{\mu}{2} \bar{\psi}_{\nu}\sigma^{\nu\rho}F_{\nu\rho}\psi_{\nu}.\end{aligned}$$
In the center-of-mass frame, we get the differential cross section [@hklyoon]: $$\begin{aligned}
\left(\frac{d \sigma }{d \Omega }\right)_{D} &=& \frac{Q^2\alpha
\mu^2}{16 \pi} \sqrt{\frac{1- \frac{m_{\nu}^2}{E^2}}{1-
\frac{m_{q}^2}{E^2}}} \left[ \right. 1+\frac{m_{\nu}^2}{E^2}+\frac{m_{q}^2
m_{\nu}^2}{E^4} \nonumber \\
&& - \left( 1-\frac{m_q^2}{E^2}\right)\left(1-
\frac{m_{\nu}^2}{E^2}\right)
\cos^2{\theta} \left. \right]\label{dsigma}\end{aligned}$$ and the total cross section $$\begin{aligned}
\sigma_{D}
= \frac{1}{6} Q^2\alpha \mu^2 \sqrt{\frac{1- \frac{m_{\nu}^2}{E^2}}{1- \frac{m_{q}^2}{E^2}}}
\left(1+ \frac{m_q^2}{2E^2} \right) \left(1+ \frac{2m_{\nu}^2}{E^2}
\right). \label{sigmat}\end{aligned}$$ Here, D denotes Dirac type neutrinos. A similar calculation has been done by Barut [*et al.*]{} [@barut], where the total cross section is calculated as $$\begin{aligned}
\sigma_{B}&=& \frac{ Q^2 \alpha \kappa^2 \sqrt{1- \frac{m_{\nu}^2}{E^2}}}{6 \sqrt{1 - \frac{m_{e}^2}{E^2}}} \left( 1 + \frac{1}{2} \frac{m_{e}^2}{E^2}
\right) \left( 1 + \frac{1}{2} \frac{m_{\nu}^2}{E^2} \right),\end{aligned}$$ where $\kappa$ is the neutrino magnetic moment in their convention. There is a small difference in the last term, which might be due to the additional $\left(1+\gamma_{5}\right)$ term in their interaction Lagrangian whereas we consider the case without chirality for the Pauli coupling. At high energy, $m_{\nu}$, $m_{q}$, $m_{e}\ll E$, which is the scale at the LHC or the UHECR of our interest, they give the same result modulo the coupling constants: $$\begin{aligned}
\left( \frac{d \sigma}{d \Omega} \right)_i = \eta_i~ \sin^2
\theta, ~~
\sigma_i
= \frac{8 \pi}{3} \eta_{i}, \label{sigmath}\end{aligned}$$ where $i$’s stand for the choice of coupling constants, $\eta_D = Q^2 \alpha \mu^2/ 16 \pi $, and $\eta_B = Q^2 \alpha \kappa^2/ 8 \pi$. The cross section becomes constant at high energy. If these are valid all the way to higher energy, then there is the problem of violation of the unitary bound. However, the Pauli coupling is an effective interaction term that is valid only up to some scale, and we assume it to be higher than the scale we are considering in this work. Now, it is interesting to note that energy dependence is quite different from that of pair production through the Z-boson channel in the standard model, where the cross section decreases with increasing colliding energy. The comparison and the possible implication will be discussed in detail in the final section.
Pair production cross section of Majorana neutrinos
====================================================
The Majorana field is basically represented by a two-component spinor, $\chi$. For a free particle, the Lagrangian of the two-component Majorana field is given by $$\begin{aligned}
\mathcal{L}_{M}=\chi^{\dagger}i\bar{\sigma} \cdot \partial \chi -\frac{m}{2}\left[\left(\chi^{C}\right)^{\dagger}\chi+\chi^{\dagger}\chi^{C}\right].\end{aligned}$$ $\chi$ can be written as $$\begin{aligned}
\chi &=& \sum_{s}\int \frac{d^{3}\vec{p}}{\left(2\pi\right)^{3/2}\left(2E_{p}\right)^{1/2}}\left[ \right. f\left(\vec{p},s\right)a\left(\vec{p},s\right) e^{-ip \cdot x} \nonumber \\
&&+g\left(\vec{p},s\right)a^{\dagger}\left(\vec{p},s\right)e^{ip \cdot x} \left. \right],\end{aligned}$$ where $f$ and $g$ are two-component spinors that satisfy the equation of motion for Majorana field, $$\begin{aligned}
i \bar{\sigma} \cdot \partial \chi - i m \sigma^{2} \chi^{\ast} = 0. \label{Majoranaeq}\end{aligned}$$ Then, it is possible to construct a four-component Majorana field: $$\begin{aligned}
\Psi_{M}
&=& \sum_{s}\int \frac{d^{3}\vec{p}}{\left(2\pi\right)^{3/2}\left(2E_{p}\right)^{1/2}}\left[u\left(\vec{p},s\right)a\left(\vec{p},s\right)
e^{-ip \cdot x} \right. \nonumber \\
&& \left. +v\left(\vec{p},s\right)a^{\dagger}\left(\vec{p},s\right)e^{ip \cdot x} \right],\end{aligned}$$ where $$\begin{aligned}
u \equiv \left( \begin{array}{c} f\left(\vec{p},s\right) \\ g^{C}\left(\vec{p},s\right) \end{array} \right), ~~ v \equiv \left( \begin{array}{c} g\left(\vec{p},s\right) \\ f^{C}\left(\vec{p},s\right) \end{array} \right).\end{aligned}$$ Now, the interaction Lagrangian for the Majorana neutrino with the Pauli interaction can be written as $$\begin{aligned}
\mathcal{L}_{int} &=& i \frac{\mu^{ij}}{2} \bar{\Psi}^{i}_{M} \sigma_{\mu\nu} \Psi^{j}_{M} F^{\mu \nu}, \label{LagM}\end{aligned}$$ where $\sigma_{\mu\nu}=\frac{i}{2}\left[\gamma_{\mu},\gamma_{\nu}\right]$ and $g_{\mu\nu} = \left(+,-,-,-\right)$. $\mu^{ij}$ is a transition magnetic moment that is antisymmetric for a Majorana neutrino, $\mu_{ij}=-\mu_{ji}$.
The differential cross section in the center of mass (CM) can be calculated in a straight forward way: $$\begin{aligned}
\label{e51-1}
&&\left( \frac{d\sigma}{d\Omega} \right)_{M} =\frac{\alpha Q^2\mu^2_{12}}{4\pi}\sqrt{\frac{E^2_1-m^2_{\nu_1}}{E^2-m^2_q}} \\
&& \times \left[ \frac{E_1 E_2}{E^2} + \frac{E_1 E_2 m^2_q}{2E^4} + \frac{m_{\nu_1}m_{\nu_2}}{E^2} + \frac{m^2_q m_{\nu_1}m_{\nu_2}}{2E^4} \right. \nonumber\\
&& \left. \quad - \frac{E_1 E_2 m^2_q \sqrt{1-\frac{m^2_{\nu_1}}{E^2_1}}\sqrt{1-\frac{m^2_{\nu_2}}{E^2_2}}}{2E^4} \right. \nonumber \\
&& \left. \quad -\frac{E_1 E_2}{E^2}\mathrm{ cos^2 \, \theta} \left(1-\frac{m^2_q}{E^2}\right)\sqrt{1-\frac{m^2_{\nu_1}}{E^2_1}}\sqrt{1-\frac{m^2_{\nu_2}}{E^2_2}} \right],\nonumber\end{aligned}$$ where $\alpha$ is a fine-structure constant and $M$ denotes the Majorana neutrino. After integrating over $d\Omega$, the total cross section is given by $$\begin{aligned}
\label{e52}
&& \sigma_{M} = \frac{\alpha Q^2 \mu^2_{12}}{6}\frac{E_1\left(2+\frac{m^2_q}{E^2}\right)}{E\left(1-\frac{m^2_q}{E^2}\right)} \sqrt{1-\frac{m^2_{\nu_1}}{E^2_1}} \nonumber \\
&& \times ( \frac{3E_1 E_2}{E^2}+ \frac{3m_{\nu_1}m_{\nu_2}}{E^2} \nonumber \\
&& -\frac{E_1 E_2}{E^2} \sqrt{1-\frac{m^2_{\nu_1}}{E^2_1}}\sqrt{1-\frac{m^2_{\nu_2}}{E^2_2}} ).\end{aligned}$$ In the high-energy region, where the particle masses are very small compared to the energy scale, $m_i \ll E$, the differential cross section and the total cross section, respectively, converge to simple expressions: $$\begin{aligned}
\left(\frac{d\sigma}{d\Omega}\right)_M&=&\frac{\alpha Q^2\mu^2_{12}}{4\pi} \mathrm{ sin^2 \, \theta}\label{e53}\end{aligned}$$ and $$\begin{aligned}
\sigma_M = \frac{2\alpha Q^2 \mu^2_{12}}{3},\label{e54}\end{aligned}$$ which are similar to the results for a Dirac neutrino with magnetic moment.
For comparison with the neutrino pair production in the SM, in the high-energy limit $E \gg m_{\nu},M_{Z} $, the differential cross section and the total cross section in the standard model are known to behave as $$\begin{aligned}
\left( \frac{d\sigma}{d\Omega} \right)_{SM} \propto (1+\cos^{2}
\theta), ~~
\sigma_{SM} \propto \left(\frac{1}{E} \right)^2.\end{aligned}$$ We can see that the total cross section behaves as $\frac{1}{E^{2}}$ in the high-energy limit. The angular distribution of the differential cross section is maximum for $\theta = 0$ and $\pi$ and minimum for $\theta \sim \pi/2$. These features are quite different from those with the Pauli interaction, Eqs. (\[e53\]) and (\[e54\]).
Discussion
==========
We calculate the high-energy behavior of the cross section for Majorana neutrino pair production, assuming that the neutrinos have non-vanishing transition magnetic moments and are interacting electromagnetically with the Pauli interaction. we found that the production cross section for Majorana-type neutrino production is similar to that of Dirac-type neutrino production. The angular distribution in the center-of-mass frame peaks at $\theta = \pi/2$ while the angular distribution in the standard model has a minimum at $\theta \sim \pi/2$. The total cross section turns out to be independent of energy: $$\begin{aligned}
\sigma_{M} & = & 1.66 \times 10^{-29} Q^{2} m^2 (\tilde{\mu})^2,
\label{sigmatn} \end{aligned}$$ where $\mu_{12} \equiv \tilde{\mu} \mu_B$. This is basically because the neutrino interaction vertex carries a momentum factor of the virtual photon, which cancels the energy dependence, which is otherwise inversely proportional to the square of the energy in the center-of-mass frame as is the case for the standard model. Hence, we can expect the neutrino production through the Pauli interaction to compete with that of the SM for high-energy collisions. The energy scale, $E_{0.1}$, for which $\sigma_{M}$ becomes $0.1 ~ \sigma_{SM}$, can be estimated as $$\begin{aligned}
E_{0.1} \sim 10^2 \left(\frac{10^{-10}}{\tilde{\mu}}\right) TeV.\end{aligned}$$ For the Majorana-type neutrino, the upper bound is somewhat less stringent (although model dependent) than it is for the Dirac neutrino. If we take $\tilde{\mu} \lesssim 10^{-9} - 10^{-10}$[@numu], we get $E_{0.1} \sim 10 - 100$ $TeV$, which can be reached in LHC and in UHECR experiments. However, if $\tilde{\mu} < 10^{-11}$, the corresponding energy scale becomes higher, beyond the GZK cutoff. One of the characteristics of neutrino production by Pauli coupling is that the differential cross section has a maximum value $\theta = \pi/2$ in the center-of-mass frame, compared to the SM, which predicts a minimum at $\theta = \pi/2$.
Since the magnetic moment for a Majorana neutrino in this process is not diagonal and can have only a transition magnetic moment, if the neutrinos produced are Majorana type, then pairs should be produced with different flavors. This difference gives us an additional way to find out which type of neutrino is produced, Majorana or Dirac. However, it should be noted that most of the present experimental detector systems are such that the neutrinos produced in high-energy collisions escape detection. Hence, for this purpose, we need detecting systems dedicated to high-energy neutrinos, for example, in high-energy cosmic ray experiments.
In summary, we discuss an observational possibility of a neutrino magnetic moment at high-energy experiments and/or high-energy cosmic-ray experiments. Although the present energy scale is found not to be sufficiently high enough for magnetic moments smaller than $\sim 10^{-11} \mu_B$, the neutrino magnetic moment with Pauli coupling can open an interesting channel in future experiments, through which the type of neutrino can be distinguished.
The authors would like to thank Byung-Gu Cheon for useful discussions. This work is supported by the World Class University (WCU) project of the Korean Ministry of Education, Science, and Technology (R33-2008-000-10087-0).
Y. Ashie *et al*. \[The Super-Kamiokande Collaboration\], Phys. Rev. Lett. [**93**]{}, 101801 (2004).
B. Aharmim *et al*. \[SNO Collaboration \], Phys. Rev. C [**72**]{}, 055502 (2005). T. Araki *et al*. \[KamLAND Collaboration\], Phys. Rev. Lett. [**94**]{}, 081801 (2005).
H. T. Wong and H-B Li, Mod. Phys. Lett. A [**20**]{}, 1103(2005).
G. G. Raffelt, Phys. Rep. [**320**]{}, 319(1999). Per Elmfors, Kari Enqvist, Georg Raffelt, and Gunter Sigl, Nucl. Phys. B [**503**]{}, 3(1997). Ashok Goyal, Sukanta Dutta, and S. R. Choudhury, Phys. Lett. B [**346**]{}, 312(1995).
N. F. Bell, Mikhail Gorchtein, Michael J. Ramsey-Musolf, Petr Vogel, and Peng Wang, Phys. Lett. B [**642**]{}, 4 (2006). K. Fujikawa and R. E. Shrock, Phys. Rev. Lett. [**45**]{}, 963 (1980); M. Dvornikov and A. Studenikin, Phys. Rev. D [**69**]{}, 073001 (2004) and references therein. H. K. Lee and Y. S. Yoon, JHEP [**0603:078**]{} (2006); H. K. Lee and Y. S. Yoon, JHEP [**0703:086**]{} (2007). W. Pauli, Rev. Mod. Phys. [**13**]{}, 203 (1941).
A. O. Barut, Z. Z. Aydin, and I. H. Duru, Phys. Rev. D [**26**]{}, 1794 (1982). L. M. Sehgal and A. Weber, Phys. Rev. D [**46**]{}, 2252 (1992). N. G. Deshpande and K. V. L. Sarma, Phys. Rev. D [**43**]{}, 943 (1991).
H. K. Lee and Y. Yoon, Mod. Phys. Lett. A [**23**]{}, 1447 (2009). $E^{GZK}_{CM}$ is the center of mass energy of UHECR ( $E^{GZK}\sim 10^{19}$ eV) and a proton at rest ($E_p=m_p$), $\sqrt{2 E^{GZK}m_p} \sim 100$ TeV.
See, for example, C. L. Ho and P. Roy, Ann. Phys. [**312**]{}, 161 (2004) and references therein.
N. F. Bell, Int. J. Mod. Phys. A [**22**]{}, 4891 (2007).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We report the identification of LP 400-22 (WD 2234+222) as a very low-mass and high-velocity white dwarf. The ultraviolet [*GALEX*]{} and optical photometric colors and a spectral line analysis of LP 400-22 show this star to have an effective temperature of $11\,080\pm140$ K and a surface gravity of $\log{g} = 6.32\pm0.08$. Therefore, this is a helium core white dwarf with a mass of $0.17\ M_\odot$. The tangential velocity of this white dwarf is $414\pm43$ km s$^{-1}$, making it one of the fastest moving white dwarfs known. We discuss probable evolutionary scenarios for this remarkable object.'
author:
- Adela Kawka
- 'Stéphane Vennes, Terry D. Oswalt, J. Allyn Smith'
- 'Nicole M. Silvestri'
title: 'LP 400-22, a Very Low-mass and High-velocity White Dwarf'
---
Introduction
============
The vast majority of white dwarfs evolve from normal main-sequence stars following normal evolutionary processes. However, ultramassive ($>1.1M_\odot$) and inframassive ($<0.40M_\odot$) white dwarfs require special evolutionary paths. The formation of low-mass helium white dwarfs ($M_{\rm WD} \la 0.4M_\odot$) has been shown to be the result of close binary evolution [@ibe1986 and references therein]. Indeed, the Galaxy is not old enough for these objects to have formed through single star evolution. The general evolutionary scenario for the formation of low-mass helium white dwarfs is that the companion stripped the white dwarf of its envelope before completing its red giant evolution [@kip1967].
Recently, several very low mass white dwarfs ($M_{\rm WD} \la 0.2 M_\odot$) have been discovered as companions to pulsars [@van2005]. The orbital periods vary from a few hours to several years. The masses of some of these white dwarfs may be determined from the Shapiro delay of radio pulses provided that the system is nearly edge on [@loh2005]. For example, @jac2003 deduced a mass of $0.20\ M_\odot$ for the companion of PSR J1909$-$3744, and obtained a spectrum which confirmed, at least qualitatively, the presence of a low mass DA white dwarf. In addition, the masses of the companions to PSR J1012$+$5307 and PSR J1911$-$5958 were determined spectroscopically to be $0.16\ M_\odot$ [@van1996; @cal1998] and $0.18 M_\odot$ [@bas2006], respectively. Finally, several low-mass white dwarf candidates were found in the Sloan Digital Sky Survey [@kle2004]. @lie2004 analyzed the brightest star in the sample, SDSS J123410.37$-$022802.9 and showed that it has a mass of $\sim 0.18\ M_\odot$ and that it does not have an obvious neutron star companion.
In this paper, we report the identification of a high velocity white dwarf with a very low mass, LP 400-22 (WD2234+222[^1], NLTT 54331). Our photometric and spectroscopic observations are presented in §2.1 and 2.2 respectively. We derive the stellar parameters in §3 and discuss our results in §4.
Observations
============
LP 400-22 was spectroscopically identified as a white dwarf as part of a survey of common-proper motion binaries with suspected white dwarf components [@osw1993]. We obtained additional high-resolution optical spectra [@sil2002] as well as new optical photometry [@smi1997] as part of the same project. More recently, LP 400-22 was observed during the [*Galaxy Evolution Explorer*]{} ([*GALEX*]{}) all-sky survey.
Photometry
----------
The $BVRI$ photometry for LP 400-21/22 were obtained with the 2.1m telescope at KPNO on 1995 July 5 UT. A Tek1K CCD (with $24\mu$m pixels) operating at the Cassegrain focus was used, providing $0\farcs305$ per pixel and a $5\farcm2$ field of view. The data for LP 400-21/22 were obtained under photometric conditions. Standard stars for this program were chosen from @lan1992.
We obtained ultraviolet (UV) photometry from [*GALEX*]{} All-Sky Survey[^2]. [*GALEX*]{} provides photometry in two bands, FUV and NUV, which are based on the AB system [@mor2005; @oke1983]. The bandwidth of FUV is 1344 to 1786 Å with an effective wavelength of 1528 Å. The bandwidth of NUV is 1771 to 2831 Å with an effective wavelength of 2271 Å.
[ccc]{} FUV & $18.38\pm0.09$ mag &\
& $18.18\pm0.08$ mag &\
NUV & $18.19\pm0.05$ mag &\
& $18.14\pm0.04$ mag &\
B & $17.338\pm0.025$ mag & $18.742\pm0.025$ mag\
V & $17.219\pm0.021$ mag & $17.177\pm0.021$ mag\
R & $17.202\pm0.023$ mag & $15.933\pm0.023$ mag\
I & $17.210\pm0.024$ mag & $14.340\pm0.023$ mag\
Table \[tbl\_phot\] presents the optical and ultraviolet photometry and Figure \[fig\_energy\_dist\] shows the energy distribution compared to a synthetic spectrum.
Spectroscopy
------------
We obtained a low-resolution spectrum of LP 400-22 using the R-C spectrograph attached to the 4 m telescope at Kitt Peak National Observatory (KPNO) on 1988 October 6. The BL250 grating (158 lines/mm) was used to obtain a spectral range of 3500 to 6200 Å with a dispersion of 5.52 Å per pixel and a resolution of 14 Å.
LP 400-22 was re-observed using the Dual Imaging Spectrogram (DIS) attached to the 3.5 m telescope at the Apache Point Observatory (APO) on 2001 July 10 and October 14. The 1200 lines/mm grating was used to obtain a spectral range of 3800 to 4600 Å with a dispersion of 1.6 Å per pixel, and the 830.8 lines/mm grating was used to obtain a spectral range of 6180 to 7210 Å with a dispersion of 1.3 Å per pixel. A $1\farcs5$ slit was used to obtain a spectral resolution of $\sim 2$ Å in the blue and $\sim 2.6$ Å in the red.
Determining the Parameters
==========================
In our analysis of LP 400-22, we used a grid of computed pure hydrogen LTE plane parallel models (see [@kaw2006] and references therein for details). The grid of models extend from $T_{\rm eff} = 7000$ to 16000 K (in steps of 1000 K), from 18000 to 32000 K (in steps of 2000 K) and from 36000 to 84000 K (in steps of 4000 K) at $\log{g} = 6.0$ to 9.5 (in steps of 0.25 dex). All our $\log{g}$ values are in cgs. We also prepared corresponding grids of synthetic spectra, one of which includes the effect of Ly$\alpha$ satellites [@all1992], and the other excludes that effect.
Photometry
----------
Using our spectral grid, we have calculated synthetic optical ($BVRI$) and ultraviolet ($FUV/NUV$) colors. Figure \[fig\_vmf\_fmn\] shows the observed photometric colors ($V-FUV$ versus $FUV-NUV$ and $B-V$ versus $V-R$) of LP 400-22 compared to synthetic white dwarf and main-sequence colors. We used Kurucz synthetic spectra [@kur1993] to calculate our main-sequence colors.
In the UV-optical diagram ($V-FUV$/$FUV-NUV$) of Figure \[fig\_vmf\_fmn\] we show two sets of WD synthetic colors. The grid shown in black includes the effect of Ly$\alpha$ satellites [@all1992] as compared to the grid in green which excludes them. A comparison of the two grids shows the significant effect that the Ly$\alpha$ satellites have on the UV colors at $T_{\rm eff} < 13\,000$ K. Comparing the UV-optical photometry of LP 400-22 to the white dwarf grid, a low surface gravity $\log{g} \sim 6$ and an effective temperature of $\sim 11\,000$ K is implied. The optical diagram ($B-V$/$V-R$) in Figure \[fig\_vmf\_fmn\] confirms the white dwarf temperature of 11000 K and the low surface gravity.
However, when comparing the photometry to main-sequence colors, a A3V spectral type is implied in the optical and a B8V spectral type in the ultraviolet. Therefore, the data are incompatible with main-sequence colors. Optical and UV colors are useful to distinguish white dwarfs from main-sequence stars.
Spectroscopy
------------
The Balmer lines of LP 400-22 were analyzed in all three available spectra using a $\chi^2$ minimization technique. The quoted uncertainties are statistical only ($1\sigma$). The Balmer lines (H$\beta$ to H9) in the KPNO spectrum were fitted with model spectra which were smoothed to the instrumental resolution of 14 Å, to obtain $T_{\rm eff} = 11\,000\pm350$ K and $\log{g} = 6.48\pm0.27$. For the two high-resolution APO spectra we fitted H$\alpha$ and H$\gamma$ to H9 with model spectra, to obtain $T_{\rm eff} = 11\,060\pm180$ K and $11\,160\pm250$ K, and $\log{g} = 6.46\pm0.13$ and $6.22\pm0.10$. The synthetic spectra used in the analysis of the APO spectra were smoothed with a gaussian profile to the instrumental resolution of 2 Å. Note that the disprepancy in the surface gravities from the 2 APO spectra are most likely the result of uncertainties in the flux calibration around the higher Balmer lines. The Balmer line fit of the KPNO spectrum is shown in Figure \[fig\_specfit\_WD2234+222\]. These measurements clearly confirm that LP 400-22 is a white dwarf with a low surface gravity. The calculated weighted average of the temperature and surface gravity is $T_{\rm eff} = 11\,080\pm140$ K and $\log{g} = 6.32\pm0.08$.
We used the evolutionary tracks for helium-core white dwarfs of @alt2001 and @ser2001 to determine a mass of $0.17\pm0.01\ M_\odot$ and a cooling age of $5\pm1\times 10^8$ years. Note that the cooling age of the white dwarf is sensitive to the mass of the hydrogen envelope left before entering the final cooling track [see @alt2001 and references therein]. Residual H-burning in a thick H-envelope causes the white dwarf to cool slower as compared to a white dwarf with a thin H-envelope.
The temperature of 11080 K places LP 400-22 near the blue edge of the ZZ Ceti instability strip [@gia2005]. Given the lack of time coverage in our data, we cannot state whether the star is variable or not. Time-series photometry is required to explore variability and place constraints on the blue edge of the instability strip at the low-mass range.
Discussion
==========
LP 400-22 was listed in the New Luyten Two-Tenths (NLTT) catalog [@luy1979] to have a common proper motion companion (LP 400-21) $338\arcsec$ away. Recently, @sal2003 have revised the coordinates and proper motions of most stars in the NLTT catalog by cross-correlating the catalog with the Two Micron All Sky Survey (2MASS) and the USNO-A catalogs. However LP 400-22 was not detected in 2MASS, and therefore they relisted Luyten’s measurement of the proper motion. They listed the proper motion of LP 400-22 to be $\mu_{\alpha} = 0\farcs1950\pm0\farcs0200$ yr$^{-1}$ and $\mu_{\delta} = 0\farcs0563\pm0\farcs0200$ yr$^{-1}$. For LP 400-21 they measured a proper motion of $\mu_{\alpha} = 0\farcs2158\pm0\farcs0055$ yr$^{-1}$ and $\mu_{\delta} = 0\farcs0283\pm0\farcs0055$ yr$^{-1}$. These proper motion measurements agree within $2\sigma$ and on this basis the two stars appear to be a common proper motion binary. However, similar measurements were reported by @lep2005, i.e., $\mu_{\alpha} = 0\farcs198$ yr$^{-1}$ and $\mu_{\delta} = 0\farcs053$ yr$^{-1}$ for LP 400-22, and $\mu_{\alpha} = 0\farcs228$ yr$^{-1}$ $\mu_{\delta} = 0\farcs020$ yr$^{-1}$ for LP 400-21. The quoted uncertainties in the @lep2005 measurements are $\sim 0\farcs007$ yr$^{-1}$ and, therefore, the diverging proper motions of the two stars appears to exclude a physical association. Another way to check whether the stars are a physical binary is to determine the distance of each star.
To estimate the distance, we calculated an absolute magnitude of $M_V = 9.1\pm0.2$ mag for LP 400-22 and a distance modulus $(V - M_V) = 8.2$ mag. This places the white dwarf at a distance of $430\pm45$ pc. Note that the Galactic extinction for this object is low and its effect was not included. @sil2005 classified LP 400-21 a dM4.5e, and using the $M_V/V-I$ relation from @rei1997 we estimate the absolute magnitude of the red dwarf as 12.7 mag. The apparent V magnitude for LP 400-21 is $V=17.177\pm0.021$ mag, and, therefore, the red dwarf is at a distance of $\sim 80$ pc. @rei1997 note a scatter of values about the relation with $\sigma = 0.46$. Even if we consider LP 400-21 at the extrema of this dispersion, it would place it at a distance of only $\sim 100$ pc. The large distance discrepancy makes LP 400-22/21 a coincidental pair rather then a wide binary as has been thought based on their proper motion alone.
The large distance and high-proper motion of LP 400-22 imply a large tangential velocity of $414\pm43$ km s$^{-1}$. Only a few white dwarfs are known to have $v_{\rm tan} > 350$ km s$^{-1}$, with most of these having halo space velocities [@ber2005]. In order to obtain the $U,V,W$ space velocity components for LP 400-22, we measured the radial velocity of the white dwarf using H$\alpha$ in the APO high-dispersion spectra to obtain a heliocentric value of $-50\pm20$ km s$^{-1}$, which is different than the velocity measured for the red dwarf (7.3 km s$^{-1}$) by @sil2002b. We calculated $U,V,W$ for LP 400-22 using @joh1987 to obtain $U = -388\pm43,\ V=-81\pm22,\ W=-83\pm22$ km s$^{-1}$. These velocity components do not agree with either disk or halo populations [@chi2000] and suggest a different origin for its peculiar motion. The Galactic orbit for LP 400-22 should be calculated.
Most white dwarfs with $M<0.2\ M_\odot$ are companions to pulsars. We searched for radio sources in the vicinity of LP 400-22 using [*VizieR*]{}[^3], and the nearest was that of the galaxy KUG2234+223 $\sim7.5\arcmin$ away. Therefore, if LP 400-22 is a companion to a neutron star, then it is probably a dead pulsar. Another possibility is that LP 400-22 has a very cool companion, which should be detectable as infrared excess. However, LP 400-22 was not detected by 2MASS. The two high-dispersion velocity measurements agree within error bars. However, a series of radial velocity measurements should be obtained to establish whether or not LP 400-22 is in a close binary system.
Yet another possibility for the origin of LP 400-22 is that it may have once been in a close double-degenerate binary, where the companion has gone through a supernova event that disrupted the binary losing the remnant of the donor star with a high-velocity and a low mass [@han2003].
Summary
=======
[llc]{} Effective Temperature & $11\,080\pm140$ K & 1\
Surface Gravity & $6.32\pm0.08$ & 1\
Mass & $0.17\pm0.01 M_\odot$ & 1\
$M_V$ & $9.1\pm0.2$ mag & 1\
Distance & $430\pm45$ pc & 1\
Proper Motion & $\mu = 0\farcs203$ yr$^{-1}$,$0\farcs205$ yr$^{-1}$ & 2,3\
& $\theta = 73.9^\circ,\ 75.0^\circ$ & 2,3\
Kinematics & $U = -388\pm43$ km s$^{-1}$ & 1\
& $V = -81\pm22$ km s$^{-1}$ & 1\
& $W = -83\pm22$ km s$^{-1}$ & 1\
We have demonstrated that LP 400-22 is a high-velocity white dwarf with a very low mass ($M=0.17\ M_\odot$) and a temperature of $11\,080$ K. Table \[tbl\_par\] summarizes the properties of LP 400-22. Since white dwarfs with masses below $0.4\ M_\odot$ must have been formed in close binary systems, radial velocity measurements and infrared photometry are required to determine whether LP 400-22 has a close companion. On the other hand, a lack of radial velocity variations would indicate that LP 400-22 lost its close massive companion following a type Ia supernova event.
This research is supported in part by a NASA/GALEX grant (NNG05GE33G). A.K. is supported by GA ČR 205/05/P186. T.D.O. acknowledges support from the NSF (AST 0206115). J.A.S. was supported on a NASA GSRP Training Grant, NGT-51086. N.M.S. acknowledges support from a NASA GSRP grant (NST 200415) and NSF AST 02-05875. We thank Paul Hintzen for acquiring the original KPNO spectrum of LP 400-22. This research has made use of the VizieR catalogue access tool, CDS, Strasbourg, France. Based on observations obtained with the APO 3.5 m telescope, which is owned and operated by the Astrophysical Research Consortium.
Allard, N.F., & Koester, D. 1992, , 258, 464 Althaus, L.G., Serenelli, A.M., & Benvenuto, O.G. 2001, , 323, 471 Bassa, C.G., van Kerkwijk, M.H., Koester, D., & Verbunt, F. 2006, , in press (astro-ph/0603267) Bergeron, P., Ruiz, M.T., Hamuy, M., Leggett, S.K., Currie, M.J., Lajoie, C.-P., & Dufour, P. 2005, , 625, 838 Callanan, P.J., Garnavich, P.M., & Koester, D. 1998, , 298, 207 Chiba, M., & Beers, T.C. 2000, , 119, 2843 Hansen, B.M.S. 2003, , 582, 915 Gianninas, A., Bergeron, P., & Fontaine, G. 2005, , 631, 1100 Iben, I.J., & Tutukov, A.V. 1986, , 311, 742 Jacoby, B.A., Bailes, M., van Kerkwijk, M.H., Ord, S., Hotan, A., Kulkarni, S.R., & Anderson, S.B. 2003, , 599, L99 Johnson, D.R.H., & Soderblom, D.R. 1987, , 93, 864 Kawka, A., & Vennes, S. 2006, , 643, in press (astro-ph/0601477) Kippenhahn, R., Kohl, K., & Weigert, A. 1967, , 66, 58 Kleinman, S.J., et al. 2004, , 607, 426 Kurucz, R.L. 1993, CD-ROM 13, ATLAS9 Stellar Atmosphere Programs and 2 km s$^{-1}$ Grid, (Cambridge: SAO) Landolt, A.U. 1992, , 104, 372 Lepine, S., & Shara, M.M. 2005, , 129, 1483 Liebert, J., Bergeron, P., Eisenstein, D., Harris, H.C., Kleinman, S.J., Nitta, A., & Krzesinski, J. 2004, , 606, L147 Löhmer, O., Lewandowski, W., Wolszczan, A., & Wielebinski, R. 2005, , 621, 388 Luyten, W.J. 1979, New Luyten Catalogue of stars with Proper Motions larger than two tenths of an arcsecond (NLTT), University of Minnesota Morrissey, P., et al. 2005, , 619, L7 Oke, J.B., & Gunn, J.E. 1983, , 266, 713 Oswalt, T. D., Smith, J. A., Shufelt, S., Hintzen, P. M., Leggett, S. K., Liebert, J., & Sion, E. M. 1993, NATO ASIC Proc. 403: White Dwarfs: Advances in Observation and Theory, 419 Reid, I.N., & Gizis, J.E. 1997, , 113, 2246 Salim, S., & Gould, A. 2003, , 282, 1011 Serenelli, A.M., Althaus, L.G., Rohrmann, R.D., & Benvenuto, O.G. 2001, , 325, 607 Silvestri, N.M. 2002, PhD Thesis, Florida Institute of Technology Silvestri, N.M., Oswalt, T.D., & Hawley, S.L. 2002, , 124, 1118 Silvestri, N.M., Hawley, S.L., & Oswalt, T.D. 2005, , 129, 2428 Smith, J.A. 1997, PhD Thesis, Florida Institute of Technology van Kerkwijk, M.H., Bergeron, P., & Kulkarni, S.R. 1996, , 467, L89 van Kerkwijk, M.H., Bassa, C.G., Jacoby, B.A., & Jonker, P.G. 2005, ASP Conf. Ser. Vol. 328: Binary Radio Pulsars, eds. Rasio, F.A., & Stairs, I.H., 357
[^1]: Online at http://www.astronomy.villanova.edu/WDCatalog/index.html
[^2]: Available from http://galex.stsci.edu/GR1/
[^3]: http://vizier.u-strasbg.fr/viz-bin/VizieR
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'P. de Laverny'
- 'D. Mékarnia'
date: 'Received 6 September 2004 /Accepted 23 October 2004'
title: 'First detection of dust clouds around R CrB variable stars [^1]'
---
Introduction
============
[*“Le R Coridini (protebbero), per qualche motivo, cingersi (....) di gas oscuranti emessi a tratti da queste stelle stesse”*]{}. In this way, Loreta ([@loreta]) was the first to propose that R Coronae Borealis type variables (R CrB, hereafter) might eject absorbing material responsible of the huge brightness declines (or fadings) characterizing these stars in visible light. Then, O’Keefe ([@okeefe]) showed that this ejected material could condense at small distances from the central star and form obscuring clouds rich in carboneous compounds. These clouds would then gradually dissipate as they are driven away by radiation pressure.
For more than 70 years, this scenario for the interpretation of the light variations of R CrB variable stars has been widely accepted: the fadings are believed to be caused by obscurations of the stellar surface by newly formed dusty and optically thick clouds. These stars indeed exhibit drastic and erratic variabilities. Their visual lightcurve is characterized by unpredicted decreases of up to 8 magnitudes, in a time-scale of weeks. The return to normal light is much slower and can last up to several months (see Clayton [@clayton96]). There are several observational evidences in favor of this scenario (see Clayton [@clayton96], for more details). Among them, the fact that (i) the dust consists primarily as amorphous carbon particles, (ii) mass-loss rates are as large as 10$^{-6}$M$_{\odot}$/year and episodic, with timescales of a few months, (iii) mass-loss is driven by fast winds (Clayton et al. [@clayton03]), (iv) the dust may form only over a small solid angle of the stellar surface (perhaps the cool regions above large convective cells, as first proposed by Wdowiak [@wdowiak]) or is ejected in some specific directions (Clayton et al. [@clayton97]), (v) polarimetric observations may reveal the presence of permanent clumpy non-spherical dust shells (Clayton et al. [@clayton97] and Yudin et al. [@yudin]), (vi) larger polarizations are seen during declines revealing more scattering particles between the star and the observer (see references in Clayton [@clayton96]), (vii) characteristic time-scales of the light variations are compatible with the formation of dust clouds close to the stellar photosphere and their dilution in the outer regions (Hartmann & Apruzese [@hartmann] and Zubko [@zubko]).
However, in spite of the above evidences, the dusty environment close to the star remains almost unknown: no direct detection of the suspected heterogeneities in the dust distribution in the circumstellar envelope of a R CrB variable have up to now been achieved. Observations of these inner layers in R CrB were reported by Ohnaka et al. ([@ohnaka03] and previous references) but no significant deviation from circular geometry was detected (perhaps because of the rather small dynamical range of their observations, inherent to the technique adopted). In this letter, we present the first direct detection of the presence of dusty clouds around RY Sgr, the brightest R CrB variable in the southern hemisphere.
Observations
============
Observations of RY Sgr were performed in service mode in May and September 2003 using the adaptive optics system NACO at the Nasmyth-B focus of the ESO/VLT fourth 8-m telescope unit Yepun located at Cerro Paranal, Chile. NACO, which is an association of the adaptive optics system NAOS (Rousset et al. [@rousset]) and the spectro-imager CONICA (Lenzen et al. [@lenzen]), provides diffraction-limited images in the near-infrared (1-5$\mu$m) spectral range (see http://www.eso.org/instruments/naco/).
RY Sgr was observed using three narrow-band filters: NB 1.04 (centered at $\lambda_c$=1.04 $\pm$ 0.015$\mu$m), NB 2.17 ($\lambda_c$=2.166 $\pm$ 0.023$\mu$m) and NB 4.05 ($\lambda_c$=4.051 $\pm$ 0.02$\mu$m). The pixel scale on CONICA was respectively 13.25 mas in the NB 1.04 and NB 2.17 filters and 27.03 mas in the NB 4.05 filter, adapted to the observing wavelengths. The seeing conditions were variable, ranging from $\sim$ 0.5 during observations at 2.17$\mu$m to $\sim$ 0.9 during observations at 4.05$\mu$m. The Auto-Jitter mode was used, i.e. that at each exposure, the telescope was moved according to a random pattern in a 6 $\times$ 6 box. The sky is then estimated from all the observations. To allow further resolution improvement through deconvolution, a PSF reference star (HD 178199) was observed immediately after each RY Sgr observation using the same configuration of the adaptive optics system. The air-masses of RY Sgr and of the PSF reference star were 1.0 and 1.1 respectively and NAOS was servoed on RY Sgr itself. The FWHM of the PSF, estimated from the reference star in the different narrow-band filters, was 0.058, 0.068 and 0.116 at 1.04$\mu$m, 2.17$\mu$m and 4.05$\mu$m respectively. Calibration files (flat fields and dark exposures) were acquired, following the ESO/VLT standard calibration plan. The total on-source integration time in the three filters were 80 mn, 33 mn and 3 mn respectively. The dynamic range of the final images goes from 3300 at 2.17$\mu$m, 12500 at 1.04$\mu$m to 54000 at 4.05$\mu$m. Table 1 summarizes the observational conditions.
Standard reduction procedures were applied using self-developed routines. The raw images were sky subtracted, then divided by the flat-field and corrected from hot pixels. In each filter, the images were cross-correlated and aligned by sub-pixel shifting, and then combined to produce the final images, eliminating discrepant points like cosmic rays. Finally, the images were deconvolved with the PSF reference star. We used the Richardson-Lucy algorithm (Richardson [@richardson] & Lucy [@lucy]), which appears to be more suitable for data with a high dynamic range, such as ours. Constancy of prominent features present in deconvolved images showed that the PSF selection and the number of iterations (10-20 iterations, typically) for the deconvolution was performed carefully and conservatively.
[l c c c ]{} Date & $\lambda_c$ & Exp. time & Seeing\
(UT) & ($\mu$m) & (on-source in min) & ()\
2003 May 17 & 1.04 & 80 & $\sim$ 0.8\
2003 May 24 & 2.17 & 33 & $\sim$ 0.5\
2003 September 17 & 4.05 & 3 & $\sim$ 0.9\
Images of the circumstellar envelope of RY Sgr
==============================================
{width="9.cm"}
{width="18.cm"}
The diffraction-limited images of RY Sgr at 1.04$\mu\mathrm{m}$, 2.17$\mu\mathrm{m}$ and 4.05$\mu\mathrm{m}$ are presented in Fig. 1 & 2. The images are displayed with a log-scale for the brightness, so that details of the morphology at all flux levels can be seen. In the 1.04$\mu\mathrm{m}$ image (Fig. 1) collected at the same epoch as the 2.17$\mu\mathrm{m}$ one (Fig. 2, left), RY Sgr is not resolved and no structures are seen in its envelope. On the other hand, RY Sgr reveals strong departures from a point-like source at longer wavelengths since several structures are seen in the K (2.17$\mu\mathrm{m}$) and L(4.05$\mu\mathrm{m}$) images (Fig. 2). In this section, we describe only the brightest detected heterogeneities, i.e. those appearing brighter than 0.2% of the peak level and thus well above the noise level defined by the dynamics of the images.
In the K-band, two elongations are clearly seen towards the NW and SW directions (at PA $\sim$330$^o$ and PA $\sim$195$^o$ respectively). These bright structures are located at about 0.1 from the central star and they appear to have a typical size similar to the one of the central star itself.
At longer wavelengths (4.05$\mu\mathrm{m}$), the departure from a point-like symmetry is even more evident. The most prominent structures are two clouds as bright as 2% of the stellar peak. They are found at about 0.2 from the center of the image (i.e. twice further than in the K-image, revealing that different clouds are seen in the L-band) at PA $\sim 20 ^o$ and PA $\sim 260^o$. Their size is of the order of 0.2.
Finally, fainter structures are seen in both images of Fig. 2. These structures could be artifacts in K-band but they are about 10 times brighter than the noise level in the L-band image. Therefore, RY Sgr might be surrounded at a given epoch by several clouds (about ten or so) located at distances up to 0.5 from its center. Nevertheless, this more complex configuration should be confirmed by new independent observations.
Discussion
==========
These images do reveal undoubtedly that large clouds are present in the vicinity of R CrB variables. They are seen in different directions and located at different distances from the central star. It has to be pointed out that the clouds seen in the K and L-images result from different ejection events. Indeed, an extremely large velocity (more than one order of magnitude above the typical escape velocity observed around R CrB variables) would be required to explain such a cloud motion (0.1in 4 months), at the estimated distance of RY Sgr ($\sim$ 2 kpc, see below). We also note that the clouds appearing at $\sim$0.2" from the center of the field in the L-band image are not seen in K-band because, despite the better spatial resolution, the dynamic range in K may not be high enough to detect these rather cold and therefore too faint clouds. On another hand, the structures seen in May 2003 in K-band are not detected in September in L-band. A possible explanation is that the angular resolution of the L image (0.116) is not high enough to resolve these features that are present at about 0.1from the center of the field. In any cases, it is very likely that these clouds are composed of dust particles since they are not detected at 1.04$\mu\mathrm{m}$ while they clearly appear at the same epoch in K-band where their emission is high enough to be detected.
Therefore, these observations do confirm, for the first time, the scenario proposed several decades ago by Loreta ([@loreta]) and O’Keefe ([@okeefe]) and now widely accepted. When a very optically thick dust cloud is ejected towards the observer, a huge brightness decline, characteristic of R CrB variable stars, is observed in visible light. On the contrary, almost no variations are seen at longer wavelengths where this cloud is optically thin.
The rather large number of clouds detected around RY Sgr also reveals a high activity for the R CrB variable stars regarding the ejection of stellar material. Such large departures from spherical symmetry around RY Sgr could explain why brightness declines of R CrB variables are not so rare (a few every 10 years, typically): a larger number of ejected clouds (if the ejection is isotropic) leads to a larger probability that one of them lies on the line of sight. R CrB stars with the most frequent brightness declines would therefore be the stars ejecting material at the larger rate. For instance, RY Sgr exhibited about 10 declines over the last 50 years (from AAVSO lightcurve). That is a frequency about half that of the star R CrB, which should be surrounded by a higher number of dust clouds and might therefore eject material at a higher rate.
In addition, not only the number of ejected clouds is large but these clouds might be very dense and optically thick close to the stellar surface. It has indeed to be noted that the mean density of the circumstellar layers, where they are detected in the NACO images, is a factor of the order of $10^{6}$ lower (assuming a density law varying as $r^{-2}$ in the envelope) with respect to the regions very close to the star where they are formed. If we assume that the density in the clouds has decreased by the same amount, it results that an impressive quantity of material is suddenly ejected from the stellar surface and eventually form the dust clouds surrounding these stars.
Furthermore, the dusty clouds are detected rather far from RY Sgr itself, shedding new light on their dilution into the interstellar medium. RY Sgr is a rather hot R CrB variables ($T_{\rm eff}$ = 7000-7500 K, following Asplund et al. [@asplund]). It could thus belong to the class of the brightest R CrB with $M_V$ = -5 as revealed by SMC & LMC R CrB variables (Alcock et al. [@alcock] and Tisserand et al. [@tisserand], assuming that galactic R CrB stars have similar properties). Adopting a bolometric correction BC = -0.15 for a G0 supergiant, the radius of RY Sgr is approximately 60 $R_\odot$. From the adopted absolute magnitude and the AAVSO photometry ($m_V \sim 6.4$ at maximum light), we estimate that RY Sgr lies at about 1.9 kpc, yielding an angular radius 0.15 mas for this star. The dust clouds shown in Fig. 2, placed at about 0.1-0.2, are therefore located between $\sim 700$-$1\,400~R_*$ and their typical radius is also close to a few hundred stellar radii. That offers new constraints on the dilution of these clouds in the outer circumstellar envelopes of R CrB variables, since they still have an important size far from the region where they were formed. The NACO images indeed reveal that the light recovery of R CrB stars in the optical would not be caused by the evaporation of the clouds close to the stellar surface since they are still seen far from the central object. These clouds are indeed rather steady because they have been ejected a few years ago (about 5-10 years, assuming a typical escape velocity of at least 200 km.s$^{-1}$, Clayton et al. [@clayton03]). We also note that, if the most distant structures seen at about 0.5 (or more than 3000 $R_*$) are confirmed, they might have been ejected a few decades ago. This reveals that dust clouds around R CrB variables move away from the star leading, if present on the line of sight, to less obscuration of the surface. Therefore, the return to maximum light might not be caused by the evaporation of the clouds very close to the star.
Finally, regarding the location of the formation of the dust clouds, we cannot discriminate from the collected data with insufficient spatial resolution between the two commonly adopted scenarios: either the dust is formed very close to the stellar surface ($\sim 2~$R$_{\star}$) or it is formed in regions more distant than $\sim 20~$R$_{\star}$ (see Clayton [@clayton96]).
Summary and conclusion
======================
We have presented the first detection of heterogeneities in the circumstellar envelope of a R Coronae Borealis variable star owing to high spatial resolution images collected with NACO.
Several large dust clouds are found in different directions at several hundred stellar radii from RY Sgr. These observations do support the current interpretation that the huge and sudden declines which characterize these objects in optical light are caused by the formation of dust clouds along the line of sight. However, the present data are not sufficiently spatially resolved to verify whether the dust is formed close to the stellar surface or in more distant regions. Future observations with higher spatial resolutions (ideally a factor 10 or more), also with very high dynamics, should help to better understand the circumstellar environment and the evolution of these variable stars.
We thank A.P.P. Recio-Blanco for her careful reading of the manuscript and her Anglo-Italian languages expertise. G. Niccolini is acknowledged for comments and discussions on this work. The referee (G.C. Clayton) and M. Asplund are thanked for their valuable suggestions. We are also greatful to the variable star observations from the AAVSO International Database, contributed by observers worldwide and used in this research.
Alcock, C., Allsman, R. A., Alves, D. R. et al. 2001, ApJ, 554, 298
Asplund, M., Gustafsson, B., Lambert, D. L., Rao, N. K. 2000, A&A, 353, 287
Clayton, G.C. 1996, PASP, 108, 225
Clayton, G.C., Bjorkman, K. S., Nordsieck, K. H., et al. 1997, ApJ, 476, 870
Clayton, G.C., Geballe, T.R., Luciana, L. 2003, ApJ, 595, 412
Hartmann, L., Apruzese, J.P. 1976, ApJ, 203, 610
Lenzen, R., Hartung, M., Brandner, et al. 2002, SPIE, 4841
Loreta, E. 1934, Astron. Nach., 254, 151
Lucy, L.B. 1974, AJ, 79, 745
Ohnaka, K., Beckmann, U., Berger, J.-P., et al. 2003, A&A, 408, 553
O’Keefe, J.A. 1939, ApJ, 90, 294
Richardson, W.H. 1972, JOSA, 62, 55
Rousset, G., Lacombe, F., Puget, P., et al. 2002, SPIE, 4007
Tisserand, P., Marquette, J.B., Beaulieu, J.P. et al. 2004, A&A, 424, 245
Wdowiak, T.J. 1975, ApJ, 198, L139
Yudin, R.V., Evans, A., Barrett, P., et al. 2003, A&A, 412, 405
Zubko, V.G. 1997, MNRAS, 289, 305
[^1]: Based on observations collected with the VLT/UT4 Yepun telescope (Paranal Observatory, ESO, Chile) using the NACO instrument (program ID 71.D-0543A)
| {
"pile_set_name": "ArXiv"
} |
---
author:
- |
Pallabi Ghosh\
[[email protected]]{}
- |
Vibhav Vineet\
[[email protected]]{}
- |
Larry S. Davis\
[[email protected]]{}
- |
Abhinav Shrivastava\
[[email protected]]{}
- |
Sudipta Sinha\
[[email protected]]{}
- |
Neel Joshi\
[[email protected]]{}
bibliography:
- 'egbib.bib'
title: 'Deep Depth Prior for Multi-View Stereo'
---
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
The value 1 problem is a decision problem for probabilistic automata over finite words: given a probabilistic automaton, are there words accepted with probability arbitrarily close to 1? This problem was proved undecidable recently; to overcome this, several classes of probabilistic automata of different nature were proposed, for which the value 1 problem has been shown decidable. In this paper, we introduce yet another class of probabilistic automata, called *leaktight automata*, which strictly subsumes all classes of probabilistic automata whose value 1 problem is known to be decidable.
We prove that for leaktight automata, the value 1 problem is decidable (in fact, PSPACE-complete) by constructing a saturation algorithm based on the computation of a monoid abstracting the behaviours of the automaton. We rely on algebraic techniques developed by Simon to prove that this abstraction is complete. Furthermore, we adapt this saturation algorithm to decide whether an automaton is leaktight.
Finally, we show a reduction allowing to extend our decidability results from finite words to infinite ones, implying that the value $1$ problem for probabilistic leaktight parity automata is decidable.
address:
- '[a]{}LIAFA, Universit[é]{} Denis Diderot - Paris 7, France, and University of Warsaw, Poland.'
- '[b]{}LaBRI, CNRS, Bordeaux, France.'
- '[c]{}LaBRI and Université de Bordeaux, France.'
- '[e]{}Université Paris-Est, LACL, France.'
author:
- Nathanaël Fijalkowa
- Hugo Gimbertb
- Edon Kelmendic
- Youssouf Oualhadje
bibliography:
- 'bib.bib'
title: |
Deciding the Value 1 Problem\
for Probabilistic Leaktight Automata
---
[^1]
Introduction {#introduction .unnumbered}
============
**Probabilistic automata.** Rabin invented a very simple yet powerful model of probabilistic machine called probabilistic automata, which, quoting Rabin, “are a generalization of finite deterministic automata” [@R63]. A probabilistic automaton has a finite set of states and reads input words from a finite alphabet. The computation starts from the initial state and consists in reading the input word sequentially; the state is updated according to transition probabilities determined by the current state and the input letter. The probability to accept a finite input word is the probability that the computation ends in one of the final states.
Probabilistic automata, and more generally partially observable Markov decision processes and stochastic games, are a widely studied model of probabilistic machines used in many fields like software verification [@BBG12; @CDHR07], image processing [@CK97], computational biology [@DEKM99] and speech processing [@M97]. As a consequence, it is crucial to understand which decision problems are algorithmically tractable for probabilistic automata. From a language-theoretic perspective, several algorithmic properties of probabilistic automata are known: while language emptiness is undecidable [@P71; @B74; @GO10], functional equivalence is decidable [@S61; @T92] as well as other properties [@CMRR08].
Our initial motivation for this work comes from control and game theory: we aim at solving algorithmic questions about partially observable Markov decision processes and stochastic games. For this reason, we consider probabilistic automata as machines controlled by a blind controller, who is in charge of choosing the sequence of input letters in order to maximize the acceptance probability. While in a fully observable Markov decision process the controller can observe the current state of the process to choose adequately the next input letter, a blind controller does not observe anything and its choice depends only on the number of letters already chosen. In other words, the strategy of a blind controller is an input word of the automaton.
**The value of a probabilistic automaton.** With this game-theoretic interpretation in mind, we define the *value* of a probabilistic automaton as the supremum acceptance probability over all input words, and we would like to compute this value. Unfortunately, as a consequence of Paz undecidability result, the value of an automaton is not computable in general. However, the following decision problem was conjectured by Bertoni [@B74] to be decidable: *Given a probabilistic automaton, does it have value $1$? In other words are there input words whose acceptance probability is arbitrarily close to $1$?*
Recently, the second and fourth authors of the present paper proved that the value $1$ problem is undecidable [@GO10].
**Our result.** We introduce a new class of probabilistic automata, called *leaktight automata*, for which the value $1$ problem is decidable. This subclass strictly subsumes all known subclasses of probabilistic automata sharing this decidability property and has good closure properties. Our algorithm to decide the value $1$ problem computes in polynomial space a finite monoid whose elements are directed graphs and checks whether it contains a certain type of elements that are value $1$ witnesses.
**Related works.** Introducing subclasses of probabilistic automata to cope with undecidability results has been a fruitful and lively topic recently. We discuss some of them here.
The first subclass which was introduced specifically to decide the value $1$ problem are the $\sharp$-acyclic automata [@GO10]. Later on, Chatterjee and Tracol [@CT12] introduced structurally simple automata, which are probabilistic automata satisfying a structural property (related to the decomposition-separation theorem from probability theory), and proved that the value $1$ problem is decidable for structurally simple automata. At the same time, a subset of the authors introduced leaktight automata, and proved a similar result. As we shall see, both $\sharp$-acyclic and structurally simple automata are leaktight, hence our results extend both [@GO10] and [@CT12].
Quite recently, Chadha, Sistla and Viswanathan introduced the subclass of hierarchical automata [@CSV11], and showed that over infinite words, they recognize exactly the class of $\omega$-regular languages. As we shall see, hierarchical automata are leaktight, hence as a consequence of our result, the value $1$ problem is decidable for hierarchical automata.
**Proof techniques.** Our proof techniques totally depart from the ones used in [@CSV11; @CT12; @GO10]. We make use of algebraic techniques and in particular Simon’s factorization forest theorem, which was used successfully to prove the decidability of the boundedness problem for distance automata [@S94], and extended models as desert automata and B-automata [@K05; @C09]
**Outline.** Basic definitions are given in Section \[sec:def\].
In Section \[sec:algo\], we introduce the Markov monoid and the Markov monoid algorithm for the value $1$ problem; since the problem is in general undecidable, the algorithm is incomplete: a positive answer implies that the automaton has value $1$, but a negative answer gives no guarantee.
In Section \[sec:leaktight\], we define the class of leaktight automata and show that the leaktight property is a sufficient condition for this algorithm to be complete; in particular, this implies that the value $1$ problem is decidable for leaktight automata.
In Section \[sec:leaktight\_properties\], we show that the Markov monoid algorithm runs in polynomial space, and obtain as a corollary that the value $1$ problem for leaktight automata is $\PSPACE$-complete. Furthermore, we extend the Markov monoid algorithm to check at the same time whether an automaton is leaktight and whether in such case it has value $1$.
In Section \[sec:leaktight\_comparisons\], we further investigate the class of leaktight automata: we provide examples of leaktight automata and show that all subclasses of probabilistic automata whose value $1$ problem is known to be decidable are leaktight.
In Section \[sec:infinite\], we give a general theorem allowing to extend the decidability results from finite words to infinite words.
Definitions {#sec:def}
===========
Probabilistic automata
----------------------
We fix $A$ a finite alphabet. A (finite) word $u$ is a (possibly empty) sequence of letters $u = a_0 a_1 \cdots a_{n-1}$, the set of finite words is denoted by $A^*$. For $i \le j$ we denote by $u[i,j]$ the subword $a_i \cdots a_{j-1}$, and $u_{< p} = u[0,p] = a_0 a_1 \cdots a_{p-1}$.
Let $Q$ be a finite set of states. A probability distribution over $Q$ is a function $\delta : Q \rightarrow [0,1]$ such that $\sum_{q \in Q} \delta(q) = 1$; we often see $\delta$ as a row vector of size $|Q|$. We denote by $\frac{1}{3} \cdot q + \frac{2}{3} \cdot q'$ the distribution that picks $q$ with probability $\frac{1}{3}$ and $q'$ with probability $\frac{2}{3}$, and by $q$ the trivial distribution picking $q$ with probability $1$. For a subset $R$ of states, the uniform distribution over $R$ picks each state in $R$ with probability $\frac{1}{|R|}$. The support of a distribution $\delta$ is the set of states picked with positive probability, *i.e.* $\supp(\delta) = \set{q \in Q \mid \delta(q) > 0}$. Finally, the set of probability distributions over $Q$ is $\DD(Q)$.
A tuple $\AA = (Q, q_0, \Delta, F)$ represents a probabilistic automaton, where $Q$ is a finite set of states, $q_0 \in Q$ is the initial state, $\Delta$ defines the transitions and $F \subseteq Q$ is the set of accepting states.
The transitions of a probabilistic automaton are given by a function $\Delta : Q \times A \rightarrow \DD(Q)$, where $\Delta(q,a)$ is the probability distribution obtained by reading the letter $a$ from the state $q$. The function $\Delta$ induces the function $\Delta' : \DD(Q) \times A \rightarrow \DD(Q)$, where $\Delta'(\delta,a) = \sum_{q \in Q} \delta(q) \cdot \Delta(q,a)$. Going further, $\Delta$ naturally extends to $\Delta^* : \DD(Q) \times A^* \rightarrow \DD(Q)$ by induction: for a letter $a \in A$, we set $\Delta^*(\delta,a) = \Delta'(\delta,a)$, and for an input word $u = a v$, we set $\Delta^*(\delta,u) = \Delta^*(\Delta'(\delta,a),v)$. Intuitively, $\Delta^*(\delta,u)$ is the probability distribution obtained by reading the word $u$ starting at the initial probability distribution $\delta$. From now on, we will make no difference between $\Delta$, $\Delta'$ and $\Delta^*$, and denote the three of them by $\Delta$.
We denote by $\pRob{\AA}(s \xrightarrow{u} t)$ the probability to go from state $s$ to state $t$ reading $u$ on the automaton $\AA$, *i.e.* $\Delta(s,u)(t)$. Then $\pRob{\AA}(s \xrightarrow{u} T)$ is defined as $\sum_{t \in T} \pRob{\AA}(s \xrightarrow{u} t)$. Finally, the *acceptance probability* of a word $u \in A^*$ by $\AA$ is $\pRob{\AA}(q_0 \xrightarrow{u} F)$, which we denote by $\pRob{\AA}(u)$.
For computational purposes, we assume that each value is a rational number given by two integers in binary decomposition.
The *value* of a probabilistic automaton $\AA$, denoted by $\val{\AA}$, is the supremum acceptance probability over all input words: $$\label{eq:value}
\val{\AA} = \sup_{u \in A^*} \pRob{\AA}(u).$$
The value 1 problem
-------------------
We are interested in the following decision problem:
\[prob:value1\] Given a probabilistic automaton $\AA$, decide whether $\val{\AA} = 1$.
The value 1 problem can be reformulated using the notion of *isolated cut-point* introduced by Rabin in his seminal paper [@R63]: an automaton has value 1 if and only if the cut-point 1 is *not* isolated.
![\[fig:x\] This automaton has value 1 if and only if $x > \frac{1}{2}$.](EPS/fig1)
The automaton depicted on figure \[fig:x\] has value 1 if and only if $x > \frac{1}{2}$ (a similar example appears in [@BBG12]). The input alphabet is $A = \set{a,b}$, the initial state is the central state $0$ and the unique final state is $L_2$.
We describe the behaviour of this automaton. After reading one $b$, the distribution is uniform over $L_1,R_1$. To reach $L_2$, one needs to read a $b$ from the state $L_1$, but on the right-hand side this leads to the non-accepting absorbing state $R_2$. In order to maximize the probability to reach $L_2$, one tries to “tip the scales” to the left.
If $x \leq \frac{1}{2}$, there is no hope to achieve this: reading a letter $a$ gives more chance to stay in $R_1$ than in $L_1$ thus all words are accepted with probability at most $\frac{1}{2}$, and $\val{\AA} = \frac{1}{2}$.
However, if $x > \frac{1}{2}$ then we show that $\AA$ has value $1$.
We have: $$\pRob{\AA}(0 \xrightarrow{b a^n} L_1) = \frac{1}{2} \cdot x^n \qquad
\textrm{ and } \qquad \pRob{\AA}(0 \xrightarrow{b a^n} R_1) = \frac{1}{2} \cdot (1-x)^n$$
We fix an integer $N$ and analyse the action of reading $(b a^n)^N \cdot b$: there are $N$ “rounds”, each of them corresponding to reading $b a^n$ from $0$. In a round, there are three outcomes: winning (that is, remaining in $L_1$) with probability $p_n = \frac{1}{2} \cdot x^n$, losing (that is, remaining in $R_2$) with probability $q_n = \frac{1}{2} \cdot (1-x)^n$, or going to the next round (that is, reaching $0$) with probability $1 - (p_n + q_n)$. If a round is won or lost, then the next $b$ leads to an accepting or rejecting sink; otherwise it goes on to the next round, for $N$ rounds. Hence: $$\begin{array}{lll}
\pRob{\AA}((b a^n)^N \cdot b) & = & \sum_{k = 1}^N (1 - (p_n + q_n))^{k-1} \cdot p_n \\[1.5ex]
& = & p_n \cdot \frac{1 - (1 - (p_n + q_n))^N}{1 - (1 - (p_n + q_n))} \\[1.5ex]
& = & \frac{1}{1 + \frac{q_n}{p_n}} \cdot \left(1 - (1 - (p_n + q_n))^N\right) \\[1.5ex]
\end{array}$$
We now set $N = 2^n$. A simple calculation shows that the sequence $((1 - (p_n + q_n))^{2^n})_{n \in \NN}$ converges to $0$ as $n$ goes to infinity. Furthermore, if $x > \frac{1}{2}$ then $\frac{1-x}{x} < 1$, so $\frac{q_n}{p_n} = \left(\frac{1-x}{x}\right)^n$ converges to $0$ as $n$ goes to infinity. It follows that the acceptance probability converges to $1$ as $n$ goes to infinity. Consequently: $$\lim_n \pRob{\AA}((b a^n)^{2^n} \cdot b) = 1.$$ This example witnesses two surprising phenomena:
- the value is discontinuous with respect to the transition probabilities, as for $x = \frac{1}{2}$ the value is $\frac{1}{2}$, and for $x > \frac{1}{2}$ the value is $1$;
- the sequence of words $((b a^n)^{2^n} \cdot b)_{n \in \NN}$ witnessing the value $1$ involves two convergence speeds: indeed, the words $a^n b$ are repeated an exponential number of times, namely $2^n$. One can show that repeating only $n$ times does not lead to words accepted with arbitrarily high probability.
Recurrent states and idempotent words
-------------------------------------
We fix $\AA$ a probabilistic automaton, and define two main notions: recurrent states and idempotent words.
Let $u$ be a finite word, it induces a Markov chain $\MM_{\AA,u}$ whose state space is $Q$ and transition matrix $M_{\AA,u}$ is defined by: $$M_{\AA,u}(s,t) = \pRob{\AA}(s \xrightarrow{u} t).$$
We rely on the classical notion of recurrent states in Markov chains.
A state $s$ is *$u$-recurrent* if it is recurrent in $\MM_{\AA,u}$.
A finite word $u$ is *idempotent* if reading once or twice the word $u$ does not change qualitatively the transition probabilities.
A Markov chain is idempotent if its transition matrix $M$ satisfies that for all states $s,t$: $$M(s,t) > 0 \iff M^2(s,t) > 0.$$
A finite word $u$ is idempotent if $\MM_{\AA,u}$ is idempotent.
In the case of idempotent words, recurrence of a state is easily characterized, relying on simple graph-theoretical arguments:
\[lem:idempotent\] Let $u$ be an idempotent word. A state $s$ is $u$-recurrent if and only if for all states $t$ we have: $$\MM_{\AA,u}(s,t) > 0 \implies \MM_{\AA,u}(t,s) > 0.
$$
An (incomplete) algorithm for the value 1 problem {#sec:algo}
=================================================
In this section, we present an algebraic algorithm for the value $1$ problem, called the Markov monoid algorithm. Since the problem is undecidable, this algorithm does not solve the problem on all instances; we will show that it is *correct*, *i.e.* if it answers that an automaton has value $1$, then the automaton does have value $1$, but not *complete*, *i.e.* the converse does not hold. In the next section, we shall show that this algorithm is *complete* for the class of leaktight automata.
The Markov monoid algorithm
---------------------------
Our algorithm for the value $1$ problem computes iteratively a set $\monoid$ of directed graphs called limit-words. Each limit-word is meant to represent the asymptotic effect of a sequence of input words, and some particular limit-words can witness that the automaton has value $1$.
A *limit-word* is a function $\limu : Q^2 \to \set{0,1}$, such that for all states $s$, there exists a state $t$ such that $\limu(s,t) = 1$.
In proofs and examples, we will adopt either of the two equivalent views for limit-words: graphs over the set $Q$ or square matrices over $Q \times Q$.
$\monoid \gets \set{\bolda \mid a\in A} \cup \set{\boldeps}$.
We now explain the algorithm in detail. For the remainder of this section, we fix $\AA$ a probabilistic automaton. Initially, $\monoid$ only contains those limit-words $\bolda$ that are induced by input letters $a \in A$ : $$\forall s,t \in Q,\ (\bolda(s,t) = 1 \iff \pRob{\AA}(s \xrightarrow{a} t) > 0),$$ plus the limit-word $\boldeps$ which is induced by the empty word: $$\forall s,t \in Q,\ (\boldeps(s,t) = 1 \iff s = t).$$
The algorithm repeatedly adds new limit-words to $\monoid$. There are two ways for that: concatenating two limit-words or iterating an idempotent limit-word.
1em **Concatenation of two limit-words** The *concatenation* of two limit-words $\limu$ and $\limv$ is the limit-word $\limu \cdot \limv$ such that: $$(\limu \cdot \limv)(s,t) = 1 \iff \exists q \in Q,\
\limu(s,q) = 1 \text{ and } \limv(q,t) = 1.$$ In other words, concatenation corresponds to the multiplication of matrices with coefficients in the boolean semiring $(\{0,1\},\vee,\wedge)$. Intuitively, the concatenation of two limit-words corresponds to the concatenation of two sequences $(u_n)_{n\in\NN}$ and $(v_n)_{n\in\NN}$ of input words into the sequence $(u_n\cdot v_n)_{n\in\NN}$.
We say that a limit-word $\limu$ is idempotent if $\limu \cdot \limu = \limu$. The following lemma gives simple properties of idempotent limit-words.
\[lem:basic\_limit\_words\] For all limit-words $\limu$:
- the limit-word $\limu^{|Q|!}$ is idempotent,
- if $\limu$ is idempotent, then for all states $r \in Q$, there exists a state $r' \in Q$ such that $\limu(r,r') = 1$ and $r'$ is $\limu$-recurrent.
The proof is omitted and relies on simple graph-theoretical arguments.
1em **Iteration of an idempotent limit-word** The *iteration* $\limu^\sharp$ of a limit-word $\limu$ is only defined when $\limu$ is idempotent. It relies on the notion of $\limu$-recurrent state.
Let $\limu$ be an idempotent limit-word. A state $s$ is $\limu$-recurrent if for all states $t$, we have: $$\limu(s,t)=1 \implies \limu(t,s) = 1.$$
Note that this echoes Lemma \[lem:idempotent\]. The *iterated limit-word* $\limu^\sharp$ removes from $\limu$ any edge that does not lead to a recurrent state: $$\limu^\sharp(s,t) = 1 \iff \limu(s,t) = 1 \text{ and } t \text{ is } \limu\text{-recurrent}.$$ Intuitively, if a limit-word $\limu$ represents a sequence $(u_n)_{n\in\NN}$ then its iteration $\limu^\sharp$ represents the sequence $\left(u_n^{f(n)}\right)_{n\in\NN}$ for some increasing function $f:\NN\to\NN$.
The Markov monoid and value 1 witnesses
---------------------------------------
The set of limit-words $\monoid$ computed by Algorithm \[algo:markov\_monoid\] is called the Markov monoid.
The Markov monoid associated with $\AA$ is the smallest set of limit-words containing $\set{\bolda \mid a \in A} \cup \set{\boldeps}$ and closed under concatenation and iteration.
Two key properties, *consistency* and *completeness*, ensure that the limit-words of the Markov monoid reflect exactly every possible asymptotic effect of a sequence of input words.
A sequence $(u_n)\nNN$ of words reifies a limit-word $\limu$ if for all states $s,t$, $(\pRob{\AA}(s \xrightarrow{u_n} t))\nNN$ converges and: $$\label{eq:consistency}
\limu(s,t) = 1 \iff \lim_n \pRob{\AA}(s \xrightarrow{u_n} t) > 0.$$
Note that if $(u_n)\nNN$ reifies $\limu$, then any subsequence of $(u_n)\nNN$ also does. We will use this simple observation several times.
\[def:consistency\] A set of limit-words $\monoid$ is *consistent* with $\AA$ if for every limit-word $\limu \in \monoid$, there exists a sequence of input words $(u_n)\nNN$ which reifies $\limu$.
\[def:complete\] A set of limit-words $\monoid$ is *complete* for $\AA$ if for each sequence of input words $(u_n)_{\nNN}$, there exists $\limu \in \monoid$ such that for all states $s,t \in Q$: $$\label{eq:completeness}
\limsup_n \pRob{\AA}(s \xrightarrow{u_n} t) = 0 \implies \limu(s,t) = 0.$$
Limit-words are useful to decide the value $1$ problem because some of these are witnesses that the automaton has value $1$.
\[def:value1witnesses\] A *value $1$ witness* is a limit-word $\limu$ such that for all states $t$: $$\label{eq:witness}
\limu(q_0,t) = 1 \implies t \in F,$$ where $q_0$ is the initial state of the automaton.
Thanks to value $1$ witnesses, the answer to the value $1$ problem can be read in a consistent and complete set of limit-words:
\[lem:criterion\] If $\monoid$ is consistent with $\AA$ and complete for $\AA$, then $\AA$ has value $1$ if and only if $\monoid$ contains a value $1$ witness.
Specifically:
- If $\monoid$ is consistent with $\AA$ and contains a value $1$ witness, then $\AA$ has value $1$,
- If $\monoid$ is complete for $\AA$ and $\AA$ has value $1$, then $\AA$ contains a value $1$ witness.
We prove the first item. Assume that $\monoid$ is consistent with $\AA$ and contains a value $1$ witness $\limu$. Since $\monoid$ is consistent, there exists a sequence $(u_n)\nNN$ reifying $\limu$. It follows from and that for $t \notin F$, we have $\lim_n \pRob{\AA}(q_0 \xrightarrow{u_n} t) = 0$. Thus $\lim_n \pRob{\AA}(u_n) = \sum_{t \in F} \lim_n \pRob{\AA}(q_0 \xrightarrow{u_n} t) = 1$, so $\AA$ has value $1$.
We now prove the second item. Assume that $\monoid$ is complete for $\AA$ and that $\AA$ has value $1$. Then there exists a sequence of words $(u_n)\nNN$ such that $\lim_n \pRob{\AA}(u_n) = 1$, *i.e.* $\lim_n \sum_{t \in F} \pRob{\AA}(q_0 \xrightarrow{u_n} t) = 1$. Since for all $n \in \NN$, we have $\sum_{q \in Q} \pRob{\AA}(q_0 \xrightarrow{u_n} q) = 1$, then for all $t \notin F$, $\limsup_n \pRob{\AA}(q_0 \xrightarrow{u_n} t) = 0$. Since $\monoid$ is complete, there exists a limit-word $\limu$ such that holds. Then $\limu$ is a value $1$ witness: let $t \in Q$ such that $\limu(q_0,t) = 1$, then according to , $\limsup_n \pRob{\AA}(q_0 \xrightarrow{u_n} t) > 0$, hence $t \in F$.
Correctness of the Markov monoid algorithm {#subsec:correctness}
------------------------------------------
\[theo:consistency\] The Markov monoid associated with $\AA$ is consistent.
This implies that if the Markov monoid algorithm outputs “true”, then for sure the input automaton has value $1$. This positive result holds for every automaton (leaktight or not).
To prove Theorem \[theo:consistency\], recall that the Markov monoid is the smallest set of limit-words containing $\set{\bolda \mid a \in A} \cup \set{\boldeps}$ and closed under concatenation and iteration, hence it suffices to prove that the initial elements form a consistent set, and the closure under the two operations.
First, $\bolda$ is reified by the constant sequence $(a)\nNN$, and $\boldeps$ by the constant sequence $(\varepsilon)\nNN$. We state the closure under the two operations in the following proposition:
\[prop:consistency\] Let $(u_n)\nNN$ and $(v_n)\nNN$ be two sequences that reify the limit-words $\limu$ and $\limv$ respectively. Then:
1. the sequence of words $(u_n \cdot v_n)\nNN$ reifies $\limu \cdot \limv$,
2. if $\limu$ is idempotent, then there exists an increasing function $f : \NN \to \NN$ such that for all increasing functions $g : \NN \to \NN$ satisfying $g \ge f$, the sequence $\left(u_{g(n)}^n\right)\nNN$ reifies the limit-word $\limu^\sharp$.
The statement about iteration is stronger than required: the existence of $f$ such that $(u_{f(n)}^n)\nNN$ reifying the limit-word $\limu^\sharp$ is enough to prove Theorem \[theo:consistency\]. However, we will use this stronger result later on (in Section \[subsec:simple\_automata\]).
1. Let $w_n = u_{n} \cdot v_{n}$. Then $(w_n)\nNN$ reifies $\limu\cdot \limv$, since: $$\pRob{\AA}(s \xrightarrow{w_n} t) =
\sum_{r \in Q} \pRob{\AA}(s \xrightarrow{u_n} r) \cdot \pRob{\AA}(r \xrightarrow{v_n} t).$$
2. Consider the Markov chain $\MM$ with state space $Q$ and transition matrix $M$ defined by $M(s,t) = \lim_n \pRob{\AA}(s \xrightarrow{u_n} t)$. Since $(u_n)\nNN$ reifies $\limu$, we have $\limu(s,t) = 1$ if and only if $M(s,t) > 0$. First observe that since $\limu$ is idempotent, the Markov chain $\MM$ is aperiodic. According to standard results about finite Markov chains, this implies that the sequence of matrices $(M^k)_{k \in \NN}$ has a limit which we denote by $M^\infty$, satisfying the following: $$\label{eq:trans}
\forall s,t \in Q,\ M^\infty(s,t) > 0 \implies t \text{ is recurrent in } \MM.$$ By definition the sequence of matrices $\left(M_{\AA,u_n}\right)\nNN$ converges to $M$. Since the matrix product operation is continuous, for every $k \in \NN$, $\left(M_{\AA,u_n}^k\right)\nNN$ converges to $M^k$. So for every $k \ge 1$, there exists $N_k \in \NN$ such that for all $p \ge N_k$, $||M^k - M_{\AA,u_p}^k ||_\infty \leq \frac{1}{k}$. We define $f : \NN \to \NN$ by induction, so that $f(k)$ is the maximum of $f(k-1) + 1$ and of $N_k$, ensuring that $f$ is increasing. Then for any increasing function $g : \NN \to \NN$ satisfying $g \ge f$, the sequence of matrices $\left(M_{\AA,u_{g(n)}}^n\right)\nNN$ converges to $M^\infty$. We prove that $\left(u_{g(n)}^n\right)\nNN$ reifies $\limu^\sharp$: $$\begin{aligned}
\limu^\sharp(s,t) = 1
& \iff \limu(s,t) = 1 \text{ and } t \text{ is } \limu\text{-recurrent}\\
& \iff M(s,t) > 0 \text{ and } t \text{ is } \text{recurrent in } \MM\\
& \iff M^\infty(s,t) > 0 \\
& \iff \lim_n \pRob{\AA}(s \xrightarrow{u_{g(n)}^n} t) > 0,\end{aligned}$$ where the first equivalence is by definition of the iteration, the second holds because $(u_n)\nNN$ reifies $\limu$, the third by definition of $M^\infty$, and the fourth because $\left(M_{\AA,u_{g(n)}^n}\right)\nNN$ converges to $M^\infty$.
This concludes the proof.
Note that completeness is not true in general; for instance, one can show that the Markov monoid of the automaton represented in figure \[fig:x\] is not complete. The next section gives a sufficient condition for completeness: the leaktight property.
Decidability of the value 1 problem for leaktight automata {#sec:leaktight}
==========================================================
In this section we establish our main result:
The value $1$ problem is decidable for leaktight automata.
The definition of leaktight automata is given in the next subsection. For now (in this section), we are only interested in decidability issues; we will actually prove in Section \[sec:leaktight\_properties\] that the value $1$ problem is $\PSPACE$-complete for leaktight automata.
Note that as observed in the literature [@BBG12; @CSV13; @Fijalkow14], the value $1$ problem for probabilistic automata over finite words is equivalent to the emptiness problem for probabilistic Büchi automata with positive semantics, hence we obtain the following corollary:
The emptiness problem is decidable for probabilistic Büchi leaktight automata with positive semantics.
The following theorem proves that the Markov monoid of a leaktight automaton is complete; since it is always consistent, by Lemma \[lem:criterion\], the Markov monoid algorithm solves the value $1$ problem for leaktight automata.
\[theo:completeness\] If a probabilistic automaton is leaktight then its Markov monoid is complete.
The remainder of this section is devoted to the proof of Theorem \[theo:completeness\]. We first define the leaktight property, and extend the Markov monoid. This extended version allows to state an algebraic characterization of the leaktight property. Then, the technical core of the proof relies on a subtle algebraic argument based on the existence of $\sharp$-factorization trees of bounded height [@S90; @S94; @C09; @T11].
Leaks {#subsec:leaks}
-----
The undecidability of the value 1 problem comes from the necessity to compare parallel convergence rates in order to track down vanishing probabilities. Comparing two convergence rates may require to compare the decimals of the rates up to an arbitrary precision, which in turn can encode a Post correspondence problem, hence the undecidability.
One of the phenomena that makes tracking vanishing probabilities difficult are *leaks*. A leak occurs in an automaton when a sequence of words turns a set of states $C\subseteq Q$ into a recurrence class $C$ *on the long run*, but *on the short run*, some of the probability of the recurrence class is “leaking” to a *different* recurrence class.
![\[fig:ex\_leak\] $(a^n \cdot b)\nNN$ is a leak from $L_1$ to $L_2$.](EPS/fig12)
Such leaks occur in the automaton depicted in the left hand side of figure \[fig:ex\_leak\] with the input sequence $(a^nb)\nNN$. As $n$ grows large, the probability to reach $L_2$ from $L_1$ while reading the input word $a^nb$ vanishes, thus the sets $\set{L_1}$ and $\set{L_2}$ are two different recurrence classes on the long run (*i.e.* asymptotically), however on the short run remains a small yet positive probability to reach $L_2$ from $L_1$.
The right hand side of figure \[fig:ex\_leak\] shows the asymptotic behaviour of reading $(a^nb)\nNN$. Since the automaton in figure \[fig:x\] contains two symmetric parts identical to figure \[fig:ex\_leak\], it features one leak on the left hand side and another in the right hand side. As a consequence, the real asymptotic behaviour is complex and depends on the compared speeds of these leaks.
An automaton without leak is called a leaktight automaton. In this section we prove that the value 1 problem is decidable when restricted to the subclass of leaktight automata.
The formal definition of a leak is as follows:
\[def:leak\] Let $(u_n)\nNN$ be a sequence of idempotent words. Assume that the sequence of matrices $\pRob{\AA}(u_n)$ converges to a limit $M$, that this limit is idempotent and denote $\MM$ the assocaited Markov chain. The sequence $(u_n)\nNN$ is a leak if there exist $r,q \in Q$ such that the following three conditions hold:
1. $r$ and $q$ are recurrent in $\MM$,
2. $\lim_n \pRob{\AA}(r \xrightarrow{u_n} q) = 0$,
3. for all $n \in \NN$, $\pRob{\AA}(r \xrightarrow{u_n} q) > 0$.
A probabilistic automaton is leaktight if it has no leak.
Several examples of leaktight automata are given in Section \[sec:leaktight\_comparisons\].
The extended Markov monoid
--------------------------
The existence of leaks can be decided by a slight extension of the Markov monoid algorithm which keeps track of strictly positive transition probabilities.
An *extended limit-word* is a couple $(\limu,\limu_+)$ of two limit-words, such that for all $s,t \in Q$, we have $\limu(s,t) = 1 \Longrightarrow \limu_+(s,t) = 1$.
As for limit-words, extended limit-words can be seen either as graphs over the set $Q$, or couples of square matrices over $Q \times Q$. Such a graph has two different kind of edges: an edge $(s,t)$ is “normal” if $\limu(s,t) = 1$, and is a $+$-edge if $\limu(s,t) = 0$ but $\limu_+(s,t) = 1$.
We define the concatenation and iteration operations for extended limit-words. The *concatenation* of two extended limit-words $(\limu,\limu_+)$ and $(\limv,\limv_+)$ is the component-wise concatenation, *i.e.* $(\limu \cdot \limv, \limu_+ \cdot \limv_+)$. The *iteration* of an extended limit-word $(\limu,\limu_+)$ is only defined when it is idempotent (*i.e.* component-wise idempotent), by $(\limu,\limu_+)^\sharp = (\limu^\sharp,\limu_+)$.
The extended Markov monoid is the smallest set of extended limit-words containing $\set{(\bolda,\bolda) \mid a \in A} \cup \set{(\boldeps,\boldeps)}$ and closed under concatenation and iteration.
Note that if $(\limu,\limu_+)$ is in the extended Markov monoid, then $\limu$ is in the Markov monoid.
The essential difference between the Markov monoid and its extended version is that the extension keeps track of those edges that are deleted by successive iteration operations. This serves two purposes: first, to characterize the leaktight property in algebraic terms, and second, to prove Theorem \[theo:completeness\].
We state a consistency result for the extended Markov monoid, extending Theorem \[theo:consistency\]. The proofs of both these results are similar and given only once.
\[lem:consistency\_extended\] For each $(\limu,\limu_+)$ in the extended Markov monoid, there exists a sequence $(u_n)_{n\in\NN}$ such that for all states $s,t \in Q$, $(\pRob{\AA}(s \xrightarrow{u_n} t))\nNN$ converges and: $$\begin{aligned}
\label{eq:appcond1}
& \limu(s,t) = 1 \iff \lim_n \pRob{\AA}(s \xrightarrow{u_n} t) > 0,\\
\label{eq:appcond2}
& \text{for all } n \in \NN,\ \left(\limu_+(s,t) = 1 \iff \pRob{\AA}(s \xrightarrow{u_n} t) > 0\right).\end{aligned}$$
Leak witnesses
--------------
\[eq:leakwitness\] An idempotent extended limit-word $(\limu,\limu_+)$ is a *leak witness* if there exist $r,q \in Q$ such that the following three conditions hold:
1. $r$ and $q$ are $\limu$-recurrent,
2. $\limu(r,q) = 0$,
3. $\limu_+(r,q) = 1$.
\[lem:leaktight\_characterization\_direct\] If a probabilistic automaton is leaktight, then its extended Markov monoid does not contain any leak witness.
Suppose that there is a leak witness $(\limu,\limu_+)$ in the extended Markov monoid: $\limu$ and $\limu_+$ are idempotent and there exists $r,q \in Q$ such that $r$ and $q$ are $\limu$-recurrent, $\limu(r,q) = 0$ and $\limu_+(r,q) = 1$. We prove that there exists a leak.
Thanks to Lemma \[lem:consistency\_extended\], there exists a sequence $(u_n)\nNN$ satisfying and . Note that since $\limu_+$ is idempotent, implies that for all $n \in \NN$, $u_n$ is idempotent.
Consider the Markov chain $\MM$ with state space $Q$ and transition matrix $M$ defined by $M(s,t) = \lim_n \pRob{\AA}(s \xrightarrow{u_n} t)$. $\MM$ is idempotent since $\limu$ is idempotent and thanks to . We show that $(u_n)\nNN$ is a leak. There are three conditions to be met.
First, $r$ and $q$ are recurrent in $\MM$: this follows from and the fact that $r$ and $q$ are $\limu$-recurrent. Second, $\lim_n \pRob{\AA}(r \xrightarrow{u_n} q) = 0$: this follows from and the fact that $\limu(r,q) = 0$. Third, for all $n \in \NN$, $\pRob{\AA}(r \xrightarrow{u_n} q) > 0$: this follows from and the fact that $\limu_+(r,q) = 1$.
As we will show in the next section, the converse of Lemma \[lem:leaktight\_characterization\_direct\] is also true, which gives an algebraic characterization of the leaktight property using the extended Markov monoid. However, the proof of the converse implication is more involved and requires the lower bound lemma (Lemma \[lem:lowerbound\]), which is the object of the next subsection.
Stabilization monoids and sharp-factorization trees
---------------------------------------------------
We now introduce the technical material required to state and prove the lower bound lemma. The key notions here are stabilization monoids and $\sharp$-factorization trees.
Factorization trees for monoids have been introduced by Simon [@S90]. Roughly speaking, Simon’s factorization theorem states that given a morphism $\phi : A^* \rightarrow M$ from the set of finite words over $A$ to a finite monoid $M$, the following holds: for all words $u$, the computation of $\phi(u)$ can be factorized in a tree whose depth is bounded independently of the length of the word.
Simon later developed the notion of decomposition trees to solve the limitedness problem for distance automata [@S94]. To this end, he defined an iteration operation $\sharp$ for monoids over the tropical semiring $(\NN \cup \set{\infty},\min,+)$. Then Kirsten extended this technique to desert automata and the nested distance desert automata [@K05]. After him, Colcombet generalized this approach by defining stabilization monoids [@C09], which are monoids equipped with an iteration operation, and proved the existence of $\sharp$-factorization trees of bounded depth. The formal definition is as follows:
A stabilization monoid $(M,\cdot,\sharp)$ is a finite monoid $(M,\cdot)$ equipped with an iteration operation $\sharp : E(M) \rightarrow E(M)$, where $E(M)$ is the set of idempotents of $M$, such that: $$\begin{aligned}
\label{eq:stab_monoid1}
& (a \cdot b)^\sharp \cdot a = a \cdot (b \cdot a)^\sharp \qquad
\text{ for } a \cdot b \in E(M) \text{ and } b \cdot a \in E(M),\\
\label{eq:stab_monoid2}
& (e^\sharp)^\sharp = e^\sharp \qquad \qquad \qquad \qquad \text{ for } e \in E(M),\\
\label{eq:stab_monoid3}
& e^\sharp \cdot e = e^\sharp \qquad \qquad \qquad \qquad \text{ for } e \in E(M).\end{aligned}$$
\[lem:stabilization\_monoid\] The extended Markov monoid is a stabilization monoid.
To start with, the extended Markov monoid is a monoid for the concatenation: $\boldeps$ is the neutral element, and the concatenation is associative.
Now, let us prove the three properties required for the iteration operation $\sharp$.
Proof of . Let $(\limu,\limu_+),(\limv,\limv_+)$ such that $(\limu \cdot \limv,\limu_+ \cdot \limv_+)$ and $(\limv \cdot \limu,\limv_+ \cdot \limu_+)$ are idempotent. By definition $\left((\limu,\limu_+) \cdot (\limv,\limv_+)\right)^\sharp \cdot (\limu,\limu_+)$ is equal to $\left((\limu \cdot \limv)^\sharp \cdot \limu, \limu_+ \cdot \limv_+ \cdot \limu_+\right)$, and $(\limu,\limu_+) \cdot \left((\limv,\limv_+) \cdot (\limu,\limu_+)\right)^\sharp$ to $\left(\limu \cdot (\limv \cdot \limu)^\sharp, \limu_+ \cdot \limv_+ \cdot \limu_+\right)$. Let $s,t \in Q$, we have the following equivalence: $\left((\limu \cdot \limv)^\sharp \cdot \limu\right)(s,t) = 1$ if and only if: $$\begin{aligned}
\label{eq:stabilization_monoid1}
\text{there exists } r,q \in Q,\ \limu(s,r) = 1 \wedge \limv(r,q) = 1 \wedge \limu(q,t) = 1 \wedge
q \text{ is } (\limu \cdot \limv) \text{-recurrent},\end{aligned}$$ and similarly, $\left(\limu \cdot (\limv \cdot \limu)^\sharp\right)(s,t) = 1$ if and only if: $$\begin{aligned}
\label{eq:stabilization_monoid2}
\text{there exists } r,q \in Q,\ \limu(s,r) = 1 \wedge \limv(r,q) = 1 \wedge \limu(q,t) = 1 \wedge
t \text{ is } (\limv \cdot \limu) \text{-recurrent}.\end{aligned}$$ We show that and are equivalent. Assume , and prove that $t$ is $(\limv \cdot \limu)$-recurrent. Let $p \in Q$ such that $(\limv \cdot \limu)(t,p) = 1$. Since $\limv$ is a limit-word, there exists $\ell \in Q$ such that $\limv(p,\ell) = 1$. Observe that $\limu(q,t) = 1$, $(\limv \cdot \limu)(t,p) = 1$ and $\limv(p,\ell) = 1$, so $(\limu \cdot \limv)^2(q,\ell) = 1$. As $\limu \cdot \limv$ is idempotent, this implies $(\limu \cdot \limv)(q,\ell) = 1$. Since $q$ is $(\limu \cdot \limv)$-recurrent, we have $(\limu \cdot \limv)(\ell,q) = 1$. Altogether, $\limv(p,\ell) = 1$, $(\limu \cdot \limv)(\ell,q) = 1$ and $\limu(q,t) = 1$ imply that $(\limv \cdot \limu)^2(p,t) = 1$. As $\limv \cdot \limu$ is idempotent, this implies $(\limv \cdot \limu)(p,t) = 1$, so $t$ is $(\limv \cdot \limu)$-recurrent, and is proved. Conversely, assume , and prove that $q$ is $(\limu \cdot \limv)$-recurrent. Note that from $\limv(r,q) = 1$, $\limu(q,t) = 1$ and the fact that $t$ is $(\limv \cdot \limu)$-recurrent, we obtain that $(\limv \cdot \limu)(t,r) = 1$. Let $p \in Q$ such that $(\limu \cdot \limv)(q,p) = 1$. Since $\limu$ is a limit-word, there exists $\ell \in Q$ such that $\limu(p,\ell) = 1$. Observe that $\limv(r,q) = 1$, $(\limu \cdot \limv)(q,p) = 1$ and $\limu(p,\ell) = 1$, so $(\limv \cdot \limu)^2(r,\ell) = 1$, and with $(\limv \cdot \limu)(t,r) = 1$ this implies $(\limv \cdot \limu)^3(t,\ell) = 1$. As $\limv \cdot \limu$ is idempotent, this implies $(\limv \cdot \limu)(t,\ell) = 1$. Since $t$ is $(\limv \cdot \limu)$-recurrent, we have $(\limv \cdot \limu)(\ell,t) = 1$. Altogether, $\limu(p,\ell) = 1$, $(\limv \cdot \limu)(\ell,t) = 1$, $(\limv \cdot \limu)(t,r) = 1$ and $\limv(r,q) = 1$ imply that $(\limu \cdot \limv)^3(p,q) = 1$. As $\limu \cdot \limv$ is idempotent, this implies $(\limu \cdot \limv)(p,q) = 1$, so $q$ is $(\limu \cdot \limv)$-recurrent, and is proved. The property follows.
Proof of . This boils down to proving $(\limu^\sharp)^\sharp = \limu^\sharp$. This is clear from the definition of $\limu^\sharp$, since the notions of $\limu$-recurrence and $\limu^\sharp$-recurrence coincide.
Proof of . This boils down to proving $\limu^\sharp \cdot \limu = \limu^\sharp$. It follows from the observation that if $r \in Q$ is $\limu$-recurrent and $\limu(r,t) = 1$, then $t$ is $\limu$-recurrent (under the assumption that $\limu$ is idempotent).
\[def:sharp\_factorization\_tree\] Let $A$ be a finite alphabet, $(M,\cdot,\sharp)$ a stabilization monoid and $\phi: A^* \to M$ a morphism into the submonoid $(M,\cdot)$. A $\sharp$-*factorization tree* of a word $u \in A^*$ is a finite unranked ordered tree, whose nodes have labels in $A^* \times M$ and such that:
1. the root is labelled by $(u,\limu)$, for some $\limu \in M$,
2. every internal node with two children (called *concatenation* nodes) labelled by $(u_1,\limu_1)$ and $(u_2,\limu_2)$ is labelled by $(u_1 \cdot u_2,\limu_1 \cdot \limu_2)$,
3. every internal node with three or more children (called *iteration* nodes) is labelled by $(u_1 \ldots u_n,\bolde^\sharp)$ for some $\bolde \in E(M)$, and its children are labelled by $(u_1,\bolde),\ldots,(u_n,\bolde)$.
4. every leaf is labelled by $(a,\bolda)$ where $a$ is a letter, or $(\varepsilon,\boldeps)$.
Note that in a factorization tree, the second label is not always the image of the first component under $\phi$; indeed, it is an element of the stabilization monoid $(M,\cdot,\sharp)$ whereas the image of a finite word under $\phi$ is an element of the submonoid $(M,\cdot)$. However, the projection of second label into this submonoid (which consists in ignoring the operation $\sharp$) is indeed the image of the first component under $\phi$.
The following theorem was stated for the tropical semiring in [@S94], and generalized in [@C09]. A simple proof can be found in [@T11].
\[theo:decomposition\] Let $A$ be a finite alphabet, $(M,\cdot,\sharp)$ a stabilization monoid and $\phi: A^* \to M$ a morphism into the submonoid $(M,\cdot)$. Every word $u \in A^*$ has a $\sharp$-factorization tree whose depth is less than $3 \cdot |M|$.
The lower bound lemma
---------------------
We are ready to state and prove the lower bound lemma, which is the central argument in the proof of completeness of leaktight Markov monoids.
\[lem:lowerbound\] Let $\AA$ be a probabilistic automaton whose extended Markov monoid contains no leak witness. Let $\pmin$ the smallest non-zero transition probability of $\AA$. Then for all words $u \in A^*$, there exists $(\limu,\limu_+)$ in the extended Markov monoid such that, for all states $s,t$: $$\begin{aligned}
\label{eq:lb1}
\limu_+(s,t) = 1 & \iff \pRob{\AA}(s \xrightarrow{u} t) > 0,\\
\label{eq:lb2}
\limu(s,t) = 1 & \implies \pRob{\AA}(s \xrightarrow{u} t) \geq \dabound.\end{aligned}$$
Consider a finite word $u \in A^*$; by Theorem \[theo:decomposition\] applied to the extended Markov monoid $\monoid_+$ associated with $\AA$ (which is a stabilization monoid thanks to Lemma \[lem:stabilization\_monoid\]) and the morphism $\phi: A \to M$ defined by $\phi(a) = (\bolda,\bolda)$, there exists a $\sharp$-factorization tree of depth at most $3 \cdot |\monoid_+|$, whose root is labelled by $(u,(\limu,\limu_+))$ for some extended limit-word $(\limu,\limu_+)$.
The depth of a node in this tree is defined in a bottom-up fashion: the leaves have depth zero, and a node has depth one plus the maximum of the depths of its children.
We prove by a bottom-up induction (on $h$) that for every node $(u,(\limu,\limu_+))$ of this tree at depth $h$, for all states $s,t$: $$\begin{aligned}
\label{eq:induc1}
\limu_+(s,t)=1 & \iff \pRob{\AA}(s \xrightarrow{u} t) > 0,\\
\label{eq:induc2}
\limu(s,t) = 1 & \implies \pRob{\AA}(s \xrightarrow{u} t) \geq \pmin^{2^h}.\end{aligned}$$
The case $h = 0$ is for leaves. Here, either $u$ is a letter $a$ and $\limu = \limu_+ = \bolda$, or $u$ is the empty word $\varepsilon$ and $\limu = \limu_+ = \boldeps$. Then both and hold.
Assume $h > 0$, there are two cases.
[**First case: a concatenation node**]{} labelled by $(u,(\limu,\limu_+))$ with two children labelled by $(u_1,(\limu_1,\limu_{+,1}))$ and $(u_2,(\limu_2,\limu_{+,2}))$. By definition $u = u_1 \cdot u_2$, $\limu = \limu_1 \cdot \limu_2$ and $\limu_+ = \limu_{+,1} \cdot \limu_{+,2}$.
We first prove that holds. Indeed, for $s,t \in Q$, $\limu_+(s,t) = 1$ if and only if there exists $r \in Q$ such that $\limu_{+,1}(s,r) = 1$ and $\limu_{+,2}(r,t) = 1$. On the other side, since: $$\pRob{\AA}(s \xrightarrow{u} t) =
\sum_{r \in Q} \pRob{\AA}(s \xrightarrow{u_1} r) \cdot \pRob{\AA}(r \xrightarrow{u_2} t),$$ then $\pRob{\AA}(s \xrightarrow{u} t) > 0$ if and only if there exists $r \in Q$ such that $\pRob{\AA}(s \xrightarrow{u_1} r) \cdot \pRob{\AA}(r \xrightarrow{u_2} t) > 0$, which is equivalent to $\pRob{\AA}(s \xrightarrow{u_1} r) > 0$ and $\pRob{\AA}(r \xrightarrow{u_2} t) > 0$. We conclude with the induction hypothesis.
Now we prove that holds. Let $s,t \in Q$ such that $\limu(s,t) = 1$. Then there exists $r \in Q$ such that $\limu_1(s,r) = 1$ and $\limu_2(r,t) = 1$. So: $$\pRob{\AA}(s \xrightarrow{u} t) \ge
\pRob{\AA}(s \xrightarrow{u_1} r) \cdot \pRob{\AA}(r \xrightarrow{u_2} t) \ge
\pmin^{2^h} \cdot \pmin^{2^h} = \pmin^{2^{h+1}},$$ where the second inequality is by induction hypothesis. This completes the proof of .
[**Second case: an iteration node**]{} labelled by $(u,(\limu^\sharp,\limu_+))$ with $k$ sons labelled by $(u_1,(\limu,\limu_{+})),\ldots,(u_k,(\limu,\limu_{+}))$. By definition, $u = u_1 \cdots u_k$, and $(\limu,\limu_+)$ is idempotent.
The proof that holds is similar to the concatenation node case.
Now we prove that holds. Let $s,t \in Q$ such that $\limu^\sharp(s,t) = 1$. Since $k \geq 3$: $$\label{eq:induction_completeness_proof}
\pRob{\AA}(s \xrightarrow{u} t) \geq
\pRob{\AA}(s \xrightarrow{u_1} t) \cdot
\sum_{q \in Q} \pRob{\AA}(t \xrightarrow{u_2 \cdots u_{k-1}} q) \cdot
\pRob{\AA}(q \xrightarrow{u_k} t).$$
To establish we prove that: $$\begin{aligned}
\label{eq:induction_completeness_proof1}
& \pRob{\AA}(s \xrightarrow{u_1} t) \geq \pmin^{2^h},\\
\label{eq:induction_completeness_proof2}
& \text{for all } q \in Q,\ \pRob{\AA}(t \xrightarrow{u_2 \cdots u_{k-1}} q) > 0 \implies
\pRob{\AA}(q \xrightarrow{u_k} t) \geq \pmin^{2^h}.\end{aligned}$$
We prove . Since $\limu^\sharp(s,t) = 1$, by definition $\limu(s,t) = 1$ and $t$ is $\limu$-recurrent. The induction hypothesis for the node $(u_1,(\limu,\limu_+))$ implies that $\pRob{\AA}(s \xrightarrow{u_1} t) \geq \pmin^{2^h}$, *i.e.* .
Now we prove . For that we use the hypothesis that $(\limu,\limu_+)$ is not a leak witness. Let $q \in Q$ such that $\pRob{\AA}(t \xrightarrow{u_2 \cdots u_{k-1}} q) > 0$. By induction hypothesis for each child, implies that $\limu_+^{k-2}(t,q) = 1$. Since $\limu_+$ is idempotent, $\limu_+(t,q) = 1$. We argue that $\limu(q,t) = 1$. Let $\ell \in Q$ a $\limu$-recurrent state such that $\limu(q,\ell) = 1$. Then $\limu_+(t,\ell) = 1$, and $t,\ell$ are $\limu$-recurrent. Since $(\limu,\limu_+)$ is not a leak witness, it follows that $\limu(t,\ell) = 1$, which implies that $\limu(\ell,t) = 1$ since $t$ is $\limu$-recurrent. Together with $\limu(q,\ell) = 1$, this implies $\limu(q,t) = 1$. Thus, by induction hypothesis and according to , $\pRob{\AA}(q \xrightarrow{u_k} t) \geq \pmin^{2^h}$, so holds.
Now, putting , and altogether: $$\begin{aligned}
\pRob{\AA}(s \xrightarrow{u} t) & \geq
\pRob{\AA}(s \xrightarrow{u_1} t) \cdot
\sum_{q \in Q} \pRob{\AA}(t \xrightarrow{u_2 \cdots u_{k-1}} q) \cdot
\pRob{\AA}(q \xrightarrow{u_k} t) \\
& \geq \pmin^{2^h} \cdot \sum_{q \in Q} \pRob{\AA}(t \xrightarrow{u_2 \cdots u_{k-1}} q) \cdot \pmin^{2^h}\\
& = \pmin^{2^{h+1}},\end{aligned}$$ where the last equality holds because $\sum_{q\in Q} \pRob{\AA}(t \xrightarrow{u_2 \cdots u_{k-1}} q) = 1$. This completes the proof of .
To conclude, note that $\monoid_+$ has less than $3^{|Q|^2}$ elements.
Completeness of the Markov monoid algorithm for leaktight automata {#subsec:completeness}
------------------------------------------------------------------
In this subsection we rely on the lower bound lemma (Lemma \[lem:lowerbound\]) to prove Theorem \[theo:completeness\]. Let $\AA$ be a leaktight automaton. By Lemma \[lem:leaktight\_characterization\_direct\], its extended Markov monoid does not contain any leak witness, hence Lemma \[lem:lowerbound\] applies.
We prove the completeness of the Markov monoid associated with $\AA$. Let $(u_n)\nNN$ be a sequence of finite words. By Lemma \[lem:lowerbound\], for each word $u_n$ there exists $(\limu_n,\limu_{+,n})$ in the extended Markov monoid such that for all states $s,t$: $$\limu_n(s,t) = 1 \implies \pRob{\AA}(s \xrightarrow{u_n} t) \geq \dabound.$$ Since the set of limit-words is finite, there exists $N \in \NN$ such that $\set{n \in \NN \mid \limu_N = \limu_n}$ is infinite. To complete the proof, we prove that $\limu_N$ satisfies, for all states $s,t$: $$\limsup \pRob{\AA}(s \xrightarrow{u_n} t) = 0 \implies \limu_N(s,t) = 0.$$ Assume $\limsup \pRob{\AA}(s \xrightarrow{u_n} t) = 0$, then $\limsup \pRob{\AA}(s \xrightarrow{u_n} t) < \dabound$ for $n$ sufficiently large. Since $\limu_N = \limu_n$ for infinitely many $n \in \NN$, this implies $\limu_N(s,t) = 0$, which completes the proof of Theorem \[theo:completeness\].
Properties of leaktight automata {#sec:leaktight_properties}
================================
In this section, we extend the algorithm presented in Section \[sec:algo\], and investigate its running complexity. The extended algorithm has two features: first, it checks *at the same time* whether an automaton is leaktight and whether it contains a value $1$ witness, second, it runs in polynomial space.
We present an algebraic characterization of the leaktight property based on the extended Markov monoid, allowing the extended algorithm to check the leaktight property. For the complexity, one needs a deeper understanding of the Markov monoid; in this section, we will show a linear bound on the $\sharp$-height, allowing to compute the extended Markov monoid in polynomial space. As a corollary, we obtain that the value $1$ problem for leaktight automata is $\PSPACE$-complete.
Characterization of the leaktight property
------------------------------------------
In this subsection, we show the converse of Lemma \[lem:leaktight\_characterization\_direct\], which implies the following theorem, characterizing the leaktight property in algebraic terms.
\[theo:leaktight\_characterization\] An automaton $\AA$ is leaktight if and only if its extended Markov monoid does not contain any leak witness.
Lemma \[lem:lowerbound\] is instrumental in the proof of this lemma.
We prove that if the extended Markov monoid of an automaton $\AA$ does not contain any leak witness, then $\AA$ is leaktight. The converse was proved in Lemma \[lem:leaktight\_characterization\_direct\].
Assume $\AA$ has a leak $(u_n)\nNN$, we show that its extended Markov monoid contains a leak witness. Consider the Markov chain $\MM$ with state space $Q$ and transition matrix $M$ defined by $M(s,t) = \lim_n \pRob{\AA}(s \xrightarrow{u_n} t)$. By assumption $M$ is idempotent.
By definition of a leak: $$\begin{aligned}
\label{eq:proof_leaktight_converse_leak1}
r \text{ and } q \text{ are recurrent in } \MM, \\
\label{eq:proof_leaktight_converse_leak2}
M(r,q) = 0,\\
\label{eq:proof_leaktight_converse_leak3}
\text{for all } n \in \NN,\ \pRob{\AA}(r \xrightarrow{u_n} q) > 0.\end{aligned}$$
Assume towards contradiction that the extended Markov monoid does not contain any leak witness, then Lemma \[lem:lowerbound\] applies. For each word $u_n$, there exists $(\limu_n,\limu_{+,n})$ in the extended Markov monoid such that for all states $s,t$: $$\begin{aligned}
\label{eq:proof_leaktight_converse1}
\limu_{+,n}(s,t) = 1 & \iff \pRob{\AA}(s \xrightarrow{u_n} t) > 0,\\
\label{eq:proof_leaktight_converse2}
\limu_n(s,t) = 1 & \implies \pRob{\AA}(s \xrightarrow{u_n} t) \geq \dabound .\end{aligned}$$ Since the extended Markov monoid is finite, there exists $N \in \NN$ such that for infinitely many $n \in \NN$, we have $(\limu_N,\limu_{+,N}) = (\limu_n,\limu_{+,n})$.
Note that since each $u_n$ is idempotent, implies that each $\limu_{+,n}$ is idempotent as well.
Let $(\limv, \limv_+) = (\limu_N, \limu_{+,N})^{|Q|!}$. The power $|Q|!$ ensures that $\limu_N^{|Q|!}$ is idempotent, by Lemma \[lem:basic\_limit\_words\]. Since $\limu_{+,N}$ is idempotent, $\limv_+ = \limu_{+,N}$. Also, since $\limv$ is idempotent, there exists $r'$ and $q'$ which are $\limv$-recurrent, such that $\limv(r,r') = 1$ and $\limv(q,q') = 1$, again thanks to Lemma \[lem:basic\_limit\_words\].
Now, we prove that $(\limv,\limv_+)$ is a leak witness: $$\begin{aligned}
\label{eq:proof_leaktight_converse_lk1}
r' \text{ and } q' \text{ are } \limv\text{-recurrent}, \\
\label{eq:proof_leaktight_converse_lk2}
\limv(r',q') = 0,\\
\label{eq:proof_leaktight_converse_lk3}
\limv_+(r',q') = 1.\end{aligned}$$
Let $\eta = \dabound$ and $K = |Q|!$.
Observe that for all states $s,t$, we have $\limv(s,t) = 1 \implies M(s,t) > 0$: $$\begin{aligned}
\notag & \limv(s,t) = 1 & \\
\notag & \implies \limu_N^K(s,t) = 1 & \text{(by definition of $\limv$)} \\
\notag & \implies \limu_n^K(s,t) = 1 \text{ for infinitely many $n$} & \text{(by definition of $N$)} \\
\notag & \implies \pRob{\AA}(s \xrightarrow{u_n^K} t) \geq \eta^K \text{ for infinitely many $n$} & \text{(by~\eqref{eq:proof_leaktight_converse2})} \\
\notag & \implies \lim_n \pRob{\AA}(s \xrightarrow{u_n^K} t) \geq \eta^K & \\
\notag & \implies M^K(s,t) > 0 & \text{(by definition of $M$)} \\
\notag & \implies M(s,t) > 0 & \text{(since $M$ is idempotent)}.\end{aligned}$$
First, is by definition of $r'$ and $q'$.
We prove . Towards contradiction, assume that $\limv(r',q') = 1$. Then $M(r',q') > 0$, so together with $M(r,r') > 0$ (which follows from $\limv(r,r') = 1$) this implies $M^2(r,q') > 0$, so $M(r,q') > 0$ as $M$ is idempotent. Since $M(q,q') > 0$ (which follows from $\limv(q,q') = 1$) and $q$ is recurrent in $M$, we have $M(q',q) > 0$. This implies $M^2(r,q) > 0$, and $M(r,q) > 0$ because $M$ is idempotent, which contradicts .
We prove . Thanks to and , we have $\limu_{+,N}(r,q) = 1$, *i.e.* $\limv_+(r,q) = 1$. Since $M(r,r') > 0$ and $r$ is recurrent in $M$, we have $M(r',r) > 0$, so implies that $\limu_{+,N}(r',r) = 1$, *i.e.* $\limv_+(r',r) = 1$. Similarly, $M(q,q') > 0$, so implies that $\limu_{+,N}(q,q') = 1$, *i.e.* $\limv_+(q,q') = 1$. The three equalities $\limv_+(r',r) = 1$, $\limv_+(r,q) = 1$ and $\limv_+(q,q') = 1$ imply $\limv_+^3(r',q') = 1$, and since $\limv_+$ is idempotent $\limv_+(r',q') = 1$.
It follows that $(\limv,\limv_+)$ is a leak witness, which completes the proof.
The immediate corollary of Theorem \[theo:leaktight\_characterization\] is that checking whether an automaton is leaktight can be done by computing the extended Markov monoid and looking for leak witnesses, hence it is decidable.
The extended Markov monoid algorithm
------------------------------------
Algorithm \[algo:extended\_markov\_monoid\] computes the extended Markov monoid, and looks for value $1$ witnesses, which in the extended Markov monoid is an extended limit-word $(\limu,\limu_+)$ such that $\limu$ is a value $1$ witness (in the Markov monoid). If there is a value $1$ witness, then the automaton has value $1$, even if it is not leaktight, thanks to Theorem \[theo:consistency\]. Otherwise, the algorithm looks for a leak witness; if there is no leak witness, then the automaton is leaktight thanks to Theorem \[theo:leaktight\_characterization\], and it does not have value $1$ thanks to Theorem \[theo:completeness\]. In case there is a leak witness, the automaton is not leaktight, and nothing can be said.
$\monoidext \gets \set{(\bolda,\bolda) \mid a\in A} \cup \set{(\boldeps,\boldeps)}$.
Parallel composition and PSPACE-hardness
----------------------------------------
The objective of this subsection is to prove the $\PSPACE$-hardness of the value $1$ problem for leaktight automata. To this end, we give a reduction from the emptiness problem of $n$ deterministic automata. To prove that the reduction indeed constructs leaktight automata, we need to show that deterministic automata are leaktight, and the closure under parallel composition.
\[prop:deterministic\] Deterministic automata are leaktight.
For all limit-words $\limu \in \set{\bolda \mid a \in A} \cup \set{\boldeps}$, for all states $s$, there exists a unique state $t$ such that $\limu(s,t) = 1$. In particular, each recurrence class is formed of only one state with a self-loop. This property is preserved by concatenation, and implies that the iteration operation is trivial, *i.e.* $\limu^\sharp = \limu$. Consequently, for all extended limit-words $(\limu,\limu_+)$ in the extended Markov monoid, we have $\limu = \limu_+$, which implies that there are no leak witnesses.
Consider two probabilistic automata, denoted $\AA = (Q^\AA,q_0^\AA,\Delta^\AA,F^\AA)$ and $\BB = (Q^\BB,q_0^\BB,\Delta^\BB,F^\BB)$. We assume that $Q^\AA$ and $Q^\BB$ are disjoint.
The parallel composition of $\AA$ and $\BB$ is: $$\AA\ ||\ \BB\ =\ (Q^\AA \uplus Q^\BB\ ,\ \delta_0\ ,\ \Delta\ ,\ F^\AA \cup F^\BB),$$ where $\delta_0 = \frac{1}{2} \cdot q_0^\AA + \frac{1}{2} \cdot q_0^\BB$, and: $$\Delta(q,a) =
\begin{cases}
\Delta_\AA(q,a) & \textrm{ if } q \in Q_\AA, \\
\Delta_\BB(q,a) & \textrm{ if } q \in Q_\BB. \\
\end{cases}$$
By definition, for $u \in A^*$, we have $\pRob{\AA||\BB}(u) = \frac{1}{2} \cdot \pRob{\AA}(u) + \frac{1}{2} \cdot \pRob{\BB}(u)$. Note that in this definition, we allowed an initial probability distribution rather than only one initial state. This could be avoided by adding a new initial state that leads to each previous initial state with probability half, but we do it here for technical convenience in the proof of the following proposition.
\[prop:parallel\_composition\] The leaktight property is stable by parallel composition.
The extended Markov monoid $\monoidext^{\AA||\BB}$ of the parallel composition embeds into the direct product $\monoidext^\AA \times \monoidext^\BB$ of the extended Markov monoids of each automaton.
Note that for $(\limu,\limu_+) \in \monoidext^{\AA||\BB}$, if $\limu(s,t) = 1$, then either $s,t \in Q^\AA$ or $s,t \in Q^\BB$, and similarly for $\limu_+$. Relying on this, we map $(\limu,\limu_+) \in \monoidext^{\AA||\BB}$ to $\left((\limu,\limu_+)[\AA]\ ,\ (\limu,\limu_+)[\BB]\right)$, where $(\limu,\limu_+)[\AA]$ is the restriction to $\AA$ and similarly for $\BB$. An easy induction on $(\limu,\limu_+)$ shows that this map is an embedding into $\monoidext^\AA \times \monoidext^\BB$.
Consequently, the extended Markov monoid of the parallel composition contains a leak witness if and only if one of the extended Markov monoid contains a leak witness.
Now that we proved that deterministic automata are leaktight, and the closure under parallel composition, the $\PSPACE$-hardness of the value $1$ problem for leaktight automata is easy.
\[prop:pspace\_hard\] The value $1$ problem for leaktight automaton is $\PSPACE$-hard.
We give a reduction from the following problem: given $n$ deterministic automata over finite words, decide whether the intersection of the languages they accept is empty. This problem is $\PSPACE$-hard [@K77].
The reduction is as follows: given $n$ deterministic automata, we construct the parallel composition of the $n$ automata, where each copy is reached with probability $\frac{1}{n}$. This automaton has value $1$ if and only if the intersection of the languages is not empty, and is leaktight by Proposition \[prop:deterministic\] and Proposition \[prop:parallel\_composition\].
Bounding the sharp-height in the Markov monoid
----------------------------------------------
We now consider the running complexity of the extended Markov monoid algorithm. A naïve argument shows that it terminates in less than $3^{|Q|^2}$ iterations, since each iteration adds a new extended limit-word in the monoid and there are at most $3^{|Q|^2}$ different limit-words. This gives an EXPTIME upper bound.
A better complexity can be achieved by looking for a value $1$ witness or a leak witness in a non-deterministic way. The algorithm guesses the witness by its decomposition into concatenations and iterations. The key observation, made by Kirsten [@K05] in the context of distance desert automata, is that the $\sharp$-height, that is the number of nested applications of the iteration operation, can be restricted to at most $|Q|$.
Note that when dealing with $\sharp$-height, it suffices to consider limit-words instead of extended limit-words, as by definition the second component of an extended limit-word does not contain any $\sharp$.
Formally, we define the $\sharp$-hierarchy inside the Markov monoid as follows: $$\begin{aligned}
\notag & S_0 = \langle \set{\bolda \mid a \in A} \cup \set{\boldeps} \rangle , \\
\notag & S_{p+1} = \langle S_p \cup \set{\limu^\sharp \mid \limu \in E(S_p)} \rangle ,\end{aligned}$$ where $\langle T \rangle$ is the set of limit-words obtained as concatenation of limit-words in $T$.
\[def:sharp\_height\] The $\sharp$-height of a limit-word $\limu$ is the minimal $p$ such that $\limu \in S_p$.
\[theo:sharp\_hierarchy\] Every limit-word has $\sharp$-height at most $|Q|$, *i.e.* the $\sharp$-hierarchy collapses at level $|Q|$.
In the following, we adapt Kirsten’s proof from [@K05] to the setting of probabilistic automata. Roughly speaking, the proof consists in associating a quantity to each idempotent element of the Markov monoid, and to show the following:
- the quantity is bounded above by $|Q|$.
- the quantity strictly decreases when iterating an unstable limit-word (*i.e.* if $\limu^\sharp \neq \limu$),
- the quantity does not increase when concatenating.
Let $\limu$ be an idempotent limit-word, we define $\sim_\limu$ the relation on $Q$ by $s \sim_\limu t$ if $\limu(s,t) = 1$ and $\limu(t,s) = 1$. Clearly, $\sim_\limu$ is symmetric, and since $\limu$ is idempotent, $\sim_\limu$ is transitive. If for some state $s$ there exists a state $t$ such that $s \sim_\limu t$, then $s \sim_\limu s$ since $\limu$ is idempotent. Consequently, the restriction of $\sim_\limu$ to the set $$Z_\limu = \set{s \in Q \mid s \sim_\limu s}$$ is reflexive, *i.e.* $\sim_\limu$ is an equivalence relation on $Z_\limu$. From now on by equivalence class of $\sim_\limu$ we mean an equivalence class of $\sim_\limu$ on $Z_\limu$. We denote by $[s]_\limu$ the equivalence class of $s$, and by $\Cl(\limu)$ the set of equivalence classes of $\sim_\limu$. The quantity associated with $\limu$ is $|\Cl(\limu)|$, the number of equivalence classes of $\sim_\limu$, that is the number of non-trivial connected components in the underlying graph of $\limu$. Note that $|\Cl(\limu)| \le |Q|$.
Here are two useful observations.
\[lem:sharp\_height\_tool\]
- Let $\limu,\limv$ be two limit-words and $s,t,r \in Q$. Then $(\limu \cdot \limv)(s,t) \geq \limu(s,r) \cdot \limv(r,t)$.
- Let $\limu$ be an idempotent limit-word and $s,t \in Q$. There exists $r \in Q$ such that $\limu(s,t) = \limu(s,r) \cdot \limu(r,r) \cdot \limu(r,t)$.
The first claim is clear and follows from the equality: $$(\limu \cdot \limv)(s,t) = \sum_{r \in Q} \limu(s,r) \cdot \limv(r,t).$$ Consider now the second claim. For all states $r \in Q$, since $\limu$ is idempotent we have: $$\limu(s,t) = \limu^3(s,t) = \sum_{p,q \in Q} \limu(s,p) \cdot \limu(p,q) \cdot \limu(q,t)
\geq \limu(s,r) \cdot \limu(r,r) \cdot \limu(r,t).$$ Since $\limu$ is idempotent, we have $\limu = \limu^{n+2}$, so there exist $s = r_0,\ldots,r_{n+2} = t$ such that $\limu(s,t) = \limu(r_0,r_1) \cdots \limu(r_{n+1},r_{n+2})$. By a counting argument, there exist $i,j$ such that $1 \leq i < j \leq (n + 1)$ and $r_i = r_j$, denote it by $r$. We have: $$\begin{aligned}
\notag & \limu(s,r) = \limu^i(s,r) \geq \limu(r_0,r_1) \cdots \limu(r_{i-1},r_i), \\
\notag & \limu(r,r) = \limu^{j-i}(r,r) \geq \limu(r_i,r_{i+1}) \cdots \limu(r_{j-1},r_j),\\
\notag & \limu(r,t) = \limu^{n+2-j}(r,t) \geq \limu(r_j,r_{j+1}) \cdots \limu(r_{n+1},r_{n+2}).\end{aligned}$$ Hence, $\limu(s,r) \cdot \limu(r,r) \cdot \limu(r,t) \geq \limu(r_0,r_1) \cdots \limu(r_{n+1},r_{n+2}) = \limu(s,t)$, and the second claim follows.
The following lemma shows that the quantity $|\Cl(\limu)|$ strictly decreases when iterating an unstable limit-word (*i.e.* if $\limu^\sharp \neq \limu$).
\[lem:sharp\_height\_decrease\] Let $\limu$ be an idempotent limit-word. $$\begin{aligned}
\label{eq:sharp_height1}
\Cl(\limu^\sharp) \subseteq \Cl(\limu), \\
\label{eq:sharp_height2}
\text{if } \limu \neq \limu^\sharp, \text{ then } \Cl(\limu^\sharp) \neq \Cl(\limu).\end{aligned}$$
We prove . Let $s \in Z_{\limu^\sharp}$; by definition we have $s \sim_{\limu^\sharp} s$. We show that $[s]_{\limu^\sharp} = [s]_\limu$. For all states $t \in Q$ such that $s \sim_{\limu^\sharp} t$, we have $s \sim_\limu t$, so $[s]_{\limu^\sharp} \subseteq [s]_\limu$. Conversely, let $t \in [s]_\limu$; we have $\limu(s,t) = 1$ and $\limu(t,s) = 1$. Since $s \sim_{\limu^\sharp} s$, we have $\limu^\sharp(s,s) = 1$. So $\limu^\sharp(s,t) = (\limu^\sharp \cdot \limu)(s,t) \geq \limu^\sharp(s,s) \cdot \limu(s,t) = 1$, and similarly $\limu^\sharp(t,s) = (\limu \cdot \limu^\sharp)(t,s) \geq \limu(t,s) \cdot \limu^\sharp(s,s) = 1$. Thus $s \sim_{\limu^\sharp} t$, *i.e.* $t \in [s]_{\limu^\sharp}$, which concludes to the equality $[s]_{\limu^\sharp} = [s]_\limu$. In other words, the equivalence classes for $\limu^\sharp$ are also equivalence classes for $\limu$, so $\Cl(\limu^\sharp) \subseteq \Cl(\limu)$.
We prove . Assume $\limu \neq \limu^\sharp$; let $s,t$ such that $\limu(s,t) = 1$ and $\limu^\sharp(s,t) = 0$. By Lemma \[lem:sharp\_height\_tool\], there exists $r$ such that $\limu(s,t) = \limu(s,r) \cdot \limu(r,r) \cdot \limu(r,t)$, so $\limu(r,r) = 1$. Towards contradiction, assume $\limu^\sharp(r,r) = 1$. It follows that: $$\limu^\sharp(s,t) = (\limu \cdot \limu^\sharp \cdot \limu)(s,t) \geq
\limu(s,r) \cdot \limu^\sharp(r,r) \cdot \limu(r,t) =
\limu(s,r) \cdot \limu(r,r) \cdot \limu(r,t) = 1,$$ *i.e.*, $\limu^\sharp(s,t) = 1$ which is a contradiction. Consequently, $\limu^\sharp(r,r) = 0$, so $r \sim_\limu r$ and $r \not\sim_{\limu^\sharp} r$. Thus, $r \in Z_\limu$ but $r \notin Z_{\limu^\sharp}$. Hence, there is a class $[r]_\limu$ in $\Cl(\limu)$, but there is no class $[r]_{\limu^\sharp}$ in $\Cl(\limu^\sharp)$.
The following lemma shows that the quantity $|\Cl(\limu)|$ is common to all idempotents in the same $\JJ$-class. The notion of $\JJ$-class is a classical notion for the theory of monoids, derived from one of the four Green’s relations called the $\JJ$-preorder (for details about Green’s relations, see [@L79; @H95], or [@C11] for its applications to automata theory).
Define $\limu \le_\JJ \limv$ if there exist $\lima,\limb$ such that $\lima \cdot \limv \cdot \limb = \limu$, and $\limu \JJ \limv$, *i.e.* $\limu$ and $\limv$ are in the same $\JJ$-class, if $\limu \le_\JJ \limv$ and $\limv \le_\JJ \limu$.
\[lem:sharp\_height\_concatenation\] Let $\limu,\limv$ be idempotent limit-words. If $\limu \leq_\JJ \limv$, then $|\Cl(\limu)| \leq |\Cl(\limv)|$.
Let $\lima,\limb$ two limit-words such that $\lima \cdot \limv \cdot \limb = \limu$. First, without loss of generality we assume that $\lima \cdot \limv = \lima$ and $\limv \cdot \limb = \limb$. Indeed, if $\lima$ and $\limb$ do not satisfy these conditions, then we consider $\lima = \lima \cdot \limv$ and $\limb = \limv \cdot \limb$.
We construct a partial surjective mapping $\beta : \Cl(\limv) \rightarrow \Cl(\limu)$, which depends on the choice of $\lima$ and $\limb$. For all states $s \in Z_\limv$ and $t \in Z_\limu$ satisfying $\lima(t,s) \cdot \limv(s,s) \cdot \limb(s,t) = 1$ we set $\beta([s]_\limv) = [t]_\limu$. To complete the proof, we have to show that $\beta$ is well defined and that $\beta$ is indeed surjective.
We show that $\beta$ is well defined. Let $s,s' \in Z_\limv$ and $t,t' \in Z_\limu$, and assume $\lima(t,s) \cdot \limv(s,s) \cdot \limb(s,t) = 1$ and $\lima(t',s') \cdot \limv(s',s') \cdot \limb(s',t') = 1$. By definition $\beta([s]_\limv) = [t]_\limu$ and $\beta([s']_\limv) = [t']_\limu$. To show that $\beta$ is well defined, we have to show that if $[s]_\limv = [s']_\limv$, then $[t]_\limu = [t']_\limu$. Assume $[s]_\limv = [s']_\limv$, *i.e.*, $s \sim_\limv s'$, so $\limv(s,s') = 1$. Since $\lima(t,s) \cdot \limv(s,s) \cdot \limb(s,t) = 1$, we have $\lima(t,s) = \limb(s,t) = 1$. Similarly, $\lima(t',s') \cdot \limv(s',s') \cdot \limb(s',t') = 1$, so $\lima(t',s') = \limb(s',t') = 1$. Consequently, $\lima(t,s) \cdot \limv(s,s') \cdot \limb(s',t') = 1$, so $\limu(t,t') = (\lima \cdot \limv \cdot \limb)(t,t') = 1$. Symmetrically, $\lima(t',s') \cdot \limv(s',s) \cdot \limb(s,t) = 1$, so $\limu(t',t) = 1$, concluding to $t \sim_\limu t'$, *i.e.* $[t]_\limu = [t']_\limu$.
We show that $\beta$ is surjective. Let $t \in Z_\limu$. We exhibit some $s$ such that $\beta([s]_\limv) = [t]_\limu$. Since $\limu = \lima \cdot \limv \cdot \limb$, there are $p, q$ such that $\lima(t,p) \cdot \limv(p,q) \cdot \limb(q,t) = \limu(t,t) = 1$, so $\lima(t,p) = \limv(p,q) = \limb(q,t) = 1$. By Lemma \[lem:sharp\_height\_tool\] there exists $s$ such that $\limv(p,s) \cdot \limv(s,s) \cdot \limv(s,q) = \limv(p,q) = 1$, so, $\limv(p,s) = \limv(s,s) = \limv(s,q) = 1$. We have $\lima(t,s) = (\lima \cdot \limv)(t,s) \geq \lima(t,p) \cdot \limv(p,s) = 1$, and $\limb(s,t) = (\limv \cdot \limb)(s,t) \geq \limv(s,q) \cdot \limb(q,t) = 1$. To sum up, $\lima(t,s) \cdot \limv(s,s) \cdot \limb(s,t) = 1$, and hence, $\beta([s]_\limv) = [t]_\limu$.
The following lemma wraps up the previous two lemma. For technical convenience, we set $S_{-1} = \emptyset$.
\[lem:sharp\_height\_conclusion\] Let $\limu$ be an idempotent limit-word and $p \ge 0$. If $\limu \in S_p \setminus S_{p-1}$, then $|\Cl(\limu)| \le |Q| - p$.
We proceed by induction on $p$. For $p = 0$, the assertion is obvious. Let $p \ge 0$, we show the claim for $p + 1$. Let $\limu$ be an idempotent limit-word such that $\limu \in S_{p+1} \setminus S_p$. By definition, $\limu = \limv_1 \cdots \limv_k$ where for all $i$, either $\limv_i \in S_p$ or $\limv_i = \limu_i^\sharp$ for $\limu_i \in S_p$ and $\limv_i \notin S_p$.
If for all $i$, $\limv_i \in S_p$, then $\limu = \limv_1 \cdots \limv_k \in S_p$, which is a contradiction. Consequently, there exists $i$ such that $\limv_i = \limu_i^\sharp$ for $\limu_i \in S_p$ and $\limv_i \notin S_p$. Since $\limu_i \in S_p$ and $\limv_i = \limu_i^\sharp \notin S_p$, we have $\limu_i^\sharp \neq \limu_i$. Towards contradiction, assume $\limu_i \in S_{p-1}$, then $p \ge 1$, and this implies $\limu_i^\sharp \in S_p$, which is a contradiction. Hence, $\limu_i \in S_p \setminus S_{p-1}$.
By induction, we have $|\Cl(\limu_i)| \le |Q| - p$. Since $\limu_i^\sharp \neq \limu_i$, by Lemma \[lem:sharp\_height\_decrease\] we have $|\Cl(\limu_i^\sharp)| < |\Cl(\limu_i)|$. Since $\limu \le_\JJ \limu_i^\sharp$, by Lemma \[lem:sharp\_height\_concatenation\] we have $|\Cl(\limu)| \le |\Cl(\limu_i^\sharp)|$. Altogether, it follows $|\Cl(\limu)| \le |Q| - (p+1)$.
It follows from Lemma \[lem:sharp\_height\_conclusion\] that $S_{|Q|} = S_{|Q| + 1}$, *i.e.* the $\sharp$-hierarchy collapses at level $|Q|$, proving Theorem \[theo:sharp\_hierarchy\].
The bound is almost tight, as shown in figure \[fig:sharpstar\]. The only value $1$ witness of this automaton is $(\cdots((\bolda_0^\sharp\ \bolda_1)^\sharp\ \bolda_2)^\sharp\ \bolda_3)^\sharp\ \cdots \bolda_{n-1})^\sharp$, whose $\sharp$-height is $|Q|-2$.
![\[fig:sharpstar\] A leaktight automaton with value $1$ and $\sharp$-height $|Q|-2$.](EPS/fig6)
Note that this automaton is leaktight, so the extended Markov monoid algorithm will find the value $1$ witness and correctly answers that it has value $1$.
Finding witnesses in the Markov monoid
--------------------------------------
In the subsection, we will prove the following complexity result.
\[prop:pspace\] There exists an algorithm which checks in polynomial space whether an automaton is leaktight and whether in such case it has value $1$.
Following Theorem \[theo:completeness\], checking whether a leaktight automaton has value $1$ boils down to finding a value $1$ witness in the Markov monoid. Similarly, following Theorem \[theo:leaktight\_characterization\], checking whether an automaton is *not* leaktight boils down to finding a leak witness in the extended Markov monoid. Note that in both cases, checking whether a given limit-word or extended limit-word is a witness is easily done in polynomial time.
Since we aim at proving that those two tasks can be computed in $\PSPACE$, which is closed under complementation, it suffices to show how to find a witness in the (extended) Markov monoid. For the sake of readability, we here only deal with the Markov monoid, but similar ideas apply to the extended Markov monoid.
We describe an algorithm to guess a witness in the Markov monoid. The key property given by Theorem \[theo:sharp\_hierarchy\] is that we can restrict ourselves to at most $|Q|$ nested iteration operations.
As the corresponding property was proved by Kirsten [@K05] in the context of distance automata, also to obtain a $\PSPACE$ algorithm, the following algorithm is also an adaptation of [@K05]. Rather than a formal proof, we here give an intuitive description of the algorithm.
A witness can be described as a tree whose nodes are labelled by limit-words, of depth at most $2 \cdot |Q| + 1$, as follows:
- a leaf is labelled either by $\bolda$ for $a \in A$ or by $\boldeps$,
- an internal node can be a *concatenation node*, then it is labelled by $\limu = \limv_1 \cdots \limv_k$ for $k \le 2^{|Q|^2}$ and has $k$ children, labelled by $\limv_1,\ldots,\limv_k$,
- an internal node can be an *iteration node*, then it is labelled by $\limu^\sharp$ and has one child labelled $\limu$.
We describe an algorithm that guesses such a tree. It starts from the root, and travels over nodes in a depth-first way: from top to bottom (and up again) and from left to right. In a node, the algorithm stores the branch that leads to this node, and for each node in the branch the limit-word obtained by concatenating all the left siblings of this node. From a node, the algorithm guesses a limit-word, and whether it will be a leaf, a concatenation node or an iteration node. In the first case, it goes up and checks the consistency of this guess. In the two other cases, it updates the value of this node by concatenating the new guess with the previous value and goes down.
Although the tree is of exponential size, in each step the algorithm only stores $2 \cdot |Q| + 1$ limit-words at most, so it runs in polynomial space.
Examples and subclasses of leaktight automata {#sec:leaktight_comparisons}
=============================================
In this section, we investigate further the class of leaktight automata, by giving examples of leaktight automata, exhibiting subclasses, and showing closure properties. In particular, we prove that hierarchical automata, $\sharp$-acylic automata and simple automata are all strict subclasses of leaktight automata. (Actually, since $\sharp$-acylic automata are already a subclass of simple automata, we do not consider them.) This implies that our decidability result extends the decidability results from [@GO10; @CT12].
Two basic examples
------------------
The automaton on figure \[fig:3\] is leaktight. As we shall see, it is not hierarchical, nor simple, hence it witnesses that leaktight automata are not subsumed by hierarchical or simple automata. Its extended Markov monoid is depicted on the right-hand side. Each of the four directed graphs represents an extended limit-word $(\limu,\limu_+)$: if $\limu(s,t) = 1$, then $(s,t)$ is an edge, and if $\limu(s,t) = 0$ but $\limu_+(s,t) = 1$, then $(s,t)$ is marked with $+$.
The initial state of the automaton is state $0$, and the unique final state is state $1$. This automaton has value $1$ and this can be checked using the extended Markov monoid: the two value $1$ witnesses are $\lima^\sharp$ and $\limb \cdot \lima^\sharp$.
![\[fig:3\] A leaktight automaton and its extended Markov monoid.](EPS/fig3)
The automaton on figure \[fig:4\] is leaktight. The initial state of the automaton is state $0$, and the unique final state is state $F$. The Markov monoid has too many elements to be represented here. This automaton does not have value $1$.
![\[fig:4\] A leaktight automaton which does not have value $1$.](EPS/fig5)
Some closure properties
-----------------------
In this subsection, we show some closure properties: synchronised product and composition with a deterministic transducer. As for Proposition \[prop:parallel\_composition\], the automata constructions reflect in algebraic constructions, allowing to use the characterization with leak witnesses given by Theorem \[theo:leaktight\_characterization\].
In this subsection, we will omit initial and final states, as we are only interested in preserving the leaktight property, which does not depend on it.
Let $\AA = (Q^\AA,\Delta^\AA)$ and $\BB = (Q^\BB,\Delta^\BB)$ be two probabilistic automata.
A synchronised product of $\AA$ and $\BB$ is: $$\AA \times \BB\ =\ (Q^\AA \times Q^\BB\ ,\ \Delta),$$ where $\Delta(q,a) = (\Delta_\AA(q,a),\Delta_\BB(q,a))$.
The leaktight property is stable by synchronized product.
The extended Markov monoid $\monoidext^{\AA \times \BB}$ of the synchronized product embeds into the direct product $\monoidext^\AA \times \monoidext^\BB$ of the extended Markov monoids of each automaton.
Let $(\limu,\limu_+)$ be an extended limit-word in $\monoidext^{\AA \times \BB}$. Define $\limu_\AA(s,t) = 1$ if there exists $s',t' \in Q^\BB$ such that $\limu((s,s'),(t,t')) = 1$, and similarly for $\limu_{+,\AA}$, $\limu_\BB$ and $\limu_{+,\BB}$. We have the following equivalence: $$\limu((s,s'),(t,t')) = 1 \iff \limu_\AA(s,t) = 1 \wedge \limu_\BB(s',t') = 1,$$ and similarly for $\limu_+$.
Relying on this, we map $(\limu,\limu_+) \in \monoidext^{\AA \times \BB}$ to $\left((\limu_\AA,\limu_{+,\AA})\ ,\ (\limu_\BB,\limu_{+,\BB})\right)$. An easy induction shows that this map is an embedding into $\monoidext^\AA \times \monoidext^\BB$.
Consequently, the extended Markov monoid of the synchronised product contains a leak witness if and only if one of the extended Markov monoid contains a leak witness.
The last closure property we prove will be useful in the next section.
Let $\AA = (Q^\AA,\Delta^\AA)$ be a probabilistic automaton, and $\MM = (Q^\MM,Q^\AA,\Delta^\MM)$ a deterministic transducer over $\AA$, *i.e.* $\Delta^\MM : Q^\MM \times Q^\AA \rightarrow Q^\MM$.
The composition of $\AA$ by $\MM$ is: $$\AA \otimes \MM\ =\ (Q^\AA \times Q^\MM\ ,\ \Delta),$$ where $\Delta(q,p,a) = (\Delta_\AA(q,a),\Delta_\MM(p,q))$.
\[prop:deterministic\_transducer\] The leaktight property is stable by composition with a deterministic transducer.
Following the same reasoning as in \[prop:deterministic\], one can show that the extended Markov monoids for $\AA$ and for $\AA \otimes \MM$ are isomorphic.
Leaktight automata strictly contain hierarchical automata
---------------------------------------------------------
The class of hierarchical automata has been defined in [@CSV11], where it was proved that they are recognize exactly the class of $\omega$-regular languages. The states $Q$ of a hierarchical automaton are sorted according to levels such that for each letter, at most one successor is at the same level and all others are at higher levels.
Formally, there exists a mapping $\rank : Q \to \NN$ such that for all $a \in A$, for all states $s,t$ such that $\pRob{\AA}(s \xrightarrow{a} t) > 0$, we have $\rank(s) \leq \rank(t)$. Furthermore, if $\pRob{\AA}(s \xrightarrow{a} t) > 0$ and $\pRob{\AA}(s \xrightarrow{a} t') > 0$ but $\rank(s) = \rank(t) = \rank(t')$, then $t = t'$.
\[prop:hierarchical\] Every hierarchical automata is leaktight.
We prove by induction that for every extended limit-word $(\limu,\limu_+)$ in the extended Markov monoid of a hierarchical automaton, for every states $s,t,t'$: $$\begin{aligned}
\label{eq:hierarchical1}
\limu_+(s,t) = 1 \implies \rank(s) \le \rank(t),\\
\label{eq:hierarchical2}
\limu_+(s,t) = 1\ \wedge\ \limu_+(s,t') = 1\ \wedge\ \rank(s) = \rank(t) = \rank(t')\ \implies \ t = t',\\
\label{eq:hierarchical3}
\limu(s,t) = 1 \implies \rank(s) \le \rank(t),\\
\label{eq:hierarchical4}
\limu(s,t) = 1\ \wedge\ \limu(s,t') = 1\ \wedge\ \rank(s) = \rank(t) = \rank(t')\ \implies \ t = t'.\end{aligned}$$ Note that and imply and , since $\limu(s,t) = 1$ implies $\limu_+(s,t) = 1$. The key property following from and is that the recurrence classes of $\limu$ and of $\limu_+$ are singletons.
This is trivial for $(\boldeps,\boldeps)$. The case of $(\bolda,\bolda)$ is the definition of hierarchical automata. The induction step for concatenation is routinely checked, and trivial for the iteration.
We now prove that the extended Markov monoid of a hierarchical automaton does not contain any leak witness. We prove a slightly stronger statement; for every extended limit-word $(\limu,\limu_+)$ in the extended Markov monoid of a hierarchical automaton, for all states $q,r$: $$\begin{aligned}
\label{eq:hierarchical5}
\limu_+(r,q) = 1\ \wedge\ r \text{ is } \limu\text{-recurrent} \implies \ \limu(r,q) = 1.\end{aligned}$$ This is clear for $(\boldeps,\boldeps)$ and for $(\bolda,\bolda)$.
Concatenation. Let $(\limu,\limu_+)$ and $(\limv,\limv_+)$ be two extended limit-words satisfying . Consider two states $r,q$ such that $(\limu_+ \cdot \limv_+)(r,q) = 1$ and $r$ is $(\limu \cdot \limv)$-recurrent.
Since $r$ is $(\limu \cdot \limv)$-recurrent, the recurrence class of $r$ for $(\limu \cdot \limv)$ is $r$ itself, so $(\limu \cdot \limv)(r,r) = 1$ and $(\limu \cdot \limv)(r,p) = 1$ implies $p = r$. Let $t$ be a state such that $\limu(r,t) = 1$ and $\limv(t,r) = 1$. Thanks to , $\rank(t) = \rank(r)$.
We argue that $\limu(r,p) = 1$ implies $p = t$. Indeed, let $p$ be a state such that $\limu(r,p) = 1$. There exists a state $p'$ such that $\limv(p,p') = 1$, so in particular $(\limu \cdot \limv)(r,p') = 1$, so $p' = r$. By , $\rank(r) \le \rank(p) \le \rank(r)$, so they are equal. By , since $\limu(r,t) = 1$, $\limu(r,p) = 1$ and $\rank(r) = \rank(t) = \rank(p)$, we have $p = t$. It follows that the state $t$ is $\limu$-recurrent: by Lemma \[lem:basic\_limit\_words\], there exists a state $p$ such that $\limu(r,p) = 1$ and $p$ is $\limu$-recurrent. The above remark implies that $p = t$.
We argue that $\limv(t,p) = 1$ implies $p = r$. Indeed, let $p$ be a state such that $\limv(t,p) = 1$. We have $(\limu \cdot \limv)(r,p) = 1$, so $p = r$. It follows that the state $r$ is $\limv$-recurrent: by Lemma \[lem:basic\_limit\_words\], there exists a state $p$ such that $\limv(t,p) = 1$ and $p$ is $\limu$-recurrent. The above remark implies that $p = r$.
Since $(\limu_+ \cdot \limv_+)(r,q) = 1$, there exists $s \in Q$ such that $\limu_+(r,s) = 1$ and $\limv_+(s,q) = 1$. By induction hypothesis for $(\limu,\limu_+)$, since $\limu_+(r,s) = 1$ and $r$ is $\limu$-recurrent, we have $\limu(r,s) = 1$. Since $r$ is $\limu$-recurrent, its recurrence class is $r$ itself, so $s = r$. By induction hypothesis for $(\limv,\limv_+)$, since $\limv_+(r,q) = 1$ and $r$ is $\limv$-recurrent, we have $\limv(r,q) = 1$.
It follows that $(\limu \cdot \limv)(r,q) = 1$.
The case of iteration is easy.
The inclusion is strict, an example is given by figure \[fig:3\].
Leaktight automata strictly contain simple automata {#subsec:simple_automata}
---------------------------------------------------
The class of simple automata has been defined in [@CT12], where it was proved that the value $1$ problem is decidable for a subset of this class (namely, for structurally simple automata). In the following we show that the class of simple automata is strictly contained in the class of leaktight automata.
We fix $\AA$ a probabilistic automaton.
A jet is a sequence $(J_n)\nNN$ where $J_k\subseteq Q$ for each $k\in \NN$.
\[def:simple\_process\] Let $w \in A^\omega$ be an infinite word. The process induced by $w$ from the state $p$ is simple if there exists $\lambda > 0$ and two jets $(A_n)\nNN$, $(B_n)\nNN$ such that:
1. for all $k\in \NN$, $A_k$ and $B_k$ are disjoint and $A_k \cup B_k = Q$,
2. for all $k \in \NN$ and all $q \in A_k$, $\pRob{\AA}(p \xrightarrow{w_{< k}} q) \ge \lambda$,
3. $\lim_{n \to \infty} \pRob{\AA} (p \xrightarrow{w_{< n}} B_n) = 0$.
\[def:simple\] $\AA$ is simple if for every infinite word $w$ and every state $p$ the process induced by $w$ from $p$ is simple.
\[theo:simple\_inclusion\_leaktight\] Every simple automaton is leaktight.
The remainder of this subsection is devoted to the proof of Theorem \[theo:simple\_inclusion\_leaktight\]. The proof is divided into two parts: first, we define non-simplicity witnesses, which are elements of the Markov monoid that witnesses the non-simplicity of an automaton, and second we show that if the Markov monoid of an automaton contains a leak, then it also contains a non-simplicity witness.
### Non-simplicity witness
\[def:nonsimp\] A triple $(\limu,\limv,\limw)$ of elements of the Markov monoid is a *non-simplicity* witness if there exist states $r,t$ such that:
1. $\limu \limv^\sharp \limw$ is idempotent,
2. r is $\limu \limv^\sharp \limw$-recurrent,
3. $\limu \limv(r,t) = 1$,
4. t is $\limv$-transient.
\[prop:non\_simplicity\_witness\_implies\_not\_simple\] If the Markov monoid of a probabilistic automaton contains a non-simplicity witness, then it is not simple.
To prove Proposition \[prop:non\_simplicity\_witness\_implies\_not\_simple\], we rely on the following two lemmata:
\[lem:rec\] Let $\limu$ be an idempotent element of the Markov monoid, $r$ be a state $\limu$-recurrent, and $(u_n)\nNN$ a sequence of words that reifies $\limu$. Then there exists a constant $\gamma > 0$ and a strictly increasing map $h : \NN \to \NN$ such that for all $n \in \NN$, we have: $$\pRob{\AA}(r \xrightarrow{u_{h(0)} \cdots u_{h(n-1)}} r) \ge \gamma.$$
Since the sequence of words $(u_n)\nNN$ reifies the limit-word $\limu$, for all states $s,t$, $\pRob{\AA}(s \xrightarrow{u_n} t)$ converges and: $$\label{eq:lim}
\limu(s,t) = 1 \iff \lim_n \pRob{\AA}(s \xrightarrow{u_n} t) > 0.$$
Define: $$\lambda = \frac{1}{2} \cdot \min \set{\lim_n \pRob{\AA}(s \xrightarrow{u_n} t) \mid \limu(s,t) = 1} .$$ Thanks to , there exists an increasing map $h : \NN \to \NN$ such that the following two conditions hold, for all states $s,t$ and for all $n \in \NN$: $$\label{eq:ln}
\text{ if } \limu(s,t) = 0, \text{ then }
\pRob{\AA}(s \xrightarrow{u_{h(n)}} t) \le \frac{1}{|Q| \cdot 2^{n+2}} ,$$ $$\label{eq:bound}
\text{ if } \limu(s,t) = 1, \text{ then }
\pRob{\AA}(s \xrightarrow{u_{h(n)}} t) \ge \lambda .$$
We now use and to prove the desired result, for $\gamma = \frac{\lambda}{2}$.
Let $w_n = u_{h(0)} u_{h(1)} \cdots u_{h(n-1)}$. Denote the $\limu$-recurrence class of $r$ by $R = \set{q \in Q \mid \limu(r,q) = 1}$. We first bound the quantity $\pRob{\AA}(r \xrightarrow{w_n} Q \setminus R)$. Note that for $q \in R$ and $t$ a state, the following holds: if $\limu(q,t) = 1$ then $t \in R$, *i.e.* $R$ is not left while following transitions consistent with $\limu$. It follows that the probability to leave $R$ from $r$ while reading $w_n$ is smaller than $\sum_{k = 0}^{n-1} \pRob{\AA}(R \xrightarrow{u_{h(k)}} Q \setminus R)$, which is smaller than $\frac{1}{2}$ by . Thus $\pRob{\AA}(r \xrightarrow{w_{n-1}} R) \ge \frac{1}{2}$. Now, since $r$ is $\limu$-recurrent, and $\limu$ is idempotent, for all $q \in R$ we have $\limu(q,r) = 1$, so using we get that $\pRob{\AA}(q \xrightarrow{u_{h(n-1)}} r) \ge \lambda$. It follows that $\pRob{\AA}(r \xrightarrow{w_n} r) \geq \frac{\lambda}{2}$, which concludes.
\[lem:nonsimp\] Let $w \in A^\omega$ be an infinite word. If there exist states $p,s,t$, $(i_n)\nNN$, $(j_n)\nNN$ and $\gamma > 0$ such that:
1. for all $n \in \NN$, $\pRob{\AA}(p \xrightarrow{w_{< i_n}} s) \ge \gamma$,
2. for all $n \in \NN$, $i_n < j_n$ and $\pRob{\AA}(s \xrightarrow{w[i_n,j_n]} t) > 0$,
3. $\lim_{n \to \infty} \pRob{\AA}(p \xrightarrow{w_{< j_n}} t) = 0$,
then the process induced by $w$ from $p$ is not simple.
Assume towards contradiction that $w$ induces a simple process from $p$ with bound $\lambda$. We first argue that for infinitely many $n \in \NN$, we have $s \in A_{i_n}$ and $t \in B_{j_n}$. Indeed, if this is not the case, then for $n \in \NN$ large enough either $s \notin A_{i_n}$ or $t \notin B_{j_n}$, so either for infinitely many $n \in \NN$ we have $s \notin A_{i_n}$, or for infinitely many $n \in \NN$ we have $t \notin B_{j_n}$. The first case is contradicted by (1), the second case by (3).
Let $n \in \NN$ such that $s \in A_{i_n}$ and $t \in B_{j_n}$, since $\pRob{\AA}(s \xrightarrow{w[i_n,j_n]} t) > 0$, along a path from $s$ to $t$ there is a transition from the jet $A$ to the jet $B$. Formally, there exists $k_n$ such that $i_n \le k_n < j_n$ and the $k_n$^th^ transition goes from $q_{k_n} \in A_{k_n}$ to $q_{k_n + 1} \in B_{k_n + 1}$. This transition is a one-step transition in the automaton $\AA$; denote by $p_{\min}$ the minimal non-zero probabilistic transition in $\AA$, we have $\pRob{\AA}(q_{k_n} \xrightarrow{w[k_n,k_n+1]} q_{k_n + 1}) \ge p_{\min}$. Now, consider $\pRob{\AA}(p \xrightarrow{w_{\leq k_n}} q_{k_n + 1})$; since $q_{k_n} \in A_{k_n}$, we have $\pRob{\AA}(p \xrightarrow{w_{< k_n}} q_{k_n}) \ge \lambda$, so $\pRob{\AA}(p \xrightarrow{w_{\leq k_n}} q_{k_n + 1}) \ge \lambda \cdot p_{\min}$. But for infinitely many $n \in \NN$, $q_{k_n + 1} \in B_{k_n + 1}$, contradicting $\lim_{n \to \infty} \pRob{\AA}(p \xrightarrow{w_{< n}} B_n) = 0$.
Now we are ready to prove Proposition \[prop:non\_simplicity\_witness\_implies\_not\_simple\].
Let $(\limu,\limv,\limw)$ be a non-simplicity witness, and $r,t\in Q$ such that $r$ is $\limu \limv^\sharp \limw$-recurrent, $\limu \limv(r,t) = 1$ and $t$ is $\limv$-transient. Let $(u_n)\nNN,(v_n)\nNN,(w_n)\nNN$ be sequences of words which reify $\limu,\limv,\limw$ respectively.
Thanks to Proposition \[prop:consistency\], there exists a strictly increasing map $f : \NN \to \NN$ such that $(v_{f(n)}^n)\nNN$ reifies $\limv^\sharp$. Note that since $(v_{f(n)})\nNN$ is a subsequence of $(v_n)\nNN$, it also reifies $\limv$. Since $\limv$ is idempotent, for all $k \in \NN$, $(v_{f(n)}^k)\nNN$ reifies $\limv^k$. By assumption $\limu \limv(r,t) = 1$, so there exists $N_k \in \NN$ such that for all $n \geq N_k$, $\pRob{\AA}(r \xrightarrow{u_n v_{f(n)}^k} t) > 0$.
Let $g(n) = \max(n,N_n)$. Since $g$ is increasing, $(u_{g(n)} v_{f(g(n))}^{g(n)} w_{g(n)})\nNN$ is a subsequence of $(u_n v_{f(n)}^n w_n)\nNN$, so it reifies $\limu \limv^\sharp \limw$ as well. By definition of the function $g$, we have: $$\label{eq:2}
\text{for all } n \in \NN,\quad \pRob{\AA}(r \xrightarrow{u_{g(n)} v_{f(g(n))}^n} t) > 0.$$
We apply Lemma \[lem:rec\] to the limit-word $\limu \limv^\sharp \limw$, the state $r$ and the sequence of words $(u_{g(n)} v_{f(g(n))}^{g(n)} w_{g(n)})\nNN$, and obtain $h$ a strictly increasing map and a constant $\gamma > 0$.
Define the new sequence of words $(z_n)\nNN = (u_{g(h(n))} \cdot v_{f(g(h(n)))}^{g(h(n))} \cdot w_{g(h(n))})\nNN$, and $(x_n)\nNN = (u_{g(h(n))} \cdot v_{f(g(h(n)))}^{h(n)})\nNN$. We have: $$\label{eq:gama}
\text{ for all } n \in \NN,\quad \pRob{\AA}(r \xrightarrow{z_0 \cdots z_{n-1}} r) \ge \gamma.$$
Let $z = z_0 z_1 \cdots$. We argue that the conditions of Lemma \[lem:nonsimp\] are met:
1. for all $n \in \NN$,$\pRob{\AA}(r \xrightarrow{z_0 \cdots z_{n-1}} r) \ge \gamma$,
2. for all $n \in \NN$, $\pRob{\AA}(r \xrightarrow{x_n} t) > 0$,
3. $\lim_{n \to \infty} \pRob{\AA}(r \xrightarrow{z_0 \cdots z_{n-1} \cdot x_n} t) = 0$,
The item (1) is , the item (2) follows from , so we consider (3).
First, note that $(x_n)\nNN$ reifies $\limu \limv^\sharp$. Indeed, we first argue that $(v_{f(g(n))}^n)\nNN$ reifies $\limv^\sharp$: it follows from Proposition \[prop:consistency\], since $f \circ g : \NN \to \NN$ is a strictly increasing map satisfying $f \circ g \ge f$, and that $(v_{f(n)}^n)\nNN$ reifies $\limv^\sharp$. Now, $(v_{f(g(h(n)))}^{h(n)})\nNN$ is a subsequence of $(v_{f(g(n))}^n)\nNN$, so it reifies $\limv^\sharp$ as well, and it follows that $(x_n)\nNN$ reifies $\limu \limv^\sharp$.
Second, let $q \in Q$, since $t$ is $\limv$-transient, $\limu \limv^\sharp(q,t) = 0$. Since $(x_n)\nNN$ reifies $\limu \limv^\sharp$, we have $\lim_{n \to \infty} \pRob{\AA}(q \xrightarrow{x_n} t) = 0$. Now, $$\pRob{\AA}(r \xrightarrow{z_0 \cdots z_{n-1} \cdot x_n} t)
= \sum_{q \in Q} \pRob{\AA}(r \xrightarrow{z_0 \cdots z_{n-1}} q)
\cdot \pRob{\AA}(q \xrightarrow{x_n} t),$$ and for each term in the sum, the second factor converges to zero, so $\lim_{n \to \infty} \pRob{\AA}(r \xrightarrow{z_0 \cdots z_{n-1} \cdot x_n} t) = 0$.
Thus Lemma \[lem:nonsimp\] applies, and $z$ induces a non-simple process from $p$, so $\AA$ is not simple.
### The presence of a leak implies the presence of a non-simplicity witness
Now we show that the presence of a leak witness implies the presence of a non-simplicity witness.
\[prop:sw\] If the extended Markov monoid of a probabilistic automaton contains a leak witness, then it also contains a non-simplicity witness.
In the following proof, we will make use of the notion of $\sharp$-height for an extended limit-word. The $\sharp$-height of a limit-word was defined in Section \[def:sharp\_height\], the $\sharp$-height of an extended limit-word $(\limu,\limu_+)$ is the $\sharp$-height of $\limu$.
Let $\AA$ be a probabilistic automaton whose extended Markov monoid contains a leak witness. Consider the subset $\CC$ of extended limit-words $(\limu,\limu_+)$ in the extended Markov monoid such that there exist states $r,t$ satisfying:
1. $(\limu,\limu_+)$ is idempotent,
2. $r$ is $\limu$-recurrent,
3. $\limu(r,t) = 0$,
4. $\limu_+(r,t) = 1$.
Note that $\CC$ is non-empty since every leak witness is in $\CC$.
Consider an element $(\limz,\limz_+)$ in $\CC$ of minimal $\sharp$-height and let $r,t \in Q$ such that $r$ is $\limz$-recurrent, $\limz(r,t) = 0$ and $\limz_+(r,t) = 1$. In particular, we have $\limz \neq \limz_+$, so in any decomposition of $(\limz,\limz_+)$ into concatenation and iteration there must be at least one iteration. Consequently, $\limz = \limu \limv^\sharp \limw$ for some $\limu,\limv,\limw$, and $\limu \limv \limw$ has a strictly smaller $\sharp$-height than $\limz$.
Let $R$ be the $\limz$-recurrence class of $r$, and $T$ the set of $\limv$-transient states. We argue that the following holds: $$\label{eq:inside}
\text{there exist } r' \in R \text{ and } t' \in T \text{ such that } \limu \limv(r',t') = 1.$$ Assume towards contradiction that does not hold, *i.e.* for all $r' \in R$ and $t' \in T$, we have $\limu \limv(r',t') = 0$, then we prove that $(\limu \limv \limw, \limz_+)^{|\monoidext |!}$ is in $\CC$, contradicting the minimality of $(\limz, \limz_+)$ as it has strictly smaller $\sharp$-height.
First observe that for all states $q$ we have $\limu \limv(r',q) = \limu\limv^\sharp(r',q)$, which implies that for all $r' \in R$ and state $q$, we have: $$\label{eq:same}
\limu \limv \limw(r',q) = \limu \limv^\sharp \limw(r',q) = \limz(r',q).$$
We check that $(\limu \limv \limw, \limz_+)^{|\monoidext |!}$ is in $\CC$, with the states $r,t$ as witnesses:
1. $(\limu \limv \limw, \limz_+)^{|\monoidext |!}$ is idempotent, this follows from Lemma \[lem:basic\_limit\_words\].
2. $r$ is $(\limu \limv \limw)^{|\monoidext |!}$-recurrent. Indeed, let $q \in Q$ such that $(\limu \limv \limw)^{|\monoidext |!}(r,q) = 1$. It follows that $(\limu \limv^\sharp \limw)^{|\monoidext |!}(r,q) = 1$, but $(\limu \limv^\sharp \limw)^{|\monoidext |!} = \limz^{|\monoidext |!} = \limz$, so $\limz(r,q) = 1$. Since $r$ is $\limz$-recurrent, we have $\limz(q,r) = 1$ and $q \in R$, so implies that $\limz(q,r) = \limu \limv \limw(q,r)$, thus $\limu \limv \limw(q,r) = 1$. Also, $\limz(r,r) = \limu \limv \limw(r,r)$, so $\limu \limv \limw(r,r) = 1$, and altogether $(\limu \limv \limw)^{|\monoidext |!}(q,r) = 1$.
3. $(\limu \limv \limw)^{|\monoidext |!}(r,t) = 0$. Indeed, assume towards contradiction that $(\limu \limv \limw)^{|\monoidext |!}(r,t) = 1$. Then there exist $q_0,q_1,\ldots,q_{|\monoidext |!}$ such that $q_0 = r$, for $i \in \set{0,\ldots,|\monoidext |!-1}$ we have $\limu \limv \limw(q_i,q_{i+1}) = 1$ and $q_{|\monoidext |!} = t$. We prove by induction on $i \in \set{0,\ldots,|\monoidext |!-1}$ that $q_i$ is in $R$ and that $\limz(q_i,q_{i+1}) = 1$. Assume $q_i$ is in $R$, then by , $\limu \limv \limw(q_i,q_{i+1}) = \limz(q_i,q_{i+1})$, so $\limz(q_i,q_{i+1}) = 1$. But $q_i$ is in $R$, which is a recurrence class for $\limz$, so $q_{i+1}$ is in $R$ as well, concluding the induction. Thus, we have $\limz^{|\monoidext |!}(r,t) = 1$, and since $\limz$ is idempotent $\limz(r,t) = 1$, a contradiction.
4. $\limz_+^{|\monoidext |!}(r,t) = 1$. It follows from the fact the $\limz_+$ is idempotent, and that $\limz_+(r,t) = 1$.
We reached a contradiction, since $(\limu \limv \limw, \limz_+)^{|\monoidext |!}$ has a strictly smaller $\sharp$-height than $(\limz, \limz_+)$. It follows from that $(\limu,\limv,\limw)$ is a non-simplicity witness for the states $r'$ and $t'$, concluding the proof.
The proof of Theorem \[theo:simple\_inclusion\_leaktight\] is now a simple combination of Proposition \[prop:non\_simplicity\_witness\_implies\_not\_simple\] and Proposition \[prop:sw\]. The inclusion is strict: figure \[fig:3\] provides an example of a leaktight automaton which is not simple.
Probabilistic omega-automata {#sec:infinite}
============================
In this section, we relate the value $1$ problem for probabilistic automata over finite words and over infinite words, as introduced and studied in [@BBG12]. We state a general theorem, showing the equivalence of the value $1$ problem for automata over finite words and the value $1$ problem for automata over infinite words with the parity condition. This theorem allows to extend the decidability results from finite words to infinite ones.
1em For the definitions of probabilistic automata over infinite words, we refer to [@BBG12]. For the sake of readability, we introduce two notations in this section. First, we denote by $\pRob{\AA,w}(E)$ the probability of the measurable event $E$ when reading the infinite word $w$ on $\AA$. Second, we denote by $\pRob{\AA}^{\delta}(u)$ the probability that the finite word $u$ is accepted by $\AA$ with $\delta$ as initial distribution. Note that in our definition, probabilistic automata have a unique initial state; however here we need to deal with automata having general initial distributions, so we sometimes consider the more general tuples $\AA = (Q, \delta, \Delta, F)$ where $\delta$ is the initial probability distribution.
\[theo:infinite\_words\] Let $\AA = (Q,\delta_0,\Delta,c)$ be a probabilistic parity automaton where $c : Q \rightarrow \NN$ is a priority function. Consider the deterministic transducer $\MM$ over $\AA$, which keeps track of the minimal priority seen: $\MM = (c(Q), Q^\AA, \Delta_\MM)$ where $\Delta_\MM(q,c) = \min(c,c(q))$.
The automaton $\AA$ over infinite words has value $1$ if and only if there exists $R \subseteq Q$, such that the two following probabilistic automata over finite words have value $1$:
- The automaton $\AA(R)$ with $R$ as set of final states;
- The automaton $\AA \otimes \MM$ with the uniform distribution over $R_c = \set{(q,c(q)) \mid q \in R}$ as initial distribution and $\set{(q,e) \mid q \in R \textrm{ and } e \textrm{ even}}$ as set of final states.
Before giving the proof, we need two lemma.
\[lem:technical\] Let $\AA = (Q,\delta,\Delta,F)$ be a probabilistic automaton, $u$ a word and $\varepsilon > 0$. Let $\mu = \min \set{\delta(q) \mid q \in \supp(\delta)}$. If $\pRob{\AA}^{\delta}(u) \ge 1 - \varepsilon \cdot \mu$, then for all $q \in \supp(\delta)$, $\pRob{\AA}^q(u) \ge 1 - \varepsilon$.
$$\pRob{\AA}^{\delta}(u) = \sum_{q \in \supp(\delta)} \delta(q) \cdot \pRob{\AA}^q(u) \ \ge \ 1 - \varepsilon \cdot \mu.$$ Let $q \in \supp(\delta)$, since the probabilities are bounded by $1$: $$\delta(q) \cdot \pRob{\AA}^q(u)\ + \ \underbrace{\sum_{p \in \supp(\delta),\ p \neq q} \delta(p)}_{1 - \delta(q)} \ \ge \
1 - \varepsilon \cdot \mu.$$ Hence: $$\pRob{\AA}^q(u)\ \ge \ 1 - \varepsilon \cdot \frac{\mu}{\delta(q)} \ \ge \ 1 - \varepsilon.$$
\[cor:value\_1\_independence\] Let $\AA = (Q,\delta,\Delta,F)$ be a probabilistic automaton. If $\AA$ has value $1$, then:
- For all distributions $\delta'$ such that $\supp(\delta') \subseteq \supp(\delta)$, it has value $1$ with $\delta'$ as initial distribution.
- For all distributions $\delta'$ such that $\sum_{q \in \supp(\delta)} \delta'(q) \ge \theta$, it has value at least $\theta$ with $\delta'$ as initial distribution.
Assume $\AA$ has value $1$. Let $\varepsilon > 0$, then there exists a word $u$ such that $\pRob{\AA}(u) \geq 1 - \varepsilon \cdot \mu$. Thanks to Lemma \[lem:technical\], this implies that for all $q \in \supp(\delta)$, we have $\pRob{\AA}^q(u) \geq 1 - \varepsilon$.
Now for $\delta'$: $$\pRob{\AA}^{\delta'}(u)\ = \ \sum_{q \in Q} \delta'(q) \cdot \pRob{\AA}^q(u) \ \ge \
\sum_{q \in \supp(\delta)} \delta'(q) \cdot \pRob{\AA}^q(u) \ \ge \
\left(\sum_{q \in \supp(\delta)} \delta'(q)\right) \cdot (1 - \varepsilon).$$ For the first item, note that if $\supp(\delta') \subseteq \supp(\delta)$, then $\sum_{q \in \supp(\delta)} \delta'(q) = 1$, so $\pRob{\AA}^{\delta'}(u) \geq 1 - \varepsilon$, so $\AA$ has value $1$ with $\delta'$ as initial distribution.
The second item follows from the last inequality, implying $\pRob{\AA}^{\delta'}(u) \geq \theta \cdot (1 - \varepsilon)$, so $\AA$ has value at least $\theta$ with $\delta'$ as initial distribution.
\[lem:seq\] For all $\varepsilon > 0$, there exists a sequence $(\varepsilon_k)_{k \ge 0}$ satisfying:
1. For all $k \ge 0$, we have $0 < \varepsilon_k < 1$;
2. $\prod_{k \ge 0} \varepsilon_k \geq 1 - \varepsilon$;
3. For all $k \ge 0$, we have $\prod_{p \le k} \varepsilon_p > \varepsilon_{k+1}$.
Define $\nu_k = 1 - \frac{1}{2^{2^k}}$, which clearly satisfies 1. It also satisfies 3.: $$\begin{array}{lll}
\prod_{p \le k} \nu_p & = & \prod_{p \le k} \frac{2^{2^p} - 1}{2^{2^p}} \\[1ex]
& = & \frac{\prod_{p \le k} 2^{2^p} - 1}{2^{2^{k+1} - 1}}
\end{array}$$ The inequality $\prod_{p \le k} \nu_p > \nu_{k+1}$ is equivalent to: $$2 \cdot \left(\prod_{p \le k} (2^{2^p} - 1)\right) < 2^{2^{k+1}} - 1,$$ which is easily proved by induction, since $(2^{2^{k+1}} - 1)^2 < 2^{2^{k+2}} - 1$.
The infinite product $\prod_{k \ge 0} \nu_k$ has a value $\nu \approx 0.350184$, in particular $0 < \nu < 1$. Let $\varepsilon > 0$, then for $\lambda = \frac{\ln(1-\varepsilon)}{\ln(\nu)}$, which satisfies $\lambda > 0$, the sequence $(\nu_k^\lambda)_{k \in \NN}$ satisfies the three conditions.
We are now fully equipped for the proof of Theorem \[theo:infinite\_words\] 1em
We begin with the left-to-right direction. Let $(w_n)_{n \in \NN}$ be a sequence of infinite words such that $\pRob{\AA,w_n}(\parc)$ converges to $1$. For each $n$, denote by $(\delta_k^n)_{k \in \NN}$ the sequence of distributions assumed by $\AA$ when reading $w_n$. Since there are finitely many possible supports, there exists a subsequence where all distributions have the same support. Since $[0,1]^Q$ is a compact space, we can extract from this subsequence another one, denoted by $(\delta_{\phi_n(k)}^n)_{k \in \NN}$, which converges to a distribution denoted by $\delta^n$. By the same compactness argument, at the expense of considering a subsequence of $(w_n)_{n \in \NN}$, we can assume that $(\delta^n)_{n \in \NN}$ converges to a distribution $\delta$, whose support is denoted by $R$.
We now prove that $\AA(R)$ as well as $\AA \otimes \MM$ have value $1$.
**The automaton $\AA(R)$.** Let $\varepsilon > 0$. Since $(\delta^n)_{n \in \NN}$ converges to $\delta$, for $n$ large enough, we have $||\delta^n - \delta||_\infty \le \frac{\varepsilon}{2 \cdot |R|}$. We fix $n$ large enough; since $(\delta_{\phi_n(k)}^n)_{k \in \NN}$ converges to $\delta^n$, for $k$ large enough, we have $||\delta_{\phi_n(k)}^n - \delta^n||_\infty \le \frac{\varepsilon}{2 \cdot |R|}$. We fix such a large $k$, altogether this implies $||\delta_{\phi_n(k)}^n - \delta||_\infty \le \frac{\varepsilon}{|R|}$.
Now consider $u$ the prefix of $w_n$ until position $\phi_n(k)$. We have: $$\pRob{\AA(R)}(u) \ =\ \sum_{q \in R} \delta_{\phi_n(k)}^n(q) \ \ge \
\sum_{q \in R} \left(\delta(q) - \frac{\varepsilon}{|R|}\right) \ = \ 1 - \varepsilon.$$ It follows that $\AA(R)$ has value $1$.
**The automaton $\AA \otimes \MM$.** Let $\nu = \min \set{\delta(q) \mid q \in \supp(\delta)}$, clearly $0 < \nu < 1$. Since $(\delta^n)_{n \in \NN}$ converges to $\delta$, we can assume that for all $n \ge 0$, $||\delta^n - \delta||_\infty \le \frac{\nu}{4}$, considering the sequence $(w_n)_{n \in \NN}$ from some index on. For each $n$, the sequence $(\delta_{\phi_n(k)}^n)_{k \in \NN}$ converges to $\delta^n$, so again by considering the sequence from some index on, we can assume that for all $k \ge 0$, $||\delta_{\phi_n(k)}^n - \delta^n||_\infty \le \frac{\nu}{4}$, hence $||\delta_{\phi_n(k)}^n - \delta||_\infty \le \frac{\nu}{2}$. As a result, for all $q \in R$, we have $\delta_{\phi_n(k)}^n(q) \ge \frac{\nu}{2} > 0$, so $R = \supp(\delta) \subseteq \supp(\delta^n) \subseteq \supp(\delta_{\phi_n(k)}^n)$.
1em Let $\varepsilon > 0$, and $n$ such that $\pRob{\AA,w_n}(\parc) \ge 1 - \varepsilon \cdot \frac{\nu}{4}$.
Let $\buchi(d)$ be the set of runs where the color $d$ is reached infinitely often and by $\cobuchi(<d)$ the set of runs where colors less than $d$ are reached only finitely often. Observe that: $$\parc\ =\ \biguplus_{d \textrm{ even color}} \buchi(d) \cap \cobuchi(<d).$$
Let $\safe(<d,k)$ be the set of runs where colors less than $d$ are not reached anymore after the position $\phi_n(k)$. Observe that: $$\buchi(d) \cap \cobuchi(<d)\ \subseteq\ \buchi(d) \cap \bigcup_{k \ge 0} \safe(<d,k),$$ so: $$\pRob{\AA,w_n}\left(\buchi(d) \cap \cobuchi(<d)\right)\ \leq\
\pRob{\AA,w_n}\left(\buchi(d) \cap \bigcup_{k \ge 0} \safe(<d,k)\right).$$ Since the sequence $\buchi(d) \cap \left(\safe(<d,k)\right)_{k \ge 0}$ is increasing with respect to set inclusion, for some $k$ we have: $$\pRob{\AA,w_n}(\buchi(d) \cap \safe(<d,k))\ \ge \
\pRob{\AA,w_n}(\buchi(d) \cap \cobuchi(<d)) - \varepsilon \cdot \frac{\nu}{4}.$$ Note that here, $k$ depends on $d$; however, since if it holds for $k$ then it holds for any bigger $k$ and that there are finitely many even colors $d$, we can assume that $k$ is uniform over all even colors $d$.
Now, let $\reach(d,k,k')$ be the set of runs where the color $d$ is reached between the positions $\phi_n(k)$ and $\phi_n(k')$. Observe that: $$\buchi(d) \cap \safe(<d,k)\ \subseteq\ \bigcup_{k' > k} \reach(d,k,k') \cap \safe(<d,k),$$ so: $$\pRob{\AA,w_n}(\buchi(d) \cap \safe(<d,k))\ \leq\
\pRob{\AA,w_n}\left(\bigcup_{k' > k} \reach(d,k,k') \cap \safe(<d,k)\right).$$ Since the sequence $\left(\reach(d,k,k') \cap \safe(<d,k)\right)_{k' > k}$ is increasing with respect to set inclusion, for some $k'$ we have: $$\pRob{\AA,w_n}(\reach(d,k,k') \cap \safe(<d,k))\ \ge \
\pRob{\AA,w_n}\left(\buchi(d) \cap \safe(<d,k)\right) - \varepsilon \cdot \frac{\nu}{4}.$$ Here again, $k'$ depends on $d$; however with the same reasoning as above, we can assume that $k'$ is uniform over all even colors $d$.
Altogether, this implies: $$\pRob{\AA,w_n}(\reach(d,k,k') \cap \safe(<d,k))\ \ge \
\pRob{\AA,w_n}(\buchi(d) \cap \cobuchi(<d)) - \varepsilon \cdot \frac{\nu}{2},$$ and summing these equalities for each even color $d$: $$\begin{array}{lcl}
\pRob{\AA,w_n}(\parity(c,k,k')) & \ge & \sum_{d \textrm{ even color}} \pRob{\AA,w_n}(\reach(d,k,k') \cap \safe(<d,k))\\[.6em]
& \ge & \pRob{\AA,w_n}(\parc) - \varepsilon \cdot \frac{\nu}{2},
\end{array}$$ where $\parity(c,k,k')$ is the set of runs where the minimal color seen between the positions $\phi_n(k)$ and $\phi_n(k')$ is even.
Let $v$ be the infix of $w_n$ between the positions $\phi_n(k)$ and $\phi_n(k')$. The distribution $\delta_{\phi_n(k)}^n$ is over the set of states $Q$; we embed it as $\widehat{\delta}$ over the set of states $Q \times c(Q)$, setting $\widehat{\delta}(q,c(q)) = \delta_{\phi_n(k)}^n(q)$ and $0$ otherwise. By construction of $v$, we have $\pRob{\AA \otimes \MM}^{\widehat{\delta}}(v) =
\pRob{\AA,w_n}(\parity(c,k,k')) \ge 1 - \varepsilon \cdot \frac{\nu}{2}$. However, the initial distribution of $\AA \otimes \MM$ is uniform over $R_c$. We apply Lemma \[lem:technical\]; by construction $\mu \ge \frac{\nu}{2}$, so $\pRob{\AA \otimes \MM}^{\widehat{\delta}}(v) \ge 1 - \varepsilon \cdot \mu$. It follows that for all $q \in \supp(\delta_{\phi_n(0)}^n)$, $\pRob{\AA \otimes \MM}^{q,c(q)}(v) \ge 1 - \varepsilon$. Since $R \subseteq \supp(\delta_{\phi_n(k)}^n)$, we have $\pRob{\AA \otimes \MM}(v) \ge 1 - \varepsilon$, hence $\AA \otimes \MM$ has value $1$.
2em We now prove the right-to-left direction. Let $R$ such that both $\AA(R)$ and $\AA \otimes \MM$ have value $1$. Let $\varepsilon > 0$, consider the sequence $(\varepsilon_k)_{k \ge 0}$ given by Lemma \[lem:seq\].
Since $\AA(R)$ has value $1$, there exists $u$ such that $\pRob{\AA(R)}(u) \ge \varepsilon_0$. We show the existence of a sequence $w_1,w_2,\ldots$ of words such that for all $k \ge 0$: $$\pRob{\AA(R)}(u \cdot w_1 \cdot w_2 \cdot \ldots \cdot w_k) \ge \prod_{p \le k} \varepsilon_p
\qquad \mathrm{ and } \qquad \pRob{\AA \otimes \MM}^{\delta'_k}(w_{k+1}) \ge \varepsilon_{k+1},$$ where $\delta_k$ is the distribution obtained by reading $u \cdot w_1 \cdot w_2 \cdot \ldots \cdot w_k$ on $\AA$; the distribution $\delta_k$ is over the set of states $Q$, we embed it as $\delta'_k$ over the set of states $Q \times c(Q)$, setting $\delta'_k(q,c(q)) = \delta_k(q)$ and $0$ otherwise.
We proceed inductively; assume $w_1,\ldots,w_k$ have been chosen. Note that by construction, $\sum_{q \in R} \delta_k(q) \ge \prod_{p \le k} \varepsilon_p$. However, the support of $\delta_k$ is not included in $R$, so the first item of Corollary \[cor:value\_1\_independence\] does not apply. Then $\sum_{q \in R_c} \delta'_k(q) \ge \prod_{p \le k} \varepsilon_p$.
Since $\AA \otimes \MM$ has value $1$, thanks to the second item of Corollary \[cor:value\_1\_independence\], it has value at least $\prod_{p \le k} \varepsilon_p$ for $\delta'_k$ as initial distribution. Together with 3., this implies that there exists $w_{k+1}$ such that $\pRob{\AA \otimes \MM}^{\delta'_k}(w_{k+1}) \ge \varepsilon_{k+1}$.
We have: $$\begin{array}{lll}
\pRob{\AA(R)}(u \cdot w_1 \cdot w_2 \cdot \ldots \cdot w_{k+1}) & \ge &
\pRob{\AA(R)}(u \cdot w_1 \cdot w_2 \cdot \ldots \cdot w_k) \cdot
\pRob{\AA(R)}^{\delta_k}(w_{k+1}) \\
& \ge & \pRob{\AA(R)}(u \cdot w_1 \cdot w_2 \cdot \ldots \cdot w_k) \cdot
\pRob{\AA \otimes \MM}^{\delta'_k}(w_{k+1}) \\
& \ge & \left(\prod_{p \le k} \varepsilon_p\right) \ \cdot \ \varepsilon_{k+1}\\
& = & \prod_{p \le k+1} \varepsilon_p,\\
\end{array}$$ which concludes the inductive construction.
Let $w = u \cdot w_1 \cdot w_2 \cdot \ldots$. We evaluate $\pRob{\AA,w}(\parc)$: $$\pRob{\AA,w}(\parc) \ \ge \ \pRob{\AA(R)}(u) \cdot \prod_{k \ge 0} \pRob{\AA \otimes \MM}^{\delta'_k}(w_{k+1})
\ \ge \ \varepsilon_0 \cdot \prod_{k \ge 1} \varepsilon_k
\ = \ \prod_{k \ge 0} \varepsilon_k \ \ge \
1 - \varepsilon.$$ It follows that $\AA$ has value $1$ as a probabilistic parity automaton, which concludes the proof of Theorem \[theo:infinite\_words\].
The value $1$ problem for leaktight automata over infinite words with the parity condition is decidable and $\PSPACE$-complete.
Indeed, observe that if $\AA$ is leaktight, then by Proposition \[prop:deterministic\_transducer\], so is $\AA \otimes \MM$.
Acknowledgment {#acknowledgment .unnumbered}
==============
We thank Thomas Colcombet for having pointed us to the work of Leung and Simon, and the anonymous reviewers for their constructive comments.
Conclusion {#conclusion .unnumbered}
==========
We introduced a subclass of probabilistic automata, called leaktight automata, for which we proved that the value $1$ problem is $\PSPACE$-complete. This subclass generalizes all subclasses of probabilistic automata whose value $1$ problem is known to be decidable.
A challenging perspective is now to find subclasses of partially observable Markov decision processes where the value 1 problem is decidable (some preliminary results were given in [@GO14]), and to extend our results to the setting of partially observable stochastic games, which is even more challenging.
[^1]: This project was supported by the french ANR project “Stoch-MC” as well as “LaBEX CPU” of Université de Bordeaux.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'SPT-3G is a detector system for the 10m diameter South Pole Telescope, comprising 16,000 millimeter-wave bolometers. It is used for a deep Cosmic Microwave Background survey of the Southern sky. This paper describes the geometry of the secondary mirror, which is a section of a prolate spheroid, in several useful coordinate systems. There is application to off-axis mirrors in general. A geometric theorem is proven, relating to the Dragone condition: the intersection of a prolate spheroid and any plane is an ellipse; the lines connecting points on that ellipse to either focus compose a right circular cone; the central axes of the two cones from the two foci intersect outside the interior of the spheroid.'
author:
- 'Antony A. Stark'
title: ' SPT-3G secondary mirror geometry'
---
This paper describes the geometry of the SPT-3G [@benson14; @pan18] offset gregorian secondary mirror on the South Pole Telescope [@carlstrom11], with respect to a global coordinate system centered on the primary mirror, and also the shape of the mirror in two coordinate systems local to the mirror itself (one referenced to the central ray, and one referenced to the edge of the mirror). This mirror has replaced the secondary mirror used in the first two generations of SPT receivers.
The telescope is oriented pointed at the horizon, toward the $+\hat x$ direction, so that the light from the star is moving in the $- x$ direction when intercepted by the primary mirror. The $\hat y$ axis points up, and the $\hat z$ axis is oriented according to the right-hand rule ($\hat x$ points right, $\hat y$ points up, and $\hat z$ points out of the page). The origin is at the vertex of the primary mirror. This coordinate system is the one used by Vertex. It differs from the one usually used in optics, by a rotation of $90{{^{\circ}}}$ about the $\hat y$ axis: $$x_{\mathrm{zemax}} = z, \,
y_{\mathrm{zemax}} = y, \,
z_{\mathrm{zemax}} = -x \, .$$
While using this memo, it may be useful to consult Nils Halverson’s Optics Dimensional Control drawing: , and A. Stark’s program and its output, which can be found under .
The Central Ray
===============
The surface of the primary is defined by: $$x = {{y^2 + z^2}\over{ 4\, f_{\mathrm{p}}}} \,$$ where $f_{\mathrm{p}} = 7000 \, \mathrm{mm}$ is the focal length of the primary mirror. Let $y_\mathrm{c} = 5300 \, \mathrm{mm}$. The central ray starts at $(+\infty, \, y_\mathrm{c}, \, 0)$, and intercepts the primary mirror at $$(y_\mathrm{c}^2 / 4 f_\mathrm{p}, \, y_\mathrm{c}, \, 0) =
(1003.214286 \, \mathrm{mm}, \, 5300 \, \mathrm{mm}, \, 0)\, ,$$ where it is reflected through a half-angle $$i_{\mathrm{p}} =
\mathrm{atan}({{y_\mathrm{c}}/{2 f_{\mathrm{p}}}}) =
20.735234{{^{\circ}}}\, ,$$ and passes through the prime focus at $$\vec{F_1} = (f_\mathrm{p}, \, 0, \, 0) =
(7000 \, \mathrm{mm}, \, 0, \, 0) \, .$$ Between the primary and the secondary, the central ray satisfies the equation: $$y = \alpha (x - f_\mathrm{p}) \,,$$ where $$\alpha \equiv {{4 f_\mathrm{p} y_\mathrm{c}}\over
{y_\mathrm{c}^2 - 4 {f_{\mathrm{p}}^2}}} =
- \mathrm{tan}( 2 i_{\mathrm{p}})
= -0.8838068
\,,$$ is the tangent of the angle from the $+\hat x$ axis to the central ray between the primary and the secondary mirrors.
The SPT3G secondary is a piece of a prolate spheroid. Let $F_1$ be the focal point at the prime focus of the primary, $F_2$ be the focus after reflection (the Gregorian focus), and $Q$ be the point where the central ray strikes the surface and is reflected. In the drawing [*Optics Dimensional Control*]{} 6/18/14 by Nils Halverson, the spheroid is defined by a semi-major axis $a \, = \, 1745 \, \mathrm{mm}$ and semi-minor axis $b \, = \, 1644.556\, \mathrm{mm}$. This immediately gives the eccentricity $e \, = \, \sqrt{1 \, - \, (b/a)^2}
\, = \, 0.334378$, conic constant $k \, = \, -e^2 \, = \, -0.111809$, and end-cap radius of curvature $R \, = \, a(1+k) \, = \, 1549.89366 \, \mathrm{mm}$. The focal distance $f_0 \, = \, \sqrt{a^2 \, - \, b^2} \, = \, ae \, = \,
583.4896 \, \mathrm{mm}$ is half the length of the line segment $F_1 F_2$, and the vertex distance $f_s \, = \, a - f_0 \, = \, 1161.5097 \, \mathrm{mm}$ is the length from vertex to focus. The surface can be expressed as: $${{(x''+ f_0)^2}\over{a^2}}
+ {{{y''}^2}\over{b^2}}
+ {{{z''}^2}\over{b^2}} = 1 \, .$$ The $(x'',\, y'', \, z'')$ coordinate system is defined with its origin at the $F_1$ focus of the spheroid and the axis of the spheroid along the $\hat x''$ direction. As shown in the [*Optics Dimensional Control*]{} diagram, the secondary can be put in place by rotating by angle $\theta_\mathrm{s} = 15.323{{^{\circ}}}$ around the $\hat z''$ axis then translating in $\hat x$ by a distance $f_{\mathrm{p}} = 7000 \, \mathrm{mm}$ to place the $F_1$ focus coincident with the prime focus: $$\begin{pmatrix}
x \\ y \\ z
\end {pmatrix}
=
\begin{pmatrix}
{ \cos\,} \theta_\mathrm{s} &{\rm -sin\,} \theta_\mathrm{s}&0\\
{\rm sin\,} \theta_\mathrm{s}&{\rm cos\,} \theta_\mathrm{s}&0\\
0&0&1
\end {pmatrix}
\begin{pmatrix}
x''\\ y''\\ z''
\end {pmatrix}
+
\begin{pmatrix}
f_\mathrm{p} \\ 0\\ 0
\end {pmatrix}
\, .$$ The inverse transform is: $$\begin{pmatrix}
x''\\ y''\\ z''
\end {pmatrix}
=
\begin{pmatrix}
{\rm cos\,} \theta_\mathrm{s} &{\rm sin\,} \theta_\mathrm{s}&0\\
{\rm -sin\,} \theta_\mathrm{s}&{\rm cos\,} \theta_\mathrm{s}&0\\
0&0&1
\end {pmatrix}
\begin{pmatrix}
x-f_\mathrm{p}\\ y\\ z
\end {pmatrix}
\,.$$ Substituting Equation 5 into Equation 3 yields an equation for the secondary in $(x, \, y, \, z)$: $${{([x-f_\mathrm{p}]\,{\mathrm{cos}}\,\theta_\mathrm{s}
+ y\, {\mathrm{sin}}\,\theta_\mathrm{s} + f_0)^2}\over{a^2}}
+
{{([x-f_\mathrm{p}]\, {\mathrm{sin}}\,\theta_\mathrm{s}
- y\, {\mathrm{cos}}\,\theta_\mathrm{s} )^2}\over{b^2}}
+ {{{z}^2}\over{b^2}} = 1 \, .$$ This equation can be solved for $y$ as a function of $x$ and $z$: $$y = {{-\beta\gamma-\eta\lambda-\sqrt{(\beta\gamma+\eta\lambda)^2-
(\beta^2+\eta^2)(\gamma^2+\lambda^2+\mu^2 -1)}}\over{\beta^2+\eta^2}} \, ,$$ where $$\begin{aligned}
\beta & = & \mathrm{sin}\,\theta_\mathrm{s}/a \, , \nonumber\\
\gamma & = & ([x - f_\mathrm{p}]\mathrm{cos}\,\theta_\mathrm{s}
+f_0)/a \, , \nonumber\\
\eta & = & -\mathrm{cos}\,\theta_\mathrm{s}/b \, , \nonumber\\
\lambda & = &([x - f_\mathrm{p}]\mathrm{sin}\,\theta_\mathrm{s})/b \, , \nonumber\\
\mu & = & z/b \, . \nonumber\end{aligned}$$ Substituting Equation 2 and $z = 0$ into Equation 6 allows us to solve for $x_\mathrm{sc}$, the $x$ coordinate of the intersection of the central ray with the secondary: $$x_\mathrm{sc} = {{-\rho\sigma+\sqrt{\rho^2\sigma^2-
(\rho^2+\tau^2)(\sigma^2-1)}}\over{\rho^2+\tau^2}} + f_\mathrm{p}
= 7981.579 \, \mathrm{mm}\, ,$$ where $$\begin{aligned}
\rho & = & (\mathrm{cos}\,\theta_\mathrm{s}
+\alpha\,\mathrm{sin}\,\theta_\mathrm{s})/a \, , \nonumber\\
\sigma & = & f_0/a \,= \, e , \nonumber\\
\tau & = & (\mathrm{sin}\,\theta_\mathrm{s}
-\alpha\,\mathrm{cos}\,\theta_\mathrm{s})/b \, . \nonumber\end{aligned}$$ Here we introduce a notation where a point in space is represented by a vector extending from the origin of the $(x, \, y, \, z)$ coordinate system to that point. Of course, these vectors can be represented in other coordinate systems, even though they depend on the $(x, \, y, \, z)$ coordinate system for their definition. The central ray intersects the secondary at the point: $$\vec{S}
= (x_\mathrm{sc} , \, \alpha [x_\mathrm{sc} - f_\mathrm{p}] , \, 0)
= (7981.579\,\mathrm{mm} , \, -867.52618,\mathrm{mm} , \, 0)
\, ,$$ and the Gregorian focus is the point: $$\vec{F_2} =
(f_\mathrm{p} - 2 f_0 \, \mathrm{cos} \, \theta_\mathrm{s} \, , \,
- 2 f_0 \, \mathrm{sin} \, \theta_\mathrm{s} \, , \,
0 )
= (5874.5032\,\mathrm{mm} , \, -308.38675 \,\mathrm{mm} , \, 0)
\, ,$$ in the $(x, \, y, \, z)$ coordinate system.
Dragone Relation
================
The distance from the Gregorian focus to point $S$ on the secondary is: $$f_2 =
\|\vec{F_2}-\vec{S}\| = 2180.00 \, \mathrm{mm}\, ,$$ while the distance from the prime focus to point $S$ on the secondary is: $$f_1 =
\|\vec{F_1}-\vec{S}\| = 1310.00\, \mathrm{mm}\, ,$$ and the ratio of these distances is the magnification of the secondary: $$M = -
{{
\|\vec{F_2}-\vec{S}\|
}\over{
\|\vec{F_1}-\vec{S}\|
}}
= - {{f_2}\over{f_1}} = -1.6641236\, ,$$ where by convention the magnification is negative for a Gregorian. The angle of incidence at the secondary is: $$i_\mathrm{s} = {{1}\over{2}}\,\mathrm{acos}
\left[{{(\vec{F_2}-\vec{S})\cdot
(\vec{F_1}-\vec{S})}}
\over{
{\|\vec{F_2}-\vec{S}\|\|\vec{F_1}-\vec{S}\|}}\right]
= 13.304411{{^{\circ}}}\, .$$ The Dragone angle [@dragone82] is then: $$\mathrm{tan}\,i_\mathrm{D} \equiv
(1-M)\mathrm{tan}\,i_\mathrm{s} +
M\mathrm{tan}\,i_\mathrm{p} = 0 \, .$$ We see that it is zero, since $\theta_\mathrm{s}$ was chosen to yield that result.
The Secondary Mirror in Local Coordinates around the Central Ray
================================================================
Define a new coordinate system $(x', y', z')$, where the origin is at $F_1$ like the $(x'', y'', z'')$ system, but rotated by angle $\theta_1$ about $\hat y''$ to place point $S$, the intersection of the central ray with the mirror surface, on the positive $\hat z'$ axis: $$\begin{pmatrix}
x'\\ y'\\ z'
\end{pmatrix}
=
\begin{pmatrix}
{\rm cos\,} \theta_1 & 0 &{\rm sin\,} \theta_1\\
0&1&0\\
{-{\rm sin\,}} \theta_1 & 0 &{\rm cos\,} \theta_1
\end{pmatrix}
\begin{pmatrix}
x'' \\ y'' \\ z''
\end{pmatrix}
\, ,$$ and the inverse transform is $$\begin{pmatrix}
x'' \\ y'' \\ z''
\end{pmatrix}
=
\begin{pmatrix}
{\rm cos\,} \theta_1 & 0 &- {\rm sin\,} \theta_1\\
0&1&0\\
{\rm sin\,} \theta_1 & 0 &{\rm cos\,} \theta_1
\end{pmatrix}
\begin{pmatrix}
x'\\ y'\\ z'
\end{pmatrix}
\, .$$ Substituting Equation 17 into Equation 3 yields the equation of the spheroid in the $(x', y', z')$ system: $${{(x'\,{\mathrm{cos}}\,\theta_1
- z'\, {\mathrm{sin}}\,\theta_1 + f_0)^2}\over{a^2}}
+ {{{y'}^2}\over{b^2}}
+
{{(x'\, {\mathrm{sin}}\,\theta_1
+ z'\, {\mathrm{cos}}\,\theta_1 )^2}\over{b^2}}
= 1 \, .$$ Setting $x' \, = \, y'\, = \, 0$, and solving the resulting quadratic for $z'$ gives the distance, $f_1$, between $F_1$ and $S$: $$f_1 = {{b^2}\over{a}}\,
(1\,+\, e\,{\mathrm{sin}\,\theta_1})^{-1} \, .$$ We see that $\theta_1$ is the complementry angle of the “true anomaly" of the point $S$. We can solve for $\theta_1$: $$\begin{aligned}
{\mathrm{sin}\,\theta_1} \, &=& \,
\left({{b^2}\over{a f_1}} \, - \, 1 \right) \,e^{-1} \, , \\
\theta_1 \, &=& \, 33.206532{{^{\circ}}}\,.\end{aligned}$$ Since $f_2$ is the distance from $Q$ to $F_2$, from the properties of ellipses we have $a \, = \, {{1}\over{2}}(f_1 \, + \, f_2)$, $b \, = \, \sqrt{{{1}\over{4}}(f_1+f_2)^2\,-\,f_0^2}$ , and we can express $\theta_1$ in terms of $f_0$, $f_1$, and $f_2$ only: $${\mathrm{sin}\,\theta_1} \, = \,
{{f_2^2\,-\,f_1^2\,-\,4f_0^2}\over{4 f_0 f_1}} \,.$$ This equation can also be derived by applying the law of cosines to the triangle $F_1 C F_2$.
In the $(x', y', z')$ system, the coordinates of point $Q$ are $(0, 0, f_1)$. Substituting into Equation 17, the coordinates of $Q$ in $(x'', y'', z'') = (- \, f_1 {\mathrm{sin}\,\theta_1} , 0,
{f_1 \mathrm{cos}\,\theta_1})$ . Solve Equation 3 for $z''$ as a function of $x''$, and differentiate to obtain the slope of the ellipse at point $C$: $${\mathrm{tan}\,\theta_2} \, = \,
{{{\mathrm{d}z''}}\over{{\mathrm{d}x''}}} \, = \,
-{{b}\over{a^2}} \, x'' \left(1 \, - \, {{{x'' \, }^2}\over{a^2}}\right)^{-1/2}
\, = \,
{{b}\over{a^2}} \,
(f_1 {\mathrm{sin}\,\theta_1}\,+\, f_0)
\left[1 \, - \, {{{
(f_1 {\mathrm{sin}\,\theta_1}\,+\, f_0)
\, }^2}\over{a^2}}\right]^{-1/2}$$
Let $i_\mathrm{s} \,= \, \theta_2 \, - \, \theta_1$. Now we can define the $(x''', y''', z''')$ system, that has its origin at $S$ in the center of the mirror, the $\hat x'''$-$\hat y'''$ plane tangent to the spheroid, and the $\hat z'''$ axis pointing into the mirror surface. $$\begin{pmatrix}
x'''\\ y'''\\ z'''
\end{pmatrix}
=
\begin{pmatrix}
{\rm cos\,} i_\mathrm{s} & 0 &{\rm sin\,} i_\mathrm{s}\\
0&1&0\\
- {\rm sin\,} i_\mathrm{s} & 0 &{\rm cos\,} i_\mathrm{s}
\end{pmatrix}
\begin{pmatrix}
x' \\ y' \\ z' - f_1
\end{pmatrix}
\, ,$$ and the inverse transform is $$\begin{pmatrix}
x' \\ y' \\ z'
\end{pmatrix}
=
\begin{pmatrix}
{\rm cos\,} i_\mathrm{s} & 0 & - {\rm sin\,} i_\mathrm{s}\\
0&1&0\\
{\rm sin\,} i_\mathrm{s} & 0 &{\rm cos\,} i_\mathrm{s}
\end{pmatrix}
\begin{pmatrix}
x'''\\ y'''\\ z'''
\end{pmatrix}
+
\begin{pmatrix}
0 \\ 0\\ f_1
\end{pmatrix}
\, .$$ Note that $i_\mathrm{s}$ is the angle of incidence at the secondary, since the incident ray lies along the $\hat z'$ axis, and $\hat z'''$ is normal to the surface. Applying the law of cosines to the triangle $F_1 S F_2$, we have: $$4 f_0^2 \, = \, f_1^2 \, + \, f_2^2 \, - \, 2 f_1 f_2 {\mathrm{cos}\,(2 i_{\mathrm s})}
\, .$$ We can now eliminate $a$, $b$, $e$, $f_0$, $\theta_1$, and $\theta_2$ in favor of $f_1$, $f_2$ and $i_\mathrm{s}$ in all of the preceeding equations. The shape of the mirror can therefore be described in terms of the two focal distances and the angle of incidence. Combine Equation 25 with Equation 26, substitute into Equation 3, and simplify: $$p z'''^2 \, + \, (q x''' \,+\, 2\, r)z''' \, + \, y'''^2 \, + \, s x'''^2 \, = \, 0 \, ,$$ where $$\begin{aligned}
p \, &=& 1\, - \,e^2\,{\mathrm{sin}}^2 \theta_2 \, = \, 0.94114858\\
q \, &=& \, e^2 \, {\mathrm{sin}} \, (2 \, \theta_2) \, = \, 0.11165324\\
r \, &=& \, f_1( {\mathrm{cos}} \, i_\mathrm{s} + e \,{\mathrm{sin}} \, \theta_2 )
\, = \, f_2( {\mathrm{cos}} \, i_\mathrm{s} - e \,{\mathrm{sin}} \, \theta_2 ) \, = \, 1592.6382 \, \mathrm{mm}\\
s \, &=& 1\, - \,e^2 \,{\mathrm{cos}}^2 \theta_2 \, = \, 0.94704272\end{aligned}$$ and $$\begin{aligned}
f_0 \, &=& \, {{1}\over{2}}\, \sqrt{f_1^2 \, + \, f_2^2 \,
- \, 2 f_1 f_2 \,{\mathrm{cos}\,(2 i_\mathrm{s})}} \, = \, 583.49067 \, \mathrm{mm}\\
\theta_2 \, &=& \, i_\mathrm{s} \,+\, {\mathrm{arcsin}}\left(
{{f_2^2\,-\,f_1^2\,-\,4f_0^2}\over{4 f_0 f_1}} \right) \, = \, 46.51091{{^{\circ}}}\\
e \, &=& \, {{2 f_0}\over{f_1\,+\,f_2}} \, = \, 0.334378 \, ,\end{aligned}$$ yielding the equation of the mirror surface sag in $(x''', y''', z''')$ coordinates: $$z''' \, = \, {{1}\over{2p}}\left[ \sqrt{
(2\, r \, + \, q x''')^2 \, - \, 4 p \, (y'''^2 \, + \, s x'''^2)}
\, - \, (2\, r \, + \, qx''' ) \, \right] \, .$$ This is the shape of the secondary mirror in a coordinate system centered on the intersection of the central ray with the surface, where the surface sag ($z'''$) is normal to the surface at that point. The value of $r$ has dimensions of length, and it is always positive (either of the expressions in equation 30 can be used, depending on the sign of $\theta_2$). It is the radius of curvature in the $\hat y'''$-$\hat z'''$ plane, but it is not the radius of curvature seen by the beam, as will be shown in Equation 36. The values of $p$ and $s$ are dimensionless and always between 0 and 1. The value of $q$ is dimensionless and always between $-1$ and 1; its sign is opposite that of the $x''$ coordinate of the point $S$. Note that $z'''$ is everywhere negative, since $p$ and $s$ are positive in Equation 35. The $\hat z'''$ direction points into the mirror, and the $\hat x'''$-$\hat y'''$ plane is tangent to the mirror at point $S$, the origin of the $(x, y, z)$ system. The direction of $\hat x'''$ is such that its dot product with the vector from $F_1$ to $F_2$ is positive.
Since we know the two focal distances $f_1$ and $f_2$, the thin lens formula gives the paraxial focal length $f$: $${{1}\over{f}} \, = \, {{1}\over{f_1}}
\, + \, {{1}\over{f_2}} \, = \,
{{2 \, \mathrm{cos}\, i_\mathrm{s}} \over {r}} \, ,$$ using Equation 30 to substitute for $f_1$ and $f_2$. So $f\, = \, r / (2 \, {\mathrm{cos}\, i_\mathrm{s}}) \,
= \, 818.28089 \, \mathrm{mm}$, and the central radius of curvature is $r / {\mathrm{cos}\, i_\mathrm{s}} \, = \,
1636.5618 \, \mathrm{mm}$.
Defining the Edge of the Secondary
==================================
Unlike the first SPT receiver and SPTpol, the edge of the secondary is not a stop of the optical system, and so the secondary need only be big enough not to significantly vingnette; the exact figure of the edge doesn’t matter optically. For convenience in manufacturing, however, we’d like the edge of the secondary to lie in a plane. This can be accomplished by defining the edge as the intersection of the secondary spheroid with a right circular cone whose vertex is at the Gregorian focus $\vec{F_2}$ and whose axis is coincident with the central ray $\vec{F_2}-\vec{Q}$. The half-angle of the cone is defined to be $\phi_2 \, = \, 23.5993{{^{\circ}}}$, and the angle between the rotation axis of the cone and the line between the focii is $\psi_2 \, = \, -30.1844{{^{\circ}}}$. These two angles, together with the major and minor axes of the spheroid ($a$ and $b$), fully define the mirror. The analysis is done in the Appendix below, and the results presented in Table 1. The calculations in the Appendix show that there are equivalent ways of defining the edge of the secondary: a right circular cone whose vertex is the prime focus, $F_1$, with half angle $\phi_1 \, = \, 38.8953{{^{\circ}}}$ and tilt $\psi_1 \, = \, -52.0529{{^{\circ}}}$ also intersects the spheroidal surface of the secondary in the same ellipse as the cone from $F_2$, although the axis of that cone does not lie on the central ray of the optics.
As a practical matter, the edge of the secondary mirror is chosen to be circular, the result of intersecting the prolate spheroid with a cylinder. The edge does not, therefore, lie in a plane. The edge of the primary is not illuminated by the detector optics, so the precise geometry does not matter.
Mirror Coordinates Relative to the Mirror Edge
==============================================
Given the definition of the edge of the secondary, we can transform into a coordinate system centered on the ellipse that is the edge of the secondary and tilted so that the plane of the mirror edge is the $x''''-y''''$ plane at $z'''' \, = \, 0$: $$\begin{pmatrix}
x''''\\ y''''\\ z''''
\end{pmatrix}
=
\begin{pmatrix}
{\rm cos\,} \theta_\mathrm{e} &{\rm sin\,} \theta_\mathrm{e}&0\\
0&0&-1\\
{-\rm sin\,} \theta_\mathrm{e}&{\rm cos\,} \theta_\mathrm{e}&0
\end{pmatrix}
\begin{pmatrix}
x-c\\ y \\ z
\end{pmatrix}
+
\begin{pmatrix}
d \\ 0 \\ 0
\end{pmatrix}
\,.$$ The inverse transform is: $$\begin{pmatrix}
x\\ y\\ z
\end{pmatrix}
=
\begin{pmatrix}
{\rm cos\,} \theta_\mathrm{e} &0&-{\rm sin\,} \theta_\mathrm{e}\\
{\rm sin\,} \theta_\mathrm{e}&0&{\rm cos\,} \theta_\mathrm{e}\\
0&-1&0
\end{pmatrix}
\begin{pmatrix}
x'''' - d\\ y''''\\ z''''
\end{pmatrix}
+
\begin{pmatrix}
c \\ 0 \\ 0
\end{pmatrix}
\, ,$$ where in this case $(x, \, y, \, z)$ is the coordinate system of Equation 41 and not the global coordinate system. The offsets $c$ and $d$ are given in Table 1, and $\theta_e \, = \, 47.9795{{^{\circ}}}$ is the angle $\theta $ defined in the Appendix ($\theta_e$, the slope of the mirror edge, differs from $\theta_2$ in Equation 33, the slope of the mirror surface at the central ray). This is the coordinate system needed to cut the mirror surface. In this coordinate system, the equation of the secondary mirror can be expressed as: $$z'''' = {{P x''''+Q+a \sqrt{J y''''^2+K x''''^2+ Lx'''' +N}}\over{J}} \,$$ where $$\begin{aligned}
c & = & (f_0 \, \cos \psi_2 \, - \, a \cos \phi_2)/\sin \, \psi_2 \, \tan \theta_\mathrm{e} \, = \, 1961.81 \, \mathrm{mm} \nonumber \\
d & = & b^2 \, c \cos \theta_\mathrm{e}/(a^2 \, \sin^2 \theta_\mathrm{e}
\, + \, b^2 \cos^2 \theta_\mathrm{e})
\, = \, 1227.92 \, \mathrm{mm}\nonumber \\
J & = & -f_0^2\, \mathrm{cos}^2 \theta_\mathrm{e}\,-\, b^2 \,
= \, -2857121.6 \, \mathrm{mm}^2 \nonumber \\
K & = & -b^2 \, = \, -2704564.4 \, \mathrm{mm}^2 \nonumber \\
L & = & 2 \, b^2\, [ d \, - \,
c \, \mathrm{cos}\,(\theta_\mathrm{e})] \, = \,
-4.614588\,\times\,10^{8} \, \mathrm{mm}^3 \nonumber \\
N & = & b^2\{b^2+f_0^2\,\mathrm{cos}^2\,\theta_\mathrm{e} \,-\,
[d \, - \, c \, \mathrm{cos}\,\theta_\mathrm{e}]^2\} \, = \,
7.7075858 \, \times\,10^{12} \, \mathrm{mm}^4 \nonumber \\
P & = & f_0^2 \, \mathrm{cos} \, \theta_\mathrm{e}\,
\mathrm{sin} \, \theta_\mathrm{e}
\, = \, 169310.4\, \mathrm{mm}^2 \nonumber \\
Q & = & -[f_0^2 \, d \, \cos \, \theta_\mathrm{e} + b^2\,c ] \,
\mathrm{sin}\, \theta_\mathrm{e} \, = \,
-4.1496374 \, \times \, 10^9 \mathrm{mm}^3 \nonumber
\, .\end{aligned}$$ The $\hat x''''$ direction is up, and the top of the mirror is $(x'''', \, y'''', \, z'''') \, = \, (869.665\, \mathrm{mm}, \, 0, \, 0)$. The point $(x'''', \, y'''', \, z'''') \, = \, (0, \, 847.599, \, 0)$ is also on the mirror edge. The point of maximum depth in the secondary mirror is not below the center of the mirror (that is, $x''''=0$, $y''''=0$), but is slightly displaced: $$\begin{aligned}
x''''_\mathrm{min}& = &{{a^2KL-L P^2-P\sqrt{(L^2-4KN)P^2+4a^2K^2N-a^2KL^2}}\over
{2(KP^2-a^2K^2)}}\\
& & = 14.2408\, \mathrm{mm} \nonumber \\
y''''_\mathrm{min}& = & 0 \nonumber \\
z''''_\mathrm{min}& = & -243.286\, \mathrm{mm} \nonumber \end{aligned}$$ is the lowest point on the mirror surface, about a half inch above the center.
Appendix: Theorems about Spheroids, Cones, and Planes
=====================================================
The coordinate systems and variables defined in this appendix are independent of the body of this memo above. Consider a prolate spheroid: $${{x^2}\over{a^2}}
+ {{y^2}\over{b^2}}
+ {{z^2}\over{b^2}} = 1 \, ,$$ where $a \, > \, b \, > \, 0$. In the $(x, \, y, \, z)$ coordinate system, the focii are $F_1 \, = \, (+ f_0 , \, 0, \, 0)$ and $F_2 \, = \, (- f_0 , \, 0, \, 0)$, where $f_0 \, \equiv \, \sqrt{a^2 \, - \, b^2}$. Cut the spheroid with an arbitrary plane, $y \, = \, (x - c)\, {\mathrm{tan}\,}\theta $, where $ - 90{{^{\circ}}}\, < \, \theta \, < \, 90{{^{\circ}}}$. By symmetry, this equation can describe any cutting of the spheroid except for the special cases of a plane parallel to the axis of the spheroid ($\theta = 0{{^{\circ}}}$ exactly) or a plane perpendicular to the axis of the spheroid ($\theta = 90{{^{\circ}}}$ exactly). These singular cases can easily be treated separately. Transform so that the plane becomes the $\hat{u} - \hat{z}$ plane, and translate in the $\hat{u}$ direction by an offset, $d$, to be determined below: $$\begin{pmatrix}
x\\ y\\ z
\end{pmatrix}
=
\begin{pmatrix}
{\rm cos\,} \theta &{\rm -sin\,} \theta&0\\
{\rm sin\,} \theta&{\rm cos\,} \theta&0\\
0&0&1
\end{pmatrix}
\begin{pmatrix}
u - d\\ v\\ z
\end{pmatrix}
+
\begin{pmatrix}
c \\ 0\\ 0
\end{pmatrix}
\, ,$$ and the inverse transform is $$\begin{pmatrix}
u\\ v\\ z
\end{pmatrix}
=
\begin{pmatrix}
{\rm cos\,} \theta &{\rm sin\,} \theta&0\\
{\rm - sin\,} \theta&{\rm cos\,} \theta&0\\
0&0&1
\end{pmatrix}
\begin{pmatrix}
x - c \\ y\\ z
\end{pmatrix}
+
\begin{pmatrix}
d \\ 0\\ 0
\end{pmatrix}
\, .$$ Substitute Equation 42 into Equation 41, and set $v \, = \, 0$, (which we see from Equation 43 gives $y \, = \, (x - c)\, {\mathrm{tan}\,}\theta $) yielding: $${{[(u-d)\,{\rm cos\,}\theta + c\,]^2}\over{a^2}}
+ {{(u-d)^2\,{\rm sin}^2 \theta}\over{b^2}}
+ {{z^2}\over{b^2}} = 1 \, .$$ Let $$g^2 \, \equiv \, a^2 \,{\rm sin}^2 \theta \, + \, b^2 \,{\rm cos}^2 \theta$$ and $$\gamma \, \equiv \, 1 \, - \, {{c^2 \,{\rm sin}^2 \theta} \over {g^2} }\, .$$ Then, choosing the arbitrary translation in the $\hat u$ direction to be: $$d \, \equiv \, {{b^2 c \,{\rm cos}\, \theta} \over {g^2} }\, ,$$ Equation 44 can be written as $${{u^2}\over{\alpha^2}}
+ {{z^2}\over{\beta^2}} = 1 \, ,$$ with $$\alpha^2 \, \equiv \, {{a^2 b^2}\over{g^2}} \gamma$$ and $$\beta^2 \, \equiv \, b^2 \, \gamma \, .$$ Equation 48 shows that the intersection of the spheroid and plane is an ellipse, provided $\gamma \, > \, 0$ or $c^2 \, < \, a^2 \, + \, b^2 \, {\rm cot}^2 \theta $ (i.e. the plane actually intersects the spheroid and doesn’t miss it). [*The intersection of a prolate spheroid and an arbitrary plane is an ellipse.*]{} (More generally, any closed figure that results from the intersection of a plane with an ellipsoid, paraboloid, or hyperboloid is an ellipse.)
In the $(u, \, v, \, z)$ coordinate system, the center of the ellipse is $C \, = \, (0, \, 0, \, 0)$, and the extrema of the ellipse are $A \, = \, (\pm \, a b g^{-1}\sqrt{ \gamma} , \, 0, \, 0)$, and $B \, = \, (0, \, 0, \, \pm\,b \sqrt{ \gamma })$; the focii are $F_1 = ( [ f_0 - c ] {\rm cos} \, \theta + d, \, - [ f_0 - c] {\rm sin} \, \theta , \, 0)$ and $F_2 = ( - [ f_0 + c ] {\rm cos} \, \theta + d, \, - [f_0 + c] {\rm sin} \, \theta , \, 0) \, .$
In the $(x, \, y, \, z)$ coordinate system, the center of the ellipse is $C = (- d \, {\rm cos}\,\theta \, + \, c , \, -d \, {\rm sin}\, \theta, \, 0)$, and the extrema of the ellipse are $$A \, = \, ([\pm \, a b g^{-1}\sqrt{ \gamma} \, - \, d\, ] \, {\rm cos}\,\theta \, + \, c, \,
[\pm \, a b g^{-1}\sqrt{\gamma}\, - \, d\, ]\, {\rm sin}\,\theta ,
\, 0) \, ,$$ and $$B \, = \, ( - d \, {\rm cos}\,\theta \, + \, c ,
\, -d \, {\rm sin}\, \theta, \,
\pm\,b \sqrt{ \gamma }) \, .$$
Consider a right circular cone whose vertex is $F_1$, and whose axis passes through the point $Q_1$. The points on cone satisfy the equation: $$(\vec{P}-\vec{F_1})\cdot
(\vec{Q_1}-\vec{F_1})=
\|\vec{P}-\vec{F_1}\|\|\vec{Q_1}-\vec{F_1}\|\,\mathrm{cos}\,\phi_1 \, ,$$ where $P$ is any point on the cone, $\phi_1$ is the opening half-angle of the cone, and $Q_1$ is a point in the $x - y$ plane that defines the axis of the cone. Without loss of generality, we take $ 0 \, < \, \phi_1 \, < \, 90{{^{\circ}}}$, so ${\rm cos} \, \phi_1 \, > 0$.
Define a coordinate system translated so that the origin is at $F_1$: $x' \, = \, x - f_0$, $y' = y$, $z' = z$, and let the angle $\psi_1$ be the angle between the unit vector in the $x'$ direction and the axis of the cone, $\vec{Q_1}-\vec{F_1}$: $$\hat x' \cdot
(\vec{Q_1}-\vec{F_1})=
\|\vec{Q_1}-\vec{F_1}\|\,\mathrm{cos}\, \psi_1 \, .$$ Without loss of generality, we take we take $ - 90{{^{\circ}}}\, < \, \psi_1 \, < \, 90{{^{\circ}}}$, so ${\rm cos} \, \psi_1 \, > 0$.
Then Equation 53 becomes: $$x' \, {\rm cos}\, \psi_1 + y' \, {\rm sin} \, \psi_1 \, = \,
\sqrt{x'^2 \, + \, y'^2 \, + \, z'^2} \, \, {\rm cos} \, \phi_1 \, .$$ Solve Equation 55 for $z'^2$, substitute into Equation 41, and solve for $y$ to yield: $$y = {{(\pm f_0 \, {\rm cos}\, \phi_1 - a \, {\rm cos}\, \psi_1) \, x
+ a\, f_0 {\rm cos}\, \psi_1 \, \mp
a^2 \, {\rm cos}\, \phi_1}\over{a \, {\rm sin}\,\psi_1}} \,.$$ We see that $${\rm tan}\,\theta \, = \,{{\pm \, e \, {\rm cos}\, \phi_1 -
{\rm cos}\, \psi_1 } \over { {\rm sin} \, \psi_1}} \, ,$$ and $$c \, = \,
{{\pm \, a \, {\rm cos}\, \phi_1 \, - \,
f_0 \, {\rm cos}\, \psi_1 }
\over
{{\rm sin}\, \psi_1 \, {\rm tan} \, \theta}} \, ,$$ where $e \, \equiv \, f_0/a$ is the eccentricity of the spheroid, describe the two planes $$y \, = \, (x - c)\, {\mathrm{tan}\,}\theta$$ whose intersection with the spheroid yields the same ellipses as the intersection of the spheroid with the two nappes of the cone that has a vertex at $F_1$, is tilted by angle $\psi_1$ with respect to the $\hat x$ axis, and has a half opening angle of $\phi_1$. Choosing the $+$ signs in Equations 57 and 58 yields the nappe of the cone to the left, where the points on its axis satisfy $x \, < \, f_0$. Note that $\psi_1$ and $\theta$ have opposite signs for the nappe to the right, and usually have opposite signs for the nappe on the left unless the spheroid is sufficiently eccentric ($e \, > \, {\cos \, \psi_1}/{\cos \, \phi_1}$). [*The intersection of a prolate spheroid with a right circular cone, whose vertex is one of the focii of the spheroid, is a planar figure, specifically an ellipse.*]{} That is not true in general of a cylinder whose axis passes through the focus — the intersection of that cylinder with the spheroid will not lie in a plane — because the intersection of the cylinder with a circular cone sharing the same axis is a circle, not an ellipse, and we know that the intersection between the cone and the spheroid is in general elliptical.
By symmetry, this result applies equally well to a cone whose vertex is at $F_2$, but if we define the tilt of such a cone with respect to the positive $\hat x$ axis: $$\hat x'' \cdot
(\vec{Q_2}-\vec{F_2})=
\|\vec{Q_2}-\vec{F_2}\|\,\mathrm{cos}\, \psi_2 \, .$$ where $x'' \, = \, x \, + \, f_0$, that breaks the symmetry and we get slightly different equations for planes: $${\rm tan}\,\theta \, = \,{{\mp \, e \, {\rm cos}\, \phi_2 -
{\rm cos}\, \psi_2 } \over { {\rm sin} \, \psi_2}} \, ,$$ and $$c \, = \,
{{\pm \, a \, {\rm cos}\, \phi_2 \, + \,
f_0 \, {\rm cos}\, \psi_2 }
\over
{{\rm sin}\, \psi_2 \, {\rm tan} \, \theta}} \, ,$$ where like before, we take $ 0 \, < \, \phi_2 \, < \, 90{{^{\circ}}}$ and $ -90{{^{\circ}}}\, < \, \psi_2 \, < \, 90{{^{\circ}}}$. Choosing the upper signs in Equations 61 and 62 yields the nappe to the left (points on the axis satisfy $x \, < \, -f_0$).
Equations 57 and 58 can be solved to yield $\phi_1$ and $\psi_1$ as a function of $\theta$ and $c$: $${\rm tan}\,\psi_1 \, = \,{{b^2 \, {\rm cot}\, \theta}
\over {c \, f_0 \, - \, a^2 }} \, ,$$ and $$\cos \phi_1 \, = \,
\left| {{\tan \, \theta \, \sin \, \psi_1 \, + \, \cos \, \psi_1}
\over{e}} \right|
\, ,$$ while Equations 61 and 62 can be similarly be inverted to yield: $${\rm tan}\,\psi_2 \, = \,{{- b^2 \, {\rm cot}\, \theta}
\over {c \, f_0 \, + \, a^2 }} \, ,$$ and $$\cos \phi_2 \, = \,
\left| {{\tan \, \theta \, \sin \, \psi_2 \, + \, \cos \, \psi_2}
\over{e}} \right|
\, .$$ For any plane defined by $\theta$ and $c$ that cuts the spheroid, there are right circular cones from each focus defined by $\phi_1$, $\psi_1$, $\phi_2$, and $\psi_2$ that generate the same ellipse as the plane. [*Any arbitrary plane cutting a prolate spheroid results in an ellipse. The set of lines connecting points on that ellipse to the focii of the spheroid comprise two right circular cones whose vertices lie on the focii.*]{}
The points $Q_1$ and $Q_2$ cannot in general be coincident with the center of the ellipse, point $C$. When an ellipse is generated by cutting a cone with a plane, the center of the ellipse does not fall on the axis of the cone, except in the singular case where the ellipse is a circle. The location of $Q_1$ could be chosen to lie anywhere on the line $y \, = \, (x - f_0)\, {\mathrm{tan}\,}\psi_1 $, while the location of $Q_2$ could be chosen to lie anywhere on $y \, = \, (x + f_0)\, {\mathrm{tan}\,}\psi_2 $. From Equations 63 and 65, we see that these two lines intersect at the point $$Q_0 \, = \, \left({{a^2}\over{c}}, \,
{{-b^2}\over{c \, \tan \, \theta}},\,0\right)
\, ,$$ and this point can serve as both $Q_1$ and $Q_2$, but in general this point does not lie on the spheroid. $Q_0$ is, of course, the intersection of the axes of the two cones, and $c$ is the intercept of the plane with the $\hat x$ axis and $- c \tan \theta$ is the intercept of the plane with the $\hat y$ axis. $Q_0$ lies on the spheroid if $c^2 \, = \, a^2 \, + \, b^2\,\cot^2 \theta$, but that implies that the plane is tangent to the spheroid, $\gamma \, = \, 0$ in Equation 46, and the ellipse has shrunk to a point — not interesting or useful. The singular on-axis case, where $\psi_1 \, = \, \psi_2 \, = \, 0$ and the ellipse is actually a circle, allows $Q_0 \, = (a, \, 0, \, 0)$. In the general case, point $Q_0$ lies outside the spheroid.
The points $Q_1$ and $Q_2$ could also lie in the plane of the ellipse or on the surface of the spheroid. The intersections of the cone axes with the line $y \, = \, (x - c)\, {\mathrm{tan}\,}\theta $, in the plane of the ellipse, in $(x,\,y,\,z)$ coordinates, are: $$Q_{1e} \, = \left({{c \, {\rm tan} \, \theta - f_0 \, {\rm tan} \, \psi_1}\over
{{\rm tan} \, \theta \, - \, {\rm tan} \, \psi_1}}, \,
{{[c \, - \, f_0] \, {\rm tan} \, \theta \tan \, \psi_1 }\over
{{\rm tan} \, \theta \, - \, {\rm tan} \, \psi_1}}\, , \, 0 \right) \, ,$$ $$Q_{2e} \, = \left({{c \, {\rm tan} \, \theta + f_0 \, {\rm tan} \, \psi_2}\over
{{\rm tan} \, \theta \, - \, {\rm tan} \, \psi_2}}, \,
{{[c \, + \, f_0] \, {\rm tan} \, \theta \tan \, \psi_2 }\over
{{\rm tan} \, \theta \, - \, {\rm tan} \, \psi_2}}\, , \, 0 \right) \, .$$ The intersections of the cone axes with the spheroid are: $$\begin{aligned}
Q_{1s} \, = \,& \left(
{{ a^2 f_0\,\tan ^2 \psi_1
\pm a b^2 \sec \psi_1}
\over{a^2\,\tan ^2\, \psi_1+b^2}}\, ,
{{
- b^2 f_0\,
\pm a b^2 \sec \psi_1}
\over{a^2\,\tan \psi_1+b^2 \, \cot \psi_1
}} ,\,
0
\right) \,,\end{aligned}$$ $$\begin{aligned}
Q_{2s} \, = \,& \left(
{{ - a^2 f_0\,\tan ^2 \psi_2
\pm a b^2 \sec \psi_2}
\over{a^2\,\tan ^2\, \psi_2+b^2}}\, ,
{{
+ b^2 f_0\,
\pm a b^2 \sec \psi_2}
\over{a^2\,\tan \psi_2+b^2\,\cot\psi_2
}} , \,
0
\right) \,.\end{aligned}$$ Choose the $+$ sign for the nappe to the right.
If the cone represents a bundle of light rays coming from $F_2$ and striking the spheroidal mirror, then $\vec Q_2 - \vec F_2$ is the central ray, and it strikes the mirror at point $Q_{2s}$. The light rays in the cone will strike the mirror and reflect, converging in a different cone onto $F_1$. [*An ellipse that is the intersection of a spheroid and a right circular cone whose vertex is one focus of the spheroid can also be described as the intersection of the spheroid with a different right circular cone whose vertex is the other focus.*]{} This second cone will have a different opening angle $\phi_1$ and tilt $\psi_1$. Also, the axes of the two cones do not, in general, intersect at the spheroidal surface, so the points $Q_{1s}$ and $Q_{2s}$ will be different. The central beam coming from $F_2$ will strike the mirror and be reflected [*before*]{} it gets to $Q_0$. This is important for optics, because if the axis of one cone is the central ray of the beam, that ray will not reflect onto the axis of the other cone, causing a skewness in the beam that can only be corrected by an appropriate choice of angles in subsequent reflections, i.e. the Dragone condition.
--------------- ----------------------------------------------- ------------------------ ------------- -------------- ------------
Variable Coordinate
System
\[0.5ex\] $a$ semimajor axis of spheroid 1745.000
$b$ semiminor axis of spheroid 1644.556
$f_0$ focal distance of spheroid 583.490
$\theta$ angle of plane $47.9795{{^{\circ}}}$
$c$ $x$ intercept of plane 1961.810
$d$ $u$ coordinate offset 1227.920
$g$ $(a^2\sin^2\theta\,+\,b^2\cos^2\theta)^{1/2}$ 1700.726
$\gamma$ $ 1-c^2\sin^2\theta/{g^2}$ 0.265634
$\alpha$ semimajor axis ellipse 869.665
$\beta$ semiminor axis ellipse 847.599
$\phi_1$ half-angle cone 1 $38.8953{{^{\circ}}}$
$\psi_1$ tilt cone 1 $-52.0529{{^{\circ}}}$
$\phi_2$ half-angle cone 2 $23.5993{{^{\circ}}}$
$\psi_2$ tilt cone 2 $-30.1844{{^{\circ}}}$
$A_t$ top of ellipse 1721.997 $-266.148$ 0 s
$A_t$ 8168.367 $44.177$ 0 p
$A_b$ bottom of ellipse 557.696 $-1558.305$ 0 s
$A_b$ 7386.922 $-1509.725$ 0 p
$B$ extrema of ellipse 1139.847 $-912.226$ $\pm847.599$ s
$B$ 7777.645 $-732.774$ $\pm847.599$ p
$Q_{0}$ intersection of axes 1552.150 $-1242.2$ 0 s
$Q_{0}$ 8262.490 $-942.058$ 0 p
$Q_{1e}$ axis 1 intercept ellipse 1222.935 $-820.0142$ 0 s
$Q_{1e}$ 7833.411 $-621.883$ 0 p
$Q_{1s}$ axis 1 intercept spheroid 1374.021 $-1013.765$ 0 s
$Q_{1s}$ 8030.327 $-768.820 $ 0 p
$Q_{2e}$ axis 2 intercept ellipse 1086.548 $-971.378$ 0 s
$Q_{2e}$ 7741.872 $-803.908$ 0 p
$Q_{2s}$ axis 2 intercept spheroid 1300.929 $-1096.073$ 0 s
$Q_{2s}$ $=$ point $S$ in Equation 9 7981.584 $-867.517 $ 0 p
\[1ex\]
--------------- ----------------------------------------------- ------------------------ ------------- -------------- ------------
Parameters of SPT3G Secondary Mirror: Variables refer to definitions in the Appendix. All dimensions in millimeters. The spheroid is the mirror surface, the ellipse is the mirror edge. Subscript 1 refers to the cone between the prime focus and the secondary. Subscript 2 refers to the cone between the secondary and the Gregorian focus. Coordinate system “s" has its origin halfway between the two focii and the $x$ axis is the line between the two focii. Coordinate system “p" is the global coordinate system with the origin at the vertex of the primary. \[table:parameters\]
/* file spt3gsecondary.c */
/* Calculate spt coordinate values from memo */
/* A. Stark */
/* 7-4-14 */
/* compile with command
cc spt3gsecondary.c -lm -o spt3gsecondary
*/
#include <math.h>
#include <stdio.h>
double fp;
double theta_s;
double xpe, ype;
double theta_e;
double theta_r;
double theta_c;
double theta_1,theta_2;
double a, b, fs, f_0;
double rtod;
double i_p;
double eslope;
double xppe, yppe, zppe;
double xdaggere, ydaggere, zdaggere;
main()
{
double norm(), dot();
void stoptoprimary();
void ptou(),utop(),utopp(),pptou(),pptop(),ptopp();
void ppptop(), pptopppp();
void daggertou();
double uonellipse(), pponellipse(), pponcone();
double daggeronellipse(), daggeroncone();
double quadp(), quadm();
double yc,xc;
double r2, k, e;
double alpha, beta, gamma, nu, lambda, mu;
double rho, sigma, tau;
double F1x,F1y,F1z ;
double Scx,Scy,Scz ;
double F2x,F2y,F2z ;
double f_1,f_2;
double M;
double i_s, i_D;
double iota;
double temp;
double xpptop, xppbot, ypptop, yppbot;
double xdaggertop, xdaggerbot, ydaggertop, ydaggerbot;
double omega;
double upsilon;
double ae, be;
double J,K,L,N,P,Q;
double p, q, r, s;
double fparaxial;
double x, y, z;
double xp, yp, zp;
double xpp, ypp, zpp;
double xdagger, ydagger, zdagger;
double xppp, yppp, zppp;
double xpppp, ypppp, zpppp;
double xppppmin, yppppmin, zppppmin;
double Tslope, Bslope;
double Atemp,Btemp,Ctemp;
rtod = 45.0/atan(1.0); /* convert radians to degrees */
yc = 5300.0; /* vertical offset of primary */
fp = 7000.0; /* focal length of primary */
i_p = atan(yc/(2*fp))*rtod; /* angle of incidence at primary */
printf("\n i_p = %12.8gdeg\t angle of incidence at primary \n", i_p);
/* F1 is vector to prime focus */
F1x = 7000.0; F1y = 0.0; F1z = 0.0;
printf("\n F1 = %12.8gmm, %12.8gmm, %12.8gmm\t prime focus",
F1x,F1y,F1z);
/* slope of central ray between primary and secondary */
alpha = (4.0*fp*yc)/(yc*yc - 4.0*fp*fp);
printf("\nalpha = %12.8g\t slope of central ray", alpha);
/* Major and minor axes are defining values for SPT3G secondary */
a = 1745.;
b = 1644.556;
/* radius of curvature and conic constant of secondary */
/* these values, chosen here, define the shape of the secondary */
e = sqrt(1.0 - b*b/(a*a));
k = -1.0*e*e;
r2 = a*(1.0+k);
/*
r2 = 1549.89366;
k = -0.111809;
e = sqrt(-k);
*/
printf("\n r2 = %12.8gmm\t secondary radius of curvature",r2);
printf("\n k = %12.8g\t conic constant of secondary",k);
printf("\n e = %12.8g\t eccentricity of secondary",e);
/* semi-major, semi-minor, focal distance of secondary */
/*
a = r2/(k+1.0);
b = r2/sqrt(k+1.0);
*/
f_0 = sqrt(a*a-b*b);
fs = a - f_0;
printf("\n a = %12.8gmm\t major axis of secondary spheroid",a);
printf("\n b = %12.8gmm\t minor axis of secondary spheroid",b);
printf("\n f_0 = %12.8gmm\t focal distance 0 of spheroid\n",
f_0);
printf("\n fs = %12.8gmm\t vertex to f_1 distance of spheroid\n",
fs);
/* rotation of secondary axis */
/* this value is chosen so that Dragone angle (calculated below) is zero */
theta_s = 15.323;
printf("\n theta_s = %12.8gdeg\t rotation secondary axis",theta_s);
/* solve for intersection of central ray with secondary */
rho = (cos(theta_s/rtod)+alpha*sin(theta_s/rtod))/a;
sigma = (a - fs)/a;
tau = (sin(theta_s/rtod)-alpha*cos(theta_s/rtod))/b;
xc = fp + (-rho*sigma+sqrt(rho*rho*sigma*sigma -
(rho*rho + tau*tau)*(sigma*sigma - 1.0)))/
(rho*rho + tau*tau ) ;
Scx = xc; Scy = alpha*(xc-fp); Scz=0.0;
printf("\n C = %12.8gmm, %12.8gmm, %12.8gmm\t"
"central ray at secondary", Scx,Scy,Scz);
/* F2 is vector to Gregorian focus */
F2x = fp-2.0*(a-fs)*cos(theta_s/rtod);
F2y = -2.0*(a-fs)*sin(theta_s/rtod);
F2z = 0.0;
printf("\n F2 = %12.8gmm, %12.8gmm, %12.8gmm\t Gregorian focus",
F2x,F2y,F2z);
/* distance from Gregorian focus to secondary along central ray */
f_2 = norm(F2x-Scx,F2y-Scy,F2z-Scz);
printf("\n f_2 = %12.8gmm\t Gregorian focus to secondary",f_2);
f_2 = 2180.0;
/* distance from prime focus to secondary along central ray */
f_1 = norm(F1x-Scx,F1y-Scy,F1z-Scz);
printf("\n f_1 = %12.8gmm\t prime focus to secondary\n",f_1);
f_1 = 1310.0;
/* magnification of secondary, is negative for Gregorians */
M = -f_2/f_1;
printf("\n M = %12.8g\tmagnification\n", M);
/* solve for angle of incidence of central ray at secondary */
i_s = 0.5*acos(dot(F2x-Scx,F2y-Scy,F2z-Scz,
F1x-Scx,F1y-Scy,F1z-Scz)
/(f_2*f_1))*rtod;
printf(" \n i_s = %12.8gdeg\t angle of"
"incidence at secondary \n", i_s);
printf("\n f_0 = %12.8g\t\n", f_0);
/*
f_0 = 0.5*sqrt(f_1*f_1+f_2*f_2-2.0*f_1*f_2*cos(2.0*i_s/rtod));
printf(" f_0 second time = %12.8g\t\n", f_0);
*/
/* solve for Dragone angle */
i_D = atan((1.0 - M)*tan(i_s/rtod)+M*tan(i_p/rtod))*rtod;
printf(" \n i_D = %12.3gdeg\t Dragone angle\n", i_D);
/* angle that rotates dagger coordinate system to p coordinate system */
theta_r = -30.1844;
printf("\n theta_r = %12.8gdeg\t ", theta_r);
theta_r = (-1.0)* (theta_s + 2.0*i_p - 2.0*i_s);
printf("\n alternate theta_r = %12.8gdeg\t ", theta_r);
theta_1 = rtod * asin((((b*b)/(a*f_1) - 1.0)/e));
printf("\n theta_1 = %12.8gdeg\t rotate pp to p ", theta_1);
theta_1 = rtod *
asin((f_2*f_2-f_1*f_1-4.0*f_0*f_0)/(4.0*f_0*f_1));
printf("\n theta_1 = %12.8gdeg\t rotate pp to p ", theta_1);
temp = f_1 * sin(theta_1/rtod) + f_0;
theta_2 = rtod * atan(b*temp/(a*a*sqrt(1.0 - temp*temp/(a*a))));
printf("\n theta_2 = %12.8gdeg ", theta_2);
e = 2*f_0/(f_1+f_2);
printf("\n e = %12.8g\t eccentricity ", e);
p = 1.0 - e*e * sin(theta_2/rtod) * sin(theta_2/rtod);
printf("\n p = %12.8g ", p);
q = e*e * sin(2.0*theta_2/rtod);
printf("\n q = %12.8g ", q);
r = f_1*(cos(i_s/rtod)+e*sin(theta_2/rtod));
printf("\n r = %12.8g ", r);
r = f_2*(cos(i_s/rtod) - e*sin(theta_2/rtod));
printf("\n r = %12.8g ", r);
s = 1.0 - e*e * cos(theta_2/rtod) * cos(theta_2/rtod);
printf("\n s = %12.8g ", s);
fparaxial = r / (2.0 * cos(i_s/rtod));
printf("\n fparaxial = %12.8g\t central radius = %12.8g\n",
fparaxial, 2.0*fparaxial);
printf("\n Is S on ellipse? %12.8g ",uonellipse(Scx,Scy,Scz));
theta_c= 23.5993;
printf("\n theta_c = %12.8gdeg\t defined"
"half angle of cone ", theta_c);
eslope = (-1.0*f_0*cos(theta_c)
-a*cos(theta_r/rtod))/(a*sin(theta_r/rtod));
printf("\n e slope Equation 44 = %12.8g ", eslope);
theta_e = rtod * atan(eslope);
printf("\t theta_e = %12.8gdeg\t angle of ellipse"
"in dagger \n", theta_e);
iota = (b*b*cos(theta_c/rtod))/(a*sin(theta_r/rtod));
printf("\n iota Equation45 = %12.8g mm \n", iota);
Tslope = tan((theta_r+theta_c)/rtod);
Bslope = tan((theta_r-theta_c)/rtod);
printf("\n Tslope = %12.8g \n", Tslope);
printf("\n Bslope = %12.8g \n", Bslope);
Atemp = a*a*Tslope*Tslope+b*b;
Btemp = -2.0*b*b*f_0;
Ctemp = f_0*f_0*b*b-a*a*b*b;
if (quadp(Atemp,Btemp,Ctemp)>quadm(Atemp,Btemp,Ctemp)) {
xdaggertop=quadp(Atemp,Btemp,Ctemp);
printf("xdaggertop positive");
} else {
xdaggertop=quadm(Atemp,Btemp,Ctemp);
printf("xdaggertop negative");
}
ydaggertop = Tslope*xdaggertop;
printf("\n xdaggertop,ydaggertop= %12.8g mm, %12.8g mm ",
xdaggertop,ydaggertop);
printf("\n Is top on ellipse pp? %12.8g ",
daggeronellipse(xdaggertop,ydaggertop,0.0));
printf("\n Is top on cone pp? %12.8g \n",
daggeroncone(xdaggertop,ydaggertop,0.0));
Atemp = a*a*Bslope*Bslope+b*b;
if (quadp(Atemp,Btemp,Ctemp)*Bslope <
quadm(Atemp,Btemp,Ctemp)*Bslope) {
xdaggerbot=quadp(Atemp,Btemp,Ctemp);
printf("xxpbot positive");
} else {
xdaggerbot=quadm(Atemp,Btemp,Ctemp);
printf("xxpbot negative");
}
ydaggerbot = Bslope*xdaggerbot;
printf("\n xdaggerbot,ydaggerbot= %12.8g mm, %12.8g mm ",
xdaggerbot,ydaggerbot);
printf("\n Is bot on ellipse dagger? %12.8g ",
daggeronellipse(xdaggerbot,ydaggerbot,0.0));
printf("\n Is bot on cone dagger? %12.8g ",
daggeroncone(xdaggerbot,ydaggerbot,0.0));
/*
xdaggertop = (a*b*b/cos((theta_c-theta_r)/rtod) - b*b*f_0)
/(a*a*Tslope*Tslope + b*b);
ydaggertop = Tslope*xpptop;
printf("\n alternate xdaggertop,ydaggertop=%12.8g mm, %12.8g mm",
xdaggertop,ydaggertop);
xdaggerbot = (a*b*b/cos((-1.0*theta_c-theta_r)/rtod)
- b*b*f_0)/(a*a*Bslope*Bslope + b*b);
ydaggerbot = Bslope*xdaggerbot;
printf("\n alternate xppbot,yppbot= %12.8g mm, %12.8g mm ",
xdaggerbot,ydaggerbot);
*/
temp = xdaggertop-xdaggerbot;
ae = temp*temp;
temp = ydaggertop-ydaggerbot;
ae += temp*temp;
ae = sqrt(ae)/2.0;
printf("\n ae= %12.8g mm ", ae);
xdaggere = (xdaggertop+xdaggerbot)/2.0;
ydaggere = (ydaggertop+ydaggerbot)/2.0;
zdaggere = b*sqrt(1.0 - (xdaggere-f_0)*(xdaggere-f_0)/(a*a)
- ydaggere*ydaggere/(b*b));
printf("\n xdaggere,ydaggere,zdaggere="
"%12.8g mm, %12.8g mm, %12.8g mm\t \n",
xdaggere,ydaggere,zdaggere);
printf("\n Is daggere on ellipse dagger? %12.8g ",
daggeronellipse(xdaggere,ydaggere,zdaggere));
printf("\n Is daggere on cone dagger? %12.8g \n",
daggeroncone(xdaggere,ydaggere,zdaggere));
eslope = (ydaggertop-ydaggerbot)/(xdaggertop-xdaggerbot);
printf("\n e slope= %12.8g ", eslope);
theta_e = rtod * atan(eslope);
printf("\t theta_e = %12.8gdeg\t angle of ellipse in dagger \n",
theta_e);
printf("\n theta_e+theta_s = %12.8gdeg", theta_e + theta_s);
iota = ydaggertop - eslope* xdaggertop;
printf("\n iota = %12.8g mm", iota);
printf("\n ydaggerbot = %12.8g mm", xdaggerbot*eslope+iota);
daggertou(xdaggertop,ydaggertop,0.0,&x,&y,&z);
printf("\n top of secondary edge = %12.8gmm, %12.8gmm, %12.8gmm",
x,y,z);
printf("\n Is top on ellipse? %12.8g \n",uonellipse(x,y,z));
daggertou(xdaggerbot,ydaggerbot,0.0,&x,&y,&z);
printf("\n bottom of secondary edge = %12.8gmm, %12.8gmm, %12.8gmm",
x,y,z);
printf("\n Is bottom on ellipse? %12.8g \n",uonellipse(x,y,z));
daggertou(xdaggere,ydaggere,zdaggere,&x,&y,&z);
printf("\n center of secondary edge = "
"%12.8gmm, %12.8gmm, %12.8gmm\t \n", x,y,z);
printf("\n Is center edge on ellipse? %12.8g \n",uonellipse(x,y,z));
printf("\nInput mirror coordinate xpppp: ");
scanf("%lg",&xpppp);
printf("Input mirror coordinate ypppp: ");
scanf("%lg",&ypppp);
J = -1.0*(f_0*f_0*cos(theta_e/rtod)*cos(theta_e/rtod)+b*b);
printf("\n J = %12.8g mm\t \n", J);
K = -1.0*b*b;
printf("\n K = %12.8g mm\t \n", K);
temp = ydaggere*sin(theta_e/rtod)+(xdaggere-f_0)*cos(theta_e/rtod);
L = -2.0*b*b*temp;
printf("\n L = %12.8g mm\t \n", L);
N = b*b*(b*b+f_0*f_0*cos(theta_e/rtod)*
cos(theta_e/rtod)-temp*temp);
printf("\n N = %12.8g mm\t \n", N);
P = f_0*f_0*cos(theta_e/rtod)*sin(theta_e/rtod);
printf("\n P = %12.8g mm\t \n", P);
Q = a*a*ydaggere*cos(theta_e/rtod)
-b*b*(xdaggere-f_0)*sin(theta_e/rtod);
printf("\n Q = %12.8g mm\t \n", Q);
zpppp = (P*xpppp+Q+a*sqrt(J*ypppp*ypppp
+K*xpppp*xpppp+L*xpppp+N))/J;
printf(" mirror sag zpppp = %12.8g\t \n", zpppp);
xppppmin = (a*a*K*L-L*P*P-P*sqrt((L*L-4.0*K*N)*P*P+
4.0*a*a*K*K*N-a*a*K*L*L))
/(2.0*(K*P*P-a*a*K*K));
xpppp = xppppmin;
ypppp = 0.0;
zpppp = (P*xpppp+Q+a*sqrt(J*ypppp*ypppp
+K*xpppp*xpppp+L*xpppp+N))/J;
printf("\n lowest point on mirror="
"%12.8gmm, %12.8gmm, %12.8gmm\t \n", xpppp,ypppp,zpppp);
utopp(7981.579,-867.52618,0,&x, &y, &z);
pptopppp(x,y,z,&xpppp, &ypppp, &zpppp);
printf("\n central ray at mirror="
"%12.8gmm, %12.8gmm, %12.8gmm\t \n",
xpppp, ypppp, zpppp);
printf("\n");
}
double norm(x,y,z) /* norm of a vector */
double x, y, z;
{
return(sqrt(x*x+y*y+z*z));
}
double dot(x1,y1,z1,x2,y2,z2) /* dot product of two vectors */
double x1,y1,z1,x2,y2,z2;
{
return(x1*x2+y1*y2+z1*z2);
}
double uonellipse(double xonel, double yonel, double zonel)
/* Is the point on the spheroid in unprimed coords? */
{
double temponel, sumonel;
temponel = ((xonel-fp)*cos(theta_s/rtod)
+yonel*sin(theta_s/rtod)+f_0)/a ;
sumonel = temponel*temponel;
temponel = ((xonel-fp)*sin(theta_s/rtod)
-yonel*cos(theta_s/rtod))/b;
sumonel += temponel*temponel;
temponel = zonel/b;
sumonel += temponel*temponel;
return(sumonel);
}
double pponellipse(double xonel, double yonel, double zonel)
/* Is the point on the spheroid in double prime coords? */
{
double temponel, sumonel;
temponel = (xonel+f_0)/a;
sumonel = temponel*temponel;
temponel = yonel/b;
sumonel += temponel*temponel;
temponel = zonel/b;
sumonel += temponel*temponel;
return(sumonel);
}
double daggeronellipse(double xonel, double yonel, double zonel)
/* Is the point on the spheroid in dagger coords? */
{
double temponel, sumonel;
temponel = (xonel-f_0)/a;
sumonel = temponel*temponel;
temponel = yonel/b;
sumonel += temponel*temponel;
temponel = zonel/b;
sumonel += temponel*temponel;
return(sumonel);
}
double daggeroncone(double xonco, double yonco, double zonco)
/* Is the point on the cone in dagger coords? */
{
double temponco;
temponco = sqrt(xonco*xonco+yonco*yonco+zonco*zonco)
*cos(theta_c/rtod);
return((xonco*cos(theta_r/rtod)+yonco
*sin(theta_r/rtod))/temponco);
}
void ptou(double xp, double yp, double zp,
double *x, double *y, double *z)
/* convert primed to unprimed coordinates */
{
*x=cos(2.0*i_p/rtod)*(xp)+sin(2.0*i_p/rtod)*yp+fp;
*y=(-1.0)*sin(2.0*i_p/rtod)*(xp)+cos(2.0*i_p/rtod)*yp;
*z=zp;
}
void ptopp(double xp, double yp, double zp,
double *xpp, double *ypp, double *zpp)
/* convert primed to double primed coordinates */
{
*xpp= cos(theta_1/rtod)*(xp) - sin(theta_1/rtod)*zp;
*ypp=yp;
*zpp= sin(theta_1/rtod)*(xp)+cos(theta_1/rtod)*zp;
}
void pptou(double xpp, double ypp, double zpp,
double *x, double *y, double *z)
/* convert double primed to unprimed coordinates */
{
*x=cos(theta_s/rtod)*(xpp)-sin(theta_s/rtod)*ypp+fp;
*y=sin(theta_s/rtod)*(xpp)+cos(theta_s/rtod)*ypp;
*z=zpp;
}
void daggertou(double xdagger, double ydagger,
double zdagger, double *x, double *y, double *z)
/* convert double primed to unprimed coordinates */
{
*x=cos(theta_s/rtod)*(xdagger - 2.0*f_0)
-sin(theta_s/rtod)*ydagger +fp;
*y=sin(theta_s/rtod)*(xdagger - 2.0*f_0)
+cos(theta_s/rtod)*ydagger;
*z=zdagger;
}
void pptop(double xpp, double ypp, double zpp,
double *xp, double *yp, double *zp)
/* convert double primed to primed coordinates */
{
*xp= cos(theta_1/rtod)*(xpp)+sin(theta_1/rtod)*zpp;
*yp=ypp;
*zp= (-1.0)*sin(theta_1/rtod)*(xpp)+cos(theta_1/rtod)*zpp;
}
void utopp(double x, double y, double z,
double *xpp, double *ypp, double *zpp)
/* convert unprimed to double primed coordinates */
{
*xpp=cos(theta_s/rtod)*(x - fp) + sin(theta_s/rtod)*y;
*ypp= (-1.0)* sin(theta_s/rtod)*(x - fp)+cos(theta_s/rtod)*y;
*zpp=z;
}
void utop(double x, double y, double z,
double *xp, double *yp, double *zp)
/* convert unprimed to primed coordinates */
{
*xp=cos(2.0*i_p/rtod)*(x - fp) - sin(2.0*i_p/rtod)*y;
*yp= sin(2.0*i_p/rtod)*(x - fp)+cos(2.0*i_p/rtod)*y;
*zp=z;
}
void pptopppp(double xpp, double ypp, double zpp,
double *x, double *y, double *z)
/* convert double primed to quadruple primed coordinates */
{
*x=cos(theta_e/rtod)*(xpp-xppe)+sin(theta_e/rtod)*(ypp-yppe);
*y=(-1.0)*zpp;
*z=(-1.0)*sin(theta_e/rtod)*(xpp-xppe)+cos(theta_e/rtod)*(ypp-yppe);
}
void stoptoprimary(double xs, double ys, double zs,
double *x, double *y, double *z)
/* project points at the stop onto primary */
{
*y = 2.0*fp*ys*(xs-fp-sqrt((xs-fp)*(xs-fp)
+ys*ys+zs*zs))/(ys*ys+zs*zs);
*x = (*y)*(xs-fp)/ys+fp;
*z = (*y)*zs/ys;
}
void ppptop(double xppp, double yppp, double zppp,
double *xp, double *yp, double *zp)
/* convert mirror coordinates to primed coordinates */
{
*xp=cos(theta_e/rtod)*xppp-sin(theta_e/rtod)*zppp+xpe;
*yp=sin(theta_e/rtod)*xppp+cos(theta_e/rtod)*zppp+ype;
*zp=-yppp;
}
double quadp(double A, double B, double C)
{
double surd;
surd = B*B-4.0*A*C;
if (surd < 0) {
printf("\nbad surd in quadp\n");
return(0.0);
} else {
return((-B+sqrt(surd))/(2.0*A));
}
}
double quadm(double A, double B, double C)
{
double surd;
surd = B*B-4.0*A*C;
if (surd < 0) {
printf("\nbad surd in quadm\n");
return(0.0);
} else {
return((-B-sqrt(surd))/(2.0*A));
}
}
[4]{} \[1\][\#1]{} \[1\][`#1`]{} urlstyle \[1\][doi: \#1]{}
B. A. [Benson]{}, P. A. R. [Ade]{}, Z. [Ahmed]{}, S. W. [Allen]{}, K. [Arnold]{}, J. E. [Austermann]{}, A. N. [Bender]{}, L. E. [Bleem]{}, J. E. [Carlstrom]{}, C. L. [Chang]{}, H. M. [Cho]{}, J. F. [Cliche]{}, T. M. [Crawford]{}, A. [Cukierman]{}, T. [de Haan]{}, M. A. [Dobbs]{}, D. [Dutcher]{}, W. [Everett]{}, A. [Gilbert]{}, N. W. [Halverson]{}, D. [Hanson]{}, N. L. [Harrington]{}, K. [Hattori]{}, J. W. [Henning]{}, G. C. [Hilton]{}, G. P. [Holder]{}, W. L. [Holzapfel]{}, K. D. [Irwin]{}, R. [Keisler]{}, L. [Knox]{}, D. [Kubik]{}, C. L. [Kuo]{}, A. T. [Lee]{}, E. M. [Leitch]{}, D. [Li]{}, M. [McDonald]{}, S. S. [Meyer]{}, J. [Montgomery]{}, M. [Myers]{}, T. [Natoli]{}, H. [Nguyen]{}, V. [Novosad]{}, S. [Padin]{}, Z. [Pan]{}, J. [Pearson]{}, C. [Reichardt]{}, J. E. [Ruhl]{}, B. R. [Saliwanchik]{}, G. [Simard]{}, G. [Smecher]{}, J. T. [Sayre]{}, E. [Shirokoff]{}, A. A. [Stark]{}, K. [Story]{}, A. [Suzuki]{}, K. L. [Thompson]{}, C. [Tucker]{}, K. [Vanderlinde]{}, J. D. [Vieira]{}, A. [Vikhlinin]{}, G. [Wang]{}, V. [Yefremenko]{}, and K. W. [Yoon]{}. . In *Society of Photo-Optical Instrumentation Engineers (SPIE) Conference Series*, volume 9153 of *Society of Photo-Optical Instrumentation Engineers (SPIE) Conference Series*, July 2014. [doi: ]{}[10.1117/12.2057305]{}.
J. E. [Carlstrom]{}, P. A. R. [Ade]{}, K. A. [Aird]{}, B. A. [Benson]{}, L. E. [Bleem]{}, S. [Busetti]{}, C. L. [Chang]{}, E. [Chauvin]{}, H.-M. [Cho]{}, T. M. [Crawford]{}, A. T. [Crites]{}, M. A. [Dobbs]{}, N. W. [Halverson]{}, S. [Heimsath]{}, W. L. [Holzapfel]{}, J. D. [Hrubes]{}, M. [Joy]{}, R. [Keisler]{}, T. M. [Lanting]{}, A. T. [Lee]{}, E. M. [Leitch]{}, J. [Leong]{}, W. [Lu]{}, M. [Lueker]{}, D. [Luong-van]{}, J. J. [McMahon]{}, J. [Mehl]{}, S. S. [Meyer]{}, J. J. [Mohr]{}, T. E. [Montroy]{}, S. [Padin]{}, T. [Plagge]{}, C. [Pryke]{}, J. E. [Ruhl]{}, K. K. [Schaffer]{}, D. [Schwan]{}, E. [Shirokoff]{}, H. G. [Spieler]{}, Z. [Staniszewski]{}, A. A. [Stark]{}, C. [Tucker]{}, K. [Vanderlinde]{}, J. D. [Vieira]{}, and R. [Williamson]{}. . **, 123:0 568–581, May 2011. [doi: ]{}[10.1086/659879]{}.
C. Dragone. A first-order treatment of aberrations in cassegrainian and gregorian antennas. *IEEE Trans. Antennas and Propagation*, AP-30:0 331, 1982.
Z. [Pan]{}, P. A. R. [Ade]{}, Z. [Ahmed]{}, A. J. [Anderson]{}, J. E. [Austermann]{}, J. S. [Avva]{}, R. B. [Thakur]{}, A. N. [Bender]{}, B. A. [Benson]{}, J. E. [Carlstrom]{}, F. W. [Carter]{}, T. [Cecil]{}, C. L. [Chang]{}, J. F. [Cliche]{}, A. [Cukierman]{}, E. V. [Denison]{}, T. [de Haan]{}, J. [Ding]{}, M. A. [Dobbs]{}, D. [Dutcher]{}, W. [Everett]{}, A. [Foster]{}, R. N. [Gannon]{}, A. [Gilbert]{}, J. C. [Groh]{}, N. W. [Halverson]{}, A. H. [Harke-Hosemann]{}, N. L. [Harrington]{}, J. W. [Henning]{}, G. C. [Hilton]{}, W. L. [Holzapfel]{}, N. [Huang]{}, K. D. [Irwin]{}, O. B. [Jeong]{}, M. [Jonas]{}, T. [Khaire]{}, A. M. [Kofman]{}, M. [Korman]{}, D. [Kubik]{}, S. [Kuhlmann]{}, C. L. [Kuo]{}, A. T. [Lee]{}, A. E. [Lowitz]{}, S. S. [Meyer]{}, D. [Michalik]{}, J. [Montgomery]{}, A. [Nadolski]{}, T. [Natoli]{}, H. [Nguyen]{}, G. I. [Noble]{}, V. [Novosad]{}, S. [Padin]{}, J. [Pearson]{}, C. M. [Posada]{}, A. [Rahlin]{}, J. E. [Ruhl]{}, L. J. [Saunders]{}, J. T. [Sayre]{}, I. [Shirley]{}, E. [Shirokoff]{}, G. [Smecher]{}, J. A. [Sobrin]{}, A. A. [Stark]{}, K. T. [Story]{}, A. [Suzuki]{}, Q. Y. [Tang]{}, K. L. [Thompson]{}, C. [Tucker]{}, L. R. [Vale]{}, K. [Vanderlinde]{}, J. D. [Vieira]{}, G. [Wang]{}, N. [Whitehorn]{}, V. [Yefremenko]{}, K. W. [Yoon]{}, and M. R. [Young]{}. . *Journal of Low Temperature Physics*, May 2018. [doi: ]{}[10.1007/s10909-018-1935-y]{}.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We study the evolution of cosmological perturbations in a non-singular bouncing cosmology with a bounce phase which has superimposed oscillations of the scale factor. We identify length scales for which the final spectrum of fluctuations obtains imprints of the non-trivial bounce dynamics. These imprints in the spectrum are manifested in the form of damped oscillation features at scales smaller than a characteristic value and an increased reddening of the spectrum at all the scales as the number of small bounces increases.'
author:
- Robert Brandenberger
- Qiuyue Liang
- 'Rudnei O. Ramos'
- Siyi Zhou
title: Fluctuations through a Vibrating Bounce
---
Introduction
============
Some bouncing cosmologies provide an alternative to cosmological inflation as a way to obtain primordial cosmological fluctuations (see, e.g., Ref. [@BP] for a recent review). Specifically, in a model which contains a matter-dominated phase of contraction, initial vacuum fluctuations in the far past which exit the Hubble radius during the matter-dominated contracting phase evolve into a scale-invariant spectrum of cosmological perturbations [@FB; @Wands]. For example, in the case of an Ekpyrotic contracting universe [@Ekp] entropy fields can source scale-invariant curvature fluctuations [@newEkp]. In all bouncing cosmologies, new physics is required in order to obtain a non-singular cosmological bounce. Such new physics could come from the matter sector (see, e.g., Refs. [@matterBounce-1; @matterBounce-2]), from modifications of the classical gravitational action (as for example in Horava-Lifshitz gravity [@HLbounce] or in the non-local gravity construction of Ref. [@Biswas]), or from quantum gravity effects. Examples of the latter are the bounce in loop quantum cosmology (see, e.g., Ref. [@LQCbounce] for reviews), in deformed AdS/CFT cosmology [@Elisa], the S-brane bounce of Ref. [@Costas] and the temperature bounce in String Gas Cosmology [@BV].
Concerning the robustness of the computations of the spectrum of cosmological fluctuations, an advantage of bouncing cosmologies (without an inflationary phase after the bounce) is that the physical length of modes which are probed in current observations remain in the far infrared throughout the cosmological evolution as long as the energy density at the bounce point is smaller than the Planck density. Hence, the computations can be done in the realm where effective field theory is well justified. This is in contrast to the situation in inflationary cosmology [@MB] where the physical wavelengths of even the largest scales which are currently observed are smaller than the Planck length at the beginning of inflation (provided that the inflationary phase lasts slightly longer than the minimal period it has to last in order to solve the horizon and flatness problems of Standard Big Bang Cosmology).
A key question is to whether the predictions for cosmological perturbations at late times in the expanding phase are sensitive to the details of the bounce phase. For simple parametrizations of the bounce phase, detailed studies have shown that the spectral shape does not change during the bounce phase provided that the duration of the bounce phase is shorter than the length scale of the fluctuations at the bounce point (see, e.g., Ref. [@matterBounce-2] in the case of matter-driven bounces, Ref. [@HLflucts] in the case of the Horava-Lifshitz bounce, Ref. [@Elisa] in the case of the AdS/CFT bounce, and Ref. [@Subodh] for the S-brane bounce). On the other hand, there are examples where the bounce phase yields dramatic changes in the spectrum [@BXue]. The reason why such dramatic changes are possible is that the Hubble radius at the bounce point is infinite, and we cannot invoke the freezing of cosmological perturbations on super-Hubble scales to argue for a constancy of the spectrum[^1].
To further analyze the sensitivity of the spectrum of cosmological fluctuations on the details of the bounce phase, we here consider a toy model where the scale factor undergoes small amplitude oscillations during the bounce phase. Such a behavior may emerge from certain models motivated by ideas from loop quantum gravity [@Alesci]. Heuristically, one would argue that those small bounces will not influence the large scale modes provided that the wavelengths of these modes are so large scale that they would not feel the small scale fluctuations of the scale factor. On the other hand, smaller scale modes whose wavelength is comparable or smaller to the total duration of the bounce phase should be sensitive to the details of the dynamics during the bounce. In this work, we would like to give a careful treatment to see if this is really the case.
In the case of a cyclic cosmology, when the time interval between cycles is larger than the wavelength of the modes being considered, it is generally sufficient [@RHBcyclic] to consider only the dominant modes in each phase (except during the bounce phase and when mode matching conditions are applied [@Durrer]). In our case however, the time scale between the small bounces is very small compared to the length scales of interest, and hence we cannot just focus on the dominant modes because for those small time durations the subdominant modes can also have an effect on the primordial power spectra. We need to keep all the contributions and give a comprehensive analysis.
This paper is organized as follows. In Sec. \[sec2\], we specify our setup for the intermediate small bounce feature and discuss about the relevant scales involved. In Sec. \[sec3\], we present the calculation of power spectrum. A detailed presentation of the required matching conditions is given and the specific results are given for the two specific small inter bounce features we considered. An analysis of these results for the power spectrum is then given in Sec. \[sec4\]. In Sec. \[sec5\], we give a generalization for the case of a large number of small bounces. In Sec. \[sec6\] we present our conclusions. An appendix is included to discuss some of the technical details.
Setup {#sec2}
=====
We will consider a spatially flat Friedmann-Lemaitre-Robertson-Walker space-time in which the metric is given by $$ds^2 \, = \, dt^2 - a(t)^2 d{\bf x}^2 \, ,$$ where $t$ is physical time, ${\bf x}$ are the comoving spatial coordinates, and $a(t)$ is the cosmological scale factor. It will be convenient to use conformal time $\tau$ related to the physical time via $dt = a(t) d\tau$. We will consider only linear cosmological perturbations (see, e.g., Ref. [@MFB] for a comprehensive review). In this case, fluctuations evolve independently in Fourier space. We will label the fluctuation modes in terms of their comoving wave number $k$.
We consider a non-singular symmetric bouncing cosmology in which the cosmological scale factor $a(\tau)$ has the form shown in Fig. \[bounceplot\], i.e., for which $a(\tau)$ has one “oscillation” between the onset and end of the bounce phase. We consider the two forms depicted in Fig. \[bounceplot\], with one peak between the time $- \tau_B$ when $a(\tau)$ reaches its first minimum, and the time $+ \tau_B$ when the second minimum of $a(\tau)$ is taken on. Specifically, we consider two specific models: Model 1, with a flat plateau about $\tau = 0$; and the Model 2, with a kink of $a(\tau)$ at $\tau = 0$, which is the limiting case of the first model when the duration of the flat plateau equals to zero. The time interval of the plateau region is $-\Delta < \tau < \Delta$ given by some conformal time $\Delta$ with $\Delta < \tau_B$. The forms shown in Fig. \[bounceplot\] are simpler enough such as to allow an analytical study, but already of sufficient complexity such as to provide the main relevant features in the power spectrum that we might also observe in some more complex setup for the inter-bounce features. In particular, our results can be easily generalized to the case of many oscillations, as we will later discuss in Sec. \[sec5\].
As seen in Fig. \[bounceplot\], for the two forms shown the time interval can be divided into five intervals. The first is the initial contracting phase (Phase I) $\tau < - \tau_B$. The second is the intermediate expanding phase (Phase II). The third is Phase III with static scale factor, the fourth (Phase IV) is the intermediate contracting phase and Phase V is the final expanding phase. Fluctuation modes exit the Hubble radius in the initial contracting phase. The top panels in Fig. \[bounceplot\] give a sketch of the scale factor, the lower panels show the corresponding time evolution of the comoving Hubble radius. The vertical axis of the lower panels can also be viewed as a label for comoving wavelength. In this way, it is easy to read off when various modes enter and exit the Hubble radius.
All scales re-enter and re-exit the Hubble radius several times since at the extrema of $a(t)$ the Hubble radius is infinite. We will treat the transitions at $-\tau_B$ and $\tau_B$ as instantaneous[^2]. More specifically, we will cut out a time interval $$-\tau_B - \epsilon < \tau < - \tau_B + \epsilon,$$ (with $\epsilon \ll \tau_B$) and correspondingly another time interval of the same length about $\tau_B$ and we will match the solutions between the neighboring phases making use of the matching conditions given in Refs. [@HV; @DM], which are the cosmological version of the Israel [@Israel] ones[^3]. Hence, the only Hubble re-entry which is important to us is the one which occurs between $-
\tau_B$ and $+ \tau_B$.
In the first model, given by the plots on the left shown in Fig. \[bounceplot\], there are two characteristic comoving length scales. The first is $k_*^{-1}$ which is defined as the length which re-enters the Hubble radius at time $- \Delta$. The second one, $k_{\Delta}^{-1}$, is the mode which undergoes one oscillation between $\tau = - \Delta$ and $\tau = + \Delta$ (we are assuming $k_{\Delta}^{-1} > k_*^{-1}$ if this is not satisfied then we recover the results for the second model, given by the plots on the right shown in Fig. \[bounceplot\]). Modes with wavelength smaller than $k_*^{-1}$ enter the Hubble radius during Phase II and exit again during Phase IV. Modes with $k_*^{-1} < k^{-1}
< k_{\Delta}^{-1}$ are inside the Hubble radius only during Phase III. For these modes the matching occurs at times $- \Delta$ and $+
\Delta$. This is also true for modes with $k^{-1} > k_{\Delta}^{-1}$. These modes, however, undergo a negligible amount of oscillations in Phase III. In the case that $k_\Delta^{-1}>k_*^{-1}$, we have three different behaviors of the power spectrum. For the very large scale modes $k^{-1}\gg k_\Delta^{-1}$, the power spectrum does not feel the influence of the small bump of the scale factor. For the modes $k_*^{-1}<k^{-1}<k_\Delta^{-1}$, there is a complicated change of the power spectrum induced by the flat plateau. For the modes $k^{-1}<k_*^{-1}$, the change of the power spectrum approaches the well known result for cyclic cosmologies [@RHBcyclic], as we will explicitly verify later on below. In the case that $k_\Delta^{-1}<k_*^{-1}$, there is only one characteristic scale $k_*^{-1}$. The mode with $k^{-1}>k_*^{-1}$ will not feel the influence of the bump, while the mode with $k^{-1}<k_*^{-1}$ will be changed by the bump according to the well known result for cyclic cosmologies. A special situation belonging to this case is the limiting case $\Delta \rightarrow 0$.
We divide the evolution of fluctuation modes into five regions as shown on Fig. \[bounceplot\] [^4]. The five regions are denoted by Region I, Region II, Region III, Region IV and Region V, respectively. Region I and Region II are separated by the time $-\tau_B$. Region IV and Region V are separated by $\tau_B$. The separation between Regions II and III, and between Regions III and IV are more complicated. Because of the existence of the flat plateau (or of the local maximum of the comoving Hubble radius in the case of Model 2), we can see that there is a clear distinction between the large scale and small scale modes separated by a characteristic scale $k_*^{-1}$. Small scale modes (i.e. $k^{-1}<k_*^{-1}$) enter the Hubble radius at time $-\tau_H(k)
\leq - \Delta$ and exit the Hubble radius at $\tau_H(k) > \Delta$, and the separation between Regions II and III and between Regions III and IV are given by $-\tau_H(k)$ and $\tau_H(k)$, respectively. For large scale modes, the separations between Regions II and III, and between Regions III and IV are given by the times $-\Delta$ and $\Delta$, respectively, because the modes enter and exit the Hubble radius at these two times. The situation in the case of Model 2 is simpler. The evolution of small scale modes is the same as the case with a flat plateau, while the large scale modes have only four regions which we denote by Regions I, II, IV and V, respectively. The separation between Regions II and IV, in this case, is the time $\tau= 0$.
The Computation of the Power Spectrum {#sec3}
=====================================
We are interested in the power spectrum of the primordial curvature perturbation $\zeta$ (see, e.g., Ref. [@MFB] for a review of the theory of cosmological perturbations). We quantize the linear fluctuations and write them in terms of the more convenient Mukhanov-Sasaki variable $v$. In the case of a constant equation of state, the relation between $\zeta$ and $v$ is $$\begin{aligned}
v = C a \zeta \, ,\end{aligned}$$ where $C$ is a constant. Thus, the equation of motion for the mode function $v$, in momentum space, is given by $$\begin{aligned}
v'' + \bigg(k^2 - \frac{a''}{a} \bigg) v = 0 \, .\end{aligned}$$ For the scale factor $a\sim \tau^q$, the solution of the Mukhanov-Sasaki equation is given by $$\begin{aligned}
v (\tau) = c_1(k) \sqrt{\tau} J_{\alpha} (k\tau) + c_2(k) \sqrt{\tau}
Y_{\alpha} (k\tau),\quad \alpha\equiv\ q-\frac{1}{2} ,\end{aligned}$$ where $J(x)$ and $Y(x)$ are Bessel functions of the first and second kind, respectively. On sub-Hubble scales, the solutions are oscillatory, on super-Hubble scales they can be approximated by a power law. To see this, we note that the expansion of the Bessel function solutions for small argument, $x\ll 1$, is $$\begin{aligned}
J_\alpha (x) &= \sum_{m = 0}^\infty
\frac{(-1)^m}{m!\Gamma(m+\alpha+1)} \left( \frac{x}{2}
\right)^{2m+\alpha}, \\ Y_\alpha (x) &=
\frac{\cos(\alpha\pi)}{\sin(\alpha\pi)} J_\alpha(x) -
\frac{1}{\sin(\alpha\pi)}J_{-\alpha}(x) \nonumber \\ &=
\frac{\cos(\alpha \pi)}{\sin(\alpha \pi)} \sum_{m = 0}^\infty
\frac{(-1)^m}{m!\Gamma(m+\alpha+1)} \left( \frac{x}{2}
\right)^{2m+\alpha} - \frac{1}{\sin(\alpha\pi)}\sum_{m = 0}^\infty
\frac{(-1)^m}{m!\Gamma(m-\alpha+1)} \left( \frac{x}{2}
\right)^{2m-\alpha},\end{aligned}$$ we can express the mode function in terms of a series expansion $$\begin{aligned}
v(\tau) = \sum_{m=0}^{\infty} d_{1m} (k) \tau^{q+2m} +
\sum_{m=0}^{\infty} d_{2m} (k) \tau^{1-q+2m} ,
\label{vtaum}\end{aligned}$$ where $d_{1m}(k)$ and $d_{2m}(k)$ are given, respectively, by $$\begin{aligned}
d_{1m}(k) & = \left[ c_1(k) + \frac{\cos(\alpha\pi)}{\sin(\alpha\pi)}
c_2(k) \right] \frac{(-1)^m}{m!\Gamma(m+\alpha+1)} \left(
\frac{k}{2} \right)^{2m+\alpha}, \\ d_{2m}(k) & =
-\frac{c_2(k)}{\sin(\alpha\pi)} \frac{(-1)^m}{m!\Gamma(m-\alpha+1)}
\left( \frac{k}{2} \right)^{2m-\alpha} .\end{aligned}$$ Since we are interested in those modes that went classical (crossed the Hubble radius), such that $k \tau \ll 1$, the higher order terms of $v(\tau)$ given by $m> 0$ are subleading. Thus, in the following, we can just focus on the $m=0$ terms in Eq. (\[vtaum\]).
The scale factors and the solutions to the mode functions of the five regions can be obtained by shifting the time coordinate. Thus, they are given, respectively, by
- Region I (contracting), where $\tau<-\tau_B$, we have that
$$\begin{aligned}
a \sim (-\tau-\tau_B)^{q_2} \sim(-t-t_B)^{p_2}, \quad v_{1} = c_{11}
(-\tau_B-\tau)^{1-q_2} + c_{12} (-\tau_B-\tau)^{q_2},\end{aligned}$$
- Region II (expanding), where $ -\tau_B<\tau<-\Delta\,\,\, {\rm
for}\,\,\, k^{-1}>k_*^{-1}, \,\,\, {\rm
and}\,\,\,-\tau_B<\tau<-\tau_H \,\,\, {\rm for}\,\,\,
k^{-1}<k_*^{-1}$, we have that
$$\begin{aligned}
a\sim (\tau+\tau_B)^{q_1}\sim(t+t_B)^{p_1},\quad v_2 =
c_{21}(\tau+\tau_B)^{1-q_1} + c_{22} (\tau+\tau_B)^{q_1},\end{aligned}$$
- Region III (intermediate), where $-\Delta<\tau<\Delta\,\,\, {\rm
for}\,\,\, k^{-1}>k_*^{-1}, \,\,\, {\rm and}\,\,\,
-\tau_H<\tau<\tau_H \,\,\, {\rm for}\,\,\, k^{-1}<k_*^{-1}$, we have that
$$\begin{aligned}
a\sim {\rm constant\,\,\,in\,\,\,-\Delta<\tau<\Delta} , \quad v_3 =
c_{31} e^{i k \tau} + c_{32} e^{-i k \tau},\end{aligned}$$
- Region IV (contracting), where $\Delta<\tau<\tau_B\,\,\, {\rm
for}\,\,\, k^{-1}>k_*^{-1}, \,\,\, {\rm and}\,\,\,
\tau_H<\tau<\tau_B \,\,\, {\rm for}\,\,\, k^{-1}<k_*^{-1}$, we have that
$$\begin{aligned}
a \sim (-\tau+\tau_B)^{q_1} \sim (-t+t_B)^{p_1},\quad v_4 = c_{41}
(-\tau+\tau_B)^{1-q_1} +c_{42} (-\tau+\tau_B)^{q_1},\end{aligned}$$
- Region V (expanding), where $\tau>\tau_B$, we have that
$$\begin{aligned}
a\sim (\tau-\tau_B)^{q_2} \sim (t-t_B)^{p_2},\quad v_5 = c_{51} (\tau
- \tau_B)^{1-q_2} + c_{52} (\tau-\tau_B)^{q_2} .\end{aligned}$$
Note that in the above, the time $\tau_H$ depends on $k$, $\tau_H\equiv \tau_H(k)$. To simplify the notation we do not write the k-dependence explicitly. Note also that $q_2$ (the power of the scale factor in conformal time) and $p_2$ (the power of the scale factor in cosmological time) are the indices of the scale factor during the initial contracting and final expanding phase, and that $q_1$ and $p_1$ are the indices in the intervening periods. In particular, the index 1 in these quantities should not be confused with the index in Region I. Note also that the $q_i$ are related to the $p_i$ through the relation
$$q_i = \frac{p_i}{1-p_i},\;\;\;i=1,2.
\label{qipi}$$
Matching conditions
-------------------
Let us now discuss the process of matching between the different regions discussed above. The matching conditions for the metric across a space-like hypersurface were derived in Hwang-Vishniac [@HV] and Deruelle-Mukhanov [@DM] and are the generalization of the Israel matching conditions [@Israel]. For cosmological fluctuations, the matching conditions say the solutions in different regions can be connected by enforcing two conditions, namely the continuity of both $v$ and its derivative across the boundary surface.
As mentioned earlier, for the two non-singular bouncing points $-
\tau_B$ and $\tau_B$ we match the solutions at times $\mp\tau_B-\epsilon$ and $\mp\tau_B+\epsilon$, neglecting any evolution in the intervening time period. This is similar to what was done in Refs. [@Elisa; @Subodh]. A second justification of this method (in addition to the one given earlier) is that for modes we are interested in, the time interval $2\epsilon$ is so small, thus the mode functions do not have enough time to oscillate inside the Hubble radius. On the other hand, in our first model (with a flat plateau for $a(t)$, model 1) we consider the case that the interval $2\Delta$ is sufficiently long such that some of the modes we are interested in have time to oscillate while the mode is inside the Hubble radius. Very large scale modes, on the other hand, still do not oscillate inside the Hubble radius.
### Between region I and region II
The matching conditions are $$\begin{aligned}
v_1(-\tau_B-\epsilon) = v_2 (-\tau_B+\epsilon), \quad
v'_1(-\tau_B-\epsilon) = v'_2 (-\tau_B+\epsilon) \, .\end{aligned}$$ We write down the equations explicitly in terms of coefficients $c_{ij}$ of the fundamental solutions of the equation of motion. The index $i$ stands for the region, the index $j$ (either $1$ or $2$) running over the two different modes:
$$\begin{aligned}
\begin{pmatrix}
\epsilon^{1-q_2} & \epsilon^{q_2} \\ -(1-q_2) \epsilon^{-q_2} &
-q_2\epsilon^{q_2-1}
\end{pmatrix} \begin{pmatrix}
c_{11} \\ c_{12}
\end{pmatrix} = \begin{pmatrix}
\epsilon^{1-q_1} & \epsilon^{q_1} \\ (1-q_1) \epsilon^{-q_1} & q_1
\epsilon^{q_1-1}
\end{pmatrix} \begin{pmatrix}
c_{21} \\ c_{22}
\end{pmatrix} .\end{aligned}$$
### Between region II and region III
To be explicit, we focus in this case on the large scale modes for which the time duration of Region III is $-\Delta<\tau<\Delta$. For the small scale modes which enter the Hubble radius before $- \Delta$, we just make the substitution $\Delta\rightarrow\tau_H(k)$. Apart from that the discussion is the same. For the next subsection the convention will be the same. The matching conditions in this case are $$\begin{aligned}
v_2(-\Delta) = v_3 (-\Delta),\quad v'_2(-\Delta) = v'_3 (-\Delta),\end{aligned}$$ which in matrix form can be expressed as
$$\begin{aligned}
\begin{pmatrix}
\delta^{1-q_1} & \delta^{q_1} \\ (1-q_1) \delta^{-q_1} & q_1
\delta^{q_1-1}
\end{pmatrix} \begin{pmatrix}
c_{21} \\ c_{22}
\end{pmatrix} = \begin{pmatrix}
e^{-ik \Delta} & e^{i k \Delta} \\ i k e^{-ik\Delta} & - i k e^{i k
\Delta}
\end{pmatrix} \begin{pmatrix}
c_{31} \\ c_{32}
\end{pmatrix},\end{aligned}$$
and where in the above equation we have defined $\delta$ as
$$\delta = \tau_B - \Delta .
\label{delta}$$
Note that for the small scale modes the definition of $\delta$ should be changed to $\delta\to \delta(k) \equiv \tau_B-\tau_H(k)$.
### Between region III and region IV
The matching conditions are $$\begin{aligned}
v_3(\Delta) = v_4 (\Delta), \quad v'_3(\Delta) = v'_4 (\Delta) \, .\end{aligned}$$ In matrix form this yields $$\begin{aligned}
\begin{pmatrix}
e^{ik\Delta} & e^{-ik\Delta} \\ i k e^{ik\Delta} & -i k e^{-ik\Delta}
\end{pmatrix} \begin{pmatrix}
c_{31} \\ c_{32}
\end{pmatrix} = \begin{pmatrix}
\delta^{1-q_1} & \delta^{q_1} \\ -(1-q_1) \delta^{-q_1} &
-q_1\delta^{q_1-1}
\end{pmatrix} \begin{pmatrix}
c_{41} \\ c_{42}
\end{pmatrix}.\end{aligned}$$
### Between region IV and region V
The matching conditions in this case are $$\begin{aligned}
v_4(\tau_B-\epsilon) = v_5 (\tau_B+\epsilon),\quad
v'_4(\tau_B-\epsilon) = v'_5 (\tau_B+\epsilon) \, .\end{aligned}$$ In matrix form this yields $$\begin{aligned}
\begin{pmatrix}
\epsilon^{1-q_1} & \epsilon^{q_1} \\ -(1-q_1) \epsilon^{-q_1} & -q_1
\epsilon^{q_1-1}
\end{pmatrix} \begin{pmatrix}
c_{41} \\ c_{42}
\end{pmatrix} = \begin{pmatrix}
\epsilon^{1-q_2} & \epsilon^{q_2} \\ (1-q_2) \epsilon^{-q_2} & q_2
\epsilon^{q_2-1}
\end{pmatrix} \begin{pmatrix}
c_{51} \\ c_{52}
\end{pmatrix}.\end{aligned}$$
Analysis and Results for the Power Spectrum {#sec4}
===========================================
Combining the results of the previous section we find that the final mode coefficients can be written in terms of the initial ones via $$\begin{aligned}
\nonumber
& \mathcal C_5 =
\begin{pmatrix}
c_{51} \\ c_{52}
\end{pmatrix},\end{aligned}$$ where[^5]
$$\begin{aligned}
c_{51}&=&\frac{c_{11}}{(1-2q_2)(1-2q_1)2k(1-2q_1)} \left[ a_{11} (1 -
2 (q_1-1) q_1 - 2 (q_2 -1 ) q_2) \right. \nonumber \\ &-&
\left. a_{12} ( (q_1 - q_2 - 1) (q_1 + q_2 - 2) \epsilon^{1 - 2 q_1}
) + a_{21} (q_1 - q_2 + 1) (q_1 + q_2) \epsilon^{2 q_1 - 1} \right],
\label{c51}
\\ c_{52}&=&\frac{c_{11}}{(1-2q_2)(1-2q_1)2k(1-2q_1)} \left[ 2 a_{11}
(1 + q_1 - q_2) (-2 + q_1 + q_2) \epsilon^{1 - 2 q_2} \right.
\nonumber \\ &+& \left. a_{12} (-2 + q_1 + q_2)^2 \epsilon^{-2 (-1 +
q_1 + q_2)} + a_{21} (1 + q_1 - q_2)^2 \epsilon^{2 (q_1 -
q_2)}\right],
\label{c52}\end{aligned}$$
and $$\begin{aligned}
\label{a11}
a_{11} & = \frac{2\left[k^2\delta^2+(q_1-1)q_1\right]\sin
(2k\Delta)}{\delta} - 2 k \cos (2 k \Delta), \\
\label{a12}
a_{12} & = \delta^{2q_1-2}
\left[-4kq_1\delta\cos(2k\Delta)-2(q_1-k\delta)(q_1+k\delta)
\sin(2k\Delta)\right],
\\
\label{a21}
a_{21} & =
\delta^{-2q_1}\left[2(q_1-k\delta-1)(q_1+k\delta-1)\sin(2k\Delta)-
4k(q_1-1)\delta\cos(2k\Delta)\right]
.\end{aligned}$$ Note that these coefficients oscillate as a function of $k$. These oscillations are important, however, only for small wavelength fluctuations. For these we will obtain oscillations in the power spectrum. The final general result for the power spectrum is given by $$\begin{aligned}
P_{\zeta} & = \zeta^2 k^{3} \sim \left( \frac{v}{z} \right)^2 k^{3}
\nonumber \\ & = \left[
\frac{c_{51}(\tau-\tau_B)^{1-q_2}+c_{52}(\tau-\tau_B)^{q_2}}{(\tau-\tau_B)^{q_2}}
\right]^2 k^{3} \nonumber \\ &= \left[ c_{51}
(\tau-\tau_B)^{1-2q_2} + c_{52} \right]^2 k^{3} \, .
\label{finalPzeta}\end{aligned}$$ Below we will analyze some of the specific cases given by our two models when applying the result given by Eq. (\[finalPzeta\]).
Limiting case of Instantaneous matching
---------------------------------------
We first consider the limit as the duration of the plateau region of $a(t)$ goes to zero, corresponding to what we have denoted by Model 2 in Sec. \[sec2\]. This is the limit $\Delta\rightarrow 0$. In this case, large scale modes $k^{-1} > k_*^{-1}$ do not enter the Hubble radius in the region near $t = 0$, and we can set $\Delta = 0$ in the matching condition equations, i.e.,
$$\begin{aligned}
\sin(2k\Delta) \rightarrow 0, \quad \cos(2 k \Delta) \rightarrow 1 \,
.\end{aligned}$$
On the other hand, small scale modes $k^{-1} < k_*^{-1}$ will enter the Hubble radius at a time given by $-\tau_H(k)$, and in the matching condition equations we must replace $\Delta$ by $\tau_H(k)$.
### Large scale modes $k^{-1} > k_*^{-1}$
Let us first consider the case for large scale modes $k^{-1} >
k_*^{-1}$. In this case we have that $$\begin{aligned}
a_{11}\rightarrow -2k, \quad a_{12}\rightarrow -4k q_1\delta^{2q_1-1},
\quad a_{21}\rightarrow -4k (q_1-1)\delta^{-2q_1+1}.\end{aligned}$$ Because we are interested in the parameter region $1/3<p<1$, then, written in terms of $q$, we have $q>1/2$. So the $c_{51}$ mode in the expression for the power spectrum Eq. (\[finalPzeta\]) is a decaying solution. Hence, we can focus on the constant mode $c_{52}$, and thus the power spectrum in this case becomes $$\begin{aligned}
\nonumber
&P_{\zeta} \sim c_{52}^2 k^{3} =\left\{
\frac{c_{11}}{(1-2q_2)(1-2q_1)2(1-2q_1)} \right. \\ \nonumber &
\left. \times \left[ -4 (1 + q_1 - q_2) (-2 + q_1 + q_2)
\epsilon^{1 - 2 q_2}-4q_1\delta^{2q_1-1} (-2 + q_1 + q_2)^2
\epsilon^{-2 (-1 + q_1 + q_2)} \right. \right. \nonumber \\ &
\left. \left. -4(q_1-1)\delta^{-2q_1+1} (1 + q_1 - q_2)^2
\epsilon^{2 (q_1 - q_2) } \right] \right\}^2 k^{3} .\end{aligned}$$
The initial power spectrum is $$\begin{aligned}
P_i = P_{\zeta} (-\tau_B-\epsilon) = \zeta^2 k^{3} = c_{11}^2
\epsilon^{2-4q_2} k^{3}\end{aligned}$$ and, thus, we can relate the final to the initial power spectrum as $$\begin{aligned}
\label{powerspectrum}
P_{\zeta} = \left(A_1 + A_2 \delta^{2q_1-1} \epsilon^{1-2 q_1 } +
A_3\delta^{-2q_1+1} \epsilon^{2q_1-1} \right)^2 P_i \, ,\end{aligned}$$ where $A_1$, $A_2$ and $A_3$ are constants that do not depend on $k$. Their explicit forms are $$\begin{aligned}
&A_1 = \frac{-2(1+q_1-q_2)(-2+q_1+q_2)}{(1-2q_2)(1-2q_1)^2},\\ &A_2 =
\frac{-2q_1(-2+q_1+q_2)^2}{(1-2q_2)(1-2q_1)^2}, \\ &A_3 =
\frac{-2(q_1-1)(1+q_1-q_2)^2}{(1-2q_2)(1-2q_1)^2}.\end{aligned}$$
For very large scale modes $k^{-1} \gg k_*^{-1}$, $\delta \to
\tau_B$ and $\delta$ can be approximated as a constant time interval. Thus, the power spectrum in this case becomes
$$\begin{aligned}
\label{powerspectrum1}
(A_1 + A_2 \tau_B^{2q_1-1} \epsilon^{1-2 q_1 } + A_3
\tau_B^{-2q_1+1} \epsilon^{2q_1-1} )^2 P_i .\end{aligned}$$
The first conclusion we draw from this result is that the shape of the spectrum for large scale modes does not change during the bounce. This agrees with the conclusions of previous work on simple bounce models [@matterBounce-2]. The amplitude, on the other hand, is amplified. Recall that $2q_i - 1 > 0$, and that $\epsilon \ll \tau_B$. Hence, it is the second term in Eq. (\[powerspectrum1\]) which dominates, and we conclude that the amplitude of the spectrum is amplified by a factor of $${\cal A} \, = \, A_2^2 \bigg(\frac{\tau_B}{\epsilon}\bigg)^{4q_1 - 2} .
\label{Amplitude}$$ This result can also be understood easily: Fluctuations grow both in the contracting and in the expanding phase. In fact, the fluctuations diverge in the limit when the scale factor becomes zero. Hence, without an effective cutoff $\epsilon$ we would get a divergence in the spectrum. With a cutoff, the enhancement factor of the amplitude of the power spectrum will be determined by the dimensionless ratio between $\tau_B$ and $\epsilon$ to a power which depends on the growth rate of the fluctuations on super-Hubble scales, i.e., on $q_1$ (see the discussion of these issues in a more general context in the review article Ref. [@RHBbounceReview]).
### Small scale modes $k^{-1} < k_*^{-1}$
For small scale modes $k^{-1}<k_{*}^{-1}$, we set $\delta$ equal to the Hubble crossing time. Thus, we can use the Hubble crossing condition $a H= k$, which from $a\sim \tau^q$ gives $$\begin{aligned}
\delta =q_1 k^{-1} .\end{aligned}$$ But we need to have $k \tau_H = k\tau_B - k\delta= k\tau_B - q_1 $. As a consequence of the oscillations in the coefficients $a_{ij}$, Eqs. (\[a11\]), (\[a12\]) and (\[a21\]), the final power spectrum of fluctuations will oscillate for small wavelengths. This is explicitly manifested when we show a numerical example for the power spectrum in Fig. \[bounceplot1\], where we chose an initial pre-bounce spectrum which is scale-invariant. We see that the scale-invariance of the spectrum is maintained on large scales, but that on small length scales there is both a change in the slope of the spectrum, and superimposed oscillations.
In the following we discuss in what range we can reproduce the results of Ref. [@RHBcyclic], which hold for a cyclic cosmology. In that work, it was found that for modes which re-enter the Hubble radius during the bounce phase, the index of the spectrum of cosmological perturbations changes during each cycle. For a matter-dominated contracting phase the change in the index $n_s$ of the power spectrum was determined to be $\Delta n_s = - 2$.
The results of Ref. [@RHBcyclic] are applicable when $k^{-1}<k_*^{-1}$, but for quite large scales such that $k^{-1}\rightarrow k_*^{-1}$. In this range we have $k\tau_H\rightarrow 0$. Using this in Eq. (\[powerspectrum\]), we obtain that $$\begin{aligned}
\label{powerspectrum_cyclic}
P_f \, = \, \left[ A_1 + A_2 (k \epsilon/q_1)^{1-2q_1} + A_3
(k\epsilon/q_1)^{2q_1-1} \right]^2 P_i.\end{aligned}$$ Since $2q_1 - 1 > 0$ it is the second term in Eq. (\[powerspectrum\_cyclic\]) which dominates. Hence, we conclude that there is a change in the index of the power spectrum by $$\Delta n_s = 2-4 q_1= - 2 \frac{3 p_1 - 1}{1 - p_1} ,
\label{tiltns}$$ which coincides with the results of Ref. [@RHBcyclic]. This is as expected because the case studied in Ref. [@RHBcyclic] corresponds to a big bounce where $\delta$ is (cosmologically) large.
Case with a flat plateau
------------------------
In the case with a flat plateau and when $\Delta$ is very small, we have just one characteristic comoving mometum. However, when $\Delta$ is big, we have two key comoving momenta which are characterized by the mode which cross the Hubble radius at $\Delta$ and $\tau_B-
\Delta$, respectively. In this subsection, we would like to analyze in detail these two cases.
First we would like to calculate the critical comoving momentum $k_*^{-1}$. We start by analyzing the Hubble parameter $H$. The corresponding comoving Hubble parameter in region II is
$$\begin{aligned}
aH =q_1 (\tau+\tau_B)^{-1}.\end{aligned}$$
The critical scale $k_{*}$, which is obtained by $k_*=aH(\tau=-\Delta)$, is therefore $$\begin{aligned}
k_{*} = q_1 (\tau_B - \Delta)^{-1}.\end{aligned}$$ The analysis here is similar to the instantaneous matching case of the previous subsection and we can obtain the power spectrum as $$\begin{aligned}
&P_{\zeta} \sim c_{52}^2 k^{3} = \left\{ \frac{ 1
}{(1-2q_2)(1-2q_1)2k(1-2q_1)} \right. \nonumber \\ & \left. \times
\left[ 2 a_{11} (1 + q_1 - q_2) (-2 + q_1 + q_2) + a_{12} (-2 +
q_1 + q_2)^2 \epsilon^{1 -2 q_1 } \right. \right. \nonumber \\ &
\left. \left. + a_{21} (1 + q_1 - q_2)^2 \epsilon^{2 q_1 -1}
\right] \right\}^2 P_i ,
\label{powerspectrumforinstantaneous}\end{aligned}$$ where $a_{11}$, $a_{12}$ and $a_{21}$ were already defined by Eqs. , and , respectively.
Model with no Region III
------------------------
Let us here consider the model with no Region III. We expect that the result we obtained in the previous Subsection will approach the result derived here in the limit when $\Delta\rightarrow 0$. The matching condition of Region I and II, Region IV and V are completely the same as in the flat plateau case, so here we only write down the matching condition between Region II and IV:
$$\begin{aligned}
v_2(0) = v_4 (0),\quad v'_2(0) = v'_4 (0) ,\end{aligned}$$
which can be written in terms of the more convenient matrix form $$\begin{aligned}
\begin{pmatrix}
\tau_B^{1-q_1} & \tau_B^{q_1} \\ (1-q_1) \tau_B^{-q_1} & q_1
\tau_B^{q_1-1}
\end{pmatrix} \begin{pmatrix}
c_{21} \\ c_{22}
\end{pmatrix} = \begin{pmatrix}
\tau_B^{1-q_1} & \tau_B^{q_1} \\ -(1-q_1) \tau_B^{-q_1} & -q_1
\tau_B^{q_1-1}
\end{pmatrix} \begin{pmatrix}
c_{41} \\ c_{42}
\end{pmatrix}.\end{aligned}$$ Combining these matching results we obtain the power spectrum completely the same as that of the instantaneous matching of the previous section.
Numerical Examples
------------------
In Fig. \[bounceplot1\] we show the form of the spectrum of cosmological perturbations and its tilt as a function of the comoving wavenumber $k$ for the two models we have considered in the previous sections.
We can clearly identify in Fig. \[bounceplot1\](a) the characteristic scales for each of the two models we have defined in Sec. \[sec2\]. For the flat plateau model (Model 1), there are two relevant comoving scales,
$$\begin{aligned}
&&k_{*,{\rm flat}}= \frac{q_1}{\tau_B-\Delta},
\label{kflat}
\\ &&k_{*,{\rm osc}}= \frac{q_1}{\tau_B}.
\label{kosc}\end{aligned}$$
In model 1 the spectrum is always evolving. On large scales $k <
k_{*,{\rm osc}}$ there is both an amplification of the spectrum and a damping evolution. On small scales $k > k_{*,{\rm osc}}$ the power spectrum shows superimposed damped oscillations.
In the instantaneous case $\Delta=0$ (Model 2), the characteristic comoving scale is $k_{*,{\rm inst}} \equiv k_{*,{\rm osc}}$, the same as Eq. (\[kosc\]). In the model 2, on large scales $k < k_{*,{\rm
inst}}$ the spectral shape is unchanged during the bounce and only the amplitude increases, as identified in Eq. (\[Amplitude\]). On smaller scales $k > k_{*,{\rm inst}}$ there is a change in the spectral index and the power spectrum, as in the case of model 1, shows superimposed damped oscillations.
The results Fig. \[bounceplot1\](b) show that for Model 1 (black dashed line) the spectral tilt always decreases with the momentum. The discontinuity at $k_{*,{\rm flat}}$ (denoted by the black dashed vertical line) is an unphysical feature that appears as a consequence of the shape we have considered and should not appear in realistic smooth shapes. The same is true for the Model 2 case (red solid line), where the discontinuity happens at the characteristic scale $k_{*,{\rm
inst}}$ in this case and comes from the kink like shape considered in this model. Other than that, the spectral index is unchanged (and null) for large scales modes $k< k_{*,{\rm inst}}$ and then decreases for small scale modes $k> k_{*,{\rm inst}}$ and agrees with that of model 1 from this point on, where the index of the power spectrum for both models acquires a large red tilt, and there are superimposed oscillations.
![\[bounceplot2\] The power spectrum for different parameters, namely $q_1= q_2 = 2/3$, $\epsilon = 0.01 \tau_B$, for the cases of $\Delta=0.7 \tau_B$ (blue dotted line), and $\Delta=0.9
\tau_B$ (red solid line) for the case of the flat plateau (Model 1) and for the instantaneous case (Model 2), where $\Delta=0$ (black dashed line). Again, the vertical lines denote the positions of the characteristic scales $k_*$ for each model.](bouncefig3.pdf){width="45.00000%"}
Figure \[bounceplot2\] shows the results for different parameter values, parameters for which the two scales $k_*$ for the models with and without a plateau for $a(t)$ are more widely separated than they are for the earlier parameter values. For both models there are oscillations of the power spectrum for $k$ values between the two critical $k_*$ values. These results in particular show that as $\Delta \to \tau_B$, the scale $k_{*,{\rm flat}}$ can occur deeper in the oscillating regime $k>k_{*,{\rm osc}}$ for the spectrum.
A study of the oscillating regime for small scales $k>k_{*,{\rm
osc}}$, and which is common for both models considered here, is given in the Appendix \[appA\]. In particular, it is shown that the envelope function of the power spectrum for the small scale modes keeps the spectral tilt $n_s = -4q_1+2$, as also seen in the previous Eq. (\[tiltns\]).
Generalization to $n$ small bounces {#sec5}
===================================
In this section, we would like to analyze the case where there are $n$ small bounces see Fig \[nbounceplot\]. Since the transfer matrix of the flat plateau case is quite involved, we would like to first consider Model 2 (no plateau interval) for illustrative purposes.
![\[nbounceplot\] An illustration of the generalization to n-vibrations (or small bounces). ](bouncefig4.pdf){width="45.00000%"}
Based on our previous calculations, we can easily write down the transfer matrices for the coefficient vector. We define the following useful matrices $$\begin{aligned}
M_1(\tau,q) = \begin{pmatrix} \tau^{1-q} & \tau^q \\ -(1-q) \tau^{-q}
& -q \tau^{q-1}
\end{pmatrix},\quad
M_2(\tau,q) = \begin{pmatrix} \tau^{1-q} & \tau^q \\ (1-q) \tau^{-q}
& q \tau^{q-1}
\end{pmatrix} .\end{aligned}$$
For large scale modes, we define the combination of matrices $$\begin{aligned}
N =
M^{-1}_2(\tau_B,q_1)M_1(\tau_B,q_1)M^{-1}_2(\epsilon,q_1)M_1(\epsilon,q_1)
,\end{aligned}$$ which becomes $$\begin{aligned}
N = \frac{1}{(1-2q_1)^2} \begin{pmatrix} 4(q_1-1)q_1
\tau_B^{2q_1-1}\epsilon^{1-2q_1} + 1 &
2q_1(\epsilon^{2q_1-1}-\tau_B^{2q_1-1})
\\ 2(q_1-1)(\tau_B^{1-2q_1}-\epsilon^{1-21_1}) &
4(q_1-1)q_1\epsilon^{2q_1-1}\tau_B^{1-2q_1}+1
\end{pmatrix}.\end{aligned}$$
Taking the two bump model as an example, we obtain the final coefficient vector to be $$\begin{aligned}
\mathcal C_F = M_2^{-1} (\epsilon ,q_2) M_1(\epsilon ,q_1) N M_2^{-1}
(\tau_B, q_1) M_1(\tau_B, q_1) M_2^{-1} (\epsilon,q_1) M_1(\epsilon,
q_2) \mathcal C_I .\end{aligned}$$ We set the initial coefficient matrix $\mathcal C_I$ to be $$\begin{aligned}
\mathcal C_I = \begin{pmatrix} c_{11} \\ 0
\end{pmatrix},\end{aligned}$$ and then we get $$\begin{aligned}
\nonumber
c_{52} =&\frac{1}{(1-2q_1)(1-2q_2^3)} \left[ 8 c_{11}
\left(q_1-1\right){}^2 q_1 \left(q_1-q_2+1\right){}^2 \epsilon
^{4 q_1-2 q_2-1} \tau _B^{2-4 q_1} \right. \\ \nonumber &
\left. -4 c_{11} \left(q_1-1\right) \left(q_1-q_2+1\right)
\left(2 q_1^2+2 q_2 q_1-5 q_1+q_2-1\right) \epsilon ^{2 q_1-2
q_2} \tau _B^{1-2 q_1} \right. \\ \nonumber & \left. -4 c_{11}
q_1 \left(q_1+q_2-2\right) \left(2 q_1^2-2 q_2 q_1+q_1+3
q_2-4\right) \epsilon ^{-2 q_1-2 q_2+2} \tau _B^{2 q_1-1}
\right. \\ \nonumber & \left. -8 c_{11} \left(q_1-1\right) q_1^2
\left(q_1+q_2-2\right){}^2 \epsilon ^{-4 q_1-2 q_2+3} \tau
_B^{4 q_1-2} \right. \\ \nonumber & \left. +2 c_{11} \left(4
q_1^4-8 q_1^3-4 q_2^2 q_1^2+16 q_2 q_1^2-10 q_1^2+4 q_2^2 q_1-16
q_2 q_1+14 q_1+2 q_2^2-5 q_2+3\right) \epsilon ^{1-2 q_2}
\right] .\end{aligned}$$ On very large scale there is no change in the spectral slope, as expected.
Now we want to deal with the small scale case. We need to define two more matrices $$\begin{aligned}
L_1 = \begin{pmatrix} e^{-ik\tau_H} & e^{i k \tau_H} \\ ik e^{-i k
\tau_H} & -i k e^{i k \tau_H}
\end{pmatrix}, \quad
L_2 = \begin{pmatrix} e^{ik\tau_H} & e^{-i k \tau_H} \\ ik e^{i k
\tau_H} & -i k e^{-i k \tau_H}
\end{pmatrix}.\end{aligned}$$ Then we have $$\begin{aligned}
\nonumber
\mathcal C_F = & M_2^{-1} (\epsilon,q_2) M_1(\epsilon,q_1) M_1^{-1}
(\delta,q_1) L_2 L_1^{-1} M_2(\delta,q_1) M_2^{-1} (\epsilon,q_1)
\\ &\times M_1(\epsilon,q_1) M_1^{-1} (\delta,q_1) L_2 L_1^{-1}
M_2(\delta,q_1) M_2^{-1} (\epsilon,q_1) M_1(\epsilon,q_2) \mathcal
C_I,\end{aligned}$$ and the general result has the form $$\begin{aligned}
P_{\zeta} = \left[\# + \# (\epsilon k)^{2-4q_1} + \# (\epsilon
k)^{4q_1-2} + \# (\epsilon k)^{1-2q_1} + \# (\epsilon
k)^{2q_1-1}\right]^2 P_i \, .\end{aligned}$$ Since we are interested in modes which exit the Hubble radius before the time $-(\tau_B + \epsilon)$, we consider values of $k$ for which $\epsilon k \ll 1$. Hence, in this range of $k$ values it is the second term above which dominates and we find the scaling $$P_{\zeta} \, \sim \, (\epsilon k)^{4 - 8q_1} P_i ~.$$ Thus, small scale modes acquire a red tilt compared to the initial spectrum. If the initial spectrum is scale invariant, then the resulting spectral index for small scale modes is $$n_s - 1 = 4 - 8q_1 ~.$$ Similarly, we can obtain the spectral index change for $n$ small bounces, which is $$n_s - 1 = (2 - 4q_1)n ~.$$
Summary and Conclusions {#sec6}
=======================
In this paper we have analyzed in detail the power spectrum of curvature fluctuations in a bouncing cosmology in which the bounce phase has small vibrations, i.e., small bounces. To be specific we have mostly considered the case of one small bounce with characteristic time scales $\tau_B$ and $\Delta < \tau_B$ which are much smaller than cosmological times. We have given a detailed study of the necessary matching conditions required to obtain the complete form for the power spectrum. The matchings connect at least five different phases for a given momentum scale which need to be treated with care.
In our study, we have adopted two simplified models for the shape of the vibrations, allowing a complete analytical study. Despite the apparent simplicity of these models, they are already of sufficient complexity to allow to extract similar features that can emerge in more realistic models. In particular, similar structures that we have considered here can appear in bounce models coming from quantum gravity, as those recently proposed in Ref. [@Alesci], which makes this study of particular importance. Our results for the power spectrum shows that there is an amplification of its amplitude and it also tends to get redder at large scales as the number of vibrations increase. At small scales the power spectrum features superimposed damped oscillations.
The reddening of the spectrum for scales which enter the small bounce agrees with the results found in Ref. [@RHBcyclic]. The oscillations in the power spectrum which are seen on small scales are reminiscent of oscillations which are obtained in some other approaches to the [*Trans-Planckian problem*]{} for cosmological fluctuations. For example, if initial conditions are set on a time-like [*new physics hypersurface*]{} [@newphysics] such that modes $k$ are initiated when the physical wavelength associated with $k$ equals a fixed physical length (e.g. the Planck length), and they are initiated in the same state (e.g. the state which locally looks like the Bunch-Davies vacuum [@BD]), then oscillations in the spectrum result.
Both the qualitative and quantitative changes in the power spectrum that we have obtained can produce observed effects in spectrum of cosmological perturbations accessible through the measurements of the cosmic microwave background radiation. These effects can manifest themselves both in pure bouncing cosmologies (no subsequent inflationary period) and in scenarios where there is a post-bounce inflationary phase. For instance, those bounce vibrations can induce particle production, changing the vacuum state such as to be different from the usual Bunch-Davis one, similar to recent pre-inflationary studies in Loop Quantum Cosmology [@Zhu:2017jew]. The modifications we have obtained in this work could then be used to put constraints on these possible features that can appear in these bounce models and which deserve further study. The results we have presented here provides then an important first step in understanding these effects and which we hope to address elsewhere.
Envelope of the Power Spectrum for Small Scale Modes {#appA}
====================================================
In this section, we would like to calculate the envelope of the power spectrum for small scale modes. Since the model without plateau is a special limit of the model with a non-vanishing flat plateau, we just focus on the latter. We can simply set $\Delta \rightarrow 0$ to get the answer for the model without a plateau.
By collecting the relevant terms in the power spectrum, we can write it in the form $$\begin{aligned}
P_{\zeta } = \left[C_1\sin(2k\Delta) + C_2 \cos(2k\Delta)\right]^2 P_i
\, .\end{aligned}$$ The envelope of the power spectrum is thus $$\begin{aligned}
P_{\zeta {\rm (env)}} = (C_1^2 + C_2^2 )P_i \, ,\end{aligned}$$ where the coefficients $C_1$ and $C_2$ are given by $$\begin{aligned}
\nonumber
C_1& = \frac{k^{-1}}{(1-2q_2)(1-2q_1)^2} \left\{ 2 \delta^{-1}
\left[k^2\delta^2+(q_1-1)q_1 \right](1+q_1-q_2)(-2+q_1+q_2)
\right. \\ & \left. - \delta^{2q_1-2} (q_1-k\delta) (q_1+k\delta)
(-2+q_1+q_2)^2 \epsilon^{1-2q_1} + \delta^{-2q_1}
(q_1-k\delta-1)(q_1+k\delta-1)(1+q_1-q_2)^2\epsilon^{2q_1-1} \right\},
\label{C1env}
\\\nonumber C_2 & = \frac{1}{(1-2q_2)(1-2q_1)^2} \left[
-2(1+q_1-q_2)(-2+q_1+q_2) \right. \\ & \left. -\delta^{2q_1-1} 2
q_1 (-2+q_1+q_2)^2 \epsilon^{1-2q_1} - \delta^{-2q_1+1} 2 (q_1-1)
(1+q_1-q_2)^2 \epsilon^{2q_1-1} \right] \, .
\label{C2env}\end{aligned}$$
We are interested in the parameter region $\epsilon / \delta \ll1$ (recall that the time scale $\epsilon$ is expected to be of the order of the Planck scale, whereas $\delta$ will be parametrically larger since it is associated with the time scale of the bounce). We are also interested in the range of values $1/3<p<1$, or equivalently, $1/2<q<+\infty$. We can then determine which are the dominant terms in $C_1$ and $C_2$, which from Eqs. (\[C1env\]) and (\[C1env\]), they are given by $$\begin{aligned}
C_1& \simeq \frac{k^{-1}}{(1-2q_2)(1-2q_1)^2} \left[ -
\delta^{2q_1-2} (q_1-k\delta) (q_1+k\delta) (-2+q_1+q_2)^2
\epsilon^{1-2q_1} \right], \\ C_2 & \simeq
\frac{1}{(1-2q_2)(1-2q_1)^2} \left[ -\delta^{2q_1-1} 2 q_1
(-2+q_1+q_2)^2 \epsilon^{1-2q_1} \right] \, .\end{aligned}$$ When $k$ is close to the $k_*$, then $\delta \rightarrow
\tau_B-\Delta$ which is a constant. In this case, taking the square of $C_1$, we get terms with different with spectral indices, but the dominant contribution is the term with the lowest power of $k$, which the gives that the slope of the envelope (for an initial spectrum which is scale-invariant) will be $$\begin{aligned}
n_s -1 = -2 ,\end{aligned}$$ because we have $k\delta<q_1$ in this range. This can be seen from the numerical results shown in Fig. \[bounceplot3\] for the two models we have considered. The change in the spectral slope is due to the matching conditions. Each time, we can get factors of $1/k$ or $k$ when we match the solution across the boundaries of Regions II and III, and of Regions III and IV.
Note that in a generic case when we have a smooth evolution of the scale factor, we expect that there will be no discontinuities in the power spectrum. Thus, in a generic case, we do not expect that we always get an interval of wavenumber with a spectrum of slope $n_s=-2$. What we expect in the case of a smoothly evolving scale factor is that on very large scales, we get a scale invariant spectrum (the actual spectrum, not just the envelope), and then it will smoothly transit to a spectrum with tilt $n_s=-4q_1+2$ when we look at the envelope only. We see oscillations with amplitude given by the envelope function on intermediate and small scales.
The coefficient $C_2^2$ gives a scale invariant power spectrum $$\begin{aligned}
n_s -1 = 0,\end{aligned}$$ but its amplitude is suppressed by $k\delta$ compared to the amplitude of $C_1$. To be a bit more precise (still in the case of constant $\delta$), we can write $$\begin{aligned}
C_1 = A_1 k^{-1} + A_2 k \, ,\end{aligned}$$ where the constants $A_1$ and $A_2$ are $$\begin{aligned}
\nonumber
A_1 & = \frac{1}{(1-2q_2)(1-2q_1)^2} \left[ 2 \delta^{-1} (q_1-1)q_1
(1+q_1-q_2) (-2+q_1+q_2) - \delta^{2q_1-2} q_1^2 (-2+q_1+q_2)^2
\epsilon^{1-2q_1} \right. \\ &\left. -\delta^{-2q_1} (q_1-1)^2
(1+q_1-q_2)^2 \epsilon^{2q_1-1} \right], \\ \nonumber A_2 & =
\frac{1}{(1-2q_2)(1-2q_1)^2} \left[ 2\delta(1+q_1-q_2)(-2+q_1+q_2) +
\delta^{2q_1} (-2+q_1+q_2)^2 \epsilon^{1-2q_1} \right. \\ &- \left.
\delta^{-2q_1+2} (1+q_1-q_2)^2 \epsilon^{2q_1-1} \right] .\end{aligned}$$ The spectral index is computed as $$\begin{aligned}
n_s - 1 = \frac{d\ln P_{\zeta {\rm (env)}}}{d\ln k} = \frac{2
P_i}{P_{\zeta {\rm (env)}}} (A_2^2 k^2 -A_1^2 k^{-2}).\end{aligned}$$ The power spectrum is hence comprised of several terms with different spectral tilts $n_s$ $$\begin{aligned}
n_s-1 = 2, 1, 0, -1, -2,\end{aligned}$$ More generally (for larger values of $k$ when $\delta$ is not constant), we have
$$\begin{aligned}
P_{\zeta } = \left\{D_1\sin[2(k\tau_B-q_1)] + D_2
\cos[2k(k\tau_B-q_1)]\right\}^2 P_i \, .\end{aligned}$$
The envelope of the power spectrum is thus $$\begin{aligned}
P_{\zeta {\rm (env)}} = (D_1^2 + D_2^2 )P_i \, ,\end{aligned}$$ where the coefficients $D_1$ and $D_2$ are given by $$\begin{aligned}
D_1 & = \frac{1}{(1-2q_2)(1-2q_1)} \left[ -2 (1+q_1-q_2) (-2+q_1+q_2)
+ k^{2q_1-1} q_1^{-2q_1} (1+q_1-q_2)^2 \epsilon^{2q_1-1} \right] ,
\\ D_2 & = \frac{1}{(1-2q_2)(1-2q_1)^2} \left[
-2(1+q_1-q_2)(-2+q_1+q_2) + k^{1-2q_1} q_1^{2q_1-1}
(-2q_1)(-2+q_1+q_2)^2\epsilon^{1-2q_1} \right. \nonumber \\ &
\left. -2 k^{2q_1-1} q_1^{-2q_1+1} (q_1-1) (1+q_1-q_2)^2
\epsilon^{2q_1-1} \right] \, .\end{aligned}$$
We now can see that this envelope function reproduces the result of Ref. [@RHBcyclic]. We have $$\begin{aligned}
D_1 &= B_1 + B_2 k^{2q_1-1},\\ D_2 &= E_1 + E_2 k^{1-2q_1} + E_3
k^{2q_1-1} \, ,\end{aligned}$$ where the constants $B_1$, $B_2$, $E_1$, $E_2$ and $E_3$ are given by $$\begin{aligned}
B_1 &= \frac{-2 (1+q_1-q_2) (-2+q_1+q_2)}{(1-2q_2)(1-2q_1)},\\ B_2
&= \frac{ q_1^{-2q_1} (1+q_1-q_2)^2 \epsilon^{2q_1-1}
}{(1-2q_2)(1-2q_1)},\\ E_1 &=
\frac{-2(1+q_1-q_2)(-2+q_1+q_2)}{(1-2q_2)(1-2q_1)^2},\\ E_2 &=
\frac{q_1^{2q_1-1}
(-2q_1)(-2+q_1+q_2)^2\epsilon^{1-2q_1}}{(1-2q_2)(1-2q_1)^2}, \\ E_3
&= \frac{-2q_1^{-2q_1+1} (q_1-1) (1+q_1-q_2)^2
\epsilon^{2q_1-1}}{(1-2q_2)(1-2q_1)^2} .\end{aligned}$$
The spectral tilt is then given by $$\begin{aligned}
n_s - 1 & = \frac{P_i}{P_{\zeta {\rm (env)}}} \left\{ 2 D_1 B_2
(2q_1-1) k^{2q_1-1} + 2 D_2 \left[E_2(1-2q_1) k^{-2q_1+1} + E_3
(2q_1-1) k^{2q_1-1} \right] \right\} \, .
\label{nsenv}\end{aligned}$$ The expression (\[nsenv\]) is comprised of several terms with spectral tilts $n_s$ given by $$\begin{aligned}
n_s-1 = 4q_1-2, 2q_1-1, 0, -4q_1+2, -2q_1+1 \, .\end{aligned}$$ Since are interested in modes with $k\epsilon<1$ and parameter values $1/3<p<1$ (or, equivalently, $1/2<q<+\infty$) we can determine the dominant terms in $D_1$ and $D_2$ and find them to be $$\begin{aligned}
D_1 \rightarrow 0,\quad D_2 \rightarrow E_2 k^{1-2q_1} \, .\end{aligned}$$ Thus, the dominant contribution to the power spectrum is $$\begin{aligned}
P_{\zeta {\rm (env)} } = E_2^2 k^{2-4q_1} P_i \, ,\end{aligned}$$ which corresponds to a spectral tilt of $$\begin{aligned}
n_s = -4q_1+2 \, .\end{aligned}$$
Acknowledgments {#acknowledgments .unnumbered}
===============
One of us (R.B.) is grateful to Emanuele Alesci and Stefano Liberati for discussions about the model of [@Alesci] which led to this project. He also thanks Stefano Liberati and the other organizers of the [*Probing the Spacetime Fabric: from Concepts to Phenomenology*]{} workshop help in July 2017 at SISSA for inviting him to participate and speak. The research at McGill was supported in part by an NSERC Discovery grant and by the Canada Research Chair program. Q.L acknowledge financial support from the University of Science and Technology of China, and from the CAST Young Elite Scientists Sponsorship Program (2016QNRC001), and by the NSFC (grant Nos. 11421303, 11653002). SZ is supported by the Hong Kong PhD Fellowship Scheme (HKPFS) issued by the Research Grants Council (RGC) of Hong Kong. R.O.R is partially supported by research grants from Conselho Nacional de Desenvolvimento Científico e Tecnológico (CNPq), grant No. 303377/2013-5 and Fundação Carlos Chagas Filho de Amparo á Pesquisa do Estado do Rio de Janeiro (FAPERJ), grant No. E - 26/201.424/2014. Q.L., R.O.R. and S.Z. are grateful for the hospitality of the Physics Department at McGill University during research visits when this work was initiated.
[999]{}
R. Brandenberger and P. Peter, “Bouncing Cosmologies: Progress and Problems,” Found. Phys. [**47**]{}, no. 6, 797 (2017). F. Finelli and R. Brandenberger, “On the generation of a scale invariant spectrum of adiabatic fluctuations in cosmological models with a contracting phase,” Phys. Rev. D [**65**]{}, 103522 (2002). D. Wands, “Duality invariance of cosmological perturbation spectra,” Phys. Rev. D [**60**]{}, 023507 (1999). J. Khoury, B. A. Ovrut, P. J. Steinhardt and N. Turok, “The Ekpyrotic universe: Colliding branes and the origin of the hot big bang,” Phys. Rev. D [**64**]{}, 123522 (2001). A. Notari and A. Riotto, “Isocurvature perturbations in the ekpyrotic universe,” Nucl. Phys. B [**644**]{}, 371 (2002);\
F. Finelli, “Assisted contraction,” Phys. Lett. B [**545**]{}, 1 (2002);\
F. Di Marco, F. Finelli and R. Brandenberger, “Adiabatic and isocurvature perturbations for multifield generalized Einstein models,” Phys. Rev. D [**67**]{}, 063512 (2003);\
J. L. Lehners, P. McFadden, N. Turok and P. J. Steinhardt, “Generating ekpyrotic curvature perturbations before the big bang,” Phys. Rev. D [**76**]{}, 103501 (2007);\
E. I. Buchbinder, J. Khoury and B. A. Ovrut, “New Ekpyrotic cosmology,” Phys. Rev. D [**76**]{}, 123503 (2007);\
P. Creminelli and L. Senatore, “A Smooth bouncing cosmology with scale invariant spectrum,” JCAP [**0711**]{}, 010 (2007). Y. F. Cai, T. Qiu, Y. S. Piao, M. Li and X. Zhang, “Bouncing universe with quintom matter,” JHEP [**0710**]{}, 071 (2007);\
C. Lin, R. H. Brandenberger and L. Perreault Levasseur, “A Matter Bounce By Means of Ghost Condensation,” JCAP [**1104**]{}, 019 (2011);\
T. Qiu, J. Evslin, Y. F. Cai, M. Li and X. Zhang, “Bouncing Galileon Cosmologies,” JCAP [**1110**]{}, 036 (2011);\
D. A. Easson, I. Sawicki and A. Vikman, “G-Bounce,” JCAP [**1111**]{}, 021 (2011);\
Y. F. Cai, D. A. Easson and R. Brandenberger, “Towards a Nonsingular Bouncing Cosmology,” JCAP [**1208**]{}, 020 (2012);\
Y. F. Cai, E. McDonough, F. Duplessis and R. H. Brandenberger, “Two Field Matter Bounce Cosmology,” JCAP [**1310**]{}, 024 (2013). Y. F. Cai, T. Qiu, R. Brandenberger, Y. S. Piao and X. Zhang, “On Perturbations of Quintom Bounce,” JCAP [**0803**]{}, 013 (2008);\
Y. F. Cai and X. Zhang, “Evolution of Metric Perturbations in Quintom Bounce model,” JCAP [**0906**]{}, 003 (2009);\
Y. F. Cai, T. t. Qiu, R. Brandenberger and X. m. Zhang, “A Nonsingular Cosmology with a Scale-Invariant Spectrum of Cosmological Perturbations from Lee-Wick Theory,” Phys. Rev. D [**80**]{}, 023511 (2009). R. Brandenberger, “Matter Bounce in Horava-Lifshitz Cosmology,” Phys. Rev. D [**80**]{}, 043516 (2009). T. Biswas, A. Mazumdar and W. Siegel, “Bouncing universes in string-inspired gravity,” JCAP [**0603**]{}, 009 (2006);\
T. Biswas, R. Brandenberger, A. Mazumdar and W. Siegel, “Non-perturbative Gravity, Hagedorn Bounce & CMB,” JCAP [**0712**]{}, 011 (2007);\
A. S. Koshelev, “Stable analytic bounce in non-local Einstein-Gauss-Bonnet cosmology,” Class. Quant. Grav. [**30**]{}, 155001 (2013). M. Bojowald, “Loop quantum cosmology,” Living Rev. Rel. [**11**]{}, 4 (2008);\
A. Ashtekar, “Singularity Resolution in Loop Quantum Cosmology: A Brief Overview,” J. Phys. Conf. Ser. [**189**]{}, 012003 (2009);\
A. Ashtekar and A. Barrau, “Loop quantum cosmology: From pre-inflationary dynamics to observations,” Class. Quant. Grav. [**32**]{}, no. 23, 234001 (2015). N. Turok, B. Craps and T. Hertog, “From big crunch to big bang with AdS/CFT,” arXiv:0711.1824 \[hep-th\];\
B. Craps, T. Hertog and N. Turok, “On the Quantum Resolution of Cosmological Singularities using AdS/CFT,” Phys. Rev. D [**86**]{}, 043513 (2012);\
C. S. Chu and P. M. Ho, “Time-dependent AdS/CFT duality and null singularity,” JHEP [**0604**]{}, 013 (2006);\
C. S. Chu and P. M. Ho, “Time-dependent AdS/CFT duality. II. Holographic reconstruction of bulk metric and possible resolution of singularity,” JHEP [**0802**]{}, 058 (2008);\
R. H. Brandenberger, E. G. M. Ferreira, I. A. Morrison, Y. F. Cai, S. R. Das and Y. Wang, “Fluctuations in a cosmology with a spacelike singularity and their gauge theory dual description,” Phys. Rev. D [**94**]{}, no. 8, 083508 (2016);\
E. G. M. Ferreira and R. Brandenberger, “Holographic Curvature Perturbations in a Cosmology with a Space-Like Singularity,” JCAP [**1607**]{}, no. 07, 030 (2016). C. Kounnas, H. Partouche and N. Toumbas, “S-brane to thermal non-singular string cosmology,” Class. Quant. Grav. [**29**]{}, 095014 (2012). R. H. Brandenberger and C. Vafa, “Superstrings in the Early Universe,” Nucl. Phys. B [**316**]{}, 391 (1989). J. Martin and R. H. Brandenberger, “The TransPlanckian problem of inflationary cosmology,” Phys. Rev. D [**63**]{}, 123501 (2001);\
R. H. Brandenberger and J. Martin, “The Robustness of inflation to changes in superPlanck scale physics,” Mod. Phys. Lett. A [**16**]{}, 999 (2001). Y. S. Piao, “Primordial Perturbation in Horava-Lifshitz Cosmology,” Phys. Lett. B [**681**]{}, 1 (2009);\
X. Gao, Y. Wang, R. Brandenberger and A. Riotto, “Cosmological Perturbations in Horava-Lifshitz Gravity,” Phys. Rev. D [**81**]{}, 083508 (2010);\
X. Gao, Y. Wang, W. Xue and R. Brandenberger, “Fluctuations in a Horava-Lifshitz Bouncing Cosmology,” JCAP [**1002**]{}, 020 (2010);\
A. Cerioni and R. H. Brandenberger, “Cosmological Perturbations in the Projectable Version of Horava-Lifshitz Gravity,” JCAP [**1108**]{}, 015 (2011). R. H. Brandenberger, C. Kounnas, H. Partouche, S. P. Patil and N. Toumbas, “Cosmological Perturbations Across an S-brane,” JCAP [**1403**]{}, 015 (2014). B. Xue and P. J. Steinhardt, “Unstable growth of curvature perturbation in non-singular bouncing cosmologies,” Phys. Rev. Lett. [**105**]{}, 261301 (2010);\
B. Xue and P. J. Steinhardt, “Evolution of curvature and anisotropy near a nonsingular bounce,” Phys. Rev. D [**84**]{}, 083520 (2011). J. Quintin, Z. Sherkatghanad, Y. F. Cai and R. H. Brandenberger, “Evolution of cosmological perturbations and the production of non-Gaussianities through a nonsingular bounce: Indications for a no-go theorem in single field matter bounce cosmologies,” Phys. Rev. D [**92**]{}, no. 6, 063532 (2015);\
Y. B. Li, J. Quintin, D. G. Wang and Y. F. Cai, “Matter bounce cosmology with a generalized single field: non-Gaussianity and an extended no-go theorem,” JCAP [**1703**]{}, no. 03, 031 (2017). E. Alesci, G. Botta, F. Cianfrani and S. Liberati, “Cosmological singularity resolution from quantum gravity: the emergent-bouncing universe,” Phys. Rev. D [**96**]{}, no. 4, 046008 (2017). R. H. Brandenberger, “Processing of Cosmological Perturbations in a Cyclic Cosmology,” Phys. Rev. D [**80**]{}, 023535 (2009). R. Durrer and F. Vernizzi, “Adiabatic perturbations in pre - big bang models: Matching conditions and scale invariance,” Phys. Rev. D [**66**]{}, 083503 (2002). V. F. Mukhanov, H. A. Feldman and R. H. Brandenberger, “Theory of cosmological perturbations. Part 1. Classical perturbations. Part 2. Quantum theory of perturbations. Part 3. Extensions,” Phys. Rept. [**215**]{}, 203 (1992). J. c. Hwang and E. T. Vishniac, “Gauge-invariant joining conditions for cosmological perturbations,” Astrophys. J. [**382**]{}, 363 (1991). N. Deruelle and V. F. Mukhanov, “On matching conditions for cosmological perturbations,” Phys. Rev. D [**52**]{}, 5549 (1995). W. Israel, “Singular hypersurfaces and thin shells in general relativity,” Nuovo Cim. B [**44S10**]{}, 1 (1966) \[Nuovo Cim. B [**48**]{}, 463 (1967)\] \[Nuovo Cim. B [**44**]{}, 1 (1966)\]. R. H. Brandenberger, “The Matter Bounce Alternative to Inflationary Cosmology,” arXiv:1206.4196 \[astro-ph.CO\]. R. Easther, B. R. Greene, W. H. Kinney and G. Shiu, “A Generic estimate of transPlanckian modifications to the primordial power spectrum in inflation,” Phys. Rev. D [**66**]{}, 023518 (2002);\
V. Bozza, M. Giovannini and G. Veneziano, “Cosmological perturbations from a new physics hypersurface,” JCAP [**0305**]{}, 001 (2003). T. S. Bunch and P. C. W. Davies, “Quantum Field Theory in de Sitter Space: Renormalization by Point Splitting,” Proc. Roy. Soc. Lond. A [**360**]{}, 117 (1978). T. Zhu, A. Wang, G. Cleaver, K. Kirsten and Q. Sheng, “Pre-inflationary universe in loop quantum cosmology,” Phys. Rev. D [**96**]{}, 083520 (2017).
[^1]: Note that there are models in which the spectrum of scalar fluctuations is boosted by a factor independent of wavelength on large scales during the bounce phase [@Jerome].
[^2]: From the point of view of string theory, one may view the time interval we are cutting out as the string time scale, the time scale where the effective field theory description will break down.
[^3]: Note that applying these matching conditions directly between a contracting phase and an expanding phase may be a bit suspect since the background does not satisfy the matching conditions (see Ref. [@Durrer] for a detailed discussion of this point). However, as long as the matching surface is unambiguously determined, the matching conditions for the fluctuations can indeed be applied.
[^4]: Note that the “regions” defined here are not the same as the “phases” defined above. The “phases” refer to particular behaviors of the scale factor, the “regions” to particular behaviors of the fluctuation modes. Phases I and V are equal to Regions I and V, but for the others there is a difference.
[^5]: Where we are neglecting the coefficient $c_{12}$ of the decaying mode in the initial phase.
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'Md. Jahoor Alam$^a$'
- Shazia Kunvar$^b$
- 'R.K. Brojen Singh$^{a}$'
title: 'The co-existence of states in $p53$ dynamics driven by $miRNA$'
---
The model biochemical network (Fig. 1) described by the twenty two reaction channels (Table 2) can be described by the following coupled ordinary differential equations (ODE) using Mass action law of chemical kinetics, $$\begin{aligned}
\label{ma}
\frac{dx_1}{dt}&=&k_4x_4-k_{10}x_1x_2+k_{12}x_5\\
\frac{dx_2}{dt}&=&k_8x_3-k_{10}x_1x_2+k_{11}x_5+k_{12}x_5
-k_{13}x_2-k_{20}x_2x_8\\
\frac{dx_3}{dt}&=&k_7x_1-k_9x_3\\
\frac{dx_4}{dt}&=&-k_2x_{10}x_4+k_5-k_6x_4\\
\frac{dx_5}{dt}&=&k_{10}x_2x_1-k_{11}x_5-k_{12}x_5\\
\frac{dx_6}{dt}&=&k_{14}-k_{15}x_6-k_{16}x_6\\
\frac{dx_7}{dt}&=&k_{16}x_6-k_{17}x_7\\
\frac{dx_8}{dt}&=&k_{18}x_7-k_{19}x_8-k_{20}x_8x_2+k_{21}x_9\\
\frac{dx_9}{dt}&=&k_{20}x_8x_2-k_{21}x_9\\
\frac{dx_{10}}{dt}&=&k_1-k_2x_{10}x_4+k_3x_{11}-k_{22}x_{10}\\
\frac{dx_{11}}{dt}&=&k_2x_{10}x_4-k_3x_{11}\end{aligned}$$ The set of ODEs can be written in compact form as in the following, $$\begin{aligned}
\label{ode}
\frac{d{\bf x}(t)}{dt}={\bf F}(x_1,x_2,\dots,x_N)\end{aligned}$$ where, ${\bf F}=\left[F_1,F_2,\dots,F_N\right]^T$ is the functional vector. The time evolution of the state vector ${\vec x}(t)$ can be obtained by numerically solving the non-linear coupled differential equations (1)-(11) using standard 4th order Runge-Kutta algorithm for numerical integration [@pre].
Stability analysis
------------------
The fixed or equilibrium points of the ODE given by equation (\[ode\]) can be obtained by putting $\frac{d{\bf x}(t)}{dt}=0$ and solving for $x_1^*$, $x_2^*$, ..., $x_N^*$ from these equations. In our model described by mathematical equations (1)-(11), we have the following equilibrium points, $$\begin{aligned}
x_1^*&=&\left[\frac{k_4k_5k_9}{k_7k_8k_{10}k_{11}}\left\{k_{13}+\frac{{\bf k_{ROS}}k_{16}k_{18}k_{20}}{k_{17}k_{18}(k_{15}+k_{16})}\right\}\frac{k_{11}+k_{12}}{k_6+\frac{k_2}{k_{22}}{\bf k_{miRNA}}}\right]^{1/2}\\
x_2^*&=&\left[\frac{k_4k_5k_7k_8(k_{11}+k_{12})}{k_9k_{10}k_{11}\left\{k_{13}+\frac{{\bf k_{ROS}}k_{16}k_{18}k_{20}}{k_{17}k_{18}(k_{15}+k_{16})}\right\}}\frac{1}{k_6+\frac{k_2}{k_{22}}{\bf k_{miRNA}}}\right]^{1/2}\\
x_3^*&=&\left[\frac{k_4k_5}{k_8k_{10}k_{11}}\left\{k_{13}+\frac{{\bf k_{ROS}}k_{16}k_{18}k_{20}}{k_{17}k_{18}(k_{15}+k_{16})}\right\}\frac{k_{11}+k_{12}}{k_6+\frac{k_2}{k_{22}}{\bf k_{miRNA}}}\right]^{1/2}\\
x_4^*&=&\frac{k_5k_{22}}{k_2+k_1k_{6}}\\
x_5^*&=&\frac{k_4k_5k_{22}}{k_{11}(k_2+k_1k_{6})}\\
x_6^*&=&\frac{{\bf k_{ROS}}}{k_{15}+k_{16}}\\
x_7^*&=&\frac{{\bf k_{ROS}}k_{16}}{k_{17}(k_{15}+k_{16})}\\
x_8^*&=&\frac{{\bf k_{ROS}}k_{16}k_{18}}{k_{17}k_{19}(k_{15}+k_{16})}\\
x_9^*&=&\left[\frac{k_4k_5k_7k_8(k_{11}+k_{12})}{k_9k_{10}k_{11}\left\{k_{13}+\frac{{\bf k_{ROS}}k_{16}k_{18}k_{20}}{k_{17}k_{18}(k_{15}+k_{16})}\right\}}\frac{1}{k_6+\frac{k_2}{k_{22}}{\bf k_{miRNA}}}\right]^{1/2}\nonumber\\
&&\times\frac{{\bf k_{ROS}}k_{16}k_{18}k_{20}}{k_{17}k_{19}k_{21}(k_{15}+k_{16})}\\
x_{10}^*&=&\frac{k_1}{k_{22}}\\
x_{11}^*&=&\frac{k_1k_2k_5}{k_3(k_2+k_1k_{6})}\end{aligned}$$ The stabilized state of p53 ($x_1^*$) and Mdm2 ($x_2^*$) are dependent on the values of the parameters $k_{miRNA}$ and $k_{ROS}$, and there is competition between these two parameters affecting stabilized states of p53 and Mdm2. Keeping $k_{miRNA}$ to a constant value, equation (13) shows that $x_1^*\propto \sqrt{1+Ak_{ROS}}$, where, $A=\frac{k_{16}k_{18}k_{20}}{k_{13}k_{17}k_{18}(k_{15}+k_{16})}$ which drives the low equilibrium state (may be normal state where $x_1^*$ is maintained minimum value) at low $k_{ROS}$ to the higher equilibrium state (may be apoptosis state where $x_1^*$ is maintained at high value) as $k_{ROS}$ increases. However, in the case of Mdm2, the scenario is opposite, where $x_2^*\propto \frac{1}{\sqrt{1+Ak_{ROS}}}$, and $k_{ROS}$ drives the higher Mdm2 equilibrium state to lower equilibrium state.
Further, if $k_{ROS}$ and other rates are kept constant, $x_1^*\propto\frac{1}{\sqrt{B+k_{miRNA}}}$ and $x_2^*\propto\frac{1}{\sqrt{B+k_{miRNA}}}$, where $B$ is a constant given by $B=\frac{k_6k_{22}}{k_2}$. This means that as $k_{miRNA}$ increases, $k_{miRNA}$ drives the higher equilibrium state of both p53 and Mdm2 to lower equilibrium states.
[99]{}
D.P. Lane, Nature [**358**]{}, 15 (1992). C. Whibley, P.D. Pharoah, and M. Hollstein, Nat. Rev. Cancer. [**9**]{}, 95 (2009). U.M. Moll and O. Petrenko, Mol. Cancer Res. [**1**]{}, 1001 (2003). G.P. Zambetti, J. Cell Physiol. [**213**]{}, 370 (2007). D. Michael and M. Oren, Semin. Cancer Biol. [**13**]{}, 49 (2003). S. Wang, Y. Zhao, D. Bernard, A. Aguilar, and S. Kumar, Top Med. Chem. [**8**]{}, 57 (2012). M.H. Kubbutat, S.N. Jones, and K.H. Vousden, Nature [**387**]{}, 299 (1997). S. Fang, J.P. Jensen, R.L. Ludwig, K.H. Vousden, and A.M. Weissman, J. Biol. Chem. [**275**]{}, 8945 (2000). M.T. Boyd, N. Vlatkovic, and D.S. Haines, J. Biol. Chem. [**275**]{}, 31883 (2000). J. Chen, X. Wu, J. Lin, and A.J. Levine, Mol. Cell. Biol. [**16**]{}, 2445 (1996). R. Amendola, M. Cervelli, E. Fratini, D.E. Sallustio, G. Tempera, T. Ueshima, P. Mariottini, and E. Agostinelli, Int. J. Oncol. [**43**]{}, 813 (2013). T.P. Devasagayam, J.C. Tilak, K.K. Boloor, K.S Sane, S.S. Ghaskadbi, and R.D. Lele, J. Assoc. Physicians India, [**52**]{}, 794 (2004). L.A. Rowe, N. Degtyareva, and P.W. Doetsch, Free Radic. Biol. Med. [**45**]{}, 1167 (2008). C.J. Proctor and D.A. Gray, Molecular Neurodegeneration [**5**]{}, 7 (2010). J.M. Thomson, M. Newman, J.S. Parker, E.M. Morin-Kensicki, T. Wright, and S.M. Hammond, Genes Dev. [**20**]{}, 2202 (2006). Y. Lee, M. Kim, J. Han, K. Yeom, S. Lee, S.H. Baek, and V.N. Kim, EMBO J. [**23**]{}, 4051 (2004). B.R. Wilfred, W.X. Wang, and P.T. Nelson, Mol. Genet. Metab. [**91**]{}, 209 (2007). Y. Zhang, J.S. Gao, X. Tang, L.D. Tucker, P. Quesenberry, I. Rigoutsos, and B. Ramratnam, FEBS Letters [**583**]{}, 3725 (2009). J. Winter, S. Jung, S. Keller, R.I. Gregory, and S. Diederichs, Nat. Cell. Biol. [**11**]{}, 228 (2009). M.T. Le, C. Teh, N. Shyh-Chang, H. Xie, B. Zhou, V. Korzh, H.F. Lodish, B. Lim, Genes Dev. [**23**]{}, 862 (2009). C.J. Proctor and D.A. Gray, BMC Syst. Biol. [**2**]{}, 75 (2008). C. Lee, B.A. Smith, K. Bandyopadhyay, and R.A. Gjerset, Cancer Res. [**65**]{}, 9834 (2005). Y. Zhang and Y. Xiong, Cell Growth Differ. [**12**]{}, 175 (2001). S. Khan, C. Guevara, G. Fujii, and D. Parry, Oncogene [**23**]{}, 6040 (2004). N. Geva-Zatorsky, N. Rosenfeld, S. Itzkovitz, R. Milo, A. Sigal, E. Dekel, T. Yarnitzky, Y. Liron, P. Polak, G. Lahav, and U. Alon, Mol Syst. Biol. [**2**]{}, 0033 (2006). M.J. Alam, G.R. Devi, Ravins, R. Ishrat, S.M. Agarwal, and R.K.B. Singh, Mol. BioSyst. [**9**]{}, 508 (2013). W.H. Press, S.A. Teukolsky, W.T. Vetterling, and B.P. Flannery Numerical Recipe in Fortran. Cambridge University Press (1992). M. Selbach, B. Schwanhäusser, N. Thierfelder, Z. Fang, R. Khanin, and N. Rajewsky, Nature [**455**]{}, 58 (2008).
[**Table 1 - List of molecular species**]{}
[|l|p[4cm]{}|p[5cm]{}|p[1.5cm]{}|]{}\
**[ S.No.]{} & **[Species Name]{} & **[Description]{} & **[Notation]{}\
1. & $p53$ & Unbounded $p53$ protein & $x_1$\
2. & $Mdm2$ & Unbounded $Mdm2$ protein & $x_2$\
3. & $Mdm2\_mRNA$ & $Mdm2$ messenger $mRNA$ & $x_3$\
4. & $p53\_mRNA$ & $p53$ messenger $mRNA$ & $x_4$\
5. & $Mdm2\_p53$ & $Mdm2$ with $p53$ complex & $x_5$\
6. & $ROS$ & Reactive Oxygen Species & $x_6$\
7. & $Dam\_DNA$ & Damage DNA & $x_7$\
8. & $ARF$ & Alternative Reading Frame protein & $x_8$\
9. & $ARF\_Mdm2$ & $ARF$ and $Mdm2$ complex & $x_9$\
10. & $mi-RNA-125b$ & Micro RNA 125b & $x_{10}$\
11. & $mi-RNA\_p53-mRNA$ & Micro RNA 125b and $p53\_mRNA$ complex & $x_{11}$\
********
[**Table 2 List of chemical reaction, Kinetic Laws and their rate constant**]{}
[|l|p[2.5cm]{}|p[4cm]{}|p[2.5cm]{}|p[3cm]{}|p[2cm]{}|]{}\
${\bf S.No}$ & ${\bf Reaction}$ & ${\bf Name~of~the~process}$ & ${\bf Kinetic~Law}$ & ${\bf Rate~Constant}$ & ${\bf References}$\
1 & $\phi\stackrel{k_{1}}{\longrightarrow}x_{10}$ & Micro RNA creation & $k_2$ & $1\times 10^{-4}{sec}^{-1}$ & [@le; @lee1; @win]\
2 & $x_{10}+x_4\stackrel{k_{2}}{\longrightarrow}x_{11}$ & Synthesis of miRNA and p53\_mRNA complex & $k_2 \langle x_{10}\rangle\langle x_4\rangle $ & $2\times 10^{-2}{sec}^{-1}$ & [@le; @lee1]\
3 & $x_{11}\stackrel{k_{3}}{\longrightarrow}x_{10}$ & $miRNA\_p53\_mRNA$ degradation & $k_3 \langle x_{11}\rangle$ & $1\times 10^{-4}{sec}^{-1}$ & [@le; @lee1]\
4 & $x_4\stackrel{k_{4}}{\longrightarrow}x_1+x_4$ & p53 mRNA translation & $k_4 \langle x_4\rangle$ & $8\times 10^{-2}{sec}^{-1}$ & [@mol; @pro]\
5 & $\phi\stackrel{k_{5}}{\longrightarrow}x_4$ & $p53\_mRNA$ synthesis & $k_5$ & $1\times 10^{-3}{sec}^{-1}$ & [@mol; @pro]\
6 & $x_4\stackrel{k_{6}}{\longrightarrow}\phi$ & $p53\_mRNA$ degradation & $k_6 \langle x_4\rangle$ & $1\times 10^{-4}{sec}^{-1}$ & [@mol; @pro]\
7 & $x_1\stackrel{k_{7}}{\longrightarrow}x_1+x_3$ & $Mdm2\_mRNA$ synthesis & $k_7 \langle x_1\rangle$ & $1\times 10^{-4}{sec}^{-1}$ & [@mol; @pro; @jah]\
8 & $x_3\stackrel{k_{8}}{\longrightarrow}x_2+x_3$ & $Mdm2$ synthesis & $k_8 \langle x_3\rangle$ & $495\times 10^{-5}{sec}^{-1}$ & [@mol; @pro; @jah]\
9 & $x_3\stackrel{k_{9}}{\longrightarrow}\phi$ & $Mdm2\_mRNA$ degradation & $k_9 \langle x_3\rangle$ & $1\times 10^{-4}{sec}^{-1}$ & [@mol; @pro; @jah]\
10 & $x_1+x_2\stackrel{k_{10}}{\longrightarrow}x_5$ & $p53\_Mdm2$ complex formation & $k_{10} \langle x_1\rangle \langle x_2\rangle$ & $1155\times 10^{-3}{sec}^{-1}$ & [@mol; @pro; @jah]\
11 & $x_5\stackrel{k_{11}}{\longrightarrow}x_2$ & Mdm2 creation & $k_{11} \langle x_5\rangle$ & $825\times 10^{-4}{sec}^{-1}$ & [@mol; @pro; @jah]\
12 & $x_5\stackrel{k_{12}}{\longrightarrow}x_1+x_2$ & Dissociation of $p53\_Mdm2$ complex & $k_{12} \langle x_5\rangle$ & $1155\times 10^{-5}{sec}^{-1}$ & [@mol; @pro; @jah]\
13 & $x_2\stackrel{k_{13}}{\longrightarrow}\phi$ & $Mdm2$ degradation & $k_{13} \langle x_2\rangle$ & $433\times 10^{-4}{sec}^{-1}$ & [@mol; @pro; @jah]\
14 & $\phi\stackrel{k_{14}}{\longrightarrow}x_6$ & ROS formation & $k_{14}$ & $1.0\times 10^{-2}{sec}^{-1}$ & [@proc; @pro]\
15 & $x_6\stackrel{k_{15}}{\longrightarrow}\phi$ & Degradation of ROS & $k_{15} \langle x_6\rangle$ & $2\times 10^{-2}{sec}^{-1}$ & [@proc; @pro]\
16 & $x_6\stackrel{k_{16}}{\longrightarrow}x_7$ & Initiation of DNA damage & $k_{16} \langle x_6\rangle$ & $2\times 10^{-2}{sec}^{-1}$ & [@pro; @proc]\
17 & $x_7\stackrel{k_{17}}{\longrightarrow}\phi$ & DNA repair & $k_{17} \langle x_7\rangle$ & $2\times 10^{-5}{sec}^{-1}$ & [@pro; @proc]\
18 & $x_7\stackrel{k_{18}}{\longrightarrow}x_8+x_7$ & Activation of ARF & $k_{18} \langle x_7\rangle$ & $33\times 10^{-5}{sec}^{-1}$ & [@zha; @pro; @kha]\
19 & $x_8\stackrel{k_{19}}{\longrightarrow}\phi$ & Degradation of ARF & $k_{19} \langle x_{8}\rangle$ & $1\times 10^{-4}{sec}^{-1}$ & [@zha; @pro; @kha]\
20 & $x_8+x_2\stackrel{k_{20}}{\longrightarrow}x_9$ & $ARF\_Mdm2$ complex formation & $k_{20} \langle x_{8}\rangle \langle x_2\rangle $ & $1\times 10^{-2}{sec}^{-1}$ & [@zha; @pro; @kha]\
21 & $x_9\stackrel{k_{21}}{\longrightarrow}x_8$ & Dissociation of $ARF\_Mdm2$ complex & $k_{21} \langle x_9\rangle$ & $1\times 10^{-3}{sec}^{-1}$ & [@zha; @pro; @kha]\
22 & $x_{10}\stackrel{k_{22}}{\longrightarrow}\phi$ & Degradation of Micro RNA & $k_{22} \langle x_{10}\rangle $ & $5\times 10^{-2}{sec}^{-1}$ & [@lee1; @win]\
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In this paper, we examine how well a rational point $P$ on an algebraic variety $X$ can be approximated by other rational points. We conjecture that if $P$ lies on a rational curve, then the best approximations to $P$ on $X$ can be chosen to lie along a rational curve. We prove this conjecture for a wide range of examples, and for a great many more examples we deduce our conjecture from Vojta’s Main Conjecture.'
address: |
Department of Pure Mathematics\
University of Waterloo\
Waterloo, ON N2L 3G1\
Canada
author:
- David McKinnon
title: A conjecture on rational approximations to rational points
---
\[section\] \[thm\][Lemma]{} \[thm\][Definition]{} \[thm\][Corollary]{} \[thm\][Conjecture]{} \[thm\][Claim]{} \[thm\][Example]{} \[thm\][Remark]{} ¶[[P]{}]{} [v]{} Ø[[O]{}]{}
[^1]
Introduction
============
The distribution of rational points on an algebraic variety $X$ is very subtle. For example, it can often happen that there is a proper subvariety $Y$ of $X$ such that the set of rational points of $Y$ has density one in the set of rational points of $X$, where density is used in the sense of Weil heights. Roughly speaking, such subvarieties $Y$ are called accumulating subvarieties. Thus, if one is interested in the arithmetic of $X$, one must first identify which rational points of $X$ lie on $Y$, and which do not.
Unfortunately, even the purely geometric problem of identifying potential accumulating subvarieties $Y$ can be very difficult. It would be helpful to have a local and arithmetic criterion to identify points $P$ which lie on accumulating subvarieties. In this paper, we do not quite manage to construct such a criterion, but we do identify an invariant, called the approximation constant of $P$ on $X$ with respect to a divisor $D$, which describes how well $P$ can be approximated by other rational points of $X$. (See Definition \[bigdef\] for details.) If $P$ lies on an accumulating subvariety, then it can be well approximated by other rational points, and therefore will have a small approximation constant.
In the course of computing this constant in many examples, it became clear that sequences which best approximate a rational point $P$ tend to lie along curves. While there are counterexamples to show that this principle cannot hold in general, we are able to formulate a conjecture which predicts that it should be true whenever $P$ lies on a rational curve defined over the field of coefficients (see Conjecture \[ratcurve\]). We prove this conjecture for a wide range of examples.
The basic technique we use for proving Conjecture \[ratcurve\] is inductive. We start by proving the conjecture for $\P^n$ (see Theorem \[ratapprox\]). We then prove a number of inductive results which enable us to take advantage of the structure of the Néron-Severi group of $X$. Using these inductive results, we prove Conjecture \[ratcurve\] for a wide range of rational surfaces by means of a careful analysis of the nef cone.
In the final section, we make several remarks about further directions in which to attack Conjecture \[ratcurve\], and drawbacks of our current techniques. We describe some cases which are unlikely to be proven using our current techniques, and we further deduce Conjecture \[ratcurve\] from Vojta’s Main Conjecture (Conjecture 3.4.3 of [@Vo]) for a generic variety of non-negative Kodaira dimension.
I am grateful to several people for helpful conversations about this material, including Doug Park, Mike Roth, and Cam Stewart. I would particularly like to thank Kevin Hare for writing some extremely useful computer programs, without which many of the theorems in this paper would still be conjectures, and the anonymous referee, whose invaluable suggestions improved the manuscript enormously.
Lines
=====
In this paper, all heights are absolute and multiplicative. We fix a number field $k$.
\[ratapprox\] Let $P\in\P^n(k)$ be any $k$-rational point. Let $\{P_i\}$ be any sequence of $k$-rational points, and let $S$ be the set of archimedean places of $k$, with the convention that pairs of complex conjugate embeddings count as one place. Suppose there is a positive real constant $c\in\R$ such that for all $i$, we have: $$\left(\sum_{v\in S}\operatorname{dist}_v(P,P_i)\right)H(P_i)\leq c$$ where the function $\operatorname{dist}_v$ denotes the local distance function on $\P^n(k)$ induced by $v$. Then the set $\{P_i\}$ is a subset of the finite union of lines through $P$ of Plücker height at most $N(P)c$, where $N(P)$ is a constant which does not depend on $c$.
[*Proof:*]{} Without loss of generality, we may assume that $P=[0:0:\ldots
:0:1]$, and that none of the $P_i$ lie on the line $x_n=0$. Fix a set $R$ of representatives of the class group of $k$.
Write $P_i = [a_{i1}:\ldots:a_{in}:b_i]$, where the $a_{ij}$ and $b_i$ are integers (that is, elements of $\O_k$) with $\gcd(a_{i1},\ldots,a_{in},b_i)\in R$. Let $v\in S$ be any archimedean place of $k$. Up to multiplication by a bounded function (which won’t affect the conclusion of the theorem), we may write the $v$-distance as: $$\operatorname{dist}_v(P,P_i) = \max_j\{|a_{ij}/b_i|_v\}$$ and the height as: $$H(P_i) = \left(\prod_{v\in S}\max\{|a_{i1}|_v,\ldots,|a_{in}|_v,|b_i|_v\}
\right)^{1/[k:\Q]}$$ Note that with our choice of representation, the non-archimedean absolute values do not contribute more than a (multiplicatively) bounded function to the height.
Now suppose that $\sum_{v\in S}\operatorname{dist}_v(P,P_i)H(P_i)\leq c$ for all $i$. Then we get: $$\left(\sum_{v\in S}\max\{|a_{ij}/b_i|_v\}\right)^{[k:\Q]}
\prod_{v\in S}\max\{|a_{i1}|_v,\ldots,|a_{in}|_v,|b_i|_v\} \leq c^{[k:\Q]}$$ and hence $$\left(\sum_{v\in S}\max\{|a_{ij}|_v\}\right)^{[k:\Q]}
\max\{|a_{i1}/b_i|_v,\ldots,|a_{in}/b_i|_v,1\} \leq c^{[k:\Q]}$$ and so [*a fortiori*]{}, we obtain: $$\sum_{v\in S}\max\{|a_{ij}|_v\}\leq c$$ In particular, there are only finitely many choices for each $a_{ij}$, and each choice corresponds to a line of height at most $c^{[k:\Q]}$.
In light of this theorem, we make the following definitions:
Let $X\subset\P^n_k$ be an algebraic variety defined over $k$, and let $S$ be the set of archimedean absolute values on $k$. Define: $$\operatorname{dist}(P,Q) = \sum_{v\in S} \operatorname{dist}_v(P,Q)$$ for any rational points $P$ and $Q$ on $X$. Note that this is a well defined distance function on the metric space $\prod_{v\in S} X(k_v)$, where $k_v$ is the completion of $k$ with respect to $v$.
In this paper, all limits will be computed with respect to the topology induced by the distance function $\operatorname{dist}$.
\[appconst\] Let $X$ be an algebraic variety defined over $k$, and let $P\in X(k)$ be any rational point. Let $D$ be any divisor on $X$, with corresponding height function $H_D$. Assume that there is a positive constant $c$ such that $H_D(Q)>c$ for all $Q$ in some Zariski open neighbourhood of $P$. (This will be satisfied if, for example, $D$ is ample, or more generally, if some multiple of $D$ is basepoint free.) For any sequence $\{P_i\}\subset X(k)-\{P\}$ with $P_i\rightarrow P$, define the [*approximation constant on $X$ of $\{P_i\}$ with respect to $P$ and $D$*]{} to be the smallest non-negative real number $\alpha$ such that: $$\limsup_{i\rightarrow\infty}\operatorname{dist}(P,P_i)^\alpha H_D(P_i)<\infty$$ If there is no such smallest non-negative real number, then the sequence does not have an approximation constant.
Notice that although the definition of $\operatorname{dist}$ depends on the choice of embedding of $X$ in projective space, the definition of the approximation constant does not. This is because two different embeddings of $X$ are diffeomorphic, and hence distances change by no more than a multiplicative function bounded away from zero and infinity. In this paper, this ambiguity will never be significant, so we will refer to $\operatorname{dist}$ as a function independent of the embedding of $X$.
\[noconst\] For example, let $P=[0\colon0\colon1]\in\P^2$, and let $f(n)=n/\log n
$. On any $\Q$-rational line $L$ through $P$, we can find an infinite sequence of rational points $\{P_i\}$ such that $\operatorname{dist}(P,P_i)
H(P_i)\leq c_L$, where $c_L$ is a constant that depends on $L$. (See for example Theorem 2.1 of [@M1].) Therefore, there is a positive constant $C$ and an infinite sequence $\{Q_n\}\subset\P^2$ of rational points such that $\operatorname{dist}(Q_n,P)f(H(Q_n))<C$ for all $n$. Moreover, we can choose the points $\{Q_n\}$ such that no two of them are collinear with $P$. Thus, by Theorem \[ratapprox\], the approximation constant of this sequence, if it exists, is greater than one. But it is clear that for every $\epsilon >0$, there exists a positive constant $C_\epsilon$ such that $\operatorname{dist}(Q_n,P)^{1+\epsilon}H(Q_n)<C_\epsilon$, so any constant of approximation for this sequence is at most one. We conclude that the sequence $\{Q_n\}$ does not have a constant of approximation.
\[bigdef\] Let $X$, $P$, $D$, and $H_D$ be as in Definition \[appconst\]. Define the [*approximation constant of $P$ on $X$ with respect to $D$*]{} to be the minimum (if it is achieved) of all approximation constants on $X$ of sequences $\{P_i\}\subset X(k)-\{P\}$ with respect to $D$. A [*sequence of best approximation to $P$ with respect to $D$*]{} is a sequence whose corresponding approximation constant is equal to the approximation constant of $P$. A [*curve of best approximation to $P$*]{} is a curve on $X$ passing through $P$ that contains a sequence of best approximation to $P$.
Notice that by this definition, sequences with no approximation constant (such as the one described in Example \[noconst\]) cannot be a sequence of best approximation to a point $P$. If Conjecture \[ratcurve\] is true, then Theorem \[curve\] makes it clear that this is a reasonable restriction.
We may therefore improve Theorem \[ratapprox\] as follows.
\[projlines\] Let $P\in\P^n(k)$ be any rational point. Then there exists a sequence of best approximation to $P$. Moreover, any such sequence is a subset of a finite set of lines and has constant of approximation equal to 1, and any line through $P$ is a curve of best approximation to $P$.
[*Proof:*]{} We first prove the theorem for $\P^1$. Without loss of generality, we may assume that $P=[0:1]$. Fix a set of representatives $R$ of the class group of $k$. Let $Q=[a\colon b]$ be any point in $\P^1(k)$, where $a$ and $b$ lie in $\O_k$ with $(a,b)\in
R$ and $a\neq 0$. Since $\{[1/i:1]\}$ is clearly a sequence with constant of approximation equal to 1, it suffices to show that $\operatorname{dist}(Q,P)H(Q)\geq C_k$, where $C_k$ is a constant depending only on $k$.
Since the arithmetic mean is always greater than the geometric mean, for $Q$ close enough to $P$ we compute: $$\operatorname{dist}(P,Q)=\sum_v|a/b|_v\geq [k:\Q]\prod_v|a/b|_v^{1/[k:\Q]}$$ $$=[k:\Q](N(a)/N(b))^{1/[k:\Q]}$$ where all sums and products are over archimedean places $v$, and $N$ denotes the norm $N_{k/\Q}$. We also compute: $$H(Q)=(\prod_v\max\{|a|_v,|b|_v\})^{1/[k:\Q]}\geq N(b)^{1/[k:\Q]}$$ where the product again ranges over archimedean places $v$. Note that since $(a,b)\in R$, if we change the representation of $Q$, then we will change the ideal $(a,b)$ by multiplication by an element of $k^*$. Since no two elements of $R$ represent the same ideal class, it follows that the representation changes by multiplication by a unit of $\O_k$, and so $H(Q)$ does not change. Since $a\in\O_k$ is nonzero, it follows that $N(a)\geq 1$, so that: $$\operatorname{dist}(P,Q)H(Q)\geq [k:\Q]$$ and we have proven the result for $\P^1$.
In general, if $P\in\P^n$ is any rational point, we can find an embedding $\phi\colon\P^1\hookrightarrow\P^n$ such that $\phi(\P^1)$ is a line, and $\phi(0:1)=P$. Since $\phi$ only changes distances and heights by multiplication by a function bounded away from zero and infinity, it follows that a sequence of best approximation to $P$ along the line $\phi(\P^1)$ has constant 1. But by Theorem \[ratapprox\], any sequence with approximation constant at most 1 must lie in a finite set of lines through $P$. The result follows.
In light of Theorem \[ratapprox\], we make the following conjecture:
\[ratcurve\] Let $V$ be an algebraic variety defined over $k$, and $D$ any ample divisor on $V$. Let $P$ be any $k$-rational point on $V$. Assume that there is some rational curve $C$ on $V$, defined over $k$, such that $P\in C(k)$. Then a sequence of best approximation to $P$ on $V$ with respect to $D$ exists, and may be chosen to lie along some rational curve through $P$.
It is easy to see that if we remove the hypothesis that $P$ lie on a rational curve, then the conjecture is false. Consider the case of a simple abelian variety $V$ of dimension greater than one, and $P$ any $k$-rational non-torsion point. Then $P$ is certainly a limit point of the set $V(k)$ in the archimedean topology, but any curve through $P$ must have geometric genus at least two, which by Faltings’ Theorem can contain only finitely many $k$-rational points.
The following theorem is quite useful:
\[curve\] Let $C\subset\P^n$ be any irreducible rational curve of degree $d$ defined over $k$, and let $P\in C(k)$ be any $k$-rational point. Let $f\colon\P^1\rightarrow C$ be the normalization map, and let $m$ be the maximum multiplicity of a branch of $f$ over $P$. Then a sequence of best approximation to $P$ on $C$ has constant of approximation $d/m$.
[*Proof:*]{} If $P$ is a smooth point of $C$, then the result follows trivially from the observation that $f$ changes distances to $P$ only by a multiplicative function bounded away from zero and infinity, and $f$ raises heights to the power $d$ (up to a multiplicative function bounded away from zero and infinity).
If $P$ is a singular point of $C$, then up to a multiplicative function bounded away from zero and infinity, we have $\operatorname{dist}(f(P),f(Q))=\operatorname{dist}(R,Q)$ near $P$, where $R$ is some point in the finite set $f^{-1}(P)$. Even though the point $R$ depends on the point $Q$ near $P$, it nevertheless provides only a finite number of alternatives for $\operatorname{dist}(f(P),f(Q))$, so that any sequence of best approximation to $P$ must have an infinite subsequence which is a sequence of best approximation to some element of $f^{-1}(P)$.
Thus, let $R\in f^{-1}(P)$ be any point, and let $m$ be the multiplicity of the branch of $f$ through $R$. Assume without loss of generality that $R=0$. Near $R$, in affine coordinates, the function $f$ can be written as $f(x)=(f_1(x),\ldots,f_n(x))$, where $C\subset\P^n$ and the $f_i$ are rational functions in $x$. The multiplicity $m_R$ of $f$ at $R$ is equal to $\min_i\{\ord_0(f_i(x))\}$, where $\ord_0(f_i)$ denotes the largest power of $x$ which divides the numerator of $f_i$ (in lowest terms). Let $i$ be some index which achieves the minimum, and write $f_i(x)=x^{m_R}+O(x^{m_R+1})$. Then as $Q=x$ approaches $R=0$, the distance between $Q$ and $R$ is $\sum_v|x|_v$, and the distance between $f(Q)$ and $f(R)$ is (up to a bounded multiplicative constant) $\sum_v|x^{m_R}|_v\ll\gg(\sum_v|x|_v)^{m_R}$. The result now follows from Theorem \[ratapprox\].
Theorem \[curve\] explains why Conjecture \[ratcurve\] does not refer specifically to rational curves of minimal $D$-degree. A rational point $P$ may be better approximated along a curve of higher degree on which it is a cusp of high multiplicity than along a smooth curve of much lower degree.
\[cusp\] Let $C$ be the plane curve $y^2z=x^3\subset\P^2$, and $P$ the cusp $[0:0:1]$. For $k=\Q$, it is not hard to see that a sequence of best approximation to $P$ along $C$ is given by the sequence $[1/B^2:1/B^3:1]$, which has constant of approximation $3/2$.
However, let $f\colon\P^1\rightarrow C$ be the normalization map. Then $f^*\O(1)$ is $D=\O(3)$. A sequence of best $D$-approximation to $P$ along $\P^1$ has constant $3$, not $3/2$, since $D$ corresponds to an embedding of $\P^1$ in $\P^3$ as a twisted cubic curve. This is because in Definition \[appconst\], we define $\operatorname{dist}$ in terms of an embedding of $\P^1$, not an arbitrary morphism, whereas in this example, we deal with the distance inherited from $\P^2$.
In other words, if $\tilde{W}\to W$ is the normalization of a subvariety of a smooth variety $V$, then the function $\operatorname{dist}\colon
\tilde{W}\times \tilde{W}\rightarrow\R$ may not agree with the pullback of the function $\operatorname{dist}\colon V\times V\rightarrow\R$ to $\tilde{W}\times \tilde{W}$. If $W$ is a curve, then the proof of Theorem \[curve\] describes how these two functions differ, but if $W$ has higher dimension, then one must be more careful.
Despite the unpleasant possibility described by Example \[cusp\], we can relate Conjecture \[ratcurve\] to the problem of finding accumulating curves through a point $P$.
Let $S\subseteq\P^n(k)$ be any set of $k$-rational points. The counting function for $S$ is defined to be: $$N_S(B)=\#\{P\in S\mid H(P)\leq B\}$$ where $H$ denotes the standard height function on $\P^n$.
Let $V\subseteq\P^n$ be an algebraic variety defined over $k$, and let $W\subset V$ be any proper closed subset. Then $W$ is said to be an accumulating subvariety of $V$ if and only if: $$N_{V-W}(B)=o(N_W(B))$$ where by $N_{V-W}$ and $N_W$ we mean $N_{V(k)-W(k)}$ and $N_{W(k)}$, respectively. Roughly speaking, an accumulating subvariety of $V$ is a proper closed subset $W$ such that asymptotically, there are more rational points of bounded height on $W$ than there are off $W$.
Let $V\subseteq\P^n$ be a smooth algebraic variety defined over $k$, and let $P\in V(k)$ be any $k$-rational point. Assume that Conjecture \[ratcurve\] is true for $P$ on $V$ — that is, assume that there is a rational curve $C$ through $P$ which contains a sequence of best approximation to $P$. Assume further that $C$ has only ordinary singularities at $P$. If $P$ lies on an accumulating curve of $V$, then that curve contains $C$ as an irreducible component.
[*Proof:*]{} First, note that since $V$ contains a rational curve $C$ through $P$, it follows that any accumulating curve containing $P$ must be a rational curve, and moreover any component containing $P$ must be a rational curve of minimal degree through $P$. By Theorem \[curve\], since $C$ has only ordinary singularities through $P$, it is clear that any rational curve through $P$ must have degree at least $\deg C$. Thus, $C$ is a rational curve of minimal degree through $P$, and therefore any accumulating curve through $P$ must contain $C$ as an irreducible component.
Other Varieties
===============
Next, we consider the question of rational approximation of rational points on more general varieties. We begin with a straightforward but surprisingly useful result on products of varieties:
\[product\] Let $X$ and $Y$ be smooth algebraic varieties defined over a number field $k$, and let $P$ and $Q$ be $k$-rational points on $X$ and $Y$, respectively. Let $L_X$ and $L_Y$ be divisors on $X$ and $Y$, respectively, whose corresponding height functions $H_X$ and $H_Y$ are bounded below by a positive constant in some Zariski open neighbourhoods of $P$ and $Q$, respectively. Let $\{P_i\}$ (respectively $\{Q_i\}$) be a sequence of best $L_X$-approximation (respectively best $L_Y$-approximation) for $P$ (respectively $Q$), with constant of approximation $\alpha$ (respectively $\beta$). Then either $\{(P_i,Q)\}$ or $\{(P,Q_i)\}$ is a sequence of best $L$-approximation for $(P,Q)$ on $X\times Y$, where $L=\pi_1^*L_X\otimes\pi_2^*L_Y$ ($\pi_i$ is the projection onto the $i$th factor). Moreover, any sequence $\{(S_i,T_i)\}$ whose constant of approximation is less than $\alpha+\beta$ must have either $S_i=P$ or $T_i=Q$ for all but finitely many $i$.
[*Proof:*]{} For any $k$-rational point $(S,T)$ on $X\times Y$, we have $H_L(S,T) = H_X(S)H_Y(T)$, where $H_X$ and $H_Y$ are heights on $X$ and $Y$ with respect to the divisors $L_X$ and $L_Y$, respectively. To prove the theorem, note that the non-negativity of $\alpha$ and $\beta$ implies that it suffices to prove the last claim.
Let $\{(S_i,T_i)\}$ be any sequence of points in $(X\times
Y)(k)-\{(P,Q)\}$ with $(S_i,T_i)\rightarrow (P,Q)$. We know that if $\{S_i\}$ is infinite then $\operatorname{dist}(P,S_i)^{\alpha-\epsilon} H_X(S_i)$ is unbounded for any $\epsilon>0$. Similarly, if $\{T_i\}$ is infinite then $\operatorname{dist}(Q,T_i)^{\beta-\epsilon} H_Y(T_i)$ is unbounded for any $\epsilon>0$. Thus, if there are infinitely many $i$ for which both $S_i\neq P$ and $T_i\neq Q$, then we deduce that the following function is unbounded: $$\max\{\operatorname{dist}(P,S_i),\operatorname{dist}(Q,T_i)\}^{\alpha+\beta-\epsilon}H_X(S_i)H_Y(T_i)$$ We may now deduce that the following function is unbounded: $$\operatorname{dist}((P,Q),(S_i,T_i))^{\alpha+\beta-\epsilon}H_L(S_i,T_i)$$ which is to say that $\{(S_i,T_i)\}$ has constant of approximation at least $\alpha+\beta$.
We have the following immediate corollary:
\[divsum\] Let $X$ be a variety defined over $k$, and let $P\in X(k)$ be any $k$-rational point. Let $D_1$ and $D_2$ be two divisors on $V$ with height functions $H_1$ and $H_2$ bounded below by a positive constant in a neighbourhood of $P\in X(k)$. Assume that $C_i$ is a curve of best $D_i$-approximation to $P$ for each $i$, and let $D=a_1D_1+a_2D_2$ be any positive linear combination of $D_1$ and $D_2$.
- If $C_1=C_2=C$, then $C$ is also a curve of best $D$-approximation to $P$.
- If $C_1.D_2=0$, then either $C_1$ is a curve of best $D$-approximation to $P$, or $C_2.D_1=0$ and $C_2$ is a curve of best $D$-approximation to $P$.
The following theorem will be used in the proofs of Theorem \[four\] and Theorem \[onefibre\].
\[addeffective\] Let $X$ be a variety defined over $k$, and let $D$ and $E$ be divisors on $X$. Let $P\in X(k)$ be a rational point, and $C$ a curve of best $D$-approximation to $P$. Assume that $H_D$ and $H_{D+E}$ are bounded below by positive constants in some neighbourhood of $P$, that $E$ is effective, and that $C\cap E=\emptyset$. Then $C$ contains a sequence of best $(D+E)$-approximation to $P$.
[*Proof:*]{} Since $C\cap E=\emptyset$, we can find a real, positive constant $\alpha$ such that $H_{E}(Q)\in [1/\alpha,\alpha]$ for all $Q\in C(k)$. Thus, the sequence of best $D$-approximation to $P$ along $C$ will have the same constant of approximation with respect to $D+E$ as with respect to $D$. However, $\log H_E$ is bounded from below away from $E$, and $P\not\in E(k)$. It follows that for any sequence $\{Q_i\}$ of rational points on $V$ with $Q_i\rightarrow P$, if the quantity $\operatorname{dist}(P,Q_i)^\gamma H_D(Q_i)$ is unbounded, then the quantity $\operatorname{dist}(P,Q_i)^\gamma H_{D+E}(Q_i)$ will also be unbounded. The theorem follows.
These theorems, surprisingly, give an immediate proof of Conjecture \[ratcurve\] for any split, geometrically minimal surface $X$ over $k$. We say that a surface $X$ is split over $k$ if and only if the inclusion of Néron-Severi groups $NS(X_k)\rightarrow
NS(X_\C)$ is an isomorphism – that is, if every algebraic cycle on $X_\C$ is numerically equivalent to some $k$-rational cycle. The proof relies crucially on the classification of minimal rational surfaces over $\C$, which can be found in, for example, [@Be].
\[minimal\] Conjecture \[ratcurve\] holds for any split, geometrically minimal rational surface defined over $k$. In particular, it holds for all the Hirzebruch surfaces $H_n$ for $n\geq 0$.
[*Proof:*]{} The classification shows that any minimal rational surface over $\C$ is either $\P^2$, $H_0=\P^1\times\P^1$, or a Hirzebruch surface $H_n$ for some integer $n>0$. The case of $\P^2$ is proven in Theorem \[ratapprox\]. The case $\P^1\times\P^1$ is immediate from Theorem \[product\]. The case of $H_n$ follows from Theorem \[product\] and Corollary \[divsum\] as follows. Let $A(H_n)$ be the closure in $N(H_n)=NS(H_n)\otimes\R$ of the cone of ample divisors on $H_n$, where $NS(H_n)$ is the Néron-Severi group of $H_n$. Then $A(H_n)$ is generated by the two divisor classes $D=S+nF$ and $F$, where $F=\pi^*(P)$ is a fibre of the map $\pi\colon
H_n\rightarrow\P^1$ and $S$ is the unique section of $\pi$ with self-intersection $-n$. The class $D$ is $f^*(L)$, where $f$ is the contraction of $S$ to the vertex of the cone $C$ over a rational normal curve of degree $n$ and $L$ is a line of the ruling of the cone.
Consider the classes $D$, $F$, and $D+F$. For any rational point $P$ on $H_n$, it is clear that a curve of best $F$-approximation to $P$ is a fibre of $\pi$, and a curve of best $D$-approximation to $P$ is the preimage of a line through $f(P)$ on the cone $C$ (if $P\not\in
S(k)$) or $S$ (if $P\in S(k)$). If $P\not\in S(k)$, then these are the same curve, so we are done by Corollary \[divsum\]. If $P\in
S(k)$, then we simply note that the divisor $D+F$ corresponds to an embedding of $H_n$ in projective space such that fibres of $\pi$ are all lines, and $S$ is also a line, so they both contain sequences of best $(D+F)$-approximation to $P$. We again conclude by Corollary \[divsum\], since every very ample divisor on $H_n$ is a positive linear combination either of $D$ and $D+F$ or $F$ and $D+F$.
\[simplefibres\] Let $n\geq 2$, and let $X$ be the blowup of the Hirzebruch surface $H_n$ at $k<n$ points, no two of which lie in the same fibre of the map to $\P^1$. Conjecture \[ratcurve\] is true for $X$.
[*Proof:*]{} First, it clearly suffices to assume that none of the blown up points lies on the $(-n)$-section on $H_n$, since blowing up points on this section will only increase $n$.
Let $S$ be the class of the strict transform of the $(-n)$-section on $H_n$, and let $F$ be the class of the total transform of a fibre of the map from $H_n$ to $\P^1$. For each of the $k$ reducible fibres, let $E_i$ and $F_i$ be the two components, of which $F_i$ is the one which intersects $S$.
Let $\alpha\in\{0,1\}^k$ be any vector, and define a divisor $D_\alpha$ on $X$ by: $$D_\alpha = S+nF-\alpha\cdot(E_1,\ldots,E_k)$$ where the $\cdot$ denotes a formal dot product.
The effective cone of $X$ is generated by the divisors $E_i$, $F_i$, and $S$. The nef cone of $X$ is generated by the divisors $D_\alpha$ and $F$, where $\alpha$ ranges over all of $\{0,1\}^k$.
[*Proof:*]{} The Picard rank of $X$ is $k+2$, so an arbitrary divisor on $X$ can be written as $D=aS+bF+\sum f_iE_i$. If $D$ is to be ample, then its intersection with $S$, $E_i$, and $F_i$ must be positive, giving: $$b-na>0\hspace*{1in}a+f_i>0\hspace*{1in}f_i<0$$ These inequalities define an open cone in $\R^{k+2}$ – let $C$ be the closure of this cone. Then $C$ is finitely generated, and the generating rays of $C$ are all intersections of $k+1$ of the hyperplanes $V=\{b-na=0\}$, $V_i=\{a+f_i=0\}$, and $W_i=\{f_i=0\}$.
Let $\v$ be a generator of $C$. If $\v\not\in V$, then for some $j$, $\v\in V_j\cap W_j$, since $\v$ must be contained in at least $k+1$ of the listed hyperplanes. Thus, $\v$ is contained in the hyperplane $a=0$. Furthermore, since $\v$ is an extremal ray of $C$, there must be some set of exactly $k+1$ of the $V_i$ and $W_i$ whose intersection is the space generated by $\v$. Since the vector $F$ is contained in the intersection of all the $V_i$ and $W_i$, it follows that $\v$ is a positive multiple of $F$.
Thus, we may assume that $\v\in V$. If $\v\in V_j\cap W_j$ for some $j$, then $\v$ is again contained in $a=0$, and thus also in $b=0$. But there must also be some $\ell$ such that $\v\not\in V_\ell\cup W_\ell$, and hence $f_\ell>0$ and $f_\ell<0$. This is clearly a contradiction, and so for all $i$, $\v\not\in V_i\cap W_i$.
Let $\alpha\in\{0,1\}^k$ be the vector whose $i$th component is 0 if $\v\in V_i$, and 1 if $\v\in W_i$. Then $D_\alpha\in V_i$ if and only if $\v\in V_i$, and similarly for $W_i$. Since $D_\alpha\in V$, the independence of the $k+1$ hyperplanes shows that $\v$ is a positive multiple of $D_\alpha$.
Thus, the two cones described in the claim are indeed dual to one another. To conclude the proof of the claim, it suffices to show that every positive linear combination of the $D_\alpha$ and $F$ is ample. First, it is clear that $F$ and $D_\alpha$ are basepoint free, since $F$ corresponds to a morphism to $\P^1$, and $D_\alpha$ corresponds to the morphism from $X$ to a cone, blowing down $S$ and exactly one component of each reducible fibre. We next note that $F^2=0$ and $F.D_\alpha=1$, and: $$\begin{aligned}
D_\alpha.D_\beta & = & (S+nF)^2+(\alpha\cdot(E_1,\ldots,E_k))(\beta\cdot
(E_1,\ldots,E_k)) \\
& = & n - (\alpha\cdot\beta) \\
& \geq & n-k\end{aligned}$$ which is positive. Therefore, by the Nakai-Moishezon criterion ([@Ha], Theorem V.1.10), any positive linear combination of $F$ and the $D_\alpha$ is ample, and thus the claim is proven.
We now prove Conjecture \[ratcurve\] for $X$. If $P$ does not lie on $S$ or any reducible fibre, then for each $\alpha$, a sequence of best $D_\alpha$-approximation to $P$ is clearly contained in the component $C$ of $F$, and certainly a sequence of best $F$-approximation is also contained in $C$. Thus, for any ample $D$, $C$ is a curve of best $D$-approximation to $P$.
If $P$ lies on exactly one $E_i$ or $F_i$ (and not $S$), then the same analysis shows that this same $E_i$ or $F_i$ is a curve of best $D$-approximation to $P$ for any ample $D$. If $P$ lies on $S$ but no $E_i$ or $F_i$, then notice that $S.D_\alpha=0$ for any $\alpha$, so we conclude that $S$ is a curve of best $D_\alpha$-approximation to $P$ for all $\alpha$. Since $S.F=1$, and since $D_\alpha+F$ contracts no curves through $P$, we conclude from Theorem \[projlines\] that $S$ and $C$ are both curves of best $(D_\alpha+F)$-approximation to $P$, for any $\alpha$. Since any ample divisor $S$ is either a positive linear combination of elements of the set $\{D_\alpha,
F+D_\alpha\}$, or a positive linear combination of elements of the set $\{F,F+D_\alpha\}$, we conclude from Corollary \[divsum\] that either $S$ or $C$ is a curve of best $D$-approximation to $P$ for any ample $D$.
If $P=S\cap E_i$ for some $i$, then a similar analysis shows that for any ample divisor $D$, either $S$ or $E_i$ is a curve of best $D$-approximation to $P$. In particular, if we divide the nef cone of $X$ into two subcones according to which of $S$ or $E_i$ has smaller degree, then it is straightforward to check that the curve of smaller degree has degree zero or one with respect to each generator of the corresponding subcone, and so by Theorem \[projlines\] and Corollary \[divsum\], we conclude that the curve of smaller $D$-degree is always a curve of best $D$-approximation to $P$. If $P=E_i\cap F_i$ for some $i$, a similar calculation shows that either $E_i$ or $F_i$ — whichever has smaller $D$-degree — is always a curve of best $D$-approximation to $P$.
Theorem \[simplefibres\] can be used to show that any $k$-split rational surface of Picard rank at most three satisfies Conjecture \[ratcurve\].
\[three\] Let $X$ be a $k$-split rational surface, of Picard rank at most three. Then Conjecture \[ratcurve\] is true for $X$.
[*Proof:*]{} If the Picard rank is one, then $X$ is isomorphic to $\P^2$. If the Picard rank is two, then either it’s geometrically minimal, or else it’s $\P^2$ blown up at a point $P$. In this case, however, $X=H_1$, and the result follows from Corollary \[minimal\].
If the Picard rank of $X$ is three, then it must be the blowup of some $H_n$ at some point $P$, with $n\geq 0$. If $n=0$, then since any blowup of $H_0=\P^1\times\P^1$ is also a blowup of $H_1$, we may instead take $n=1$. If $n>1$, then we can apply Theorem \[simplefibres\] directly. Thus, we assume that $n=1$, and therefore that $X$ is the blowup of $\P^2$ at two different points.
We begin by finding generators for the effective cone of $X$. Let $\pi\colon X\rightarrow\P^2$ be the blowing down map, and let $E_1$ and $E_2$ be the two exceptional divisors. Let $S$ be the strict transform of the line in $\P^2$ which joins the two blown up points. If we write $L=S+E_1+E_2$, then $L$ is the class of $\pi^*\O(1)$. Then $L-E_1$ and $L-E_2$ correspond to morphisms $\psi_1$ and $\psi_2$ to $\P^1$. We have the following well known description of the effective and nef cones of $X$.
The curves $S$, $E_1$, and $E_2$ generate the closed cone $\overline{NE}(X)$ of effective divisors on $X$, and the classes $L$, $L-E_1$, and $L-E_2$ generate the nef cone of $X$.
We now prove Conjecture \[ratcurve\] for $X$. Let $P\in X(k)$ be any point, and let $D$ be any ample divisor. If $P\neq S\cap E_i$, let $C_i$ be the unique irreducible component of the fibre of $\psi_i\colon X\to\P^1$ through $P$. Each of $L$, $L-E_i$, and $A=(L-E_1)+(L-E_2)$ either contracts $C_i$, or else contracts no curves through $P$ and maps $C_i$ to a line. (Note that if $P$ lies on $S$, then $C_i=S$.) Thus, by Theorem \[projlines\] and Corollary \[divsum\], we conclude that $C_i$ is a curve of best $D$-approximation to $P$ for any $D$ in the cone generated by $L$, $L-E_i$, and $A$.
This covers all cases except $P=S\cap E_i$ and $D$ is a positive linear combination of $L$, $L-E_i$, and $A$. In this case, notice that by Theorem \[product\], both $S$ and $E_i$ are curves of best approximation to $P$ with respect to $2L-E_i$ and $L+A$. Since $L$ contracts $E_i$, this means that if $D$ lies in the cone generated by $L$, $2L-E_i$, and $L+A$, then $E_i$ is a curve of best $D$-approximation to $P$ by Corollary \[divsum\]. And since $L-E_i$ and $A$ both contract $S$, we conclude again by Corollary \[divsum\] that $S$ is a curve of best $D$-approximation to $P$ if $D$ is a positive linear combination of $L-E_i$, $2L-E_i$, $A$, and $L+A$. This concludes the proof of Conjecture \[ratcurve\] for $X$.
Proceeding to the other extreme from Theorem \[simplefibres\], we now prove a theorem about the case in which the map to $\P^1$ has exactly one reducible fibre.
\[onefibre\] Let $X$ be a smooth rational surface obtained by a succession of blowups of the Hirzebruch surface $H_n$, and let $\pi\colon
X\rightarrow\P^1$ be the associated map. Assume that $\pi$ has only one reducible fibre, with $m$ components, and assume $m<n$. If every multiple component of the reducible fibre intersects at least two other components then Conjecture \[ratcurve\] is true for every point of $X$.
[*Proof:*]{} If any of the blown-up points of $H_n$ lie on the $(-n)$-section $S$, then the strict transform of $S$ will have strictly smaller self-intersection than $-n$, and $X$ can be obtained as a blowup of $H_r$ for some $r>n$. Thus, if we choose $n$ large enough, we may therefore assume that none of the blown-up points of $H_n$ lie on $S$.
Let us establish some notation:
- The components of the reducible fibre of $\pi$ are denoted by $E_1,
\ldots,E_m$, and we write $E_0=S$ for the unique $(-n)$-section of $\pi$.
- The classes $E_0,\ldots,E_m$ are a basis for the Néron-Severi group $NS(X)$, and we denote by $D_0,\ldots,D_m$ the dual basis with respect to the intersection pairing. That is, $D_i.E_j=\delta_{ij}$.
- We write $F=m_1E_1+\ldots +m_mE_m$, so that $m_i$ is the multiplicity of $E_i$ as a component of the reducible fibre of $\pi$.
We proceed with a technical lemma.
\[inductivestep\] Assume without loss of generality that $E_m^2=-1$ and $E_m.S=0$, and let $f\colon X\rightarrow Y$ be the map which blows down $E_m$. Let $\phi\colon Y\rightarrow\P^1$ be the map satisfying $\pi=\phi\circ f$, let $E'_0$ be the $(-n)$-section of $\phi$, and let $E_1',\ldots,E_{m-1}'$ be the components of the unique reducible section of $\phi$, ordered so that $E_i'=f_*(E_i)$.
Let $\{D_0',\ldots,D_{m-1}'\}$ be the dual basis to $\{E'_0,E_1',\ldots,E_{m-1}'\}$. Then $D_i=f^*(D_i')$ if $0\leq i<m$.
Furthermore, $E_m$ can intersect either one or two other components $E_i$. If $E_m$ intersects one component $E_i$, then $D_m=f^*(D_i')-E_m$. If $E_m$ intersects $E_i$ and $E_j$, then $D_m=f^*(D_i')+f^*(D_j')-E_m$.
[*Proof of Lemma:*]{} The fact that $E_m$ can intersect at most two components $E_i$ follows from the fact that no point of $Y$ can lie on more than two components of the reducible fibre. The $D_i$ are defined by the property that $D_i.E_j=\delta_{ij}$, where we define $E_0=S$. It is a straightforward matter to verify this equality for all $i$ and $j$.
The next step is to compute the effective cone of $X$.
\[effectcone\] Assume that for all $i$, some multiple of the divisor $D_i$ is basepoint free. Then the effective cone of $X$ is generated by the components $\{E_1,\ldots,E_m\}$ of the reducible fibre and the $(-n)$-section $S$ of $\pi$. Furthermore, we have $D_0=F$ and $D_1=S+nF$.
[*Proof of Lemma:*]{} Let $E_1,\ldots,E_m$ be the set of components of the reducible fibre, where $E_1$ is the unique component which intersects $S$. Let $A=(a_{ij})_{i,j=0}^m$ be the intersection matrix of the $S$ and $E_i$; that is, let $a_{ij}=E_i.E_j$, where $E_0=S$. The only entry of $A$ which is not $0$ or $\pm 1$ is $a_{00}=-n$. We therefore may regard $A$ as a specialization of the matrix $A_N$ with entries in $\Z[N]$, where $(A_N)_{ij}=a_{ij}$ except that $(A_N)_{00}=-N$. Specializing $N=n$ reduces $A_N$ to $A$.
To check that $\{S,E_i\}$ generates the effective cone, it suffices to check that the dual cone is the nef cone. Since $\{S,E_i\}$ form a basis for the vector space $NS(X)\otimes\R$, the dual cone will be generated by the vectors $D_j = \sum b_{ij}E_i$, where $B_N=(b_{ij}(N))_{i,j=0}^m$ is the inverse of the matrix $A_N$, $b_{ij}=b_{ij}(n)$, and $E_0=S$. To show that this dual is the ample cone, it suffices to show that every positive linear combination of the $D_j$ is ample. By the Nakai-Moishezon criterion, since we assume that some multiple of each $D_i$ is basepoint free, it suffices to show that $b_{ij}=D_i.D_j\geq 0$ for all $i$ and $j$, and that for each $i$, there is some $j$ for which $D_i.D_j>0$.
First, note that $D_0=F$, the class of a fibre of $\pi$. To see this, note that $D_0.E_i=0$ for all $i>0$, and therefore $D_0=\lambda F$ for some $\lambda$ by the non-degeneracy of the intersection pairing. Since $D_0.S=1$, we conclude that $D_0=F$. In particular, the first row and column of $B$ are independent of $N$, and $b_{0j}>0$ for all $j>0$.
Now let $i>0$. Since the only entry of $A_N$ depending on $N$ is $(A_N)_{00}=-N$, Cramer’s Rule implies that the $i$th row $\r_i$ of $A_N$ is $\r_i=N\r_i^\prime+\r_i^\circ$, where $\r_i^\prime$ and $\r_i^\circ$ are independent of $N$. Specializing to $N=n$, we obtain $D_i=nD_i^\prime+D_i^\circ$. (Note that $F.E_j=0$ for all $j>0$, so that the $(0,0)$-cofactor matrix of $A_N$ has nontrivial kernel, and therefore zero determinant. We therefore conclude that the determinant of $A_N$ is $\pm1$, and in particular independent of $N$.) Furthermore, we know that if $\r^*_j$ denotes that $j$th column of $A_N^{-1}$, then $\r_i\cdot\r^*_j=\pm D_i.E_j=\delta_{ij}$ is independent of $N$ for all $j$, and hence $D_i^\prime.E_j=0$ for all $j>0$. Thus, $D_i^\prime=\lambda_i F$ for some $\lambda_i$.
Clearly $D_1=S+nF$, since $(S+nF).S=0$, $(S+nF).E_j=0$ if $j>1$, and $(S+nF).E_1=1$. Thus, $\lambda_1=1>0$. Furthermore, since $S+nF=D_1=\sum_j b_{j1}E_j$, this means that $b_{1j}=b_{j1}=\lambda_j
n$ is just $n$ times the multiplicity of $E_j$ in the fibre $F$ — in particular, $\lambda_j$ is a positive integer. It therefore remains only to show that the coefficient of $E_j$ in $D_i^\circ$ is at least $-m\lambda_j$, and that at least one coefficient is strictly greater than $-m\lambda_j$. (Recall that $m$ is the number of components in the reducible fibre.)
We proceed by induction on $m$. If $m\leq 1$, the result follows immediately from the preceding calculations, since $D_0$ and $D_1$ will be the full list of $D_i$’s. For general $m$, note that there must be some $E_i$ with $E_i^2=-1$ and $i\neq 1$ — without loss of generality, we may assume that $i=m$. Let $f\colon X\rightarrow Y$ be the map that blows down $E_m$. Then $Y$ is a smooth rational surface which admits a map $\phi \colon Y\rightarrow\P^1$ such that $\phi\circ
f=\pi$, and the unique singular fibre of $\phi$ has $m-1$ components.
Let $E^\prime_i=f_*(E_i)$ for $i<m$, and let $S'=f_*(S)$. Denote the dual basis of $\{S',E_i^\prime\}$ by $\{D_i^\prime\}$. By induction, the coefficient of $E^\prime_j$ in $D_i^{\prime\circ}$ is at least $-(m-1)\lambda_j$, since the multiplicity of $E^\prime_j$ in $f_*(F)$ is also $\lambda_j$. If $i<m$, then by Lemma \[inductivestep\], the coefficient of $E_j$ in $D_i^\circ$ is also at least $-(m-1)\lambda_j\geq -m\lambda_j$. If $i=m$, then the coefficient of $E_j$ in $D_i$ is at least $-(m-1)\lambda_j-1\geq -m\lambda_j$.
Thus, if $m\leq n$, then $D_i.D_j\geq 0$ for all $i$ and $j$. Since the pairing is non-degenerate, for each $i$ there is some $j$ for which $D_i.D_j>0$. This concludes the proof of the lemma.
\[strictremark\] Note that the proof of Lemma \[effectcone\] also shows that $D_i.D_j>0$ for all $i$ and $j$ except $i=j=0$.
The next step is to prove that for each $i$, some multiple of $D_i$ is basepoint free. Together with Lemma \[effectcone\], this will complete the computation of the effective cone of $X$.
Define a graph $G=(V,E)$ with vertex set $V=\{E_1,\ldots,E_m\}$ and edge set $E=\{(E_i,E_j)\mid E_i.E_j>0\}$. Since $E_i^2<0$, this defines a simple graph, and moreover, it is a tree. Root this tree at $E_1$; that is, define a partial order on $V$ by $E_j\succeq E_i$ if and only if there is a simple path (that is, a path with no repeated vertices) from $E_1$ to $E_j$ which contains the vertex $E_i$. Call $E_i$ a leaf of $G$ if and only if $E_j\succeq E_i$ implies $i=j$. The hypothesis of the theorem is equivalent to demanding that the leaves of $G$ have multiplicity one in the fibre $F$.
\[leafbasepointfree\] Let $E_i$ be a leaf of $G$. Then $D_i$ is basepoint free.
[*Proof:*]{} We proceed by double induction on $n$ and $m$. If $n=1$ there there is nothing to prove. If $m=1$, then $X$ is the Hirzebruch surface $H_n$ and the lemma is clear. If $n=2$, then we must have $m\leq 1$, and the result is again clear. Note also that if $m>1$, then the sum of the multiplicities of the $E_i$ with $E_i^2=-1$ is always at least two, and that this property will be preserved by blowing up.
Now consider a general $n>2$, and let $E_i$ be a leaf of $G$. Since $E_i$ has multiplicity one by hypothesis, there must be some other $(-1)$-curve $E_j$. (It is possible that $j=1$.) Let $f\colon
X\rightarrow Y$ be the map that blows down $E_j$. Then $Y$ will also be a rational surface with a map $\phi\colon Y\rightarrow\P^1$ with at most one reducible fibre. The unique $(-\ell)$-section of $\phi$ will satisfy $\ell\leq n$, and the reducible fibre will have $m-1$ components. Therefore, by induction, the dual divisor $D_i^\prime$ to $f_*(E_i)$ is basepoint free. But $D_i=f^*(D_i^\prime)$, so $D_i$ is basepoint free as well.
\[multiplegens\] Assume that $E_i\succeq E_j$. Then $m_iD_j-m_jD_i$ is an effective divisor supported on the reducible fibre of $\pi$. Furthermore, if $E_i$ and $E_j$ intersect, then $m_iD_j-m_jD_i=\sum_{E_t\succeq
E_i}m_tE_t$.
[*Proof:*]{} Since $F.(m_iD_j-m_jD_i)=0$, it follows that $m_iD_j-m_jD_i$ is supported on the reducible fibre of $\pi$. Thus, it remains only to show that it is effective. It suffices to prove the result in the case that $E_i.E_j=1$; a simple induction will prove the general case from there. Furthermore, if we blow down a $(-1)$-curve $E_\ell$ on $X$ to obtain $Y$, then by Lemma \[inductivestep\], the truth of the lemma for $E_i$ and $E_j$ on $X$ will be equivalent to the truth of the lemma for the images of $E_i$ and $E_j$ on $Y$, unless $i$ or $j$ equals $\ell$. Thus, by induction on the number $m$ of components of the reducible fibre, we may assume that $E_i$ or $E_j$ is a $(-1)$-curve on $X$, that any $(-1)$-curve in the reducible fibre is either $E_i$ or $E_j$, and that $i\neq 1$ (because $E_i\preceq E_j$ and $i\neq j$).
Say $E_i$ is a $(-1)$-curve of multiplicity one. Then it must be a leaf of $G$, so $E_j$ must be the unique other component which intersects $E_i$. In that case, $E_j$ must also have multiplicity one, so let $Y$ be the surface obtained by blowing down $E_i$. By applying Lemma \[inductivestep\], we see that $D_j-D_i =E_i$, and the lemma follows.
Now assume that $E_i$ is a $(-1)$-curve of multiplicity $m_i>1$. Then by hypothesis $E_i$ cannot be a leaf, so it must intersect two curves $E_j$ and $E_\ell$, with $m_i=m_j+m_\ell$. We have $E_\ell\succeq
E_i\succeq E_j$, so let $f\colon X\rightarrow Y$ be the map that blows down $E_i$. Then $Y$ is also a rational surface with $(-n)$-section and a single reducible fibre with $m-1$ components, so $\{S',E_1',\ldots,E_{m-1}'\}$ are generators of the effective cone of $Y$, with $S'=f_*(S)$ and $E_r'=f_*(E_r)$ for any $r$. Let $\{D_0',\ldots,D_{m-1}'\}$ be the dual basis of $\{S',E'_1,\ldots,E'_{m-1}\}$ in $\mbox{NS}(Y)$. By Lemma \[inductivestep\], we have $D_i=D_j+D_\ell-E_i$. By induction on $m$, we have: $$m_\ell D_j' - m_j D_\ell'=\sum_{E_t\succeq E_\ell} m_tE_t'$$ since the multiplicity of $E_r$ is the same as the multiplicity of $E_r'$, for any $r\neq i$. By Lemma \[inductivestep\] again, pulling back to $X$ via $f$ gives: $$m_\ell D_j - m_j D_\ell =(\sum_{E_t\succeq E_\ell} m_tE_t)+m_\ell E_i$$ and therefore: $$\begin{aligned}
m_iD_j-m_jD_i & = & m_iD_j-m_jD_j-m_jD_\ell +m_jE_i \\
& = & m_\ell D_j-m_jD_\ell + m_jE_i \\
& = & (\sum_{E_t\succeq E_\ell} m_tE_t) + m_iE_i \\
& = & \sum_{E_t\succeq E_i} m_tE_t\end{aligned}$$ as desired.
Finally, we treat the case that $E_j$ is the only $(-1)$-curve in the reducible fibre. It is not a leaf, since $E_i\succeq E_j$. If $j=1$, then since $E_1=E_j$ has multiplicity one, then since $m>1$ ($i\neq
j$), there must be at least one other $(-1)$-curve $E_t$ on $X$. Therefore $j\neq 1$, and let $E_i$ and $E_\ell$ be the two components which intersect $E_j$. (Note that $j\neq 1$ because $E_i\preceq
E_j$.) We have $E_i\succeq E_j\succeq E_\ell$, $D_j=D_i+D_\ell -E_j$, and $m_j=m_i+m_\ell$. By induction and Lemma \[inductivestep\], we compute: $$m_iD_\ell - m_\ell D_i = (\sum_{E_t\succeq E_i} m_tE_t)+m_iE_j$$ and therefore: $$\begin{aligned}
m_iD_j-m_jD_i & = & m_iD_i+m_iD_\ell-m_iE_j-m_jD_i \\
& = & (m_iD_\ell - m_\ell D_i) - m_iE_j \\
& = & \sum_{E_t\succeq E_i} m_tE_t\end{aligned}$$ and the lemma is proven.
\[basepointfree\] For any $i$, some positive multiple of $D_i$ is basepoint free.
[*Proof of lemma:*]{} We have already proven this in the case that $i=0$ (proof of Lemma \[effectcone\]) or $E_i$ is a leaf (see Lemma \[leafbasepointfree\]). By Lemma \[multiplegens\], we know that some multiple of each $D_i$ is effective, and in particular, the base locus of $D_i$ is supported on the set $\cup_{E_t\succ E_i} E_t$. It therefore suffices to find an effective divisor $E$, linearly equivalent to some multiple of $D_i$, supported on the set $S\cup(\bigcup_{E_t\not\succeq E_i} E_t)$. We claim that for all $i$, there is an effective divisor $C_i$ and a positive rational number $\alpha_i$ such that $D_i=(m_iD_1-\alpha_iF)+C_i$, $C_i$ is supported on the set $\cup_{E_t\not\succeq E_i} E_t$, and the base locus of a suitable multiple of $m_iD_1-\alpha_iF$ is a subset of $S$. In particular, we will show that $\alpha_i/m_i\leq\ell_i$, where $\ell_i$ denotes the number of edges in the shortest path in the graph $G$ from $E_1$ to $E_i$.
If $i=1$, this is trivial. We proceed by induction on $\ell_i$, the number of edges in the shortest path from $E_1$ to $E_i$ in the rooted tree. Let $E_j$ be the parent of $E_i$ — that is, let $E_j$ be the unique component with $E_j\prec E_i$ and $E_j.E_i=1$. Then by induction we can write $D_j=(m_jD_1-\alpha_jF)+C_j$, where $C_j$ is effective and supported on the set $\cup_{E_t\not\succeq E_j} E_t$, and the base locus of a suitable multiple of $m_jD_1-\alpha_jF$ is a subset of $S$. By Lemma \[multiplegens\], we can write: $$m_iD_j-m_jD_i=\sum_{E_t\succeq E_i} m_tE_t$$ We therefore deduce: $$\begin{aligned}
D_i & = & (1/m_j)(m_iD_j-\sum_{E_t\succeq E_i}m_tE_t) \\
& = & (1/m_j)(m_im_jD_1-m_i\alpha_jF+m_iC_j-\sum_{E_t\succeq E_i}m_tE_t) \\
& = & m_iD_1-(m_i\alpha_j/m_j+1)F+m_iC_j+\sum_{E_t\not\succeq E_i}m_tE_t\end{aligned}$$ Set $\alpha_i=m_i\alpha_j/m_j+1$. Since $\alpha_j/m_j\leq\ell_j$, it follows that $\alpha_i/m_i=\alpha_j/m_j+1/m_i \leq\ell_i=\ell_j+1$. We have $\ell_i\leq m<n$ for all $i$, so a suitable multiple of $D_1-\alpha_i/m_iF=S+(n-\alpha_i/m_i)F$ will be basepoint free away from $S$, as desired. This means that for each $i$, the base locus of a suitable multiple of $D_i$ is contained in the union of $S$ and $\cup_{E_t\not\succeq E_i}E_t$. But we have already proven that the base locus of a suitable multiple of $D_i$ is contained in the union $\cup_{E_t\succ E_i} E_t$. Since these two sets are disjoint, we conclude that a large enough multiple of $D_i$ is basepoint free, as desired.
We now conclude the proof of Theorem \[onefibre\]. Choose any $P\in
X(k)$.
\[vertexproof\] For any $i$, a sequence of best $D_i$-approximation to $P$ can be found on the fibre of $\pi$ through $P$ (this includes possibly along some $E_j$), or else on $S$.
[*Solution:*]{} If $i=0$, then $D_i=F$, so the claim is trivial. If $i=1$, then $D_i=S+nF$, and against the claim is clearly true. For general $i$, if $P$ lies on a curve $C$ with $C.D_i=0$, then clearly $C$ will be a curve of best $D_i$-approximation to $P$. However, such a curve $C$ must be linearly equivalent to a nonnegative linear combination of $\{E_j\mid j\neq i\}$, where as usual $E_0=S$. If $C$ is not $S$ or some $E_i$, then $C$ must be nef, and so must be a positive linear combination of the $D_i$. Since $m<n$, Remark \[strictremark\] shows that this is impossible. Thus, if $C.D_i=0$, then $C$ is either $S$ or some $E_i$ for $i\geq 1$.
Assume that $E_i$ is a leaf with $i>1$. Then $D_i$ is basepoint free, and $D_i.F=1$, so if $\phi_i\colon X\to\P^n$ is the morphism associated to $D_i$, then $\phi_i$ maps the fibre of $\pi$ through $P$ to a line. By Theorem \[projlines\], this means that a sequence of best $D_i$-approximation to $P$ either lies along the fibre of $\pi$ through $P$, or else along some irreducible curve $C$ with $C.D_i=0$, which as noted must either be $S$ or some $E_i$.
If $E_i$ is not a leaf, then there is some $j$ for which $E_j$ is a leaf and $E_j\succeq E_i$. By Lemma \[multiplegens\], we can write $D_i=m_i D_j+\sum_{E_t\succeq E_j} m_tE_t$. If $P$ does not lie on any $E_t$ with $E_t\succeq E_j$, then Theorem \[addeffective\] implies that either the fibre of $\pi$ through $P$ or $S$ is a curve of best $D_i$-approximation to $P$. If $P$ lies on some $E_t$ for $t\neq i$, then $E_t$ is a curve of best $D_i$-approximation to $P$ because $E_t.D_i=0$ and some multiple of $D_i$ is basepoint free. If $P$ lies on $E_i$ but not on any other $E_t$, then as in the proof of Lemma \[basepointfree\], we can write $D_i=\alpha F+E$, where $\alpha$ is a positive rational number and $E$ is an effective divisor whose support does not include $E_i$. Since $E_i$ is a curve of best $F$-approximation to $P$, Lemma \[multiplegens\] implies that $E_i$ is also a curve of best $D_i$-approximation to $P$. In all cases, the claim is proven.
The claim, together with Corollary \[divsum\], shows that if $D$ is any ample divisor on $X$, then $D$ is a nonnegative linear combination of the $D_i$, and so a sequence of best $D$-approximation to $P$ can be found on the fibre of $\pi$ which contains $P$, unless $P$ lies on $S$. If $P$ lies on $S$, then $S$ is a curve of best $D_i$-approximation to $P$ for all $i$, and so $S$ is also a curve of best $D$-approximation to $P$. This concludes the proof of Theorem \[onefibre\].
These same techniques will also prove Conjecture \[ratcurve\] for split rational surfaces of Picard rank four.
\[four\] Conjecture \[ratcurve\] is true for any split rational surface $X$ of Picard rank four.
[*Proof:*]{} Every such rational surface is a blowup of a Hirzebruch surface, and so $X$ admits a surjective morphism $\pi\colon
X\rightarrow\P^1$ whose generic fibre is irreducible. The Néron-Severi group has rank four, generated by the class $F$ of a fibre of this morphism, the class $S$ of a section, and two more classes, corresponding to irreducible components of reducible fibres, or equivalently, exceptional divisors of the blowing down map to some Hirzebruch surface.
There are three possible configurations of reducible fibres (see Table \[fibretable\]):
---------- ---------- ----------
Case (1) Case (2) Case (3)
---------- ---------- ----------
: Configurations of reducible fibres[]{data-label="fibretable"}
1. Two reducible fibres, each with two components. These components intersect each other transversely in a single point.
2. One reducible fibre, with three components, configured like a letter F, where the vertical component is the one which intersects the section $S$.
3. One reducible fibre, with three components, configured like a letter H, where the leftmost vertical component is the one which intersects the section $S$.
In the proof, the first case will generate a further subcase, corresponding to $\P^2$ blown up at three points in general position, and the third case will generate a further subcase corresponding to a double component in the reducible fibre. However, in all cases, the combinatorial description of the reducible fibres, combined with the self-intersection of the section $S$, will completely determine the intersection product on $X$. Since our proof relies almost completely on the intersection product, this will suffice to prove Theorem \[four\].
Case (1): The fibration $\pi\colon X\rightarrow\P^1$ admits two reducible fibres, each having two components. (See Table \[fibretable\].)
In addition to the classes $S$ and $F$ described above, let $E_1$ and $E_2$ be the components of one irreducible fibre, and let $F_1$ and $F_2$ be the components of the other. Let $E_2$ and $F_2$ be the components which do not intersect $S$. Since the Picard rank is four, the four classes $S$, $F$, $E_2$, and $F_2$ are a basis for the vector space $NS(X)\otimes\R$. The intersection matrix for $X$ is: $$\begin{array}{r|rrrr}
& F & E_2 & F_2 & S \\ \hline
F & 0 & 0 & 0 & 1 \\
E_2 & 0 & -1 & 0 & 0 \\
F_2 & 0 & 0 & -1 & 0 \\
S & 1 & 0 & 0 & -n
\end{array}$$ where $n=-S^2$ is a positive integer.
The case $n>2$ is covered by Theorem \[simplefibres\]. We first treat the case $n=2$.
The effective cone of $X$ is generated by the classes of $E_i$, $F_i$, and $S$, and the nef cone of $X$ is generated by the classes of $F$, $D_2=2F+S$, $D_1=2F-E_2+S$, $D'_1=2F-F_2+S$, and $D_0=2F-E_2-F_2+S$.
[*Proof:*]{} A straightforward calculation shows that the two cones in the claim are dual to one another, and the first cone is clearly contained in the effective cone, so it suffices to show that the interior of the cone generated by $F$, $D_i$ and $D'_1$ is contained in the ample cone. We will use the Nakai-Moishezon Criterion again, and note that the intersection properties of the given five divisors are as follows: $$\begin{array}{r|rrrrr}
& F & D_2 & D_1 & D'_1 & D_0 \\ \hline
F & 0 & 1 & 1 & 1 & 1 \\
D_2 & 1 & 2 & 2 & 2 & 2 \\
D_1 & 1 & 2 & 1 & 2 & 1 \\
D'_1 & 1 & 2 & 2 & 1 & 1 \\
D_0 & 1 & 2 & 1 & 1 & 0
\end{array}$$ Each of these five divisors are basepoint free: $F$ and $D_0$ correspond to morphisms to $\P^1$, $D_1$ and $D'_1$ correspond to birational maps to $\P^2$, and $D_2=D_1+E_2=D_1'+F_2$. It is therefore clear that any positive linear combination of these five divisors has positive self-intersection, and intersects any irreducible curve $C$ positively, so by the Nakai-Moishezon Criterion, must be ample. The claim is therefore proven.
We will now prove Case (1) of Theorem \[four\] for the case $n=2$. Let $P\in X(k)$ be any rational point — we first assume that $P$ does not lie on $S$, $E_i$, or $F_i$. Let $\psi\colon
X\rightarrow\P^1$ be the morphism corresponding to $D_0$. A sequence of best $D_0$-approximation to $P$ clearly lies along the fibre of $\psi$ through $P$, while a sequence of best $D$-approximation to $P$ for any positive linear combination $D$ of the other four generators lies along the component of $\pi$ through $P$. We therefore define two subcones of the nef cone of $X$: subcone $A$, generated by $F$, $D_2$, $D_1$, $D'_1$, and $D_0+F$, and subcone $B$, generated by $D_0$, $D_0+F$, $D_1$, and $D'_1$. For each generator of subcone $A$, a sequence of best approximation to $P$ lies along the component of $F$ of minimal degree through $P$. If $P$ does not lie on a reducible fibre or $S$, then we can conclude from Corollary \[divsum\] that if $D$ lies in the cone $A$, then the fibre of $\pi$ through $P$ is a curve of best $D$-approximation to $P$. For each generator of subcone $B$, a similar argument shows that a sequence of best approximation to $P$ lies along the component of the fibre of $\psi$ through $P$. Since the union of $A$ and $B$ is the entire nef cone of $X$, we conclude that Conjecture \[ratcurve\]is true for $P$.
If $P$ does lie on $S$, $E_i$, or $F_i$, then we divide the nef cone into five subcones, one for each of the five curves $S$, $E_i$, and $F_i$. The subcone $A_C$ corresponding to a curve $C$ is the set of nef divisors $D$ for which $C$ has minimal $D$-degree amongst $S$, $E_i$, and $F_i$. A straightforward calculation shows that generators for these subcones can always be found from the set $\{F,D_0,D_1,D'_1,D_2, F+D_0,F+D_1,F+D'_1,F+D_2+D_0\}$, and that each divisor in that set intersects each of $S$, $E_i$, and $F_i$ in either 0 or 1. It then follows from Corollary \[divsum\] that for each curve $C\in\{S,E_1,E_2,F_1,F_2\}$, the curve $C$ is a curve of best $D$-approximation to $P$ for every divisor $D\in A_C$. Since the nef cone is the union of the subcones $A_C$, we conclude that Conjecture \[ratcurve\] is true in case (1) with $n=2$.
We now move to the case $n=1$. In this case, $X$ is the blowup of the first Hirzebruch surface $H_1$ at two points, possibly ([*a priori*]{}) infinitely near. Thus, $X$ is also the blowup of $\P^2$ at three points (again, possibly infinitely near). If any of the three points are infinitely near, then there will be a curve $S$ of self-intersection $-2$ or less, and a morphism from $X$ to $\P^1$ of which $S$ is a section, so $X$ is covered by the already-treated case $n\geq 2$. Thus, the three points must be distinct. If they are collinear, then no morphism from $X$ to $\P^1$ will have more than one reducible section, so we cannot be in Case (1). Thus, $X$ must be isomorphic to the blowup of $\P^2$ at three points in general position.
Let $\phi\colon X\rightarrow\P^2$ be the blowing down map. Let $L$ be the divisor class corresponding to the invertible sheaf $\phi^*\O(1)$, and let $E_i$ for $i=1,2,3$ be the three exceptional divisors of $\phi$. These four divisors are a basis of the vector space $NS(X)\otimes\R$, and their intersection matrix is as follows: $$\begin{array}{r|rrrr}
& L & E_1 & E_2 & E_3 \\ \hline
L & 1 & 0 & 0 & 0 \\
E_1 & 0 & -1 & 0 & 0 \\
E_2 & 0 & 0 & -1 & 0 \\
E_3 & 0 & 0 & 0 & -1
\end{array}$$
The effective cone of $X$ is generated by the six curves $E_i$ and $L-E_i-E_j$ for $i\neq j$. The nef cone of $X$ is generated by the five curves $L$, $L-E_i$, and $F=2L-E_1-E_2-E_3$.
[*Proof:*]{} This can be found in, for example, [@Ma].
The divisors $L$ and $F$ correspond to birational morphisms to $\P^2$, and $L-E_i$ corresponds to a morphism to $\P^1$. Thus, each of the divisors is basepoint free, and their intersection matrix is: $$\begin{array}{r|ccccc}
& L & L-E_1 & L-E_2 & L-E_3 & F \\ \hline
L & 1 & 1 & 1 & 1 & 2 \\
L-E_1 & 1 & 0 & 1 & 1 & 1 \\
L-E_2 & 1 & 1 & 0 & 1 & 1 \\
L-E_3 & 1 & 1 & 1 & 0 & 1 \\
F & 2 & 1 & 1 & 1 & 1
\end{array}$$
We now conclude the proof of Conjecture \[ratcurve\] in Case (1). Let $P\in X(k)$ be any $k$-rational point, and let $D$ be any ample divisor on $X$. Let $A_1$ be the cone generated by the following divisor classes: $$\label{conegens}
L,L-E_1,(L-E_1)+(L-E_2),(L-E_1)+(L-E_3),F$$ By permuting the indices on the $E_i$ in the above list, we obtain two more cones $A_2$ and $A_3$ such that the nef cone of $X$ is the union of the subcones $A_i$. Assume that $D$ lies in $A_1$’ by symmetry, it suffices to consider this case.
If $P$ does not lie on $E_i$ or $L-E_i-E_j$ for $i\neq j$, then Theorem \[projlines\] and Corollary \[divsum\] imply that the component of $L-E_m$ through $P$ is a curve of best $D$-approximation to $P$. (To see this, note that each of the generators of $A_1$ is basepoint free, contracts no curves through $P$, and intersects $L-E_1$ either one or zero times.) Thus, if $P$ does not lie on any $E_i$ or $L-E_i-E_j$, then we have proven Conjecture \[ratcurve\] for $P$.
Now say that $P$ does lie on some $E_i$ or $L-E_i-E_j$. The incidence graph of these six divisors is a hexagon: $E_i$ intersects precisely $L-E_i-E_k$ for $k\neq i$, and $L-E_i-E_j$ intersects precisely $E_i$ and $E_j$. Thus, $P$ either lies on exactly one or exactly two of these six curves.
If $P$ lies only on $E_1$, then consider the subcone $B_1$ of $A_1$, generated by: $$L,L+(L-E_1),(L-E_1)+(L-E_2),(L-E_1)+(L-E_3),F$$ The generator $L$ contracts $E_1$, and so $E_1$ is clearly a curve of best $L$-approximation to $P$. The remaining generators are all basepoint free, intersect $E_1$ once, and contract no curves through $P$, and so by Theorem \[projlines\], $E_1$ is a curve of best approximation to $P$ with respect to all the generators of $B_1$. Corollary \[divsum\] now implies that for any $D$ in $B_1$, $E_1$ is a curve of best approximation to $P$.
If $D$ does not lie in $B_1$, then $D$ must lie in the cone $B'_1$, generated by: $$L+(L-E_1),L-E_1,(L-E_1)+(L-E_2),(L-E_1)+(L-E_3),F$$ The generator $L-E_1$ contracts the component $C$ of $L-E_1$ through $P$, and so $C$ is a curve of best $(L-E_1)$-approximation to $P$. Each of the other generators is basepoint free, does not contract any curve through $P$, and intersects $C$ once, so by Theorem \[projlines\], $C$ is a curve of best approximation to $P$ with respect to all the generators of $B'_1$. Corollary \[divsum\] now implies that $C$ is a curve of best $D$-approximation to $P$. In summary, if $P$ lies on $E_1$ and no other generator of the effective cone of $X$, then we have proven Conjecture \[ratcurve\] for $P$.
If $P$ lies only on $E_2$, $E_3$, $L-E_1-E_2$, or $L-E_1-E_3$, then by a similar argument to the previous paragraph, Theorem \[projlines\] and Corollary \[divsum\] imply that for any $D$ in $A_1$, one of the curves $E_2$, $E_3$, $L-E_1-E_2$, or $L-E_1-E_3$ will be a curve of best $D$-approximation to $P$, and so Conjecture \[ratcurve\] is proven for $P$.
If $P$ lies only on $L-E_2-E_3$, then define the cone $B_{23}$ to be that generated by: $$F,L,(L-E_1)+(L-E_2),(L-E_1)+(L-E_3),F+(L-E_1)$$ Since $F$ contracts $L-E_2-E_3$, it’s clear that $L-E_2-E_3$ is a curve of best $F$-approximation to $P$. All the other generators are basepoint free, do not contract any curves through $P$, and intersect $L-E_2-E_3$ once. Thus, by Theorem \[projlines\] and Corollary \[divsum\], we conclude that if $D$ lies in $B_{23}$, $L-E_2-E_3$ is a curve of best $D$-approximation to $P$. If $D$ does not lie in $B_{23}$, then $D$ must lie in $B'_{23}$, generated by: $$L-E_1,L,(L-E_1)+(L-E_2),(L-E_1)+(L-E_3),F+(L-E_1)$$ A similar argument to the previous paragraph shows that the component of $L-E_1$ through $P$ is a curve of best $D$-approximation to $P$. We conclude that if $P$ lies only on one of the six generators of the effective cone of $X$, then Conjecture \[ratcurve\] is true for $P$.
It only remains to consider the possibility that $P$ is the intersection of two of the curves $E_i$ or $L-E_i-E_j$. First, assume that $P=E_1\cap
L-E_1-E_i$ for $i=2$ or $i=3$. Let $B_{11i}$ be the cone generated by: $$L+(L-E_1),L+F,L+(L-E_1)+(L-E_i),(L-E_1)+(L-E_{5-i}),$$ $$L-E_1,(L-E_1)+(L-E_i),F$$ The last three generators of $B_{11i}$ all contract $L-E_1-E_i$, and so $L-E_1-E_i$ is a curve of best approximation to $P$ for each of these generators. The first four generators are all basepoint free, contract no curves through $P$, and map $L-E_1-E_i$ to a line, and so by Theorem \[projlines\] and Corollary \[divsum\] we conclude that $L-E_1-E_2$ is a curve of best $D$-approximation to $P$ if $D$ lies in $B_{11i}$.
If $D$ does not lie in $B_{11i}$, then $D$ must lie in the cone $B'_{11i}$, generated by: $$L,L+(L-E_1),L+(L-E_1)+(L-E_i),(L-E_1)+(L-E_{5-i}),L+F$$ (Recall that $D$ is assumed to lie in the cone $A_1$, defined above.) The divisor $L$ contracts $E_1$, and the other four generators are basepoint free, contract no curves through $P$, and map $E_1$ to a line. We conclude from Theorem \[projlines\] and Corollary \[divsum\] that $E_1$ is a curve of best $D$-approximation to $P$.
Next, assume that $P$ lies on the intersection of $L-E_1-E_i$ and $E_i$ for $i=2$ or $i=3$. Define the cone $B_{1ii}$ to be the cone generated by: $$L,L-E_1,(L-E_1)+(L-E_{5-i}),L+(L-E_1)+(L-E_i),L+F,$$ $$F+(L-E_1)+(L-E_{5-i})$$ All of these classes are basepoint free. The first three contract $E_i$, and the last three contract no curves through $P$ and map $E_i$ to a line, and so we conclude by Theorem \[projlines\] and Corollary \[divsum\] that $E_i$ is a curve of best $D$-approximation to $P$ for any $D$ in $B_{1ii}$. Thus, Conjecture \[ratcurve\] is true for $P$ if $D$ lies in $B_{1ii}$.
If $D$ does not lie in $B_{1ii}$, then $D$ must lie in the cone $B'_{1ii}$, generated by: $$F,L-E_1,(L-E_1)+(L-E_i),L+F,L+(L-E_1)+(L-E_i),$$ $$F+(L-E_1)+(L-E_{5-i})$$ All generators are basepoint free. The first three generators contract $L-E_1-E_i$, and the last three contract no curves through $P$ but map $L-E_1-E_i$ to a line. We conclude by Theorem \[projlines\] and Corollary \[divsum\] that $L-E_1-E_i$ is a curve of best $D$-approximation to $P$ for all $D$ in $B'_{1ii}$. We conclude that Conjecture \[ratcurve\] is true for $P=E_i\cap(L-E_1-E_i)$ and all $D$.
Finally, let $P$ be the intersection of $E_i$ and $L-E_i-E_{5-i}$ for $i=2$ or $i=3$. Define $B_{123}$ to be the cone generated by: $$L,L-E_1,(L-E_1)+(L-E_{5-i}),(L-E_1)+(L-E_i),F+L,F+(L-E_1),$$ $$F+(L-E_1)+(L-E_{5-i})$$ All of these classes are basepoint free. The first three contract $E_i$, and the last four do not contract any curve through $P$, and map $E_i$ to a line. Thus, by Theorem \[projlines\] and Corollary \[divsum\], we conclude that $E_i$ is a curve of best $D$-approximation to $P$, provided $D$ lies in $B_{123}$.
If $D$ does not lie in $B_{123}$, then it must lie in the cone $B'_{123}$, generated by: $$F,F+L,(L-E_1)+(L-E_i),F+(L-E_1),F+(L-E_1)+(L-E_{5-i})$$ All five of these are basepoint free. The last four contract no curves through $P$, and map $L-E_2-E_3$ to a line, and $F$ contracts $L-E_2-E_3$ to a point. By Theorem \[projlines\] and Corollary \[divsum\], we conclude that $L-E_i-E_j$ is a curve of best $D$-approximation to $P$ if $D$ lies in $B'_{123}$. Conjecture \[ratcurve\] is thus proven for $X$.
Case (2): One reducible fibre, with three components, configured like a letter F, where the vertical component is the one which intersects the section $S$. (See Table \[fibretable\].)
Let $F_1$ be the component of the reducible fibre which intersects $S$, and let $E_1$ and $E_2$ be the other two components. Let $F$ be the class of a fibre. These four classes are a basis of $\mbox{NS}(X)$, with intersection matrix: $$\begin{array}{r|rrrr}
& S & E_1 & E_2 & F \\ \hline
S & -n & 0 & 0 & 1 \\
E_1 & 0 & -1 & 0 & 0 \\
E_2 & 0 & 0 & -1 & 0 \\
F & 1 & 0 & 0 & 0
\end{array}$$
The curves $S$, $F_1$, $E_1$, and $E_2$ generate the effective cone of $X$. The nef cone of $X$ is generated by the divisors $F$, $D_1=nF+S$, $D_2=nF+S-E_1$, and $D_3=nF+S-E_2$.
[*Proof:*]{} A straightforward calculation shows that the two cones described in the claim are dual to one another. Thus, it suffices to show that $F$, $D_1$, $D_2$, and $D_3$ generate the nef cone of $X$. To do this, it further suffices to show that every positive linear combination of $F$, $D_1$, $D_2$, and $D_3$ is ample, since the converse inclusion is clear from the effectivity of $S$, $F_1$, $E_1$, and $E_2$.
The intersections of $F$ and the $D_i$ are tabulated as follows: $$\begin{array}{r|cccc}
& F & D_1 & D_2 & D_3 \\ \hline
F & 0 & 1 & 1 & 1 \\
D_1 & 1 & n & n & n \\
D_2 & 1 & n & n-1 & n \\
D_3 & 1 & n & n & n-1
\end{array}$$
Since $n\geq 1$, it is clear that any positive linear combination of $F$ and the $D_i$ must have positive self-intersection. It’s clear that $F$ is basepoint free, since it corresponds to a morphism to $\P^1$, and similarly each $D_i$ corresponds to a morphism to a cone over a rational normal curve (see Remark \[actualcones\]). In particular, each of these four divisors is nef, so by the Nakai-Moishezon criterion for ampleness, we conclude that they must generate the nef cone of $X$, as desired.
\[actualcones\] Let $Y$ be the cone over a smooth rational normal curve of degree $n$. It is easy to see that if we blow up $Y$ at the vertex, we obtain the Hirzebruch surface $H_n$. Moreover, the exceptional curve of the blowup $\pi\colon H_n\to Y$ is precisely the unique $(-n)$-section $S'$ of $H_n$, and if $L$ is the class of a hyperplane section of $Y$, then $\pi^*L=S'+nF'$, where $F'$ is a fibre of the morphism $H_n\to\P^1$.
Our rational surface $X$ admits several birational maps to Hirzebruch surfaces. It is a straightforward calculation to see that each $D_i$ is the pullback of $\pi^*L$ via one of these maps. For example, we might blow down $E_1$ and then $F_1$ to obtain $H_n$, in which case pulling back $\pi^*L$ to $X$ gives $D_3$. This technique will be used frequently in what follows.
We now prove Conjecture \[ratcurve\] for Case (2). We first assume that $n>1$. Let $P\in X(k)$ be any rational point not lying on $S$, $F_1$, or either $E_i$. Let $C$ be the (irreducible) component of $F$ through $P$. Then $F$ contracts $C$, and all of the other generators of the nef cone map $C$ to a line, and so by Theorem \[projlines\] and Corollary \[divsum\], we conclude that $C$ is a curve of best $D$-approximation to $P$ for any ample divisor $D$.
If $P$ lies on $S$ but not $F_1$, then we divide the nef cone of $X$ into two subcones: a cone $A$ generated by $D_1$, $D_2$, $D_3$, $D_1+F$, $D_2+F$, and $D_3+F$, and a cone $B$ generated by $D_1+F$, $D_2+F$, $D_3+F$, and $F$. The classes $D_i$ each contract $S$, and $F+D_i$ is basepoint free, contracts no curves through $P$ but maps $S$ to a line, and so for each of these six divisor classes, $S$ is a curve of best approximation to $P$. Thus, by Corollary \[divsum\], Conjecture \[ratcurve\] is true for $P$ with respect to any ample divisor in $A$. On the other hand, $F+D_i$ also maps the component $C$ of $F$ through $P$ to a line, and $F$ contracts $C$, so similar reasoning shows that Conjecture \[ratcurve\] is true for $P$ with respect to any ample divisor in $B$. We conclude that Conjecture \[ratcurve\] is true for $P$.
If $P$ lies on $E_i$ but not $F_1$, or if $P$ lies on $F_1$ but not $S$ or either $E_i$, then let $C$ be the unique irreducible component of $F$ through $P$ (that is, either $F_1$ or $E_i$). Every generator of the nef cone either contracts $C$ or else maps $C$ to a line and contracts no curves through $P$, and so Theorem \[projlines\] and Corollary \[divsum\] imply as usual that $C$ is a curve of best $D$-approximation to $P$ for any ample divisor $D$.
If $P=S\cap F_1$, then consider the cone $A'$ generated by: $$F,F+D_1,D_2,D_3$$ Note that all four of these classes are basepoint free. All but $F+D_1$ contract $F_1$, and $F+D_1$ maps $F_1$ to a line and contracts no curve through $P$. We conclude from Theorem \[projlines\] and Corollary \[divsum\] that $F_1$ is a curve of best $D$-approximation to $P$ for any $D$ in $A'$. If $D$ is not in $A'$, then $D$ must lie in $B'$, generated by: $$D_1,F+D_1,D_2,D_3$$ Every generator is basepoint free, and all but $F+D_1$ contracts $S$ The class $F+D_1$ maps $S$ to a line, and contracts no curves through $P$. We conclude by Theorem \[projlines\] and Corollary \[divsum\] that $S$ is a curve of best $D$-approximation to $P$ for every $D$ in $B'$. Thus, Conjecture \[ratcurve\] is true for $P$.
Finally, if $P=F_1\cap E_i$ for some $i$, then consider the cone $\tilde{A}$ generated by $F$, $D_1+D_{i+1}$, $D_2$, and $D_3$. Each generator is basepoint free, and all but $D_1+D_{i+1}$ contract $F_1$. Furthermore, the class $D_1+D_{i+1}$ maps $F_1$ to a line and contracts no curves through $P$. Theorem \[projlines\] and Corollary \[divsum\] now imply that $F_1$ is a curve of best $D$-approximation to $P$ for any $D$ in $\tilde{A}$.
If $D$ does not lie in $\tilde{A}$, then it must lie in $\tilde{B}$, generated by $F$, $D_1$, $D_1+D_{i+1}$, and $D_{4-i}$. Each generator is basepoint free, and all but $D_1+D_{i+1}$ contract $E_i$. Furthermore, the class $D_1+D_{i+1}$ maps $E_i$ to a line and contracts no curves through $P$. Theorem \[projlines\] and Corollary \[divsum\] now imply that $E_i$ is a curve of best $D$-approximation to $P$ for any $D$ in $\tilde{B}$. We conclude that Conjecture \[ratcurve\] is true for $X$ in the case that $n>1$.
(Note that if $n>3$, this result also follows immediately from Theorem \[onefibre\].)
If $n=1$, then $D_2^2=D_3^2=0$, so these both correspond to morphisms to $\P^1$. In fact, if we cannot describe $X$ with $n>1$, then $X$ must be a blowup of $\P^1\times\P^1$ at two (possibly infinitely near) points. In fact, the points cannot be infinitely near, since otherwise the exceptional divisor of the first blowup would have self-intersection $-2$, and we could choose it to be the section $S$, with $n=2>1$. And if the configuration of exceptional divisors is to be as in Case (2), the two blown up points must lie on the same fibre of at least one of the canonical projections $\P^1\times\P^1\rightarrow\P^1$.
In other words, we have shown that $X$ must be isomorphic to $\P^2$ blown up at three different points on a straight line. The exceptional curves of this blowup are $S$, $E_1$, and $E_2$, while $F_1$ is the strict transform of the line joining the blown up points. The divisors $D_2$ and $D_3$ correspond to the maps $\pi_i\colon
X\rightarrow\P^1$ ($i=2,3$) which factor through the blowup of $\P^2$ at a single point, while $F$ corresponds to the third map $\pi_1\colon
X\rightarrow\P^1$, induced in the same way. The divisor $D_1$ corresponds to the map $\phi\colon X\rightarrow\P^2$ that blows down $S$, $E_1$, and $E_2$.
Let $P\in X(k)$ be a rational point which does not lie on $S$, $F_1$, or either $E_i$, and let $D$ be an ample divisor on $X$. By symmetry, we may assume that $D.F\leq D.D_i$ for $i=2,3$, since $F$, $D_2$, and $D_3$ are all conjugate under the automorphism group of $X$. In that case, $D$ lies in the cone $A$ generated by: $$F,D_1,F+D_2,F+D_3,F+D_2+D_3$$ All these divisors are basepoint free. The last four contract no curves through $P$ and map the component $C$ of $F$ through $P$ to a line, and $F$ contracts $C$. Thus, by Theorem \[projlines\] and Corollary \[divsum\], we conclude that $C$ is a curve of best $D$-approximation to $P$, and that Conjecture \[ratcurve\] is true for $P$.
If $P$ lies on one or more of $S$, $E_1$, $E_2$, or $F_1$, then we must subdivide $A$ further. If $P$ lies on $S$ but not $F_1$, then assume that $D$ lies in the cone $A_S$ generated by: $$D_1,F+D_1,F+D_2,F+D_3,F+D_2+D_3$$ Each of these divisors is basepoint free. The last four contract no curves through $P$, and map $S$ to a line, and $D_1$ contracts $S$, so by Theorem \[projlines\] and Corollary \[divsum\], we conclude that $S$ is a curve of best $D$-approximation to $P$. If $D$ does not lie in $A_S$, then $D$ must lie in $A'_S$, generated by: $$F,F+D_1,F+D_2,F+D_3$$ All of these divisors are basepoint free. The last four contract no curves through $P$ and map the component $C$ of $F$ through $P$ to a line, and $D_1$ contracts $C$. We conclude from Theorem \[projlines\] and Corollary \[divsum\] that $C$ is a curve of best $D$-approximation to $P$. In all cases, if $P$ lies on $S$ but not $F_1$, then Conjecture \[ratcurve\] is true for $P$.
If $P$ lies on exactly one of $E_1$, $E_2$, or $F_1$, then a quick check (using Theorem \[projlines\] and Corollary \[divsum\]) shows that $E_i$ is a curve of best $D$-approximation for every $D$ in the cone $A$. The same is true if $P=S\cap F_1$.
The only remaining case is thus $P=F_1\cap E_i$. Consider the cone $A_i$ generated by: $$F,D_1,F+D_{4-i},F+D_1+D_{i+1},F+D_1+D_2+D_3$$ All these divisors are basepoint free. The first three generators contract $E_i$, while the last two contract no curves through $P$ but map $E_i$ to a line. By Theorem \[projlines\] and Corollary \[divsum\], we conclude that $E_i$ is a curve of best $D$-approximation to $P$ for any $D$ in $A_i$. If $D$ does not lie in $A_i$, then $D$ must lie in the cone $A'_i$, generated by: $$F,F+D_2,F+D_3,F+D_2+D_3,F+D_1+D_2,F+D_1+D_2+D_3$$ All of these divisors are basepoint free. The first four contract $F_1$, and the last two contract no curves through $P$ but map $F_1$ to a line, so we conclude by Theorem \[projlines\] and Corollary \[divsum\] that $F_1$ is a curve of best $D$-approximation to $P$.
The sum of all these calculations is that Conjecture \[ratcurve\] is true for $X$ in Case (2).
Case (3): One reducible fibre, with three components, configured like a letter H, where the leftmost vertical component is the one which intersects the section $S$. (See Table \[fibretable\].)
We first deal with the case in which the reducible fibre has no multiple components. Let $E_1$ be the component of the reducible fibre which intersects the section $S$. Let $E_2$ be the component which intersects both other components, and let $E_3$ be the remaining component. Let $F=E_1+E_2+E_3$ be the class of a fibre. These four classes are a basis of $\mbox{NS}(X)$, with intersection matrix: $$\begin{array}{r|rrrr}
& S & E_2 & E_3 & F \\ \hline
S & -n & 0 & 0 & 1 \\
E_2 & 0 & -2 & 1 & 0 \\
E_3 & 0 & 1 & -1 & 0 \\
F & 1 & 0 & 0 & 0
\end{array}$$
The curves $S$, $E_1$, $E_2$, and $E_3$ generate the effective cone of $X$. The nef cone of $X$ is generated by the divisors $F$, $D_1=nF+S-E_2-2E_3$, $D_2=nF+S-E_2-E_3$, and $D_3=nF+S$.
[*Proof:*]{} A straightforward calculation shows that the two cones described in the claim are dual to one another. Thus, it suffices to show that $F$, $D_1$, $D_2$, and $D_3$ generate the nef cone of $X$. To do this, it further suffices to show that every positive linear combination of $F$, $D_1$, $D_2$, and $D_3$ is ample, since the converse inclusion is clear from the effectivity of $S$, $F_1$, $E_1$, and $E_2$.
The intersections of $F$ and the $D_i$ are tabulated as follows: $$\begin{array}{r|cccc}
& F & D_1 & D_2 & D_3 \\ \hline
F & 0 & 1 & 1 & 1 \\
D_1 & 1 & n-2 & n-1 & n \\
D_2 & 1 & n-1 & n-1 & n \\
D_3 & 1 & n & n & n
\end{array}$$
We first note that $n$ can be chosen to be at least two, since if $n=0$ or $n=1$, then $X$ is a blowup of $\P^2$ at two infinitely near points, and is therefore also a blowup of the second Hirzebruch surface $H_2$, in which case we can take $n=2$.
Now assume that $n\geq 2$. It’s clear that $F$ and each of the $D_i$ are basepoint free: $F$ corresponds to a morphism to $\P^1$, $D_1$ corresponds to a morphism to a cone over a smooth rational curve of degree $n-2$ (if $n=2$ we take this cone to be $\P^1$), $D_2$ corresponds to a morphism to a cone over a smooth rational curve of degree $n-1$, and $D_3$ corresponds to a morphism to a cone over a smooth rational curve of degree $n$ (see Remark \[actualcones\]). For each $D_i$, the ruling of the cone pulls back to $F$. In particular, every positive linear combination of $F$ and the $D_i$ has positive intersection with every curve on $X$, and has positive self-intersection. Therefore, by the Nakai-Moishezon criterion, every such positive linear combination is ample, and we have proven the claim.
We now prove Conjecture \[ratcurve\] for $X$. Let $P\in X(k)$ be any rational point. The case $n>3$ is immediate from Theorem \[onefibre\].
Assume $n=3$, and let $P\in X(k)$ be any $k$-rational point. If $P$ does not lie on $S$ or any $E_i$, then every generator of the nef cone either contracts the component $C$ of $F$ through $P$, or else contracts no curves through $P$ and maps $C$ to a line. Thus, Theorem \[projlines\] and Corollary \[divsum\] imply that for any ample divisor $D$, $C$ is a curve of best $D$-approximation to $P$.
If $P$ lies on $S$ but not $E_1$, then consider the cone $A_S$, generated by: $$F,F+D_1,F+D_2,F+D_3$$ Each divisor is basepoint free. The last three divisors contract no curves through $P$, but map the component $C$ of $F$ through $P$ to a line, while $F$ contracts $C$. By Theorem \[projlines\] and Corollary \[divsum\], we conclude that $C$ is a curve of best $D$-approximation to $P$ for any $D$ in $A_S$. If $D$ does not lie in $A_S$, then it must lie in $A'_S$, generated by: $$D_1,F+D_1,D_2,F+D_2,D_3,F+D_3$$ Each $D_i$ contracts $S$, while $F+D_i$ contracts no curves through $P$ and maps $S$ to a line. Thus, by Theorem \[projlines\] and Corollary \[divsum\], we conclude that $S$ is a curve of best $D$-approximation to $P$ for every $D\in A'_S$. Conjecture \[ratcurve\] is thus proven for $P$.
If $P$ lies on exactly one of the $E_i$ but not $S$, then for each generator $G$ of the nef cone, we see that either $G$ contracts $E_i$ or else maps $E_i$ to a line and contracts no curves through $P$. Theorem \[projlines\] and Corollary \[divsum\] now imply that $E_i$ is a curve of best $D$-approximation to $P$.
If $P=S\cap E_1$, then consider the cone $A_1$, generated by: $$F,D_1,D_2,F+D_3$$ The first three all contract $E_1$, and the last contracts no curves through $P$ but maps $E_1$ to a line. By Theorem \[projlines\] and Corollary \[divsum\], we conclude that $E_1$ is a curve of best $D$-approximation to $P$ for all $P\in A_1$. If $D$ does not lie in $A_1$, then $D$ must lie in $A'_1$, generated by: $$D_1,D_2,D_3,F+D_3$$ The first three curves all contract $S$, while the last contracts no curves through $P$ but maps $S$ to a line. Theorem \[projlines\] and Corollary \[divsum\] now imply that $S$ is a curve of best $D$-approximation to $P$ for any $D$ in $A'_1$.
If $P=E_1\cap E_2$, then consider the cone $A_2$, generated by: $$F,D_1,D_3,D_2+D_3$$ The first three classes all contract $E_2$, and the last contracts no curves through $P$ but maps $E_2$ to a line. We conclude by Theorem \[projlines\] and Corollary \[divsum\] that $E_2$ is a curve of best $D$-approximation to $P$ for all $D$ in $A_2$. If $D$ does not lie in $A_2$, then $D$ must lie in $A'_2$, generated by: $$F,D_1,D_2,D_2+D_3$$ The first three contract $E_1$, and the last contracts no curve through $P$ but maps $E_1$ to a line. By Theorem \[projlines\] and Corollary \[divsum\], we conclude that $E_1$ is a curve of best $D$-approximation to $P$ for all $D$ in $A'_2$. Conjecture \[ratcurve\] is thus proven for $P$.
Finally, if $P=E_2\cap E_3$, then consider the cone $A_3$, generated by: $$F,D_2,D_3,D_1+D_2$$ The first three all contract $E_3$, while the last contracts no curves through $P$ but maps $E_3$ to a line. We conclude by Theorem \[projlines\] and Corollary \[divsum\] that $E_3$ is a curve of best $D$-approximation to $P$ for any $D$ in $A_3$. If $D$ does not lie in $A_3$, then $D$ must lie in $A'_3$, generated by: $$F,D_1,D_3,D_2+D_3$$ The first three contract $E_2$, while that last contracts no curves through $P$ but maps $E_2$ to a line. Theorem \[projlines\] and Corollary \[divsum\] now imply that $E_2$ is a curve of best $D$-approximation to $P$ for all $D$ in $A'_3$. We conclude that Conjecture \[ratcurve\] is true for this $P$. Since this exhausts the possibilities for $P$, we conclude that Conjecture \[ratcurve\] is proven for $X$ in Case (3) with $n=3$.
If $n=2$, then both $F$ and $D_1$ correspond to maps to $\P^1$, so that $X$ is realized as the blowup of $\P^1\times\P^1$ at two infinitely near points. Let $P$ be any $k$-rational point on $X$ which does not lie on $S$ or any $E_i$. Let $C_1$ and $C_F$ be the components of $D_1$ and $F$, respectively, through $P$. Consider the cone $A$, generated by: $$F,F+D_1,D_2,D_3,D_1+D_3$$ All of these classes are basepoint free, and all but the last either contract $C_F$ or else contract no curves through $P$ and map $C_F$ to a line. Thus, if we can show that $C_F$ is a curve of best $(D_1+D_3)$-approximation to $P$, then Theorem \[projlines\] and Corollary \[divsum\] will imply that $C_F$ is a curve of best $D$-approximation to $P$ for any $D$ in $A$.
To prove this, note that $C_F.D_1=C_F.D_3=1$, so that a sequence of best $D_1$- or $D_3$-approximation to $P$ along $C.F$ has constant of approximation equal to 1. Thus, by Theorem \[product\], since $D_1$ and $D_3$ are both basepoint free, it follows that any sequence with a better constant of $(D_1+D_3)$-approximation to $P$ than $1+1=2$ must have all but finitely many points contained in a union of curves through $P$ which are contracted by $D_1$ or $D_3$. The only such curve is $C_1$, which is only contracted by $D_1$. Since $C_1.(D_1+D_3)=2$, it follows that no sequence of points on $C_1$ can have a constant of $(D_1+D_3)$-approximation to $P$ of better than 2. Thus, $C_F$ is a curve of best $(D_1+D_3)$-approximation to $P$, and since $C_1.(D_1+D_3)=2$, we see that $C_1$ is also a curve of best $(D_1+D_3)$-approximation to $P$. (Recall that $P$ is assumed not to lie on $S$ or any $E_i$.) We conclude that $C_F$ is a curve of best $D$-approximation to $P$ for any $D$ in $A$.
If $D$ does not lie in $A$, then it must lie in the cone $B$, generated by: $$D_1,F+D_1,D_2,D_1+D_3$$ Each divisor is basepoint free, and the first three each either contract $C_1$ or else contract no curves through $P$ and map $C_1$ to a line, and we have already seen that $C_1$ is a curve of best $(D_1+D_3)$-approximation to $P$. Thus, by Corollary \[divsum\], we conclude that $C_1$ is a curve of best $D$-approximation to $P$ for any $D$ in $B$. It follows that Conjecture \[ratcurve\] is true for $P$.
If $P$ lies on $S$ but not any $E_i$, then for any $D$ in the cone $B$, it is easy to check that $S$ is a curve of best $D$-approximation to $P$. If $D$ lies in $A$, then consider the cone $A_S$, generated by: $$D_1+D_3,D_2,D_3,F+D_1,F+D_2,F+D_3$$ The first three of these contract $S$, and the last three contract no curves through $P$ but map $S$ to a line. By Theorem \[projlines\] and Corollary \[divsum\], it follows that $S$ is a curve of best $D$-approximation to $P$ for all $D$ in $A_S$. If $D$ does not lie in $A_S$ or $B$, then it must lie in the cone $A'_S$, generated by: $$F,F+D_1,F+D_2,F+D_3$$ The first of these contracts $C_F$, and the last three contract no curves through $P$ but map $C_F$ to a line. By Theorem \[projlines\] and Corollary \[divsum\], it follows that $C_F$ is a curve of best $D$-approximation to $P$ for all $D$ in $A'_S$. We conclude that Conjecture \[ratcurve\] is true for $P$.
Next, assume that $P$ lies on a curve $C$ which is either $E_1$ or $E_2$, and assume that $P$ does not lie on $S$ or $E_3$, and is not the point $E_1\cap E_2$. It is easy to check that each generator of the nef cone of $X$ either contracts $C$ or else contracts no curves through $P$ and maps $C$ to a line. We conclude that for any ample divisor $D$, the curve $C$ is a curve of best $D$-approximation to $P$.
If $P$ lies on $E_3$ but not $E_2$, then it is easy to check that every generator of the cone $A$ either contracts $E_3$ or else contracts no curves through $P$ but maps $E_3$ to a line. Thus, if $D$ lies in $A$, then $E_3$ is a curve of best $D$-approximation to $P$. If $D$ does not lie in $A$, then consider the cone $A_3$, generated by: $$D_2,D_1+D_3,F+D_1,D_1+D_2,2D_1+D_3$$ It is clear that $E_3$ is contracted by $D_2$ and mapped to a line by the next three divisors (which contract no curves through $P$). Moreover, $E_3$ is a curve of best $D_1$-approximation to $P$, and therefore is also a curve of best $2D_1$-approximation to $P$, and by Theorem \[addeffective\], since $E_3.D_3=0$, it follows that $E_3$ is also a curve of best $(2D_1+D_3)$-approximation to $P$. By Corollary \[divsum\], it follows that $E_3$ is a curve of best $D$-approximation to $P$ for any $D$ in $A_3$. If $D$ does not lie in either $A_3$ or $A$, then it must lie in $B_3$, generated by: $$D_1,F+D_1,D_1+D_2,2D_1+D_3$$ The first of these contracts $C_1$, the next two contract no curves through $P$ but map $C_1$ to a line, and since $(2D_1+D_3).C_1=2=(2D_1+D_3).E_3$, it follows that $C_1$ is a curve of best $(2D_1+D_3)$-approximation to $P$ as well. Thus, by Corollary \[divsum\], we deduce that $C_1$ is a curve of best $D$-approximation to $P$ for any $D$ in $B_3$, and hence for any $D$ in the nef cone.
This leaves just three points $P$ to check. If $P=S\cap E_1$, then consider the cone $A_1$, generated by: $$F+D_3,D_1,D_2,D_3$$ All of these are basepoint free, and all but the first of them contract $S$. Since $F+D_3$ contracts no curves through $P$ but maps $S$ to a line, we conclude by Theorem \[projlines\] and Corollary \[divsum\] that $S$ is a curve of best $D$-approximation to $P$ for all $D$ in $A_1$. If $D$ does not lie in $A_1$, then $D$ must lie in $B_1$, generated by: $$F,D_1,D_2,F+D_3$$ The first three of these contract $E_1$, and the last maps $E_1$ to a line but does not contract any curve through $P$. By Theorem \[projlines\] and Corollary \[divsum\], we conclude that $E_1$ is a curve of best $D$-approximation to $P$ for any $D$ in $B_1$, and hence that Conjecture \[ratcurve\] is true for $P$.
If $P=E_1\cap E_2$, then consider the cone $A_2$, generated by: $$F,D_1,D_2,D_2+D_3$$ The first three of these contract $E_1$, and the last contracts no curves through $P$ but maps $E_1$ to a line. We conclude by Theorem \[projlines\] and Corollary \[divsum\] that $E_1$ is a curve of best $D$-approximation to $P$ for all $D$ in $A_2$. If $D$ does not lie in $A_2$, then $D$ must lie in $B_2$, generated by: $$F,D_1,D_3,D_2+D_3$$ By an exactly similar argument to that used for $A_2$, we conclude that $E_2$ is a curve of best $D$-approximation to $P$ for any $D$ in $B_2$, and therefore that Conjecture \[ratcurve\] is true for $P$.
Finally, we consider the case that $P=E_2\cap E_3$. Let $A'_2$ be the cone generated by: $$F,D_1,D_3,D_1+D_2$$ The first three divisors contract $E_2$, and the last contracts no curves through $P$ but maps $E_2$ to a line. Theorem \[projlines\] and Corollary \[divsum\] thus imply that $E_2$ is a curve of best $D$-approximation to $P$ for any $D$ in $A'_2$. If $D$ does not lie in $A'_2$, then it must lie in $B'_2$, generated by: $$F,D_2,D_3,D_1+D_2$$ Again, each of the first three divisors contracts $E_3$, while the last contracts no curves through $P$ but maps $E_3$ to a line. We conclude from Theorem \[projlines\] and Corollary \[divsum\] that $E_3$ is a curve of best $D$-approximation to $P$ for any $D$ in $B'_2$, and therefore that Conjecture \[ratcurve\] is true for $P$. This completes the proof of Conjecture \[ratcurve\] for Case (3), provided that the reducible fibre has no multiple components.
To finish the proof, it remains only to consider the case of a multiple component in the reducible fibre. The only way this can occur is if the crossbar $E_2$ of the H has multiplicity two in the fibre.
The effective cone of $X$ is generated by the classes of $S$, $E_1$, $E_2$, and $E_3$. The nef cone of $X$ is generated by the classes of $F$, $D_1=S+nF$, $D_2=2S+2nF-2E_2-E_3$, and $D_3=S+nF-E_2-E_3$.
[*Proof of claim:*]{} A simple calculation shows that the two cones are dual, so to prove the claim it suffices to show that every positive linear combination of $F$, $D_1$, $D_2$, and $D_3$ is ample. The classes $D_1$ and $D_3$ are basepoint free, since they correspond to morphisms to cones (see Remark \[actualcones\]), and $F$ is also basepoint free because it corresponds to a morphism to $\P^1$. Finally, note that $D_2$ is basepoint free because it is linearly equivalent both to $2D_3+E_3$ and $D_1+S+(n-1)F+E_1$.
Thus, to prove the claim, we may invoke the Nakai-Moishezon Criterion for ampleness so that it suffices to show that every positive linear combination of $F$, $D_1$, $D_2$, and $D_3$ has positive self-intersection. These four divisors have intersection numbers as follows: $$\begin{array}{r|cccc}
& F & D_1 & D_2 & D_3 \\ \hline
F & 0 & 1 & 2 & 1 \\
D_1 & 1 & n & 2n & n \\
D_2 & 2 & 2n & 4n-2 & 2n-1 \\
D_3 & 1 & n & 2n-1 & n-1
\end{array}$$ Thus, provided that we choose $n\geq 1$ – which we may do without loss of generality – the claim is proven.
We now prove Conjecture \[ratcurve\] in the case that the reducible fibre has a multiple component. First, assume that $P$ does not lie on $S$ or the reducible fibre. The component $C$ of $F$ through $P$ is, by Theorem \[projlines\], a curve of best $D_1$- and $D_3$-approximation to $P$ (since these divisors contract no curves through $P$), and it’s clear that $C$ is a curve of best $F$-approximation to $P$. Finally, if we write $D_2=2D_3+E_3$, then Corollary \[addeffective\] implies that $C$ is a curve of best $D_2$-approximation to $P$. Thus, by Corollary \[divsum\], we see that $C$ is a curve of best $D$-approximation to $P$ for all ample divisors $D$.
Next, assume that $P$ lies on $S$ but not any $E_i$, and consider the cone $A_S$, generated by: $$D_1,D_2,D_3,F+D_1,F+D_3,2F+D_2$$ All of these divisors are basepoint free. The first three contract $S$, the next two contract no curves through $P$ and map $S$ to a line, so by Theorem \[projlines\], $S$ is a curve of best approximation to $P$ with respect to any of these five divisors. Moreover, $S$ is a curve of best $F$-approximation to $P$, and since $D_2.S=0$ and $D_2$ is basepoint free, it follows from Corollary \[addeffective\] that $S$ is also a curve of best $(2F+D_2)$-approximation to $P$. Thus, by Corollary \[divsum\], it follows that $S$ is a curve of best $D$-approximation to $P$ for all $D$ in $A_S$. If $D$ does not lie in $A_S$, then it must lie in the cone $B_S$, generated by: $$F,F+D_1,F+D_3,2F+D_2$$ Since $F$ contracts the component $C$ of $F$ through $P$, we see that $C$ is a curve of best $F$-approximation to $P$. Furthermore, $F+D_1$ and $F+D_3$ both contract no curves through $P$ but map $C$ to a line, so $C$ is a curve of best approximation to $P$ with respect to those two divisors as well. Finally, since $C.(2F+D_2)=2=S.(2F+D_2)$, the fact that $S$ is a curve of best $(2F+D_2)$-approximation to $P$ implies immediately that $C$ is also a curve of best $(2F+D_2)$-approximation to $P$. By Corollary \[divsum\], we conclude that Conjecture \[ratcurve\] is true for $P$.
If $P$ lies on exactly one $E_i$ but not $S$, then $F$ and $D_j$ for $j\neq i$ contract $E_i$, and $D_i$ contracts no curves through $P$ but maps $E_i$ to a line. Thus, by Theorem \[projlines\] and Corollary \[divsum\], $E_i$ is a curve of best $D$-approximation to $P$ for any ample divisor $D$.
If $P=S\cap E_1$, consider the cone $A_1$, generated by: $$F+D_1,D_2,D_3,F$$ The last three divisors contract $E_1$, and the first contracts no curves through $P$ but maps $E_1$ to a line. We conclude by Theorem \[projlines\] and Corollary \[divsum\] that $E_1$ is a curve of best $D$-approximation to $P$ for any $D$ in $A_1$. If $D$ does not lie in $A_1$, then it must lie in $B_1$, generated by: $$D_1,D_2,D_3,F+D_1$$ The first three divisors contract $S$, while the last contracts no curves through $P$ but maps $S$ to a line. By Theorem \[projlines\] and Corollary \[divsum\], it follows that $S$ is a curve of best $D$-approximation to $P$ for any $D$ in $B_1$. Conjecture \[ratcurve\] is therefore proven for $P$.
If $P=E_1\cap E_2$, then consider the cone $A_2$, generated by: $$F,D_1,D_3,D_1+D_2$$ The first three contract $E_2$, and the last contracts no curves through $P$ but maps $E_2$ to a line. Theorem \[projlines\] and Corollary \[divsum\] now imply that $E_2$ is a curve of best $D$-approximation to $P$ for any $D$ in $A_2$. If $D$ does not lie in $A_2$, then it must lie in $B_2$, generated by: $$F,D_2,D_3,D_1+D_2$$ The first three contract $E_1$, while the last contracts no curves through $P$ but maps $E_1$ to a line. Theorem \[projlines\] and Corollary \[divsum\] now imply that $E_1$ is a curve of best $D$-approximation to $P$ for any $D$ in $B_2$. Conjecture \[ratcurve\] is therefore proven for $P$.
Finally, assume that $P=E_2\cap E_3$. Assume that $D$ lies in the cone $A_3$, generated by: $$F,D_1,D_2,D_2+D_3$$ The first three of these contract $E_3$, and the last contracts no curves through $P$ but maps $E_3$ to a line. By Theorem \[projlines\] and Corollary \[divsum\], it follows that $E_3$ is a curve of best $D$-approximation to $P$ for any $D$ in $A_3$. If $D$ does not lie in $A_3$, then it must lie in $B_3$, generated by: $$F,D_1,D_3,D_2+D_3$$ The first three contract $E_2$, and the last contracts no curves through $P$ but maps $E_2$ to a line. By Theorem \[projlines\] and Corollary \[divsum\], we conclude that $E_2$ is a curve of best $D$-approximation to $P$ for any $D$ in $B_3$. Conjecture \[ratcurve\] is therefore proven for $X$, for Case (3), and for any split rational surface of Picard rank at most four.
When the Picard rank is larger than four, the number of different cases to consider becomes much larger, so we will prove only a few cases, hopefully representative of the general flavour. We first turn our attention to the blowup of $\P^2$ at four $k$-rational points in general position. The surface $X$ admits a morphism $\pi_1\colon
X\rightarrow \P^1$ whose fibres are the strict transforms of the conics through the four blown up points, so that in particular $\pi_1^*\O(1) = 2L-E_1-E_2-E_3-E_4$. It also admits a morphism $\pi_2\colon X\rightarrow \P^2$, which is the blowing down map. Let $E_1$, $E_2$, $E_3$, and $E_4$ be the four exceptional divisors of $\pi_2$, and let $L=\pi_2^*\O(1)$ be the pullback of a line in $\P^2$.
Let $X$ be the blowup of $\P^2$ at four $k$-rational points in general position, and let $P\in X(k)$ be any $k$-rational point. Then a sequence of best approximation to $P$ can be chosen to lie as a subset of the rational curve of minimal degree through $P$.
[*Proof:*]{} The geometry of $X$ is well understood (see for instance Example 2.1.2 of \[Tsch\], or [@Ma]). The nef cone of $X$ is generated by the ten classes $L$, $L_i=L-E_i$, $D=2L-E_1-E_2-E_3-E_4$, and $D_i=D+E_i$ for $i=1,2,3,4$. Each of these is basepoint free: $D$ and each $L_i$ is the fibre of a morphism to $\P^1$, while $L$ and each $D_i$ corresponds to a birational morphism to $\P^2$, each one blowing down four pairwise disjoint smooth rational curves. There are exactly ten $(-1)$-curves on $X$, namely $E_i$ and $L-E_i-E_j$ for $i=1,2,3,4$ and $i\neq j$.
We have the following table of intersection numbers: $$\begin{array}{r|llllllllll}
& L & L_1 & L_2 & L_3 & L_4 & D_1 & D_2 & D_3 & D_4 & D \\ \hline
L & 1 & 1 & 1 & 1 & 1 & 2 & 2 & 2 & 2 & 2 \\
L_1 & 1 & 0 & 1 & 1 & 1 & 2 & 1 & 1 & 1 & 1 \\
L_2 & 1 & 1 & 0 & 1 & 1 & 1 & 2 & 1 & 1 & 1 \\
L_3 & 1 & 1 & 1 & 0 & 1 & 1 & 1 & 2 & 1 & 1 \\
L_4 & 1 & 1 & 1 & 1 & 0 & 1 & 1 & 1 & 2 & 1 \\
D_1 & 2 & 2 & 1 & 1 & 1 & 1 & 2 & 2 & 2 & 1 \\
D_2 & 2 & 1 & 2 & 1 & 1 & 2 & 1 & 2 & 2 & 1 \\
D_3 & 2 & 1 & 1 & 2 & 1 & 2 & 2 & 1 & 2 & 1 \\
D_4 & 2 & 1 & 1 & 1 & 2 & 2 & 2 & 2 & 1 & 1 \\
D & 2 & 1 & 1 & 1 & 1 & 1 & 1 & 1 & 1 & 0
\end{array}$$
Let $P\in X(k)$ be any $k$-rational point which does not lie on a $(-1)$-curve. Define $B$ to be the cone of nef divisors $A$ such that $A.D\leq A.L_i$ for all $i$. Similarly, for each $i$ between 1 and 4, define the cone $B_i$ as the cone of all nef divisors $A$ such that $A.L_i\leq A.D$ and $A.L_i\leq A.L_j$ for $j\neq i$. A short calculation shows that $B$ is generated by the ten classes: $$D, D+L, D_i, D+L_i$$ for $i=1,2,3,4$, and that $B_i$ is generated by the ten classes: $$L, D+L, L_i, L_i+L_j, D+L_i, D_j$$ for $j\neq i$.
Consider the cone $B_i$. Let $C_i$ be the component of $L_i$ through $P$, and let $C_D$ be the component of $D$ through $P$. Each generator of $B_i$ except $L+D$ either contracts $C_i$, or else it contracts no curves through $P$ and maps $C_i$ to a line. Thus, by Theorem \[projlines\], $C_i$ is a curve of best approximation to $P$ with respect to all the generators of $B_i$ except possibly $L+D$. For $L+D$, note that $C_i.L=C_i.D=1$, so that a sequence of best $L$- or $D$-approximation to $P$ along $C_i$ has constant of approximation equal to 1. Since $L$ and $D$ are basepoint free, it follows from Theorem \[product\] and Theorem \[projlines\] that any sequence with a better constant of $(L+D)$-approximation to $P$ than $1+1=2$ must have all but finitely many points contained in a union of curves through $P$ which are contracted by $L$ or $D$. The only such curve is $C_D$, which is contracted by $D$ but has $L$-degree 2, and so a sequence of best $(L+D)$-approximation to $P$ along $C_D$ also has constant of approximation equal to 2. Thus, we conclude that both $C_i$ and $C_D$ are curves of best $(L+D)$-approximation to $P$. By Corollary \[divsum\], we conclude that Conjecture \[ratcurve\] is true for any divisor $A$ in $B_i$.
Now consider the cone $B$. Each generator of $B$ except $L+D$ either contracts $C_D$, or else contracts no curves through $P$ and maps $C_D$ to a line. Thus, by Theorem \[projlines\], $C_D$ is a curve of best approximation to $P$ with respect to all the generators of $B$, except possibly $L+D$. We have already shown that $C_D$ is a curve of best $(L+D)$-approximation to $P$, so by Corollary \[divsum\], it follows that Conjecture \[ratcurve\] is true for any ample divisor $A$ in $B$, and thus for any ample divisor $A$.
There remains the possibility that $P$ might lie on some $(-1)$-curve. Assume first that $P$ lies on exactly one $(-1)$-curve. By applying a suitable automorphism of $X$, we may assume without loss of generality that this curve is $E_i$. It is easy to check that any generator of the cone $B_j$ ($j\neq i$) either contracts $E_i$ or else contracts no curves through $P$ and maps $E_i$ to a line. Thus, by Theorem \[projlines\] and Corollary \[divsum\], it follows that $E_i$ is a curve of best $A$-approximation to $P$ for any $A$ in $B_j$ for $j\neq i$.
If $A$ lies in the cone $B_i$, then consider the subcone $M_i$ of $B_i$ generated by: $$L+L_i,L_i,L_i+L_j,D_j,D+L_i$$ Each generator is basepoint free, and either contracts $C_i$ or else contracts no curves through $P$ and maps $C_i$ to a line. Thus, by Theorem \[projlines\] and Corollary \[divsum\], we conclude that $C_i$ is a curve of best $A$-approximation to $P$ for any $A$ in $M_i$. If $A$ lies in $B_i$ but not $M_i$, then it must lie in $N_i$, generated by: $$L,L+L_i,L_i+L_j,D_j,D+L$$ Each generator is basepoint free, and either contracts $E_i$ or else contracts no curves through $P$ and maps $E_i$ to a line. Thus, by Theorem \[projlines\] and Corollary \[divsum\], we conclude that $E_i$ is a curve of best $A$-approximation to $P$ for any $A$ in $N_i$, and therefore for any $A$ in $B_i$.
The only other possibility is that $A$ lies in the cone $B$. Consider the cone $J_i$, generated by: $$D_i,D_j,L+D,L_j+D,D_i+D$$ where $j\neq i$. All divisors are basepoint free, and either contract $E_i$ or else contract no curves through $P$ and map $E_i$ to a point. By Theorem \[projlines\] and Corollary \[divsum\], we conclude that $E_i$ is a curve of best $A$-approximation to $P$ if $A$ lies in $J_i$. If $A$ lies in $B$ but not in $J_i$, then $A$ is in the cone $J'_i$, generated by: $$D,D_j,L_i+D,L_j+D,D_i+D$$ where $j\neq i$. All divisors are basepoint free, and either contract $D$ or else contract no curves through $P$ and map $D$ to a point. By Theorem \[projlines\] and Corollary \[divsum\], we conclude that $D$ is a curve of best $A$-approximation to $P$ if $A$ lies in $J'_i$. This concludes the proof of Conjecture \[ratcurve\] for $P$ lying on a single $(-1)$-curve.
Finally, assume that $P=E_i\cap L-E_i-E_j$ for some $i$ and $j$. By symmetry, we may assume that $i=1$ and $j=2$. Let $A$ be any element of the nef cone. Consider the cone $M_1$, generated by: $$L,L+L_1,L_2,L_3,L_4,L_1+L_3,L_1+L_4,D_1,D_2,D_3+L_4,D_4+L_3,$$ $$D+L_3,D+L_4,D+D_1$$ Each generator of $M_1$ is basepoint free, and either contracts $E_1$ or else contracts no curves through $P$ and maps $E_1$ to a line. Theorem \[projlines\] and Corollary \[divsum\] now imply that $E_1$ is a curve of best $A$-approximation to $P$ for any $A$ in $M_1$. If $A$ does not lie in $M_1$, then it must lie in the cone $N_1$, generated by: $$L_1,L_2,L+L_1,L_1+L_3,L_1+L_4,D_2,D_3,D_4,D,D_3+L_4,D_4+L_3,D+L_3,$$ $$D+L_4,D+D_1$$ Each generator of $N_1$ is basepoint free, and either contracts $L-E_1-E_2$ or else contracts no curves through $P$ and maps $L-E_1-E_2$ to a line. By Theorem \[projlines\] and Corollary \[divsum\], we conclude that $L-E_1-E_2$ is a curve of best $A$-approximation to $P$ for any $A$ in $N_1$, and hence for any ample divisor $A$.
This concludes the proof of Conjecture \[ratcurve\] for $X$.
Let $X$ be the blowup of $\P^2$ at five $k$-rational points in general position. Conjecture \[ratcurve\] is true for $X$.
[*Proof:*]{} The geometry of $X$ is well understood (see, for example, [@Ma] or [@Dr]). The surface $X$ admits a blowing-down map $\pi\colon X\rightarrow\P^2$. Let $L$ be the class of the preimage of a line, and let $E_1$, $E_2$, $E_3$, $E_4$, and $E_5$ be the exceptional curves of $\pi$. The classes $L$ and $E_i$ generate the Néron-Severi group of $X$. The effective cone of $X$ is generated by the 16 $(-1)$-curves on $X$, namely, $E_i$, $L-E_i-E_j$ for $i\neq j$, and $E=2L-E_1-E_2-E_3-E_4-E_5$. The ample cone is dual to the effective cone, and is generated by the following 26 divisors: $$\begin{array}{ccccc}
L, & L_i=L-E_i, & L_{ij}=E+E_i+E_j, & C_i=E+E_i, & B_i=E+L_i
\end{array}$$ Let us adopt the convention that the notation $L_{ij}$ implies that $i<j$. We have the following table of intersection numbers:
$L$ $L_i$ $L_{ij}$ $C_i$ $B_i$
------------- ----- -------------------------------- -------------------------------- -------------------------------- --------------------------------
$L$ 1 1 2 2 3
$L_k$ 1 $1-\delta_{ik}$ $1+\delta_{ik}+\delta_{jk}$ $1+\delta_{ik}$ $2-\delta_{ik}$
$L_{k\ell}$ 2 $1+\delta_{ik}+\delta_{i\ell}$ $3-\delta_{ik}-\delta_{j\ell}$ $2-\delta_{ik}-\delta_{i\ell}$ $2+\delta_{ik}+\delta_{i\ell}$
$C_k$ 2 $1+\delta_{ik}$ $2-\delta_{ik}-\delta_{jk}$ $1-\delta_{ik}$ $1+\delta_{ik}$
$B_k$ 3 $2-\delta_{ik}$ $2+\delta_{ik}+\delta_{jk}$ $1+\delta_{ik}$ $2-\delta_{ik}$
Let $P$ be any point on $X$. We first assume that $P$ does not lie on any of the 16 $(-1)$-curves on $X$. In that case, we will show that for any ample divisor $D$, a sequence of best $D$-approximation to $P$ lies along a curve in the class $|C_i|$ for some $i$, or in the class $|L_i|$ for some $i$.
To do this, we introduce 10 subcones of the nef cone of $X$. For $i$ between 1 and 5, we define the cone $M_i$ to be the cone of ample divisors for which the divisor $L_i$ has minimal degree amongst the classes $L_j$ and $C_j$, and we define the cone $N_i$ to be the cone of ample divisors for which the divisor $C_i$ has minimal degree amongst the classes $L_j$ and $C_j$. We will show that for every divisor in $M_i$, a sequence of best approximation to $P$ lies along a curve in $|L_i|$, and similarly that for any divisor in $N_i$, a sequence of best approximation to $P$ lies along a curve in $|C_i|$.
We first consider $M_i$. For simplicity, we will consider only $M_1$; the other four cases differ from this one only by a permutation of the indices. The cone $M_1$ has 26 generators: $$L,\,\,L_1,\,\,L_1+L_i,\,\,L_1+C_i,\,\,L_{ij},\,\,B_1,\,\,L+C_i,
\,\,L_1+C_1,\,\,B_1+L_i$$ where in all cases we assume that $i\neq 1$ and $j\neq 1$. Of these generators, only $L_1$ is not big. Let $F$ be the unique curve through $P$ satisfying $F\in |L_1|$. Since we have assumed that $P$ does not lie on a $(-1)$-curve, it follows that if a curve through $P$ has intersection zero with one of these generators, then the curve in question must be $F$.
It suffices to show that for any generator of $M_1$, a sequence of best approximation to $P$ lies on $F$. For $L$, $L_1$, $L_1+L_i$, $L_1+C_i$, $L_{ij}$, and $B_1$, this follows from Theorem \[projlines\] because $F$ has degree one or zero. In all other cases, $F$ has degree two, and we can write them as follows: $$\begin{aligned}
L+C_2 & = & L_3+L_{23} \\
L+C_3 & = & L_2+L_{23} \\
L+C_4 & = & L_2+L_{24} \\
L+C_5 & = & L_2+L_{25} \\
L_1+C_1 & = & L_2+C_2 \\
B_1+L_2 & = & C_3+L_{45} \\
B_1+L_3 & = & C_2+L_{45} \\
B_1+L_4 & = & C_2+L_{35} \\
B_1+L_5 & = & C_2+L_{34}\end{aligned}$$ For each generator $D$ in the above list, we have written $D=D_1+D_2$, where $D_i.F=1$. Thus, by Theorem \[product\], $F$ contains a sequence of best $D$-approximation to $P$ unless $P$ lies on some curve of smaller $D$-degree. Since we have assumed that $P$ does not lie on a $(-1)$-curve, and since $D$ lies in $M_1$, it follows that $P$ cannot lie on any curve of smaller $D$-degree than $L_1$. (Note that the only candidates for a curve of smaller degree than two are those which intersect one of $D_1$ or $D_2$ trivially, and such curves are either $(-1)$-curves or else of the form $L_i$ or $C_i$ for some $i$.) We conclude that a sequence of best $D$-approximation to $P$ lies on $F$.
Now to $N_i$. Again, we will consider only $N_1$, the cone of divisors with respect to which $C_1$ has minimal degree amongst the $L_j$ and $C_j$. Like $M_1$, $N_1$ has 26 generators: $$C_1,\,\,C_1+L_i,\,\,C_1+L_{1i},\,\,L_{1i},\,\,B_i,\,\,L+C_1,\,\,
L_1+C_1,\,\,C_1+L_{ij},\,\,B_2+C_2$$ where again we assume that $i\neq 1$ and $j\neq 1$. Let $F$ be the curve in $|C_1|$ containing $P$. As before, it suffices to show that for any generator of $N_1$, a sequence of best approximation to $P$ lies on $F$. For $C_1$, $C_1+L_i$, $C_1+L_{1i}$, $L_{1i}$, and $B_i$, this follows from Theorem \[projlines\] because $F$ has degree one or zero, and all the generators contract no curves through $P$ other than $F$. For the remaining nine generators, we can write: $$\begin{aligned}
L+C_1 & = & L_2+L_{12} \\
L_1+C_1 & = & L_2+C_2 \\
B_2+C_2 & = & B_2+C_2 \\
C_1+L_{45} & = & L_{15}+C_4 \\
C_1+L_{35} & = & L_{15}+C_3 \\
C_1+L_{34} & = & L_{14}+C_3 \\
C_1+L_{25} & = & L_{15}+C_2 \\
C_1+L_{24} & = & L_{14}+C_2 \\
C_1+L_{23} & = & L_{13}+C_2\end{aligned}$$ In each case, we have written the generator $D$ as $D=D_1+D_2$, where $D_i.F=1$. Thus, by Theorem \[product\], we conclude that since $P$ does not lie on a $(-1)$-curve, a sequence of best $D$-approximation to $P$ lies on $F$ for any generator $D$ of $N_1$, and by Corollary \[divsum\] also for any $D$ in $N_1$.
It remains only to treat the case where $P$ lies on a $(-1)$-curve. Since no three $(-1)$-curves meet at a point, $P$ lies either on exactly one or exactly two $(-1)$-curves. Let’s first assume that $P$ lies on exactly one $(-1)$-curve. Since the automorphism group of $X$ acts transitively on the set of $(-1)$-curves (see for example Theorem 2.1 of [@Ho]), we may assume that this curve is $E$.
Consider the cone $M_i$. We will subdivide $M_i$ into two subcones. The first is $R_i$, generated by the following twenty-two divisor classes: $$L,L_i,L_i+L_j,L_{j\ell},L_i+C_j,L_i+B_i,L+C_j,L+L_i+C_i$$ where $j$ and $\ell$ are assumed to be different from $i$. The classes $L$, $L_i$, $L_i+L_j$, $L_{j\ell}$, $L+C_j$, and $L_i+B_i$ all either contract the component $F_i$ of $L_i$ through $P$, or else contract no curves through $P$ and map $F_i$ to a line. Thus, by Theorem \[projlines\] $F_i$ is a curve of best approximation to $P$ with respect to any of these divisor classes. For $L+C_j$, we see that $L.L_i=C_j.L_i=1$, and so Theorem \[product\] implies that if $F_i$ is not a curve of best $(L+C_j)$-approximation to $P$, then a sequence of best approximation to $P$ can be found on a curve contracted by either $L$ or $C_j$. The only such curve through $P$ is $E$, and we check that $E.(L+C_j)=2$, so $F_i$ is a curve of best $(L+C_j)$-approximation to $P$ (as is $E$). This only leaves $L+L_i+C_i$, which can be rewritten $L+L_i+C_i=L+L_j+C_j$ for any $j$. Since $L.L_i=L_j.L_i=C_j.L_i=1$ for any $j\neq i$, we may apply Theorem \[product\] to argue that if $F_i$ is not a curve of best $(L+L_i+C_i)$-approximation to $P$, then by Theorem \[product\] a sequence of best approximation to $P$ must lie on some curve contracted by $L$, $L_j$, or $C_j$. The only such curves through $P$ are $E$ and the component of $L_j$ through $P$. None of these have $(L+L_i+C_i)$-degree less than $3$ (which is the degree of $F_i$), so we conclude that $F_i$ is a curve of best $(L+L_i+C_i)$-approximation to $P$. Thus, by Corollary \[divsum\], it follows that $F_i$ is a curve of best $A$-approximation to $P$ for any $A$ in $R_i$.
If $A$ lies in $M_i$ but not $R_i$, then it must lie in the cone $R'_i$, generated by: $$L_{j\ell},L_i+C_i,L_i+C_j,B_i,B_i+L_j,B_i+L_i,L+L_i+C_i,L+C_j$$ where $j$ and $\ell$ range over every index different from $i$. All but the last two divisors in this list either contract $E$, or else contract no curves through $E$ and map $P$ to a line. For the last two, we see that $E$ has the same degree as $F_i$, and we already know that $F_i$ is a curve of best $(L+L_i+C_i)$- and $(L+C_j)$-approximation to $P$. Thus, we conclude that $E$ is a curve of best $A$-approximation to $P$ for every $A$ in $R'_i$.
Consider the cone $N_i$. It is straightforward to check that except for $L+C_i$, every generator of $N_i$ either contracts $E$ or else contracts no curves through $P$ and maps $E$ to a line. We can rewrite $L+C_i=L_j+L_{ij}$ for any $j\neq i$, and note that $L_j.E=L_{ij}.E=1$, so that by Theorem \[product\], $E$ is a curve of best $(L+C_i)$-approximation to $P$ unless $P$ lies on some curve of smaller $(L+C_i)$-degree which is contracted by $L$ or $C_i$. Since this is not the case, we conclude by Theorem \[projlines\] and Corollary \[divsum\] that $E$ is a curve of best $A$-approximation to $P$ for any $A$ in $N_i$.
To finish the proof of Conjecture \[ratcurve\] for $X$, we just need to consider the case that $P$ is the intersection point of two $(-1)$-curves. Up to automorphisms of $X$, we may assume that one of these $(-1)$-curves is $E$. Since the only $(-1)$-curves intersecting $E$ are the $E_i$, we may assume that $P=E\cap E_i$ for some $i$.
Consider the cone $J_1$, generated by the following divisor classes: $$L+L_i,L_i,L_i+L_j,L_{j\ell},L_i+C_j,L_i+B_i$$ where $j$ and $\ell$ range over all values different from $i$. Each of these divisors either contracts the component $F_i$ of $L_i$ through $P$, or else contracts no curves through $P$ and maps $F_i$ to a line. By Theorem \[projlines\] and Corollary \[divsum\], we conclude that Conjecture \[ratcurve\] is true for any $A$ in $J_1$.
Next, consider the cone $J_2$, generated by: $$L,L_j,L+L_i,C_i,L_i+L_j,L_{ij},L_{ji},L_{j\ell},L_j+C_\ell,L_{ij}+C_j,
L_j+B_j$$ where $j$ and $\ell$ range over all indices different from $i$ and each other. Thus, the cone $J_2$ is generated by 41 divisor classes. Each of these divisors either contracts $E_i$, or else contracts no curves through $P$ and maps $E_i$ to a line. By Theorem \[projlines\] and Corollary \[divsum\], we conclude that Conjecture \[ratcurve\] is true for all $A$ in $J_2$.
Finally, if $A$ does not lie in $J_1$ or $J_2$, then it must lie in $J_3$, generated by: $$L_i+B_i,C_i,L_{j\ell},L_i+C_j,L_j+C_\ell,B_j,L_{ij}+C_j,L_j+B_j$$ where again $j$ and $\ell$ are different from each other and $i$, for a total of 41 divisor classes in the list. Each class either contracts $E$, or else contracts no curves through $P$ and maps $E$ to a line. Thus, it follows from Theorem \[projlines\] and Corollary \[divsum\] that Conjecture \[ratcurve\] is true for any divisor $A$ in $J_3$.
This concludes the proof of Conjecture \[ratcurve\] for $X$.
Further Remarks
===============
These techniques will probably be able to prove Conjecture \[ratcurve\] for more split rational surfaces, but they will probably not suffice to prove the conjecture for a general rational surface. For example, let $X$ be a blowup of $\P^2$ at six $k$-rational points in general position, embedded in $\P^3$ as a smooth cubic surface. The family $\mathcal{F}$ of plane cubic curves passing through the set of blown up points in $\P^2$ is a three-dimensional (projective) linear subspace of the $\P^{9}$ of plane cubic curves. The Zariski closure $Z$ of the set of cuspidal cubics is a closed subset of dimension 7. Therefore, $Z\cap\mathcal{F}$ has dimension at least one, so if the six points are chosen to lie on some cuspidal cubic whose cusp is not one of the six blown up points (which will generically be the case), there will be a one-dimensional family $\mathcal{C}$ of cuspidal cubic curves on $X$.
Let $P$ be the cusp of a cuspidal cubic $C$, and assume that $P$ is $k$-rational. Then one can find a sequence on $C$ which approximates $P$ with constant of approximation $3/2$. In particular, no rational curve of minimal degree through $P$ is a curve of best approximation to $P$, since that curve will generally be a conic ($P$ cannot lie on any of the 27 lines on $X$). This suggests that any proof of Conjecture \[ratcurve\] for $X$ will be beyond the techniques of this paper.
Note also that Vojta’s Main Conjecture ([@Vo], Conjecture 3.4.3) implies Conjecture \[ratcurve\] for many varieties. Vojta’s Main Conjecture is as follows:
\[vstrong\] Let $X$ be a smooth algebraic variety defined over a number field $k$, with canonical divisor $K$. Let $S$ be a finite set of places of $k$. Let $L$ be a big divisor on $X$, and let $D$ be a normal crossings divisor on $X$. Choose height functions $h_K$ and $h_L$ for $K$ and $L$, respectively, and define a proximity function $m_S(D,P) =
\sum_{v\in S} h_{D,v}(P)$ for $D$ with respect to $S$, where $h_{D,v}$ is a local height function for $D$ at $v$. Choose any $\epsilon>0$. Then there exists a nonempty Zariski open set $U=U(\epsilon)\subset X$ such that for every $k$-rational point $P\in U(k)$, we have the following inequality: $$M_S(D,P)H_K(P) \leq H_L(P)^{\epsilon}$$
(Note that our notation is multiplicative, rather than the additive notation used in [@Vo]. Thus, Vojta’s $h$ is our $\log H$, and Vojta’s $m$ is our $\log M$.)
Assume that $X$ is as in the conjecture, and has non-negative Kodaira dimension. Let $P$ be any fixed $k$-rational point, lying on a rational curve $C\subset X$, and let $D$ be any normal crossings divisor containing $P$. Since $X$ has non-negative Kodaira dimension, there is a non-empty open set $U_1\subset X$ such that there is a constant $c$ satisfying $H_K(Q)>c>1$ for all $Q\in U_1(k)$. Furthermore, if $S$ is the set of archimedean places of $k$, we may write: $$M_S(D,Q)\gg\left(\operatorname{dist}(P,Q)\right)^{-1}$$ where the implied constant is independent of $Q$. (If $Q$ lies on the support of $D$, then $M_S(D,Q)$ is infinite, but this will not affect our proof.)
Let $d$ be the constant of approximation for $P$ on $C$. For any $\epsilon>0$, let $U_\epsilon$ be the intersection of $U_1$ with the non-empty open set provided by Conjecture \[vstrong\]. Then for any point $Q\in U_\epsilon(k)$, Conjecture \[vstrong\] implies: $$\operatorname{dist}(P,Q)H_L(P)^\epsilon>c$$ If we choose $\epsilon<d$, then it’s clear that any sequence of best approximation to $P$ must eventually lie in the complement of $U_\epsilon$. If $X$ is a surface, then this complement is a union of curves, in which case the truth of Conjecture \[ratcurve\] is clear. If $X$ is of higher dimension, then a simple Noetherian induction shows that it suffices to assume that $X$ contains no subvariety $V$ whose Kodaira dimension is negative but whose dimension is at least two. Slightly more generally, we have therefore proven the following theorem:
Let $V$ be a smooth algebraic variety with non-negative Kodaira dimension. Let $P\in V(k)$ be any rational point with a rational curve $C/k$ on $V$ through $P$, and assume that every subvariety of $V$ with negative Kodaira dimension satisfies Conjecture \[ratcurve\] for $P$. Assume that Conjecture \[vstrong\] is true for $V$. Then Conjecture \[ratcurve\] is true for $P$ on $V$.
One can also use our techniques to prove Conjecture \[ratcurve\] for certain special points on some varieties of nonnegative Kodaira dimension. For example, let $X$ be a smooth quartic in $\P^3$ containing a line $L$. This $X$ is a K3 surface, and $L$ is a smooth rational curve on $X$ with self-intersection $-2$. For a general such $X$, the effective cone of $X$ is spanned by an elliptic curve $E$ and $L$. (Indeed, one can choose $E$ so that $E\cup L$ is a hyperplane section of $X$. See section 5 of [@Ko] for details.)
The closure of the nef cone of $X$ is therefore spanned by the divisors $E$ and $D=2E+3L$. The former corresponds to a morphism from $X$ to $\P^1$ giving $X$ the structure of an elliptic surface, and $D$ corresponds to the contraction of the $-2$ curve $L$. Let $P$ be a point on $L$. Then clearly any sequence of best $D$-approximation to $P$ lies on $L$, and by Theorem \[ratapprox\], a sequence of best $(E+L)$-approximation to $P$ also lies along $L$. Therefore, for any ample divisor $A$ in the positive span of $D$ and $E+L$, a sequence of best $A$-approximation to $P$ must lie along $L$, by Theorem \[divsum\].
[Tsch]{}
Beauville, A., [*Complex Algebraic Surfaces*]{}, Cambridge University Press, 1996.
Drozd, E., “Curves on a nonsingular Del Pezzo Surface in $P^4_k$”, preprint, arXiv.org/math.AG/0410518, 2004.
Hartshorne, R., [*Algebraic Geometry*]{}, Springer Verlag, New York, 1977.
, Hosoh, T., “Automorphism groups of quartic del Pezzo surfaces”, J. Algebra [**185**]{} (1996), 374-389.
Kovács, S., “The cone of curves of a K3 surface”, Math. Annalen 300 (1994), no. 4, 681-691.
Manin, Yu., [*Cubic Forms*]{} (trans. M. Hazewinkel), Elsevier Science Publishers, 1986.
McKinnon, David, “Counting Rational Points on Ruled Varieties”, Canad. Math. Bull., [**47**]{} (2004), no. 2, 264-270.
Tschinkel, Yu., “Fujita’s Program and Rational Points”, in [*Higher Dimensional Varieties and Rational Points*]{}, Springer-Verlag, 2003.
Vojta, P., [*Diophantine Approximations and Value Distribution Theory*]{}, Springer Lecture Notes in Mathematics, 1239, Springer-Verlag, 1987.
[^1]: This research supported in part by a generous grant from NSERC
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In this paper, we study the asymptotics of the discrete Chebyshev polynomials $t_{n}(z,N)$ as the degree grows to infinity. Global asymptotic formulas are obtained as $n\rightarrow\infty$, when the ratio of the parameters $n/N=c$ is a constant in the interval $(0,1)$. Our method is based on a modified version of the Riemann-Hilbert approach first introduced by Deift and Zhou.'
author:
- |
Y. Lin$^{\,a,b}$ and R. Wong$^{\,b}$\
\
title: Global Asymptotics of the Discrete Chebyshev Polynomials
---
Introduction
============
The discrete Chebyshev polynomials were introduced by Chebyshev in the study of least-squares data fitting. They are explicitly given by $$t_{n}(z,N)
=
n!\Delta^{n} \left\{ {z \choose n} {z-N \choose n}
\right\},\qquad n=0,1,\cdots,N-1
,
\label{t}$$ where $\Delta$ is the forward difference operator with unit spacing on the variable $z$; see [@Szego (2.8.1)]. These polynomials are orthogonal on the discrete set $\{ 0,1,\cdots,N-1 \}$ with respect to the weight function $w(k)=1$, $k=0,1\cdots N-1$. More precisely, we have $$\sum_{k=0}^{N-1}t_{n}(k,N)t_{m}(k,N)
=
\frac{ (N+n)! }{ (2n+1)(N-n-1)!}
\delta_{n,m}
,
\qquad
n,m=0,1,\cdots,N-1
.
$$ Like all other classical orthogonal polynomials, these polynomials also satisfy a three-term recurrence relation. Indeed, we have $$(n+1)t_{n+1}(z,N)
=
2(2n+1)(z-\tfrac12 (N-1) )
t_{n}(z,N)
-
n(N^2-n^2)
t_{n-1}(z,N)
;
\label{recurrence relations}$$ see Gautschi [@Gautschi].\
The discrete Chebyshev polynomials are also known as a special case of the Hahn polynomials [@BealsWong p.174] $$Q_n(z;\a,\b,N):={}_3F_2(-n,-z,n+\a+\b+1;-N,\a+1;1);$$ see also [@KarlinMcgregor]. Recall the weight function of the Hahn polynomials $$w(z;\a,\b,N):=
\frac{ (\a+1)_z }{ z! } \frac{ (\b+1)_{N-1-z} }{ (N-1-z)! },
\qquad
z=0,1,\cdots,N-1.$$ Note that by taking $\a=\b=0$, the weight function becomes 1. In this case, the Hahn polynomials reduce to the discrete Chebyshev polynomials except for a constant factor.
For further properties and applications, we refer to Szegö [@Szego] and Hildebrand [@Hildebrand]. Like other discrete orthogonal polynomials, these polynomials do not satisfy a second-order linear differential equation. Hence, the asymptotic theory for differential equations can not be applied. By using the steepest descent method for double integrals, Pan and Wong [@PanWong] have recently derived infinite asymptotic expansions for the discrete Chebyshev polynomials when the variable $z$ is real and the ratio $n/N$ lies in the interval $(0,1)$; their results involve a confluent hypergeometric function and its derivative. For more information about the integral methods, we refer to Wong [@WongBook]. On the other hand, Baik et al. [@BaikBook] have studied the asymptotics of discrete orthogonal polynomials with respect to a general weight function by using the Riemann-Hilbert approach. The results in [@BaikBook] are very general, but it is difficult to use them to write out explicit formulas for specific polynomials. In the case of the Hahn polynomials (1.4), only the situation when both parameters $\a$ and $\b$ are large has been considered. More precisely, the authors of [@BaikBook] considered the case when $\a=NA$ and $\b=NB$, where $A$ and $B$ are fixed positive numbers, thus excluding the special case of the discrete Chebyshev polynomials (i.e., $\a=\b=0$). Moreover, the results in [@BaikBook] are more local in nature; that is, more formulas are needed to describe the behavior of the orthogonal polynomials in the complex plane.
In this paper, we investigate the asymptotics of the discrete Chebyshev polynomials as $n\rightarrow \infty$, when the ratio of the parameters $n/N$ is a constant $c\in(0,1)$. Global asymptotic formulas are obtained in the complex $z$-plane when $n$ goes to infinity. Our approach is based on a modified version of the steepest-descent method for Riemann-Hilbert problems introduced by Deift and Zhou in [@Deift]; see [@DaiWong], [@OuWong] and [@X.S.Wang]. More precisely, we derive two Airy-type asymptotic formulas for the discrete Chebyshev polynomials. The regions of validity of these formulas overlap, and together they cover the whole complex plane.
The presentation of this paper is arranged as follows: In Section 2, we first present the basic Riemann-Hilbert problem associated with the discrete Chebyshev polynomials. Then, we transform the basic Riemann-Hilbert problem into a continuous one, which is similar to that in [@DaiWong; @X.S.Wang]. In Section 3, we introduce some auxiliary functions, known as the $g$-function and the $\phi$-function. Moreover, we define a function $D(z)$, which is analogous to the function first used in [@X.S.Wang] to remove the jumps of the continuous Riemann-Hilbert problem near the endpoints of the interval of orthogonality. In Section 4, we construct our parametrix by using Airy functions and their derivatives, and prove that this parametrix is asymptotically equal to the solution of the Riemann-Hilbert problem associated with the discrete Chebyshev polynomials formulated in Section 2. Our main result is given in Section 5. In Section 6, we compare our formulas with those given by Pan and Wong [@PanWong].
Riemann-Hilbert Problem
=======================
We start with the fundamental Riemann-Hilbert problem (RHP) for the discrete orthogonal polynomials. Note that the leading coefficient of $t_{n}(z,N)$ is $(2n)!/n!^2$. Hence, the monic discrete Chebyshev polynomials are given by $$\pi_{N,n}(z):=\frac{n!^2}{(2n)!} t_{n}(z,N)
.
\label{pi_definition}$$ Clearly, they satisfy the new orthogonality relation $$\sum_{k=0}^{N-1}\pi_{N,n}(k)\pi_{N,m}(k)
=
\delta_{n,m}/\gamma_{N,n}^2
,$$ where $$\gamma_{N,n}^2
=
\frac{ (2n+1)\Gamma(N-n)\Gamma^2(2n+1) }{ \Gamma(N+n+1)\Gamma^4(n+1) }
.$$ Consider the following RHP for a $2\times2$ matrix-value function $Y(z)$:
1. $Y(z)$ is analytic for $z\in \mathbb{C}\setminus \mathbb{N}$;
2. at each $z=k\in \mathbb{N}$, the first column of $Y$ is analytic and the second column of $Y$ has a simple pole where the residue is $$\Res\limits_{z=k} Y(z)=\lim_{z\rightarrow k} Y(z)
\begin{pmatrix} 0 & 1 \\ 0 & 0 \end{pmatrix};$$
3. as $z\rightarrow \infty$, $$Y(z)
=
\left( I + O\left(\frac1z \right) \right)
\begin{pmatrix}
z^n & 0\\
0 & z^{-n}
\end{pmatrix}.$$
By a well-known theorem of Fokas, Its and Kitaev [@FokasIts] (see also Baik et al. [@BaikBook]), we have
The unique solution to the above RHP is given by $$Y(z)
:=
\begin{pmatrix}
\pi_{N,n}(z) & \sum\limits_{k=0}^{N-1}\dfrac{ \pi_{N,n}(k) }{z-k}\\
\\
\gamma_{N,n-1}^2\pi_{N,n-1}(z) & \sum\limits_{k=0}^{N-1}
\dfrac{\gamma_{N,n-1}^2\pi_{N,n-1}(k) }{ z-k }
\end{pmatrix}
.
\label{Y}$$
Let $X_N$ denote the set defined by $$X_N:=\{x_{N,k} \}_{k=0}^{N-1}, \qquad \text{where\ \ } x_{N,k}:=\frac{k+1/2}{N}.$$ The $x_{N,k}$’s are called *nodes* and they all lie in the interval $(0,1)$. Following Baik et al. [@BaikBook], we introduce the first transformation $$\begin{aligned}
H(z)
&:=
\begin{pmatrix}
N^{-n} & 0 \\
0 & N^n
\end{pmatrix}
\, Y(Nz-1/2)
\begin{pmatrix}
\prod\limits_{j=0}^{N-1}(z-x_{N,j})^{-1} & 0 \\
0 & \prod\limits_{j=0}^{N-1}(z-x_{N,j})
\end{pmatrix}
\nonumber
\\
&\ =
N^{-n\sigma_3}\, Y(Nz-1/2)\left[
\prod\limits_{j=0}^{N-1}(z-x_{N,j}) \right]^{-\sigma_3},
\label{H}\end{aligned}$$ where $\sigma_3:=\begin{pmatrix} 1 & 0\\ 0 & -1 \end{pmatrix}$ is a Pauli matrix.\
\
A straightforward calculation shows that $H(z)$ is a solution of the RHP:
1. $H(z)$ is analytic for $z\in \mathbb{C}\setminus X_N$;
2. at each $x_{N,k}\in X_N$, the second column of $H(z)$ is analytic and the first column of $H(z)$ has a simple pole where the residue is $$\Res\limits_{z=x_{N,k}} H(z)=\lim_{z\rightarrow
x_{N,k}} H(z)
\begin{pmatrix} 0 & 0\ \\ \prod\limits^{N-1}_{j=0\atop j\neq k}(z-x_{N,j} )^{-2} & 0\ \end{pmatrix};$$
3. as $z\rightarrow \infty$, $$H(z)=\left( I + O\left(\frac1z\right) \right)
\begin{pmatrix}
z^{n-N} & 0\\
0 & z^{-n+N}
\end{pmatrix}.$$
We next consider the second transformation which will remove the poles as well as transform the discrete RHP into a continuous one. Let $\delta>0$ be a sufficiently small number, and we define $$R(z)
:= H(z)
\begin{pmatrix}
1 & \ 0 \ \ \\\\
\dfrac{\mp ie^{ \pm N \pi iz } \cos(N\pi z) }{ N\pi \prod_{j=0}^{N-1}( z-x_{N,j} )^2 } & \ 1 \
\end{pmatrix}
\label{R_H-a}$$ for $z\in\Omega_{\pm}$, and $$R(z):=H(z)
\label{R_H}$$ for $z\notin\Omega_{\pm}$, where the domains $\Omega_{\pm}$ are shown in Figure \[Figure\_R\]. Let $\Sigma_+$ be the boundary of $\Omega_+$ in the upper half-plane, and $\Sigma_-$ be the mirror image of $\Sigma_+$ in the lower half-plane. For the contour $\Sigma=(0,1)\cup\Sigma_\pm$, see also Figure \[Figure\_R\].
![The domains $\Omega_{\pm}$ and the contour $\Sigma$.[]{data-label="Figure_R"}](h2r_revised2)
For each $x_{N,k}\in X_N$, the singularity of $R(z)$ at $x_{N,k}$ is removable; that is, $\Res\limits_{z=x_{N,k}}R(z)=0$.
For $z\in \Omega_{\pm}$, it follows from (\[R\_H-a\]) that $$R_{11}(z)=H_{11}(z)+H_{12}(z)
\dfrac{\mp ie^{ \pm N \pi iz } \cos(N\pi z) }{ N\pi \prod_{j=0}^{N-1}( z-x_{N,j} )^2 }
,
\qquad
R_{12}(z)=H_{12}(z)
.
\label{R_11}$$ Since $H_{12}(z)$ is analytic by ($H_b$), the residue of $R_{12}(z)$ at $x_{N,k}$ is zero. From ($H_b$), we observe that the residue of $H_{11}(z)$ at $z=x_{N,k}$ is $$\Res\limits_{z=x_{N,k}}H_{11}(z)=H_{12}(x_{N,k})\prod\limits^{N-1}_{j=0\atop j\neq k}(x_{N,k}-x_{N,j} )^{-2}.
\label{H_11}$$ On the other hand, it is readily seen that $$\Res\limits_{z=x_{N,k}}
\dfrac{\mp ie^{ \pm N \pi iz } \cos(N\pi z) }{ N\pi \prod_{j=0}^{N-1}( z-x_{N,j} )^2 }
=
-\prod\limits^{N-1}_{j=0\atop j\neq k}(x_{N,k}-x_{N,j} )^{-2}.
\label{H_12}$$ Thus, applying (\[H\_11\]) and (\[H\_12\]) to (\[R\_11\]) shows that the residue of $R_{11}(z)$ at $z=x_{N,k}$ is zero. The entries in the second row of the matrix $R(z)$ can be studied similarly. This completes the proof of the lemma.
Note that this transformation makes $R_+(x)$ and $R_-(x)$ continuous on the interval (0,1). Therefore, $R(z)$ is a solution of the following RHP:
1. $R(z)$ is analytic for $z\in \mathbb{C}\setminus \Sigma$;
2. the jump conditions on the contour $\Sigma$: for $x\in (0,1)$, $$R_+(x)
=
R_-(x)
\begin{pmatrix}
1 & 0\\
r(x) & 1
\end{pmatrix}
,
\label{R_real}$$ where $$r(x)= \frac{4\cos^2(N\pi x)}{ W
\prod\limits_{j=0}^{N-1}(x-x_{N,j})^2 };
\label{r}$$ for $z\in \Sigma_{\pm}$, $$R_+(z)=R_-(z)
\begin{pmatrix}
1 & 0\\
\widetilde{r}_{\pm}(z) & 1
\end{pmatrix},
$$ where $$\widetilde{r}_{\pm}(z)=\frac{\mp 2 e^{ \pm N\pi iz }
\cos(N\pi z )}{ W \prod\limits_{j=0}^{N-1}(
z-x_{N,j} )^2 } ;
\label{rtilde}$$
3. as $z\rightarrow \infty$, $$R(z)=\left( I + O\left(\frac1z\right) \right)
\begin{pmatrix}
z^{n-N} & 0\\
0 & z^{-n+N}
\end{pmatrix}.
\label{R_inf}$$
For simplicity in the following section, here we have introduced the notation $W:=2N\pi i$.
The Auxiliary Functions
=======================
For our subsequent analysis, we need some auxiliary functions. Consider the monic orthogonal polynomials $p_{N,n}(z):=N^{-n}\pi_{N,n}(Nz-1/2)$. From (\[recurrence relations\]) and (\[pi\_definition\]), we have $$p_{N,n+1}(z)=(z-a_n)p_{N,n}(z)-b_{n}p_{N,n-1}(z),$$ where $$a_n=\frac12, \qquad \qquad b_n=
\frac{ 1-(n/N)^2 }{ 16-4/n^2 }.$$
Using the method introduced by Kuijlaars and Van Assche [@Arno], we can derive the equilibrium measure corresponding to the discrete Chebyshev polynomials in our case. Recall $c:=n/N$, and let $$a:=\frac12 -\frac12 \sqrt{1-c^2},
\qquad \qquad
b:=\frac12+\frac12 \sqrt{1-c^2}.
\label{MRS}$$ The density function is given by $$\mu(x)=\left\{\begin{array}{ll}
\dfrac{2}{\pi c} \arcsin( \dfrac{c}{2\sqrt{ x-x^2 } } ), & x\in[a,b],\\
\\
\dfrac1c, & x\in[0,a] \cup [b,1].
\end{array}\right.
\label{density function}$$
As usual, we now define the auxiliary $g$-function and $\phi$-function.
The $g$-function is the logarithmic potential defined by $$g(z):=\int_{0}^{1} \log(z-s) \mu(s)\,ds, \qquad \qquad z\in \mathbb{C} \setminus (-\infty,1],
\label{g}$$ and the so-called $\phi$-function is defined by $$\phi(z):=\frac l 2 -g(z)
\label{phi}$$ for $z\in \mathbb{C} \setminus (-\infty,1]$, where $l:=2\int_0^1\log|b-s|\mu(s) ds$ is called the Lagrange multiplier.
An explicit formula for the $g$-function is given in (\[gexplicit\]) in Section 6. On account of (\[density function\]), the derivative of $g(z)$ is given by $$g'(z)
=
\frac2c \log\left( \frac{z+\sqrt{ (z-a)(z-b) }+ c/2 }{ (z-1)+\sqrt{ (z-a)(z-b) } -c/2 } \right)
+
\frac1c \log \left( \frac{ z-1 }{ z } \right)
.
\label{g1}$$ Introduce the ${\phi}^*$-function $$\phi^*(z)
:=
\int_b^{z} \left(-g'(s)\mp \frac{1}{c} \pi i \right) ds
=
\phi(z)\pm\frac1c \pi i (1-z)
\label{phistar}$$ for $z\in \mathbb{C}_{\pm}$. It is readily seen that $$\begin{aligned}
{\phi}^*(z)
&=
-\int_b^{z}
\left(\frac2c \log\left( \frac{s+\sqrt{ (s-a)(s-b) }+c/2 }{ 1-s-\sqrt{ (s-a)(s-b) } +c/2 } \right)
+
\frac1c \log \left( \frac{ 1-s }{ s } \right) \right)
ds
\label{phistar_integral}\end{aligned}$$ for $z\in \mathbb{C}\backslash (-\infty,b] \cup [1,\infty)$.
Similarly, we also set $$\begin{aligned}
\widetilde{\phi}(z)
&:=
\int_a^{z} \left(-g'(s)\mp \frac{1}{c} \pi i \right) ds
=
\phi\pm \pi i (1-\frac{1}{c}z )
\label{phitilde}\end{aligned}$$ for $z\in \mathbb{C}_{\pm}$. Note that for $x\in(0,a)$, we have from (\[g\]) $$g_{\pm}(x)
=
\int_0^1 \log|x-s|\mu(s)ds \pm \pi i \int_x^1 \mu(s)ds.$$ Thus, from (\[phitilde\]) and (\[phi\]), it follows that $\pt_+(x)=\pt_-(x)$ for $x\in(0,a)$; that is, $\pt(z)$ can be analytically continued to the interval $(0,a)$.
The functions $\widetilde{\phi}(z)$ and $\phi^*(z)$ play an important role in our argument, and the following are some of their properties.
For $x\in (a,b)$, we have $$\phi^*_+(x)+\phi^*_-(x)=0
,\qquad
\widetilde\phi_+(x)+\widetilde\phi_-(x)=0
.
\label{phi_jump_ab}$$ Furthermore, we have $$\phi^*(x)<0
\qquad
\text{for } x\in (b,1)
\qquad
\text{and}
\qquad
\widetilde\phi(x)<0
\qquad
\text{for } x\in (0,a)
.
\label{phi_less_0}$$ For any $x\in(b,1)$ and sufficiently small $\delta>0$, we have $$\text{\upshape Re\,}\phi^*(x\pm i\delta )
=
\phi^*(x)+O(\delta^2)
,\qquad
\text{\upshape Re\,}\phi(x\pm i\delta )
=
\phi^*(x)- \frac1c\pi \delta +O(\delta^2)
.
\label{phistar_re}$$ For any $x\in(0,a)$ and sufficiently small $\delta>0$, we have $$\text{\upshape Re\,}\widetilde\phi(x\pm i\delta )
=
\widetilde\phi(x)+O(\delta^2)
,\qquad
\text{\upshape Re\,}\phi(x\pm i\delta )
=
\widetilde\phi(x)- \frac1c\pi \delta +O(\delta^2)
.
\label{phitilde_re}$$
Since $$\phi^*_{\pm}(x)
=
-\int_b^{x}
\left(
\frac2c \log\left( \frac{s\pm i\sqrt{ (s-a)(b-s) }+c/2 }{ 1-s\mp i\sqrt{ (s-a)(b-s) } +c/2 } \right)
+
\frac1c \log \left( \frac{ 1-s }{ s } \right)
\right)
ds$$ for $x\in(a,b)$, it is easy to verify that $$\phi^*_+(x)+\phi^*_-(x)
=
-\int_b^{x}
\left(
\frac2c \log\left( \frac{s+sc }{ (1-s)(1+c)} \right)
+
\frac2c \log \left( \frac{ 1-s }{ s } \right)
\right)
ds
=
0
.
\label{e315}$$ Similarly, we also have $\widetilde\phi_+(x)+\widetilde\phi_-(x)=0$ for $x\in(a,b)$, thus proving (\[phi\_jump\_ab\]). Together with (\[phi\]) and (\[phistar\]), this implies that $\int_0^1 \log|x-s|\mu(s) ds \equiv l/2$ for $x\in(a,b)$.
For any small $\varepsilon>0$ and $z\in U(b,\varepsilon):=\{z\in \mathbb{C}: |z-b|<\varepsilon \}$, we have from (\[phistar\_integral\]) $$\begin{aligned}
\phi^*(z)
&=
-\int_b^{z}
\left[
\frac2c
\left(
\frac{ \sqrt{ (b-a)(s-b) } }{ b+c/2 } +
\frac{ \sqrt{ (b-a)(s-b) } }{ 1-b+c/2 }
\right)
+O(\varepsilon)
\right]
ds
\nonumber
\\
&=
-\frac{8 }{3c^2 } (1-c^2)^{1/4} (z-b)^{3/2}
+
O(\varepsilon^2)
.\end{aligned}$$ Here, we have used the fact that $a+b=1$, $b-a=(1-c^2)^{1/2}$ and $ab=c^2/4$; see (\[MRS\]). By straightforward calculation, it is readily seen from that $$\begin{aligned}
{\phi^*}'(x)
&=
-
\frac2c \log\left( \frac{x+\sqrt{ (x-a)(x-b) }+c/2 }{ 1-x-\sqrt{ (x-a)(x-b) } +c/2 } \right)
-
\frac1c \log \left( \frac{ 1-x }{ x } \right)
\notag
\\
&=
-\frac1c \log
\left[
\left(
\frac{ x+\sqrt{ (x-a)(x-b) }+c/2 }
{ 1-x-\sqrt{ (x-a)(x-b) } +c/2 }
\right)^2
\left( \frac{1-x}{x} \right)
\right]
<0
\label{e317}\end{aligned}$$ for $b<x<1$. Thus, we obtain $\phi^*(x)<\phi(b+\varepsilon)<0$.
In the same manner, if $z\in U(a,\varepsilon):=\{z\in \mathbb{C}: |z-a|<\varepsilon \}$, then $$\begin{aligned}
\widetilde\phi(z)
&=
\int_a^{z}
\left[
\frac2c
\left(
\frac{ \sqrt{ (a-s)(b-a) } }{ a+c/2 }
+
\frac{ \sqrt{ (a-s)(b-a) } }{ 1-a+c/2 }
\right)
+O(\varepsilon)
\right]
ds
\nonumber
\\
&=
-\frac{8 }{3c^2 } (1-c^2)^{1/4} (a-z)^{3/2}
+
O(\varepsilon^2)
\label{phitilde_experssion}\end{aligned}$$ and $$\begin{aligned}
\widetilde\phi '(x)
&=
-
\frac2c \log\left( \frac{x-\sqrt{ (a-x)(b-x) }+c/2 }{ 1-x+\sqrt{ (a-x)(b-x) } +c/2 } \right)
-
\frac1c \log \left( \frac{ 1-x }{ x } \right)
\notag
\\
&=
-\frac1c \log
\left[
\left(
\frac{ x+\sqrt{ (x-a)(x-b) }+c/2 }
{ 1-x-\sqrt{ (x-a)(x-b) } +c/2 }
\right)^2
\left( \frac{1-x}{x} \right)
\right]
>0
\label{e319}\end{aligned}$$ for $0<x<a$. Consequently, $\widetilde\phi(x)<\widetilde\phi(a-\varepsilon)<0$ for $0<x<a$, thus proving (\[phi\_less\_0\]).
For any $x\in(b,1)$ and sufficiently small $\delta>0$, we have from the Taylor expansion $$\begin{aligned}
\text{\upshape Re\,}\phi^*(x\pm i\delta )
&=
\text{\upshape Re\,}\phi^*(x )
\mp
\delta\,\text{\upshape Im\,}{\phi^*}'(x )+O(\delta^2)
=
\text{\upshape Re\,}\phi^*(x )+O(\delta^2)
\label{phistar_re1}
.\end{aligned}$$ The last equality follows from the fact that $\po{}'(x)$ is real; see (\[e317\]). Coupling this with (\[phistar\]) gives $$\text{\upshape Re\,}\phi(x\pm i\delta )
=
\text{\upshape Re\,}\phi^*(x )
-
\frac1c\pi\delta +O(\delta^2)
;$$ thus (\[phistar\_re\]) holds.
Similarly, for any $x\in(0,a)$ and sufficiently small $\delta>0$, we have $$\begin{aligned}
\text{\upshape Re\,}\widetilde\phi(x\pm i\delta )
&=
\text{\upshape Re\,}\widetilde\phi(x )
\mp
\delta\,\text{\upshape Im\,}\widetilde\phi'(x )+O(\delta^2)
=
\text{\upshape Re\,}\widetilde\phi(x )+O(\delta^2)
.
\label{phitilde_re1}\end{aligned}$$ From (\[phitilde\]), it follows that $$\text{\upshape Re\,}\phi(x\pm i\delta )
=
\text{\upshape Re\,}\widetilde\phi(x )
-
\frac1c\pi\delta +O(\delta^2)
.$$ This ends the proof of (\[phitilde\_re\]).
![The upper half-plane under the transformations $\phi$, $\widetilde\phi$ and $\phi^*$.[]{data-label="Figure_phi"}](original2_map "fig:")\
(a) $z$-plane
![The upper half-plane under the transformations $\phi$, $\widetilde\phi$ and $\phi^*$.[]{data-label="Figure_phi"}](phistar_map2 "fig:")\
(b) $\po$-plane
\
\
\
![The upper half-plane under the transformations $\phi$, $\widetilde\phi$ and $\phi^*$.[]{data-label="Figure_phi"}](phitilde_map2 "fig:")\
(c) $\pt$-plane
![The upper half-plane under the transformations $\phi$, $\widetilde\phi$ and $\phi^*$.[]{data-label="Figure_phi"}](phi_map "fig:")\
(d) $\phi$-plane
From the definition in (\[phitilde\]), it is easy to see that $\widetilde{\phi}(a)=0$ and $\widetilde{\phi}_+(b)=(1-\frac1c)\pi i $. Furthermore, from (\[phistar\_integral\]) it follows that $\po(b)=0$, and from (\[phistar\]) and (\[phitilde\]) we have $\po_+(a)= (\frac1c -1)\pi i$. Let us now give a brief outline of the argument used in establishing the following mapping properties of the $\phi$-function.
The images of the upper half of the $z$-plane under the mappings $\phi^*$, $\pt$ and $\phi$ are depicted in Figure \[Figure\_phi\].
For $x\in(b,1)$, $\po(x)$ is real and negative in view of (\[phi\_less\_0\]). Furthermore, (\[e317\]) implies that $\po(x)$ is a monotonically decreasing function in $(b,1)$. Since $\po(b)=0$, $\po(1)$ is negative. For $x\in(a,b)$, we have from (\[phistar\]), (\[phi\]) and (\[g\]) $$\po_+(x)
=
\frac{l}{2}
-\int_0^1 \log|x-s| \mu(s) ds
+\pi i \int_x^1 \left(\frac1c -\mu(s) \right)ds.$$ On account of a statement following (\[e315\]), the first two terms on the right-hand side cancel. Hence, we obtain $$\po_+(x)=\pi i \int_x^1 \( \frac1c-\mu(s) \) ds$$ for $x\in(a,b)$. By (\[density function\]), $\frac1c > \mu(s)$ for $s\in(0,1)$. Thus, Re$\, \po_+(x)=0$ and Im$\, \po_+(x)>0$ for $x\in(a,b)$. Furthermore, since $(\text{Im}\, \po_+(x) )'=-\pi (\frac1c -\mu(x))<0$, Im$\, \po_+(x)$ is monotonically decreasing in $(a,b)$. By a similar argument, it can be shown that for $x\in(1,+\infty)$, Im$\, \po_+(x)$ is negative and decreasing; also Re$\, \po_+(x)<$ Re$\, \po_+(1)$ and Re$\, \po_+(x)$ is decreasing. Again, by a similar argument, it can be shown that for $x\in(0,a)$, Re$\, \po_+(0)=\,$Re$\, \po_+(1)$, Im$\, \po_+(0)=\,$Im$\, \po_+(a)=\pi(\frac1c -1)>0$ and Im$\, \po_+(x)=\pi(\frac1c -1)=\,$Im$\, \po_+(a)$. Finally, one can show that for $x\in(-\infty,0)$, Im$\,\po_+(x)>\pi(\frac1c-1)>0$, Im$\, \po_+(x)$ is decreasing and Re$\, \po_+(x)$ is increasing.
Coupling (\[phistar\]) and (\[phitilde\]), we have $$\pt(z)=\po(z)\pm \pi i (1-\frac1c),$$ from which it is easy to verify the image of the upper half-plane under the mapping $\pt$ shown in Figure 2c.
To verify the graph in Figure 2d, we first show that for $x\in(-\infty, 0)$, Re$\, \phi_+(x)<\,$Re$\,\phi_+(0)<0$ and Im$\, \phi_+(x)=-\pi$. Then, we show that for $x\in(0,1)$, Im$\, \phi_+(x)<0$ and Im$\, \phi_+(x)$ is an increasing function. To deal with Re$\,\phi_+(x)$ in $(0,1)$, we consider three intervals $(0,a)$, $(a,b)$ and $(b,1)$, separately. For $x\in(b,1)$, we show that $(\text{Re}\,\phi_+(x))'=(\po(x))'<0$. Thus, Re$\, \phi_+(x)$ is a decreasing function. For $x\in(a,b)$, we show that Re$\, \phi_+(x)=0$. For $x\in(0,a)$, we first show Re$\, \phi_+(x)=\pt(x)$. By (\[phi\_less\_0\]) and (\[e319\]), Re$\, \phi_+(x)<0$ and $(\text{Re}\,\phi_+(x))'=(\pt(x))'>0$. Thus, Re$\, \phi_+(x)$ is an increasing function. This ends the verification of Figure 2d, and completes the proof of the proposition.
To construct our global parametrix, we first define the function $$N(z)
=
\begin{pmatrix}
\
\dfrac{ \sqrt{z-a}+\sqrt{z-b} }{ 2(z-a)^{1/4}(z-b)^{1/4} } &
-i\dfrac{ \sqrt{z-a}-\sqrt{z-b} }{ 2(z-a)^{1/4}(z-b)^{1/4} } \ \\\\
\ i\dfrac{ \sqrt{z-a}-\sqrt{z-b} }{ 2(z-a)^{1/4}(z-b)^{1/4} } &
\dfrac{ \sqrt{z-a}+\sqrt{z-b} }{ 2(z-a)^{1/4}(z-b)^{1/4} }\
\end{pmatrix}.
\label{N}$$ It is readily verified that $N(z)$ is analytic in $\mathbb{C} \setminus [a,b]$ and $$\begin{aligned}
N_+(x)
=
N_-(x)
\begin{pmatrix}
0 & -1 \\
1 & 0
\end{pmatrix},
\qquad \qquad x\in(a,b).
\label{N_jump}\end{aligned}$$
Next, we introduce the Airy parametrix $$A(z)
:=
\left\{
\begin{array}{ll}
\begin{pmatrix}
\text{Ai}(z) & \omega^2 \text{Ai}(\omega^2 z) \\
i \text{Ai}'(z) & i \omega \text{Ai}'(\omega^2 z)
\end{pmatrix}
& z\in \mathbb{C}_+;\\\\
\begin{pmatrix}
\text{Ai}(z) & -\omega \text{Ai}(\omega z) \\
i \text{Ai}'(z) & -i \omega^2 \text{Ai}'(\omega z)
\end{pmatrix}
& z\in \mathbb{C}_- ,
\end{array}
\right.
\label{A}$$ where $\om=e^{2\pi i/3}$. In view of the well-known identity [@Handbook (9.2.12)] $$\begin{aligned}
\Ai(z)+\om\Ai(\om z)+\om^2\Ai(\om^2z)=0
,
\label{Ai_relation}\end{aligned}$$ it is clear that $$A_+(z)
=
A_-(z)
\begin{pmatrix}
1 & -1 \\
0 & 1
\end{pmatrix},
\qquad x\in\mathbb{R}.
\label{A_jump}$$ By (\[Ai\_relation\]), we also have $$\begin{aligned}
\label{A2}
\A(z)\begin{pmatrix}
1&0 \\
\pm1&1
\end{pmatrix}
=
\begin{cases}
\begin{pmatrix}
-\om\Ai(\om z)&\om^2\Ai(\om^2z) \\
-i\om^2\Ai'(\om z)&i\om\Ai'(\om^2z)
\end{pmatrix} &z\in\mathbb{C}_+;\\
\\
\left(\begin{matrix}
-\om^2\Ai(\om^2z)&-\om\Ai(\om z) \\
-i\om\Ai'(\om^2z)&-i\om^2\Ai'(\om z)
\end{matrix}\right)&z\in\mathbb{C}_-.
\end{cases}\end{aligned}$$ Recall the asymptotic expansions of the Airy function and its derivative [@Handbook 9.7(ii)] $$\text{Ai}(z)
\sim
\frac{z^{-1/4}}{ 2\sqrt{\pi} } e^{ -\frac23 z^{3/2} }
\sum_{s=0}^{\infty} \frac{ (-1)^s u_s }{ (\frac23 z^{3/2})^s },
\ \ \ \
\text{Ai}'(z)
\sim
- \frac{ z^{1/4} }{ 2\sqrt{\pi} } e^{- \frac23 z^{3/2}}
\sum_{s=0 }^{\infty } \frac{ (-1)^s v_s }{ (\frac23 z^{3/2} )^s }
\label{Ai_asy}$$ as $z\rightarrow \infty$ in $|\text{arg }z |<\pi $, where $u_s, v_s$ are constants with $u_0=v_0=1$. From (\[A\]) and (\[Ai\_asy\]), we obtain $$\A(z)
=
\frac{ z^{ -\sigma_3/4 } }{ 2\sqrt{\pi} }
\begin{pmatrix}
1 & -i \\
-i & 1
\end{pmatrix}
(I+ O(|z|^{-\frac32})) e^{-\frac23 z^{3/2} \sigma_3}\label{A_asy}$$ as $z\rightarrow \infty$ in $|\text{arg }z |<\pi $, $$\A(z)
\begin{pmatrix}
1 & 0 \\
\pm 1 & 1
\end{pmatrix}
=
\frac{ z^{ -\sigma_3/4 } }{ 2\sqrt{\pi} }
\begin{pmatrix}
1 & -i \\
-i & 1
\end{pmatrix}
(I+ O(|z|^{-\frac32})) e^{-\frac23 z^{3/2} \sigma_3} \label{A_asy-other}$$ as $z\rightarrow \infty$ with $\text{arg\,}z\in (\pm\pi/3, \pm\pi]$. Finally, we introduce the $D$-functions: $$D(z):= \dfrac{ e^{Nz} \Gamma(Nz+1/2) }{ \sqrt{2\pi} (Nz)^{Nz} }
\label{D}$$ and $$\widetilde{D}(z)
:=
\dfrac{ \sqrt{2\pi} e^{Nz}(-Nz)^{-Nz} }{ \Gamma(-Nz+1/2) },
\label{Dtilde}$$ which were first used in [@X.S.Wang] to construct global asymptotic formulas without any cut in the complex plane. For simplicity, we also introduce the notations $$D^*(z):=D(1-z)
,\qquad \qquad
\widetilde D^*(z) := \widetilde D(1-z)
.
\label{D_star}$$ By Eulers reflection formula, we have $$\widetilde{D}(z)= 2\cos(N\pi z )e^{\pm N\pi i z} D(z)
\label{D-Dtilde}$$ for $z\in \mathbb{C}_{\pm}$. The reason why we consider $D(z)$ here is to make sure that the jump matrix $J_S(z)$ in Lemma 4.1 is asymptotically equal to the identity matrix. As $n\rightarrow \infty$, by applying Stirling’s formula to (\[D\]) and (\[Dtilde\]), we obtain $$D(z)= 1+O(1/n)
\label{D_asy}$$ for $|\text{arg\,}z|<\pi$, and $$\widetilde{D}(z)=1+O(1/n)
\label{Dtilde_asy}$$ for $|\text{arg}(-z)|<\pi$.
Construction of Parametrix
==========================
![The domains I, II, III and IV.[]{data-label="Figure_Rtilde"}](Rtilde)
First, we select an arbitrary fixed point $x_0 \in(a,b)$. Let $\Gamma$ be the line $\text{Re\,} z= x_0$; see Figure \[Figure\_Rtilde\]. To facilitate our discussion below, we divide the complex plane into four regions by using $\Gamma$ and the real line. With a similar technique as used in [@DaiWong] and [@X.S.Wang], we construct the parametrix of the RHP for $R$ in these four regions. Define $$\widetilde{R}(z):=(-1)^n \sqrt{\pi} (We^{nl} )^{\sigma_3/2} N(z)
\begin{pmatrix}
1 & i \\
i & 1
\end{pmatrix}
\widetilde{f}(z)^{-\sigma_3/4}\sigma_1 \A(\widetilde{f})\sigma_1
\left[ \frac{ 4\cos^2(N\pi z) D(z)^2 }{ W \prod\limits_{j=0}^{N-1}(
z-x_{N,j} )^2 } \right]^{\sigma_3/2}
\label{Rtilde}$$ for $z\in $ I and III, $$\begin{aligned}
\widetilde{R}(z):=
(-1)^N \sqrt{\pi} (We^{nl} )^{\sigma_3/2} N(z)
\begin{pmatrix}
1 & -i \\
-i & 1
\end{pmatrix}
&f^*(z)^{-\sigma_3/4}\sigma_2 \A(f^*)\sigma_2^{-1}
\nonumber
\\
&\qquad \times
\left[ \frac{ 4\cos^2(N\pi z) D^*(z)^2 }{ W \prod\limits_{j=0}^{N-1}(
z-x_{N,j} )^2 } \right]^{\sigma_3/2}
\label{Rstar}\end{aligned}$$ for $z\in $ II and IV, where $\widetilde{f}(z)$ and $f^*(z)$ are given by $$\widetilde{f}(z):= \left( -\frac32 n\widetilde{\phi}(z) \right)^{2/3}, \qquad \qquad
f^*(z) :=\left( -\frac32 n{\phi}^*(z) \right)^{2/3}
\label{f}$$ and $
\sigma_1:=
\begin{pmatrix}
0 & 1\\
1 & 0
\end{pmatrix},
\sigma_2:=
\begin{pmatrix}
0 & 1\\
-1 & 0
\end{pmatrix}.
$ To show that $\widetilde{R}(z)$ has the same behavior as $R(z)$ as $z\rightarrow \infty$, we only consider the case when $z\in$ I. The other cases can be handled in a similar manner. A combination of (\[A\_asy\]) and (\[f\]) gives $$\sqrt{\pi}
\begin{pmatrix}
1 & i \\
i & 1
\end{pmatrix}
\widetilde{f}(z)^{-\sigma_3/4}\sigma_1 \text{A}(\widetilde{f})\sigma_1
\sim
(-1)^ne^{-n(l/2-g) \sigma_3 }
e^{N\pi iz \sigma_3}.$$ Here, we have made use of (\[phi\]) and (\[phitilde\]). Thus, we obtain from (\[D-Dtilde\]) and (\[Rtilde\]) $$\widetilde{R}(z)
\sim
(We^{nl} )^{\sigma_3/2} N(z)e^{-n(l/2-g)\sigma_3}
\widetilde{D}^{\sigma_3}
W^{-\sigma_3/2} \prod\limits_{j=0}^{N-1}(z-x_{N,j} )^{-\sigma_3}
.$$ Also note that $\widetilde{D}(z) \sim 1$ and $\prod_{j=0}^{N-1}(z-x_{N,j} ) \sim z^N$ as $z\rightarrow \infty$. From (\[g\]) and the asymptotic behavior of $N(z)$, it follows that $$\begin{aligned}
\widetilde{R}(z)
\sim
(We^{nl} )^{\sigma_3/2} N(z)
(We^{nl} )^{-\sigma_3/2}e^{ng\sigma_3}z^{-N\sigma_3}
=
(I+O(1/z))z^{(n-N)\sigma_3}
\label{Rtilde_infty}\end{aligned}$$ as $z\rightarrow\infty$. Define the matrix $$S(z):=(W e^{nl})^{-\sigma_3/2}R(z) \widetilde{R}(z)^{-1} (We^{nl })^{\sigma_3/2}
.\label{S}$$ It is easy to see that $$S(z)=I+O(1/z)$$ as $z\rightarrow \infty$. The jump matrix of $S(z)$ is given by $$J_{S}(z):=S_-(z)^{-1}S_+(z)=
(W e^{nl})^{-\sigma_3/2} \widetilde{R}_-(z) J_{R}(z)\widetilde{R}_+(z)^{-1} (W e^{nl})^{\sigma_3/2},
\label{S_jump}$$ and the contour associated with the matrix $J_S(z)$ is $\Sigma_S=\Sigma+\Gamma+\mathbb{R}$; see Figures 1 and 3.
$J_S(z)=I+O(\frac1n) $ and $S(z) =I+O\left(\frac1n\right) $ as $n\rightarrow \infty.$
For $z\in\Gamma$ and $\pm\,$Im$z\in(0,\delta)$, we have $J_R(z)=I$. This together with (\[Rtilde\]), (\[Rstar\]) and (\[S\_jump\]) gives $$\begin{aligned}
J_S(z)
&=
(-1)^N N(z)
\left[
\sqrt{\pi}
\begin{pmatrix}
1 & -i \\
-i & 1
\end{pmatrix}
f^*(z)^{-\sigma_3/4}\sigma_2 A(f^*)\sigma_2^{-1}
\right]
(D^*(z)/D(z))^{\sigma_3}
\nonumber
\\
&\quad \times
\left[
\sqrt{\pi}
\begin{pmatrix}
1 & i \\
i & 1
\end{pmatrix}
\widetilde{f}(z)^{-\sigma_3/4}\sigma_1 A(\widetilde{f})\sigma_1
\right]^{-1} N(z)^{-1} (-1)^n\
.
\label{S_x0}\end{aligned}$$ On account of (\[f\]), we have from (\[A\_asy-other\]) $$\sqrt{\pi}
\begin{pmatrix} 1 & -i \\ -i & 1 \end{pmatrix}
f^*(z)^{-\sigma_3/4}\sigma_2 A(f^*)\sigma_2^{-1}
=
(I+O(1/n))
e^{-n{\phi}^* \sigma_3}
\begin{pmatrix}1 & \pm 1 \\0 & 1\end{pmatrix}$$ and $$\left[
\sqrt{\pi}
\begin{pmatrix}1 & i \\i & 1\end{pmatrix}
\widetilde{f}(z)^{-\sigma_3/4}\sigma_1 A(\widetilde{f})\sigma_1
\right]^{-1}
=
\begin{pmatrix}1 & \mp 1\\0 & 1\end{pmatrix}
e^{n\widetilde{\phi}\sigma_3}(I+O(1/n))$$ as $n\rightarrow \infty$. Applying the above two equations to (\[S\_x0\]) gives $$\begin{aligned}
J_S(z)
&=
N(z)
\begin{pmatrix}
\frac{D^*(z)}{D(z)} & \left(\mp\frac{D^*(z)}{D(z)}\pm\frac{D(z)}{D^*(z)} \right)e^{-2n\widetilde{\phi} }\
\\\\
0 & \frac{D(z)}{D^*(z)}
\end{pmatrix}
N(z)^{-1}(I+O(1/n))
\\
&=
I+O(1/n)
.\end{aligned}$$ Here, we have used the fact that $D(z)=1+O(1/n)$, $D^*(z)=1+O(1/n)$ and Re$\, \widetilde{\phi}> 0$; see Figure \[Figure\_phi\]c. (A corresponding drawing can be given for the image of the lower half-plane under the mapping $\pt$.)
For $z\in \Gamma$ and $\pm\,$Im$z\notin(0,\delta)$, we have from (\[A\_asy\]) and (\[f\]) $$\sqrt{\pi}
\begin{pmatrix}
1 & -i \\
-i & 1
\end{pmatrix}
f^*(z)^{-\sigma_3/4}\sigma_2 A(f^*)\sigma_2^{-1}
=
e^{-n{\phi}^*\sigma_3}(I+O(1/n))$$ and $$\left[
\sqrt{\pi}
\begin{pmatrix}
1 & i \\
i & 1
\end{pmatrix}
\widetilde{f}(z)^{-\sigma_3/4}\sigma_1 A(\widetilde{f})\sigma_1
\right]^{-1}
=
e^{n\widetilde{\phi}\,\sigma_3}(I+O(1/n)).$$ Applying the last two equations to (\[S\_x0\]) yields $$\begin{aligned}
J_S(z)
&=
(-1)^{N-n} N(z)
e^{-n{\phi}^* \sigma_3}
\left( \frac{D^*(z)}{D(z)} \right)^{\sigma_3}
e^{n\widetilde{\phi}\sigma_3}
N(z)^{-1}(I+O(1/n))
\\
&=
I+O(1/n)
.\end{aligned}$$
Let us now consider $z$ in the left half-plane of $\Gamma$. For $z=x\in (-\infty,0)$, we have $J_R(x)=I$. Coupling (\[Rtilde\]) and (\[S\_jump\]) gives $$\begin{aligned}
J_S(z)
&=
(-1)^n N(z)
\left[
\begin{pmatrix}
1 & i \\
i & 1
\end{pmatrix}
\widetilde{f}_-(z)^{-\sigma_3/4}\sigma_1 A(\widetilde{f}_-)\sigma_1
\right]
\left( D_-(z)/D_+(z) \right)^{\sigma_3}
\nonumber
\\
&\quad\times
\left[
\begin{pmatrix}
1 & i \\
i & 1
\end{pmatrix}
\widetilde{f}_+(z)^{-\sigma_3/4}\sigma_1 A(\widetilde{f}_+)\sigma_1
\right]^{-1}
N(z)^{-1}
(-1)^n
.
\label{S_0}\end{aligned}$$ Note that arg$\, \widetilde f_{\pm}(x)\in (-\pi,\pi)$ in this case; see Figure 2c. Hence, we obtain from (\[A\_asy\]) and (\[f\]) $$\begin{pmatrix}
1 & i \\
i & 1
\end{pmatrix}
\widetilde{f}_{\pm}(z)^{-\sigma_3/4}\sigma_1 A(\widetilde{f}_{\pm})\sigma_1
=
\frac{1 }{ \sqrt{ \pi} }e^{-n\widetilde{\phi}_{\pm} \sigma_3 }(I+O(1/n))
.$$ Substituting the above equation in (\[S\_0\]) yields $$\begin{aligned}
J_S(z)
&=
N(z) e^{-n\widetilde{\phi}_- \sigma_3} e^{2N \pi iz \sigma_3 } e^{ n\widetilde{\phi}_+ \sigma_3} N(z)^{-1}
(I+O(1/n))
\\
&= I+O(1/n).\end{aligned}$$ Here, we have used the facts $D_-/D_+= e^{2N \pi i z}$ and $e^{ n(\widetilde{\phi}_+-\widetilde{\phi}_-) } = e^{-2N \pi i z }$.
For $z=x\in [0,x_0 ]$, we have from (\[R\_real\]) and (\[r\]) $$J_R(x)
=
\begin{pmatrix}
1 & 0 \ \\\\
\dfrac{4\cos^2(N\pi x )}{ W
\prod\limits_{j=0}^{N-1}(x-x_{N,j})^2 } & 1 \
\end{pmatrix}.$$ Note that $\widetilde{f}(z) \in \mathbb{C}_{\mp}$ when $z\in \mathbb{C}_{\pm}$. This together with (\[Rtilde\]) and (\[S\_jump\]) gives $$\begin{aligned}
J_S(x)
&=
\left[ N(x) \begin{pmatrix} 1 & i \\ i & 1 \end{pmatrix} \widetilde{f}(x)^{-\sigma_3/4} \sigma_1 A(\widetilde{f}_-) \sigma_1 \right]
\begin{pmatrix}
1 & 0 \\
D^{-2} & 1
\end{pmatrix}
\nonumber
\\
&\quad \times
\left[ N(x) \begin{pmatrix} 1 & i \\ i & 1 \end{pmatrix} \widetilde{f}(x)^{-\sigma_3/4} \sigma_1 A(\widetilde{f}_+) \sigma_1 \right]^{-1}
$$ From (\[A\_jump\]), it follows that $$\begin{aligned}
J_S(x)
&=
\left[ N(x) \begin{pmatrix} 1 & i \\ i & 1 \end{pmatrix} \widetilde{f}(x)^{-\sigma_3/4} \sigma_1 A(\widetilde{f}_+) \sigma_1 \right]
\left[ \sigma_1 \begin{pmatrix} 1 & -1 \\ 0 & 1 \end{pmatrix} \sigma_1 \right]
\begin{pmatrix}
1 & 0 \\
D^{-2} & 1
\end{pmatrix}
\nonumber
\\
&\quad \times
\left[ N(x) \begin{pmatrix} 1 & i \\ i & 1 \end{pmatrix} \widetilde{f}(x)^{-\sigma_3/4} \sigma_1 A(\widetilde{f}_+) \sigma_1 \right]^{-1}
\nonumber
\\
&=
\left[ N(x) \begin{pmatrix} 1 & i \\ i & 1 \end{pmatrix} \widetilde{f}(x)^{-\sigma_3/4} \sigma_1 A(\widetilde{f}_+) \sigma_1 \right]
\begin{pmatrix}
1 & 0 \\
D^{-2}-1 & 1
\end{pmatrix}
\nonumber
\\
&\quad\times
\left[ N(x) \begin{pmatrix} 1 & i \\ i & 1 \end{pmatrix} \widetilde{f}(x)^{-\sigma_3/4} \sigma_1 A(\widetilde{f}_+) \sigma_1 \right]^{-1}
.
\label{S_0-x_0}\end{aligned}$$ Note that $N(z)$ and $\widetilde{f}(z)^{-1/4}$ are both discontinues on the interval $(a,x_0]$. From Figure \[Figure\_phi\]c and (\[f\]), we observe that arg$\, \widetilde{\phi}_{\pm}(x)=\mp \pi/2$ and arg$\, \widetilde{f}_{\pm}(x)=\mp \pi$ for $x\in(a,x_0]$. Thus, we have $\widetilde{f}_+(x)^{ -\sigma_3/4 }= \widetilde{f}_-(x)^{-\sigma_3/4 } e^{ i \pi \sigma_3/2 }$. By (\[N\_jump\]), it is readily seen that $$N(z)
\begin{pmatrix}
1 & i \\
i & 1
\end{pmatrix}
\widetilde{f}(z)^{-\sigma_3/4}$$ has no jump on the interval $(a,x_0]$. Applying (\[D\_asy\]) to (\[S\_0-x\_0\]) gives $J_S(x)=I+O(1/n)$.
For $z\in \Sigma_{\pm}$ and Re$\, z<x_0$, the jump matrix $J_R(z)$ is $$J_R(z)
= \begin{pmatrix}
1 & 0 \ \\
\\
\dfrac{\mp 2 e^{ \pm i\pi Nz }
\cos(N\pi z) }{ W \prod\limits_{j=0}^{N-1}(
z-x_{N,j} )^2 } & 1
\end{pmatrix};$$ see (2.12). This together with (\[Rtilde\]) gives $$\begin{aligned}
J_S(z)
&=
N(z)
\left[
\begin{pmatrix}
1 & i \\
i & 1
\end{pmatrix}
\widetilde{f}(z)^{-\sigma_3/4}\sigma_1 A(\widetilde{f})\sigma_1
\right ]
\begin{pmatrix}
1 & 0 \ \ \\
\\
\dfrac{\mp e^{\pm i\pi Nz} }{ 2\cos(N\pi z) D^2 } & 1
\end{pmatrix}
\nonumber
\\
&\quad \times
\left[
\begin{pmatrix}1 & i \\i & 1\end{pmatrix}
\widetilde{f}(z)^{-\sigma_3/4}\sigma_1 A(\widetilde{f})\sigma_1
\right]^{-1}
N^{-1}(z).
\label{S_l-a}\end{aligned}$$ From (\[A\_asy\]) and(\[f\]), we then obtain $$\begin{pmatrix}
1 & i \\
i & 1
\end{pmatrix}
\widetilde{f}(z)^{-\sigma_3/4}\sigma_1 A(\widetilde{f})\sigma_1
=
\frac{e^{-n\widetilde{\phi} \sigma_3}}{ \sqrt{\pi} }(I+O(1/n)).$$ Applying the above formula to (\[S\_l-a\]) yields $$\begin{aligned}
J_S(z)
&=
N(z)
\begin{pmatrix}
1 & 0 \ \ \\\\
\dfrac{ \mp e^{ 2n\widetilde{\phi} } e^{\pm 2i\pi Nz } }
{ 2\cos(N\pi z) e^{\pm i\pi N z}D(z)^2 } & 1
\end{pmatrix}
N(z)^{-1}
(I+O(1/n))
\\
&=
N(z)
\begin{pmatrix}
1 & 0 \ \ \\\\
\dfrac{ \mp e^{2n\phi} }{ \widetilde{D}(z)D(z) } & 1
\end{pmatrix}
N(z)^{-1}
(I+O(1/n))
.\end{aligned}$$ One can show that $\text{Re}\,\phi <0$ in this case; see Figure \[Figure\_phi\]d. Thus, we have $J_S(z)=I+O(1/n)$ as $n\rightarrow \infty$.
In a similar manner, we can prove that for $z$ in the right half-plane of $\Gamma$, the corresponding jump matrix $J_S(z)$ on $\Sigma$ and on the real line tends to the identity matrix. Hence, we have shown that $J_S(z)=I+O(1/n)$ as $n\rightarrow\infty$ on the contour $\Sigma_S$. By the main result in [@QiuWong], we conclude that $S(z)=I+O(\frac1n)$ as $n\rightarrow \infty$.
Main Results
============
Let $l$, $D(z)$ and $\widetilde{f}(z)$ be defined as in (\[phi\]), (\[D\]) and (\[f\]), respectively. We have $$\begin{aligned}
\pi_{N,n}(Nz-1/2)
&=
(-N)^n\sqrt{\pi} e^{nl/2}
\nonumber
\\
&\ \times
\Bigg\{
\left[
\sin(N\pi z) \text{\upshape Ai}(\widetilde{f}(z))+\cos(N\pi z)D(z) \text{\upshape Bi}(\widetilde{f}(z))
\right]
\widetilde{A}(z,n)
\nonumber
\\
&\qquad\ +
\left[
\sin(N\pi z) \text{\upshape Ai}'(\widetilde{f}(z))+\cos(N\pi z)D(z) \text{\upshape Bi}'(\widetilde{f}(z))
\right]
\widetilde{B}(z,n)
\Bigg\}
\label{pi_left}\end{aligned}$$ for $\, z \in$ [I and III]{}, where $$\widetilde{A}(z,n)=\frac{(z-b)^{1/4} }{ (z-a)^{1/4} }\widetilde{f}(z)^{1/4}\left[1+O(1/n)\right],
\qquad
\widetilde{B}(z,n)=\frac{ (z-a)^{1/4} }{ (z-b)^{1/4} }\widetilde{f}(z)^{-1/4}[1+O(1/n)].$$ Similarly, with $D^*(z)$ and $f^*(z)$ defined in (\[D\_star\]) and (\[f\]), $$\begin{aligned}
\pi_{N,n}(Nz-1/2)
&=
(-1)^N N^n\sqrt{\pi} e^{nl/2}
\nonumber
\\
&\ \times
\Bigg\{
\Big[
\cos(N\pi z)D^*(z) \text{\upshape Bi}(f^*(z))-\sin(N\pi z) \text{\upshape Ai}(f^*(z))
\Big]
{A}^*(z,n)
\nonumber
\\
&\qquad +
\Big[
\cos(N\pi z)D^*(z) \text{\upshape Bi}'(f^*(z))-\sin(N\pi z) \text{\upshape Ai}'(f^*(z))
\Big]
{B}^*(z,n)
\Bigg\}
\label{pi_right}\end{aligned}$$ for $\, z \in$ [II and IV]{}, where $${A}^*(z,n)=\frac{(z-a)^{1/4} }{ (z-b)^{1/4} }f^*(z)^{1/4}[1+O(1/n)],
\qquad
{B}^*(z,n)=\frac{ (z-b)^{1/4} }{ (z-a)^{1/4} }f^*(z)^{-1/4}[1+O(1/n)].$$
From the definition of $S(z)$ in (\[S\]), we have $$R(z)=(W e^{nl})^{\sigma_3/2} S(z) (We^{nl })^{-\sigma_3/2}\widetilde{R}(z).$$ For any matrix $X$, we denote its $(i,j)$ element by $X_{ij}$. The above formula then gives $$R_{11}(z)=S_{11}(z)\widetilde{R}_{11}(z)
+S_{12}(z)\widetilde{R}_{21}(z)W e^{nl}$$ and $$R_{12}(z)=S_{11}(z)\widetilde{R}_{12}(z)
+S_{12}(z)\widetilde{R}_{22}(z)W e^{nl}
.$$
First, let us restrict $z$ to the region I indicated in Figure \[Figure\_Rtilde\]. A combination of (\[N\]) , (\[A\]) and (\[Rtilde\]) gives $$R_{11}(z)
=
(-1)^n\sqrt{\pi}e^{nl/2}
\left[
-i\omega\text{Ai}(\omega \widetilde{f})
\widetilde{a}(z,n)
-i\omega^2 \text{Ai}'(\omega \widetilde{f})
\widetilde{b}(z,n)
\right]
\times
\frac{ 2\cos(N\pi z) D(z) }{ \prod_{j=0}^{N-1}(
z-x_{N,j} ) }
$$ and $$R_{12}(z)
=
(-1)^n\sqrt{\pi}e^{nl/2}
\left[
i\text{Ai}(\widetilde{f})
\widetilde{a}(z,n)
+i\text{Ai}'(\widetilde{f})
\widetilde{b}(z,n)
\right]
\times
\frac{ W \prod_{j=0}^{N-1}(
z-x_{N,j} ) }{ 2\cos(N\pi z) D(z) }
,$$ where $$\widetilde{a}(z,n)=\frac{(z-b)^{1/4} }{ (z-a)^{1/4} }\widetilde{f}(z)^{1/4}(S_{11}(z)-iS_{12}(z))$$ and $$\widetilde{b}(z,n)=\frac{ (z-a)^{1/4} }{ (z-b)^{1/4} }\widetilde{f}(z)^{-1/4}(S_{11}(z)+iS_{12}(z)).$$
From (\[R\_H-a\]) and (\[R\_H\]), we know that $H_{11}(z)$ has different expressions in different parts of region I. Let us first consider the regions I$\,\cap\, \Omega_+ $ and III$\,\cap\, \Omega_-$. For $z\in$ I$\,\,\cap\,\Omega_+ $, we have from (\[R\_H-a\]) $$\begin{aligned}
H_{11}(z)
&=
R_{11}(z)
-
\dfrac{\mp ie^{ \pm N \pi iz } \cos(N\pi z) }{ N\pi \prod_{j=0}^{N-1}( z-x_{N,j} )^2 }R_{12}(z)
\nonumber
\\
&=
(-1)^n\sqrt{\pi} e^{nl/2}\prod_{j=0}^{N-1}( z-x_{N,j} )^{-1}
\nonumber
\\
&\quad\times
\Bigg\{
\cos(N\pi z)D(z)
\left[
-2i\omega\text{Ai}(\omega\widetilde{f})
\,\widetilde{a}(z,n)
-2i\omega^2
\text{Ai}'(\omega\widetilde{f})
\,\widetilde{b}(z,n)
\right]
\nonumber
\\
&\qquad \quad
-i e^{ i\pi N z}D(z)^{-1}
\left[
\text{Ai}(\widetilde{f})
\,\widetilde{a}(z,n)
+
\text{Ai}'(\widetilde{f})
\,\widetilde{b}(z,n)
\right]
\Bigg\}
.
\label{H_1p}\end{aligned}$$ The terms in curly brackets in (\[H\_1p\]) can be rewritten as $$\begin{aligned}
\cos(N\pi z)&D(z)
\Bigg[
\left(-2i\omega\text{Ai}(\omega\widetilde{f})
-i\text{Ai}(\widetilde{f})
-i(D(z)^{-2}-1)\text{Ai}(\widetilde{f})
\right)
\,\widetilde{a}(z,n)
\nonumber
\\
&\quad\qquad +
\left(
-2i\omega^2
\text{Ai}'(\omega\widetilde{f})
-i\text{Ai}'(\widetilde{f})
-i(D(z)^{-2}-1)\text{Ai}'(\widetilde{f})
\right)
\,\widetilde{b}(z,n)
\Bigg]
\nonumber
\\
&\qquad
+\sin(N\pi z)D(z)^{-1}
\Bigg[
\text{Ai}(\widetilde{f})
\,\widetilde{a}(z,n)
+
\text{Ai}'(\widetilde{f})
\,\widetilde{b}(z,n)
\Bigg].
\label{H_112}\end{aligned}$$ Recall the well-known formula of the Airy functions [@Handbook (9.2.11)] $$\begin{aligned}
\text{Bi}(z)
=
\pm i \left[ 2 e^{\mp \pi i/3} \text{Ai}(\omega^{\pm 1} z)-\text{Ai}(z) \right]
.
\label{Bi}\end{aligned}$$ Also, note that $D(z)\sim 1$ for $z\neq 0$ and $\text{Ai}(\widetilde{f})$ and $\text{Ai}'(\widetilde{f})$ are exponentially small as $n\rightarrow \infty$ when $z$ is close to the origin. Hence, we can always neglect the terms $(D(z)^{-2} -1) \text{Ai}(\widetilde{f})$ and $(D(z)^{-2} -1) \text{Ai}'(\widetilde{f})$ in (\[H\_112\]). This together with Lemma 4.1 and (\[Bi\]) gives $$\begin{aligned}
H_{11}(z)
&=
(-1)^n\sqrt{\pi} e^{nl/2}\prod_{j=0}^{N-1}( z-x_{N,j} )^{-1}
\nonumber
\\
&\ \times
\Bigg\{
\left[
\sin(N\pi z) \text{Ai}(\widetilde{f}(z))+\cos(N\pi z)D(z) \text{Bi}(\widetilde{f}(z))
\right]
\widetilde{A}(z,n)
\nonumber
\\
&\qquad\ +
\left[
\sin(N\pi z) \text{Ai}'(\widetilde{f}(z))+\cos(N\pi z)D(z) \text{Bi}'(\widetilde{f}(z))
\right]
\widetilde{B}(z,n)
\Bigg\}
.
\label{H_1}\end{aligned}$$
Similarly, for $z\in$ III$\,\cap\,\Omega_- $ $$\begin{aligned}
H_{11}(z)
&=
(-1)^n\sqrt{\pi} e^{nl/2}\prod_{j=0}^{N-1}( z-x_{N,j} )^{-1}
\\
& \ \times
\Bigg\{
\cos(N\pi z)D(z)
\left[
2i\omega^2\text{Ai}(\omega^2\widetilde{f})
\,\widetilde{a}(z,n)
+2i\omega
\text{Ai}'(\omega^2\widetilde{f})
\,\widetilde{b}(z,n)
\right]
\nonumber
\\
& \qquad
+i e^{ -i\pi N z}D(z)^{-1}
\left[
\text{Ai}(\widetilde{f})
\,\widetilde{a}(z,n)
+
\text{Ai}'(\widetilde{f})
\,\widetilde{b}(z,n)
\right]
\Bigg\}
.\end{aligned}$$ Again, by (\[Bi\]), we get exactly the same formula given in (\[H\_1\]). Using Lemma 4.1, we can obtain the desired result (\[pi\_left\]). Now, we show that the asymptotic formula of $H_{11}(z)$ in (\[H\_1\]) holds not only for $z$ in $\Omega_\pm$, but also for $z$ in the whole I and III. From the relation between ${R}(z)$ and $H(z)$ in (\[R\_H\]), we know that $H_{11}(z)$ =$R_{11}(z)$ for $z\in$ I $ \setminus\Omega_+$. In contrast to the expansion in (\[H\_1p\]), for $z\in$ I $ \setminus\Omega_+$ the term $$(-1)^n\sqrt{\pi} e^{nl/2}
\prod_{j=0}^{N-1}(z-x_{N,j})^{-1}
\times
\Big\{
-i e^{ i\pi N z}D(z)^{-1}
\left[
\text{Ai}(\widetilde{f})
\,\widetilde{a}(z,n)
+
\text{Ai}'(\widetilde{f})
\,\widetilde{b}(z,n)
\right]
\Big\}$$ does not appear, since the quantity $e^{i\pi Nz}\Ai(\widetilde f)$ is exponentially small as $n$ goes to infinity, in comparison with the other term in (\[H\_1p\]). This suggests that the region of validity of the expansion given in (\[H\_1\]) can be extended to $z \in\text{I}\cup\text{III}$. As a consequence, (\[pi\_left\]) holds for $z \in\text{I}\cup\text{III}$.
In a similar manner, we have $$\begin{aligned}
H_{11}(z)
&=
(-1)^N\sqrt{\pi} e^{nl/2}\prod_{j=0}^{N-1}( z-x_{N,j} )^{-1}
\\
& \ \times
\Bigg\{
\cos(N\pi z)D^*(z)
\left[
2i\omega^2\text{Ai}(\omega^2f^*)
\,{a}^*(z,n)
+2i\omega
\text{Ai}'(\omega^2f^*)
\,{b}^*(z,n)
\right]
\\
& \qquad
+i e^{ i\pi N z}D^*(z)^{-1}
\left[
\text{Ai}(f^*)
\,{a}^*(z,n)
+
\text{Ai}'(f^*)
\,{b}^*(z,n)
\right]
\Bigg\},\end{aligned}$$ where $$\,{a}^*(z,n)=\frac{(z-a)^{1/4} }{ (z-b)^{1/4} }f^*(z)^{1/4}(S_{11}+iS_{12})$$ and $$\,{b}^*(z,n)=\frac{ (z-b)^{1/4} }{ (z-a)^{1/4} }f^*(z)^{-1/4}(S_{11}-iS_{12})$$ for $z\in$ II$\,\cap\, \Omega_+$. Following the same argument as given above, one can obtain (\[pi\_right\]) for $z\in$ II and IV. This completes the proof of Theorem 5.1.
Comparison with Earlier Results
===============================
To compare our formulas in Theorem 5.1 with those given in [@PanWong], we first derive from (\[pi\_left\]) two simple asymptotic formulas for $\pi_{N,n}(Nz-1/2)$ when $z$ is real and less than $a$; cf. (3.2).
Let $l$, $\pt(z)$, $D(z)$ and $\wt D(z)$ be defined as in (\[phi\]), (\[phitilde\]), (\[D\]) and (\[Dtilde\]), respectively. We have $$\begin{aligned}
\pi_{N,n}(Nz-1/2)
=
(-N)^n e^{nl/2}
&\bigg\{
e^{ -n \pt(z) }D(z)\cos(N\pi z)
\frac{ (z-a)^{1/2}+(z-b)^{1/2} }{ (z-a)^{1/4}(z-b)^{1/4} }
\left[
1+O(1/n)
\right]
\notag
\\
&\
+O(e^{ n\text{Re}\,\pt(z) })
\bigg\}
\label{apprPL}\end{aligned}$$ for $0<z\leq a-\delta<a$, and $$\begin{aligned}
\pi_{N,n}(Nz-1/2)
=
N^n e^{nl/2}
\bigg\{
e^{ -n\phi(z) }\wt D(z)
\frac{ (z-a)^{1/2}+(z-b)^{1/2} }{ 2(z-a)^{1/4}(z-b)^{1/4} }
\left[
1+O(1/n)
\right]
+O(e^{ n\text{Re}\,\pt(z) })
\bigg\}
\label{apprPS} \end{aligned}$$ for $z<0$.
From the well-known asymptotic expansions of the Airy function [@Handbook (9.7.5)-(9.7.8)], we have $$\begin{aligned}
&\Ai(\ft(z))\sim \frac{\ft(z)^{-1/4}}{2\sqrt{\pi}}
e^{ n\pt(z) }
\sum_{k=0}^{\infty}
\frac{ (-1)^k u_k }{ (-n\pt(z))^{k} }
,
&&\Ai'(\ft(z))\sim -\frac{\ft(z)^{1/4}}{2\sqrt{\pi}}
e^{ n\pt(z) }
\sum_{k=0}^{\infty}
\frac{ (-1)^k v_k }{ (-n\pt(z))^{k} }
,
\label{Aift}\end{aligned}$$ and $$\begin{aligned}
&\Bi(\ft(z))\sim \frac{\ft(z)^{-1/4}}{\sqrt{\pi}}
e^{ -n\pt(z) }
\sum_{k=0}^{\infty}
\frac{ u_k }{ (-n\pt(z))^{k} },
&&
\Bi'(\ft(z))\sim \frac{\ft(z)^{1/4}}{\sqrt{\pi}}
e^{ -n\pt(z) }
\sum_{k=0}^{\infty}
\frac{ v_k }{ (-n\pt(z))^{k} }
,
\label{Bift}\end{aligned}$$ where $\ft(z)$ is defined in (\[f\]). Note that $\pt(z)$ is analytic for real $z\in(0,a)$; see the paragraph containing (\[phitilde\]). Applying (\[Aift\])-(\[Bift\]) to (\[pi\_left\]) gives $$\begin{aligned}
\pi_{N,n}(Nz-1/2)
=
(-N)^n e^{nl/2}
&\left\{
e^{n\pt(z)}\sin(N\pi z)
\frac{ (z-b)^{1/2}-(z-a)^{1/2} }{ 2(z-a)^{1/4}(z-b)^{1/4} }\left[
1+O(1/n)
\right]
\right.
\notag
\\
&\quad +
\left.
e^{-n\pt(z)}D(z)\cos(N\pi z)
\frac{ (z-a)^{1/2}+(z-b)^{1/2} }{ (z-a)^{1/4}(z-b)^{1/4} }
\left[
1+O(1/n)
\right]
\right\}
\label{apprLd}
.\end{aligned}$$ Since $\pt(z)<0$ in this case on account of (\[phi\_less\_0\]), $e^{n\pt(z)}$ is exponentially small as $n$ goes to infinity. Thus, (\[apprPL\]) follows from (\[apprLd\]).
Note that $\pt(z)$ has a jump across the negative real axis. For real $z<0$, we obtain from (\[Aift\])-(\[Bift\]) and (\[pi\_left\]) $$\begin{aligned}
\pi_{N,n}^+(Nz-1/2)
=
(-N)^n e^{nl/2}
&\left\{
e^{n\pt_+(z)}\sin(N\pi z)
\frac{ (z-b)^{1/2}-(z-a)^{1/2} }{ 2(z-a)^{1/4}(z-b)^{1/4} }\left[
1+O(1/n)
\right]
\right.
\notag
\\
&\ \
+
\left.
e^{-n\pt_+(z)}D_+(z)\cos(N\pi z)
\frac{ (z-a)^{1/2}+(z-b)^{1/2} }{ (z-a)^{1/4}(z-b)^{1/4} }
\left[
1+O(1/n)
\right]
\right\}
\label{apprD1}\end{aligned}$$ as $n\rightarrow \infty$, where $\pi_{N,n}^+(Nz-1/2)$ denotes the limiting value of $\pi_{N,n}(Nz-1/2)$ as $z$ approaches the real line from above. Observe from Figure 2c that Re$\, \pt_+(z)<0$. Thus, it follows from (\[apprD1\]) that $$\begin{aligned}
\pi_{N,n}^+(Nz-1/2)
=
(-N)^ne^{nl/2}
&\bigg\{
e^{ - n\pt_+(z) } D_+(z) \cos(N\pi z)
\frac{ (z-a)^{1/2}+(z-b)^{1/2} }{ (z-a)^{1/4}(z-b)^{1/4} }
\left[
1+O(1/n)
\right]
\notag
\\
&\ \
+O(e^{ n\text{Re}\,\pt(z) })
\bigg\}
.
\label{apprD11}\end{aligned}$$ Similarly, one can see that $\pi_{N,n}^-(Nz-1/2)$ is given by $$\begin{aligned}
(-N)^ne^{nl/2}
\bigg\{
e^{ - n\pt_-(z) } D_-(z)\cos(N\pi z)
\frac{ (z-a)^{1/2}+(z-b)^{1/2} }{ (z-a)^{1/4}(z-b)^{1/4} }
\left[
1+O(1/n)
\right]
+O(e^{ n\text{Re}\,\pt(z) })
\bigg\}
\label{apprD12}
.\end{aligned}$$ From the definition of $\phi$ in (\[phi\]), we note that $e^{n\phi(z)}$ can be analytically extended to $(-\infty,0)$. Thus, from (\[phitilde\]) and (\[D-Dtilde\]) it follows that $$\begin{aligned}
e^{- n\pt_\pm(z) }D_\pm(z)\cos(N\pi z)
=\frac{ (-1)^n }{ 2 }e^{-n\phi(z) }\wt D(z).
\label{D11}\end{aligned}$$ In view of (\[D11\]), the two asymptotic formulas (\[apprD11\]) and (\[apprD12\]) are exactly the same. Hence, $$\begin{aligned}
\pi_{N,n}(Nz-1/2)
&=
N^n e^{nl/2}
\bigg[
e^{ -n\phi(z) }\wt D(z)
\frac{ (z-a)^{1/2}+(z-b)^{1/2} }{ 2(z-a)^{1/4}(z-b)^{1/4} }
\left[
1+O(1/n)
\right]
+O(e^{ n\text{Re}\,\pt(z) })
\bigg]
,\end{aligned}$$ for real $z<0$. This gives (\[apprPS\]).
We will now show that our asymptotic formulas for the discrete Chebyshev polynomials are the same as those given by Pan and Wong [@PanWong]. First, we introduce the notation $$\begin{aligned}
x:=Nz-1/2
.
\label{fixedx}\end{aligned}$$ Two different asymptotic approximations for $t_n(x,N+1)$ are given in [@PanWong (8.13) and (8.6)], one for $x$ negative and the other for $x$ positive. Changing $N+1$ to $N$, they read $$\begin{aligned}
t_n(x,N) &=
\frac{ (-1)^{n+1}\Gamma(n+N+1)(N-1)^x n^{-2x-2} \Gamma(x+1) }
{ \Gamma(N)\pi }
\notag
\\
&\qquad\times
\left\{
\sin(\pi x)
\left[
1+ O\(\frac1N\)
\right]+O(e^{N\eta})
\right\}
\label{PanL}\end{aligned}$$ for fixed $x>0$ and $$\begin{aligned}
t_n(x,N) &=
\frac{ (-1)^{n}\Gamma(n+N+1)(N-1)^x n^{-2x-2} }
{ \Gamma(N)\Gamma(-x) }
\left[
1+ O\(\frac1N\)
\right]
\label{PanS}\end{aligned}$$ for fixed $x<0$.
We next derive from (\[apprPL\]) and (\[apprPS\]) asymptotic formulas for $t_{n}(x,N)$ when $x$ is fixed (i.e., $z=O(1/N)$). For $x>0$, i.e., $z>0$, substituting (\[fixedx\]) in (\[D\]) gives $$\begin{aligned}
D(z)\cos(N\pi z)=\frac{ e^{Nz} \Gamma(Nz+1/2) }{ \sqrt{2\pi}(Nz)^{Nz} }\cos(N\pi z)=-\frac{ e^{x+1/2} \Gamma(x+1) }{ \sqrt{2\pi}(x+1/2)^{x+1/2} }
\sin(\pi x)
.
\label{Dx}\end{aligned}$$ Moreover, it is readily verified that $$\begin{aligned}
\frac{ (2n)! }{ n!^2 }
\sim
\frac{ 1 }{ \sqrt{\pi} }2^{2n}n^{-1/2}
\label{LGamma}\end{aligned}$$ as $n\rightarrow \infty$ and $$\begin{aligned}
\frac{ (z-a)^{1/2}+(z-b)^{1/2} }{ (z-a)^{1/4}(z-b)^{1/4} }
\sim
\frac{a^{1/2}+b^{1//2} }{ (ab)^{1/4} }
=
\frac{(a+b+2a^{1/2}b^{1/2})^{1/2} }{ (ab)^{1/4} }
=
\sqrt{2}(1+c)^{1/2}c^{-1/2}
\label{cosx}\end{aligned}$$ as $z \rightarrow 0$. Here, we have made use of the fact that $a+b=1$ and $ab=c^2/4$; see (3.2).
Note that $z\rightarrow 0$ as $n\rightarrow \infty$; a combination of (2.1), (\[Dx\])-(\[cosx\]) and (\[apprPL\]) yields $$\begin{aligned}
t_n(x,N)
\sim
(-1)^{n+1}
N^n 2^{2n}n^{-1/2}
\frac{ e^{x+1/2} \Gamma(x+1) }{ \pi (x+1/2)^{x+1/2} }
(1+c)^{1/2}c^{-1/2}\sin(\pi x)
e^{n(l/2-\pt(z)) }
\label{apprTLD}\end{aligned}$$ as $n\rightarrow \infty$. We now derive an explicit formula for $l/2-\pt(z)$; see (\[Pt\]) below. From (\[phi\]) and (\[phitilde\]), one has $$\begin{aligned}
l/2-\pt(z)
=l/2-\phi_+(z)-\pi i(1-\frac1c z)
=
\text{Re}\,g_+(z)
\label{ptexplicit}\end{aligned}$$ for $0<z\leq a-\delta$. This fact evokes us to calculate $g(z)$ first. From (\[density function\]) and (\[g\]), it is easily seen that $$\begin{aligned}
g(z)=\frac1c
\int_0^a \log(z-s)ds +\frac1c \int_b^1 \log(z-s)ds
+\int_a^b\log(z-s) \mu(s)ds
.\end{aligned}$$ By integration by parts twice, we obtain $$\begin{aligned}
\int_a^b \log(z-s) \mu(s) ds
&=\mu(s)\log(z-s) s \Big |_{a}^{b}
-z\mu(s)\log(z-s) \Big |_{a}^{b}
-\int_a^b \mu(s) ds
\notag
\\
&\qquad
+z\int_a^b \mu(s) ' \log(z-s) ds
-
\int_a^b s \log(z-s) \mu(s) ' ds
.
\label{temp1}\end{aligned}$$ By the definition of $\mu(x)$, we have $$\begin{aligned}
\mu(x)'
=
-\frac {1}{2\pi} \frac{1-2x}
{ (x-x^2) \sqrt{ (x-a)(b-x) } }
.\end{aligned}$$ Here, we have again used the relations $a+b=1$ and $ab=c^2/4$. Moreover, one can show that $$\begin{aligned}
\int_a^b \log(z-s) \mu(s) ds
&=
\bigg[
\frac1c (z-a)\log(z-a)-
\frac1c (z-b)\log(z-b)
\bigg]
-(1-\frac{2a}{c})
\notag
\\
&\quad
-\frac{ z }{ 2\pi } \int_a^b \frac{\log(z-s)}
{ s \sqrt{ (s-a)(b-s) } }ds
+\frac{ 1 }{ 2\pi } (z-1) \int_a^b \frac{\log(z-s)}
{ (1-s) \sqrt{ (s-a)(b-s) } }ds
\notag
\\
&\quad
+
\frac {1}{\pi} \int_a^b \frac{\log(z-s)}
{ \sqrt{ (s-a)(b-s) } }ds
.\end{aligned}$$ Let the integrals on the right-hand side of the equality be denoted by $I_1$, $I_2$ and $I_3$, respectively. To evaluate $I_1$, we note that from [@WangWong_Stieltjes-Wigert] we have $$\begin{aligned}
\int_a^b \frac{ \log s }{ (s-z) \sqrt{ (s-a)(b-s) } }ds
=\frac{ 2\pi }{ (z-a)^{1/2}(z-b)^{1/2} }
\log \frac{ z+\sqrt{ ab }+(z-a)^{1/2}(z-b)^{1/2} }
{ (\sqrt a +\sqrt b)z }
.
\label{wangintegral}\end{aligned}$$ Making the change of variable $s=z-x$, (\[wangintegral\]) gives $$\begin{aligned}
I_1
&=
\int_{z-a}^{z-b} \frac{\log x }
{ (z-x) \sqrt{ (z-x-a)(b-z+x) } }d(z-x)
\hh
-\frac{4\pi}{c} \log
\frac{ z+\sqrt{ (z-a)(z-b) }+c/2 }
{ (\sqrt{z-a}+\sqrt{z-b})z }.\end{aligned}$$ Similarly, we can evaluate $I_2$ and $I_3$. Thus, a straightforward calculation shows that $$\begin{aligned}
g(z)
&=
-1-2\log 2
+\frac1c (z-1)\log(z-1)-\frac1c z\log z
+(2-\frac2c)\log\left[ (z-a)^{1/2}+ ({z-b})^{1/2} \right]
\notag
\\
&\ +
\frac2c z
\log\left[ z+ (z-a)^{1/2}(z-b)^{1/2}+c/2 \right]+\frac2c (1-z)
\log
\left[ z-1+ (z-a)^{1/2}(z-b)^{1/2}-c/2 \right]
.
\label{gexplicit}\end{aligned}$$ By (\[ptexplicit\]) and (\[gexplicit\]), $l/2-\pt(z)$ can be explicitly given by $$\begin{aligned}
&-1-2\log 2
+\frac1c (z-1)\log(1-z)-\frac1c z\log z
+(2-\frac2c)\log\left[ (a-z)^{1/2}+ ({b-z})^{1/2} \right]
\notag
\\
&\quad+
\frac2c z
\log\left[ z- (a-z)^{1/2}(b-z)^{1/2}+c/2 \right]+\frac2c (1-z)
\log
\left[ 1-z+ (a-z)^{1/2}(b-z)^{1/2}+c/2 \right]
\label{Pt}
.\end{aligned}$$ Note that $n=cN$ and $z\rightarrow 0$ as $n\rightarrow \infty$. From (\[Pt\]), we have $$\begin{aligned}
l/2-\pt(z)=
\left[
-1-2\log 2 +(1+\frac1c)\log(1+c)
\right]
+z
\left[
\frac1c \log(z) -\frac2c \log c -\frac1c
\right]+O(z^2)\end{aligned}$$ and so $$\begin{aligned}
e^{n(l/2-\pt(z))}
\sim
e^{-n}2^{-2n} (1+c)^{n+N}(x+1/2)^{x+1/2}e^{-x-1/2}N^{x+1/2}n^{-2x-1}
.
\label{apprept}\end{aligned}$$ Substituting (\[apprept\]) in (\[apprTLD\]) yields $$\begin{aligned}
t_n(x,N)
\sim
\frac{ (-1)^{n+1}\Gamma(x+1)n^{-2x-2} }{ \pi } \sin(\pi x)
N^{n+x+1} (1+c)^{n+N+1/2} e^{-n}
\label{apprTL}\end{aligned}$$ which will agree with (\[PanL\]) to leading order if we can show that $$\begin{aligned}
\frac{ \Gamma(n+N+1)(N-1)^x }{ \Gamma(N) }
\sim N^{n+x+1}(1+c)^{n+N+1/2}e^{-n}.
\label{PGamma} \end{aligned}$$ The last result is a direct consequence of Stirling’s formula.
Next, we consider the case when $x$ is negative, but for a moment let’s restrict $-\infty<x<-\frac12$ (i.e., $z<0$). Inserting (\[fixedx\]) in (\[Dtilde\]), we have $$\begin{aligned}
\wt D(z)= \frac{ \sqrt{2\pi}e^{x+1/2} }{ (-x-1/2)^{x+1/2}\Gamma(-x) }
.
\label{Dtildex}\end{aligned}$$ A combination of (\[apprPS\]), (\[LGamma\])-(\[cosx\]) and (\[Dtildex\]) gives $$\begin{aligned}
t_n(x,N)
\sim
N^n 2^{2n}n^{-1/2}
\frac{ e^{x+1/2} }{ (-x-1/2)^{x+1/2} \Gamma(-x) }
(1+c)^{1/2}c^{-1/2}
e^{n(l/2-\phi(z)) }
.
\label{apprTSD}\end{aligned}$$ Since $e^{n\phi(z)}$ is analytic in the interval $(-\infty,0)$, we obtain from (\[phi\]) that $e^{n(l/2-\phi(z) )}=e^{n(l/2-\phi_+(z) )}= e^{ng_+(z)}$. On account of (\[gexplicit\]), it can be shown that $$\begin{aligned}
g_+(z)=\left[\pi i
-1-2\log 2 +(1+\frac1c)\log(1+c)
\right]
+z
\left[
\frac1c \log(-z) -\frac2c \log c -\frac1c
\right]+O(z^2)
\label{Pp}\end{aligned}$$ as $z\rightarrow 0$. Thus, letting $n\rightarrow \infty$, we have from (\[Pp\]) $$\begin{aligned}
e^{n(l/2-\phi(z))}
\sim
(-1)^n 2^{-2n}(1+c)^{n+N}e^{-n}
c^{-2x-1}N^{-x-1/2} (-x-1/2)^{x+1/2} e^{-x-1/2}
.
\label{t22}\end{aligned}$$ From (\[apprTSD\]) and (\[t22\]), we obtain $$\begin{aligned}
t_n(x,N)\sim
(-N)^n \frac{ n^{-2x-2} }{ \Gamma(-x) }
(1+c)^{n+N+1/2}N^{x+1}e^{-n}
.
\label{temp1}\end{aligned}$$ In the case $-\frac12<x<0$, formula (\[temp1\]) follows directly from (\[apprTL\]) on an appeal to the reflection formula $\Gamma(1+x)\Gamma(-x)=-\pi/\sin(\pi x)$. On the other hand, coupling (\[PanS\]) and (\[PGamma\]) gives $$\begin{aligned}
t_n(x,N)\sim
(-N)^n \frac{ n^{-2x-2} }{ \Gamma(-x) }
(1+c)^{n+N+1/2}N^{x+1}e^{-n}\end{aligned}$$ as $n\rightarrow \infty$. Therefore, our results (\[apprTL\]) and (\[temp1\]) agree with those of Pan and Wong [@PanWong] stated in (\[PanL\]) and (\[PanS\]).
[99]{}
J. Baik, T. Kriecherbauer, K. T.-R. McLaughlin and P. D. Miller, *Discrete Orthogonal Polynomials: Asymptotics and Applications*, Annals of Mathematics Studies, Vol. 164, Princeton University Press, Princeton, NJ, 2007.
R. Beals and R. Wong, *Special Functions: A Graduate Text*, Cambridge University Press, Cambridge, 2010.
D. Dai and R. Wong, *Global asymptotics of Krawtchouk polynomials - a Riemann-Hilbert approach*, Chin. Ann. Math. Ser. B **28** (2007), 1-34.
P. Deift and X. Zhou, *A steepest descent method for oscillatory Riemann-Hilbert problems, Asymptotics for the MKdV equation*, Ann. Math. **137** (1993), 295-368.
A. S. Fokas, A. R. Its and A. V. Kitaev, *The isomonodromy approach to matrix models in 2D quantum gravity*, Comm. Math. Phys. **147** (1992), 395-430.
W. Gautschi, *Orthogonal Polynomials: Computation and Approximation*, Oxford University Press, Oxford, 2004.
F. B. Hildebrand, *Introduction to Numerical Analysis*, 2nd ed., McGraw-Hill, New York, 1974.
S. Karlin and J. L. McGregor, *The Hahn polynomials, formulas and an application*, Scripta Math. **26** (1961), 33-46.
A. B. J. Kuijlaars and W. Van Assche, *The asymptotic zero distribution of orthogonal polynomials with varying recurrence coefficients*, J. Approx. Theory **99** (1999), 167-197.
F. W. J. Olver, D. W. Lozier, R. F. Boisvert and C. W. Clark, *NIST Handbook of Mathematical Functions*, Cambridge University Press, Cambridge, 2010.
Chunhua Ou and R. Wong, *The Riemann-Hilbert approach to global asymptotics of discrete orthogonal polynomials with infinite nodes*, Anal. Appl. **8** (2010), 247-286.
J. H. Pan and R. Wong, *Uniform asymptotic expansions for the discrete Chebyshev polynomials*, Stud. Appl. Math. **128** (2012), 337-384.
W.-Y. Qiu and R. Wong, *Asymptotic expansions for Riemann-Hilbert problems*, Anal. Appl. **6** (2008), 269-298.
G. Szegö, *Orthogonal Polynomials*, 4th ed., AMS Colloquium Publications, Vol. 23, Amer. Math. Soc., Providence R.I., 1975.
X. S. Wang and R. Wong, *Global asymptotics of the Meixner polynomials*, Asymptotic Analysis **75** (2011), 211-231.
Z. Wang and R. Wong, *Uniform asymptotics of the Stieltjes-Wigert polynomials via the Riemann-Hilbert approach*, J. Math. Pures Appl. **85** (2006), 698-718.
R. Wong, *Asymptotic Approximations of Integrals*, Academic Press, Boston, MA, 1989. Reprinted by SIAM, Philadelphia, PA, 2001.
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'A.C. Deshpande$^{1}$[^1], T.D. Kitching$^{1}$, V.F. Cardone$^{2,3}$, P.L. Taylor$^{1,4}$, S. Casas$^{5}$, S. Camera$^{6,7,8}$, C. Carbone$^{9,10}$, M. Kilbinger$^{5,11}$, V. Pettorino$^{5}$, Z. Sakr$^{12,13}$, D. Sapone$^{14}$, I. Tutusaus$^{13,15,16}$, N. Auricchio$^{17}$, C. Bodendorf$^{18}$, D. Bonino$^{8}$, M. Brescia$^{19}$, V. Capobianco$^{8}$, J. Carretero$^{20}$, M. Castellano$^{3}$, S. Cavuoti$^{19,21,22}$, R. Cledassou$^{23}$, G. Congedo$^{24}$, L. Conversi$^{25}$, L. Corcione$^{8}$, F. Dubath$^{26}$, S. Dusini$^{27}$, G. Fabbian$^{28}$, M. Fumana$^{10}$, B. Garilli$^{10}$, F. Grupp$^{18}$, H. Hoekstra$^{29}$, F. Hormuth$^{30}$, H. Israel$^{31}$, K. Jahnke$^{32}$, S. Kermiche$^{33}$, B. Kubik$^{34}$, M. Kunz$^{35}$, S. Ligori$^{8}$, P.B. Lilje$^{36}$, I. Lloro$^{15,16}$, E. Maiorano$^{17}$, O. Marggraf$^{37}$, R. Massey$^{38}$, S. Mei$^{39}$, M. Meneghetti$^{17}$, G. Meylan$^{40}$, L. Moscardini$^{17,41,42}$, C. Padilla$^{20}$, S. Paltani$^{26}$, F. Pasian$^{43}$, S. Pires$^{5}$, G. Polenta$^{44}$, M. Poncet$^{23}$, F. Raison$^{18}$, J. Rhodes$^{4}$, M. Roncarelli$^{17,41}$, R. Saglia$^{18}$, P. Schneider$^{37}$, A. Secroun$^{33}$, S. Serrano$^{16,45}$, G. Sirri$^{46}$, J.L. Starck$^{5}$, F. Sureau$^{5}$, A.N. Taylor$^{24}$, I. Tereno$^{47,48}$, R. Toledo-Moreo$^{49}$, L. Valenziano$^{17,46}$, Y. Wang$^{50}$, J. Zoubian$^{33}$'
bibliography:
- 'ref.bib'
date: 'Received December 16, 2019/ Accepted –'
title: '*Euclid*: On the reduced shear approximation and magnification bias for Stage IV cosmic shear experiments[^2]'
---
[Stage IV weak lensing experiments will offer more than an order of magnitude leap in precision. We must therefore ensure that our analyses remain accurate in this new era. Accordingly, previously ignored systematic effects must be addressed.]{} [In this work, we evaluate the impact of the *reduced shear approximation* and *magnification bias*, on the information obtained from the angular power spectrum. To first-order, the statistics of reduced shear, a combination of shear and convergence, are taken to be equal to those of shear. However, this approximation can induce a bias in the cosmological parameters that can no longer be neglected. A separate bias arises from the statistics of shear being altered by the preferential selection of galaxies and the dilution of their surface densities, in high-magnification regions.]{} [The corrections for these systematic effects take similar forms, allowing them to be treated together. We calculate the impact of neglecting these effects on the cosmological parameters that would be determined from *Euclid*, using cosmic shear tomography. We also demonstrate how the reduced shear correction can be calculated using a lognormal field forward modelling approach.]{} [These effects cause significant biases in $\Omega_{\rm m}$, $n_{\rm s}$, $\sigma_8$, $\Omega_{\rm DE}$, $w_0$, and $w_a$ of $-0.51\sigma$, $-0.36\sigma$, $0.37\sigma$, $1.36\sigma$, $-0.66\sigma$, and $1.21\sigma$, respectively. We then show that these lensing biases interact with another systematic: the intrinsic alignment of galaxies. Accordingly, we develop the formalism for an *intrinsic alignment-enhanced lensing bias* correction. Applying this to *Euclid*, we find that the additional terms introduced by this correction are sub-dominant.]{}
\[sec:level1\]Introduction
==========================
The constituent parts of the Lambda Cold Dark Matter ($\Lambda$CDM) model, and its extensions, are not all fully understood. In the current framework, there is no definitive explanation for the physical natures of dark matter and dark energy. Today, there are a variety of techniques available to better constrain our knowledge of the $\Lambda$CDM cosmological parameters. Cosmic shear, the distortion in the observed shapes of distant galaxies due to weak gravitational lensing by the large-scale structure of the Universe (LSS), is one such cosmological probe. By measuring this distortion over large samples of galaxies, the LSS can be explored. Given that the LSS depends on density fluctuations, and the geometry of the Universe, this measurement allows us to constrain cosmological parameters. In particular, it is a powerful tool to study dark energy [@DETFrep]. A three-dimensional, redshift-dependent, picture can be obtained using a technique known as tomography. In this technique, the observed galaxies are divided into different tomographic bins; each covering a different redshift range.
Since its debut at the turn of the millennium [@Bacon2000; @Kaiser2000; @VanWaerbeke2000; @Wittman2000; @Rhodes2000], studies of cosmic shear have evolved to the point where multiple independent surveys have carried out precision cosmology [@DESpap; @cfhtmain; @kids450]. Now, with the impending arrival of Stage IV [@DETFrep] dark energy experiments like *Euclid*[^3] [@EuclidRB], *WFIRST*[^4] [@WFIRSTpap], and LSST[^5] [@LSSTpap], we are poised for a leap in precision. For example, even a pessimistic analysis of *Euclid* weak lensing data is projected to increase precision by a factor of $\sim$25 over current surveys [@SellentinStarck19].
To ensure that the accuracy of the analysis keeps up with the increasing precision of the measurements, the impact of assumptions in the theory must be evaluated. In cosmic shear a wide range of scales are probed, so the non-linear matter power spectrum must be precisely modelled. This can be accomplished through model fitting to N-body simulations [@Smith03; @Takahashi12]. A robust understanding of how baryonic physics affects the matter power spectrum is also necessary [@RuddBaryon; @Owlsbisup]. Furthermore, spurious signals arising from intrinsic alignments (IAs) [@IA1; @IA2; @IA3] in observed galaxy shapes need to be taken into account.
Additionally, assumptions in the theoretical formalism must also be relaxed. The effects of several such extensions on a *Euclid*-like experiment have been investigated. These include: the impacts of relaxing the Limber, Hankel transform and flat-sky approximations [@limitsofshear17], of using unequal-time correlators [@UTCpap], and of making the spatially-flat Universe approximation [@Spatflatpap].
The formalism to correct for the effect of measuring reduced shear, rather than shear itself, is known [@Shapiro09; @KrauseHirataRS]. However, its impact on impending surveys has not yet been quantified. The correction to the two-point cosmic shear statistic for magnification bias is also known. While the impact of this on Stage IV experiments has been quantified in [@LSSTMagBias], the approach taken here risks underestimating the bias for surveys covering the redshift range of *Euclid*. Rather than assuming that the magnification bias at the survey’s mean redshift is representative of the bias at all covered redshifts, a tomographic approach is required. Conveniently, the magnification bias correction takes a mathematically similar form to that of reduced shear; meaning these corrections can be treated together [@RSMBcombpap]. Within this work, we calculate the bias on the predicted cosmological parameters obtained from *Euclid*, when these two effects are neglected. We further extend the existing correction formalism to include the impact of IAs, and recompute the bias for this case.
In Sect. \[sec:2\], we establish the theoretical formalism. We begin by summarising the standard, first-order, cosmic shear power spectrum calculation. We then review the basic reduced shear correction formalism of [@Shapiro09]. Following this, the correction for magnification bias is explained. Next, the theory used to account for the IAs is examined. We then combine the discussed schemes, in order to create a description of an *IA-enhanced lensing bias* correction to the cosmic shear power spectrum. We also explain how we quantify the uncertainties and biases induced in the measured cosmological parameters.
In Sect. \[sec:3\], we describe how we calculate the impact of the aforementioned corrections for *Euclid*. Our modelling assumptions and choice of fiducial cosmology are stated, and computational specifics are given.
Finally, in Sect. \[sec:4\], our results are presented, and their implications for *Euclid* are discussed. The biases and change in confidence contours of cosmological parameters, resulting from the basic reduced shear and magnification bias corrections, are presented. We also present the biases from the IA-enhanced lensing bias correction.
\[sec:2\]Theoretical formalism
==============================
Here, we first review the standard cosmic shear calculation. We then explain the corrections required to account for the reduced shear approximation, and for magnification bias. We further consider the effects of IAs, and construct an IA-enhanced lensing bias correction. The formalism for accounting for the shot noise is then stated. Our chosen framework for predicting uncertainties and biases is also detailed.
\[sec:21\]The standard cosmic shear calculation
-----------------------------------------------
When a distant galaxy is weakly lensed, the change in its observed ellipticity is proportional to the reduced shear, *g*: $$\label{eq:redshear}
g^\alpha(\boldsymbol{\theta})= \frac{\gamma^\alpha(\boldsymbol{\theta})}{1-\kappa(\boldsymbol{\theta})},$$ where $\boldsymbol{\theta}$ is the galaxy’s position on the sky, $\gamma$ is the shear, which is an anisotropic stretching that turns circular distributions of light elliptical, and $\kappa$ is the convergence, which is an isotropic change in the size of the image. The superscript, $\alpha$, encodes the fact that the spin-2 shear has two components. Since $|\gamma|$, $|\kappa|$ $\ll$ 1 for individual galaxies in weak lensing, Eq. (\[eq:redshear\]) is typically approximated to first-order as $g^\alpha(\boldsymbol{\theta}) \approx \gamma^\alpha(\boldsymbol{\theta})$. This is known as the *reduced shear approximation*.
The convergence of a source being weakly lensed by the LSS, in a tomographic redshift bin *i*, is given by: $$\label{eq:convergence}
\kappa_i(\boldsymbol{\theta})=\int_{0}^{\chi_{\rm lim}} {\rm d}\chi\:\delta[d_{\rm A}(\chi)\boldsymbol{\theta},\, \chi]\:W_i(\chi).$$ It is the projection of the density contrast of the Universe, $\delta$, over the comoving distance, $\chi$, along the line-of-sight, to the limiting comoving distance of the observed sample of sources, $\chi_{\rm lim}$. The function *$d_{\rm A}(\chi)$* accounts for the curvature of the Universe, $K$, depending on whether it is flat, open, or closed: $$\label{eq:dA}
d_{\rm A}(\chi) = \begin{cases}
|K|^{-1/2}\sin(|K|^{-1/2}\chi) & \text{\small{$K>0$ (Closed)}}\\
\chi & \text{\small{$K=0$ (Flat)}}\\
|K|^{-1/2}\sinh(|K|^{-1/2}\chi) & \text{\small{$K<0$ (Open)},}
\end{cases}$$ and *W$_{i}(\chi)$* is the lensing kernel for sources in bin *i*, with the definition $$\label{eq:Wi}
W_i(\chi) = \frac{3}{2}\Omega_{\rm m}\frac{H_0^2}{c^2}\frac{d_{\rm A}(\chi)}{a(\chi)}\int_{\chi}^{\chi_{\rm lim}}{\rm d}\chi'\:n_i(\chi')\:\frac{d_{\rm A}(\chi'-\chi)}{d_{\rm A}(\chi')}.$$ Here, $\Omega_{\rm m}$ is the dimensionless present-day matter density parameter of the Universe, $H_0$ is the Hubble constant, $c$ is the speed of light in a vacuum, $a(\chi)$ is the scale factor of the Universe, and $n_i(\chi)$ is the probability distribution of galaxies within bin $i$.
Meanwhile, the two shear components, for a bin $i$, when caused by a lensing mass distribution, can be related to the convergence in a straightforward manner in frequency space: $$\label{eq:fourier}
\widetilde{\gamma}_i^\alpha(\boldsymbol{\ell})=\frac{1}{\ell(\ell+1)}\sqrt{\frac{(\ell+2)!}{(\ell-2)!}}\:T^\alpha(\boldsymbol{\ell})\,\widetilde{\kappa}_i(\boldsymbol{\ell}),$$ where $\boldsymbol{\ell}$ is the spherical harmonic conjugate of $\boldsymbol{\theta}$. Here, the small-angle limit is used. However, we do not apply the ‘prefactor unity’ approximation [@limitsofshear17], in which the factor of $1/\ell(\ell+1)\sqrt{(\ell+2)!/(\ell-2)!}$ is taken to be one, despite the fact that the impact of making the approximation is negligible for a *Euclid*-like survey [@Kilbingerellfactors]. This is done to allow consistent comparison with the spherical-sky reduced shear and magnification bias corrections. The trigonometric weighting functions, $T^\alpha(\boldsymbol{\ell})$, of the two shear components are defined as: $$\begin{aligned}
\label{eq:Trigfunc1}
T^1(\boldsymbol{\ell}) &= \cos(2\phi_\ell),\\
\label{eq:Trigfunc2}
T^2(\boldsymbol{\ell}) &= \sin(2\phi_\ell),\end{aligned}$$ with $\phi_l$ being the angular component of vector $\boldsymbol{\ell}$ which has magnitude $\ell$. Then, for an arbitrary shear field (e.g. one estimated from data), the two shear components can be linearly combined to be represented as a curl-free $E$-mode, and a divergence-free $B$-mode: $$\begin{aligned}
\label{eq:Emode}
\widetilde{E}_i(\boldsymbol{\ell})&=\sum_\alpha T^\alpha\:\widetilde{\gamma}_i^\alpha(\boldsymbol{\ell}),\\
\label{eq:Bmode}
\widetilde{B}_i(\boldsymbol{\ell})&=\sum_\alpha \sum_\beta \varepsilon^{\alpha\beta}\,T^\alpha(\boldsymbol{\ell})\:\widetilde{\gamma}_i^\beta(\boldsymbol{\ell}),\end{aligned}$$ where $\varepsilon^{\alpha\beta}$ is the two-dimensional Levi-Civita symbol, so that $\varepsilon^{12}=-\varepsilon^{21}=1$ and $\varepsilon^{11}=\varepsilon^{22}=0$. The $B$-mode vanishes in the absence of any higher-order systematic effects. Therefore, we can then define the $E$-mode angular auto-correlation and cross-correlation spectra, $C_{\ell;ij}^{\gamma\gamma}$, as: $$\label{eq:powerspecdef}
\braket{\widetilde{E}_i(\boldsymbol{\ell})\widetilde{E}_j(\boldsymbol{\ell'})} = (2\pi)^2\,\delta_{\rm D}^2(\boldsymbol{\ell}+\boldsymbol{\ell'})\,C_{\ell;ij}^{\gamma\gamma},$$ where $\delta_{\rm D}^2$ is the two-dimensional Dirac delta function. From here, an expression is derived for $C_{\ell;ij}^{\gamma\gamma}$: $$\label{eq:Cl}
C_{\ell;ij}^{\gamma\gamma} = \frac{(\ell+2)!}{(\ell-2)!}\frac{1}{\ell^4}\int_0^{\chi_{\rm lim}}{\rm d}\chi\frac{W_i(\chi)W_j(\chi)}{d^{\,2}_{\rm A}(\chi)}P_{\delta\delta}(k, \chi),$$ where $P_{\delta\delta}(k, \chi)$ is the three-dimensional matter power spectrum. Obtaining Eq. (\[eq:Cl\]) relies on making the Limber approximation, i.e. assuming that only $\ell$-modes in the plane of the sky contribute to the lensing signal. Under the Limber approximation, $k=(\ell+1/2)/d_{\rm A}(\chi)$. In this equation, the factors of $(\ell+2)!/(\ell-2)!$ and $1/\ell^4$ come once again from the fact that the prefactor unity approximation is not used. For a comprehensive review, see [@Kilbinger15].
\[sec:22\]The reduced shear correction
--------------------------------------
We account for the effects of the reduced shear approximation by means of a second-order correction to Eq. (\[eq:Cl\]) [@Shapiro09; @KrauseHirataRS; @DodelsonRS]. This can be done by taking the Taylor expansion of Eq. (\[eq:redshear\]) around $\kappa=0$, and keeping terms up to second-order: $$\label{eq:gexpan}
g^\alpha(\boldsymbol{\theta})=\gamma^\alpha(\boldsymbol{\theta})+(\gamma^\alpha\kappa)(\boldsymbol{\theta})+\mathcal{O}(\kappa^3).$$ By substituting this expanded form of $g^\alpha$ for $\gamma^\alpha$ in Eq. (\[eq:Emode\]) and then recomputing the $E$-mode ensemble average, we obtain the original result of Eq. (\[eq:powerspecdef\]), plus a correction: $$\begin{aligned}
\label{eq:ecorr}
\delta\braket{\widetilde{E}_i(\boldsymbol{\ell})\widetilde{E}_j(\boldsymbol{\ell'})} &= (2\pi)^2\,\delta_{\rm D}^2(\boldsymbol{\ell}+\boldsymbol{\ell'})\:\delta C^{\rm RS}_{\ell;ij} \nonumber\\ &= \sum_\alpha \sum_\beta T^\alpha(\boldsymbol{\ell})T^\beta(\boldsymbol{\ell'})\braket{\widetilde{(\gamma^\alpha\kappa)}_i(\boldsymbol{\ell})\:\widetilde{\gamma}_j^\beta(\boldsymbol{\ell'})} \nonumber\\ &+ T^\alpha(\boldsymbol{\ell'})T^\beta(\boldsymbol{\ell})\braket{\widetilde{(\gamma^\alpha\kappa)}_j(\boldsymbol{\ell'})\:\widetilde{\gamma}_i^\beta(\boldsymbol{\ell})},\end{aligned}$$ where $\delta C^{\rm RS}_{\ell;ij}$ are the resulting corrections to the angular auto and cross-correlation spectra. Applying the Limber approximation once more, we obtain an expression for these: $$\begin{aligned}
\label{eq:dCl}
\delta C^{\rm RS}_{\ell;ij} &= \ell(\ell+1)\frac{(\ell+2)!}{(\ell-2)!}\frac{1}{\ell^6}\int_0^\infty\frac{{\rm d}^2\boldsymbol{\ell'}}{(2\pi)^2}\cos(2\phi_{\ell'}-2\phi_\ell) \nonumber\\
&\times B_{ij}^{\kappa\kappa\kappa}(\boldsymbol{\ell}, \boldsymbol{\ell'}, -\boldsymbol{\ell}-\boldsymbol{\ell'}).\end{aligned}$$ The factors of $\ell(\ell+1)(\ell+2)!/(\ell-2)!$ and $1/\ell^6$ arise from foregoing the three-point equivalent of the prefactor unity approximation. As in the case of Eq. (\[eq:fourier\]), the product of these factors can be well approximated by one. However, we do not make this approximation for the sake of completeness, and as the additional factors do not add any significant computational expense. $B_{ij}^{\kappa\kappa\kappa}$, is the two-redshift convergence bispectrum; which takes the form: $$\begin{aligned}
\label{eq:bispecK}
B_{ij}^{\kappa\kappa\kappa}(\boldsymbol{\ell_1}, \boldsymbol{\ell_2}, \boldsymbol{\ell_3}) &= B_{iij}^{\kappa\kappa\kappa}(\boldsymbol{\ell_1}, \boldsymbol{\ell_2}, \boldsymbol{\ell_3}) + B_{ijj}^{\kappa\kappa\kappa}(\boldsymbol{\ell_1}, \boldsymbol{\ell_2}, \boldsymbol{\ell_3})\nonumber\\ &=\int_0^{\chi_{\rm lim}}\frac{{\rm d}\chi}{d^{\,4}_{\rm A}(\chi)}W_i(\chi)W_j(\chi)[W_i(\chi)+W_j(\chi)]\nonumber\\
&\times B_{\delta\delta\delta}(\boldsymbol{k_1},\boldsymbol{k_2},\boldsymbol{k_3},\chi),\end{aligned}$$ where $B_{iij}^{\kappa\kappa\kappa}$ and $B_{ijj}^{\kappa\kappa\kappa}$ are the three-redshift bispectra, $k_x$ is the magnitude and $\phi_{\ell;x}$ is the angular component of $\boldsymbol{k_x}$ (for $x=1, 2, 3$). Under the Limber approximation, $k_x=(\ell_x+1/2)/d_{\rm A}(\chi)$. Here, we also approximate our photometric redshift bins to be infinitesimally narrow. In reality, because these bins would have a finite width, the product of lensing kernels in Eq. \[eq:bispecK\] would be replaced by a single integral over the products of the contents of the integral in Eq. \[eq:Wi\]. Accordingly, the values of the bispectrum would be slightly larger. However, given that *Euclid* will have high quality photometric redshift measurement, we expect this difference to be negligible. Consequently, in our calculations we proceed with the narrow-bin approximation; which allows us to use the same lensing kernels as used in the power spectrum calculation.
Analogous to the first-order power spectra being projections of the three-dimensional matter power spectrum, the two-dimensional convergence bispectra are a projection of the three-dimensional matter bispectrum, $B_{\delta\delta\delta}(\boldsymbol{k_1},\boldsymbol{k_2},\boldsymbol{k_3},\chi)$. The analytic form of the matter bispectrum is not well known. Instead, a semi-analytic approach starting with second-order perturbation theory [@Fryperturb], and then fitting its result to N-body simulations, is employed. We use the fitting formula of [@Scoccimarro01]. Accordingly, the matter bispectrum can be written: $$\begin{aligned}
\label{eq:bispecFF}
B_{\delta\delta\delta}(\boldsymbol{k_1},\boldsymbol{k_2},\boldsymbol{k_3},\chi) &= 2F_2^{\rm eff}\,(\boldsymbol{k_1},\boldsymbol{k_2})\,P_{\delta\delta}(k_1, \chi)P_{\delta\delta}(k_2, \chi) \nonumber\\
&+ \text{cyc.} ,\end{aligned}$$ where $F_2^{\rm eff}$ encapsulates the simulation fitting aspect, and is defined as: $$\begin{aligned}
\label{eq:Feff}
F_2^{\rm eff}(\boldsymbol{k_1},\boldsymbol{k_2}) &= \frac{5}{7}\,a(n_{\rm s},k_1)\,a(n_{\rm s},k_2) \nonumber\\
&+ \frac{1}{2}\frac{\boldsymbol{k_1}\cdot\boldsymbol{k_2}}{k_1k_2}\,\bigg(\frac{k_1}{k_2}+\frac{k_2}{k_1}\bigg)\,b(n_{\rm s},k_1)\,b(n_{\rm s},k_2) \nonumber\\
&+\frac{2}{7}\,\bigg(\frac{\boldsymbol{k_1}\cdot\boldsymbol{k_2}}{k_1k_2}\bigg)^2c(n_{\rm s},k_1)\,c(n_{\rm s},k_2),\end{aligned}$$ where $n_{\rm s}$ is the scalar spectral index, which indicates the deviation of the primordial matter power spectrum from scale invariance ($n_{\rm s}=1$), and the functions $a,b$, and $c$ are fitting functions, defined in [@Scoccimarro01].
There are no additional correction terms of form $\widetilde{E}\widetilde{B}$ or $\widetilde{B}\widetilde{B}$, and it has been shown that higher-order terms are sub-dominant [@KrauseHirataRS], so further terms in Eq. (\[eq:gexpan\]) can be neglected for now.
\[sec:22.5\]The magnification bias correction
---------------------------------------------
The observed overdensity of galaxies on the sky is affected by gravitational lensing, in two competing ways [@MBorig]. Firstly, individual galaxies can be magnified (or demagnified), which results in their flux being increased (or decreased). At the flux limit of a survey, this can cause fainter sources (which in the absence of lensing would be excluded) to be included in the observed sample. Conversely, the density of galaxies in the patch of sky around this source appears reduced (or increased) due to the patch of sky being magnified (or demagnified) similarly to the source. Accordingly, the net effect of these depends on the slope of the intrinsic, unlensed, galaxy luminosity function, at the survey’s flux limit. This net effect is known as *magnification bias*. Additionally, galaxies can also be pulled into a sample because their effective radius is increased as a consequence of magnification, such that they pass a resolution factor cut. In this work, we do not consider this effect as it is more important for ground-based surveys than space-based ones such as *Euclid*.
In the case of weak lensing, where $|\kappa| \ll 1$, and assuming that fluctuations in the intrinsic galaxy overdensity are small on the scales of interest, the observed galaxy overdensity in tomographic bin $i$ is [@MBcorssource; @MBorig]: $$\label{eq:galover}
\delta^g_{{\rm obs}; i}(\boldsymbol{\theta}) = \delta^g_i(\boldsymbol{\theta}) + (5s_i-2)\kappa_i(\boldsymbol{\theta}),$$ where $\delta^g_i(\boldsymbol{\theta})$ is the intrinsic, unlensed, galaxy overdensity in bin $i$, and $s_i$ is the slope of the cumulative galaxy number counts brighter than the survey’s limiting magnitude, $m_{\rm lim}$, in redshift bin $i$. This slope is defined as: $$\label{eq:slope}
s_i = \frac{\partial{\rm log}_{10}\,\mathfrak{n}(\bar{z_i}, m)}{\partial m}\bigg|_{m_{\rm lim}},$$ where $\mathfrak{n}(\bar{z_i}, m)$ is the true distribution of galaxies, evaluated at the central redshift of bin $i$, $\bar{z_i}$. It is important to note that, in practice, this slope is determined from observations, and accordingly depends on the wavelength band within which the galaxy is observed in addition to its redshift.
Operationally, magnification bias causes the true shear, $\gamma^\alpha_i$, to be replaced, within the estimator used to determine the power spectrum from data, by an ‘observed’ shear: $$\label{eq:MBshear}
\gamma^\alpha_{{\rm obs}; i} \xrightarrow{} \gamma^\alpha_i+\gamma^\alpha_i\delta^g_{{\rm obs}; i} = \gamma^\alpha_i+\gamma^\alpha_i\delta^g_i + (5s_i-2)\gamma^\alpha_i\kappa_i.$$ Now, we can evaluate the impact of magnification bias on the two-point statistic by substituting $\widetilde{\gamma}^\alpha_{{\rm obs}; i}$ for $\widetilde{\gamma}^\alpha_i$ in Eq. (\[eq:Emode\]), and recomputing. As source-lens clustering terms of the form $\gamma^\alpha_i\delta^g_i$ are negligible [@RSMBcombpap], we recover the standard result of Eq. (\[eq:powerspecdef\]), with an additional correction term: $$\begin{aligned}
\label{eq:MBEcorr}
\delta\langle\widetilde{E}_i(\boldsymbol{\ell})\widetilde{E}_j(\boldsymbol{\ell'})\rangle &= \sum_\alpha \sum_\beta T^\alpha(\boldsymbol{\ell})T^\beta(\boldsymbol{\ell'})(5s_i-2)\braket{\widetilde{(\gamma^\alpha\kappa)}_i(\boldsymbol{\ell})\,\widetilde{\gamma}_j^\beta(\boldsymbol{\ell'})}\nonumber\\
&+ T^\alpha(\boldsymbol{\ell'})T^\beta(\boldsymbol{\ell})(5s_j-2)\braket{\widetilde{(\gamma^\alpha\kappa)}_j(\boldsymbol{\ell'})\,\widetilde{\gamma}_i^\beta(\boldsymbol{\ell})}.\end{aligned}$$ Analogously to the reduced shear case, we then obtain corrections to the auto and cross-correlation angular spectra of the form: $$\begin{aligned}
\label{eq:dClMB}
\delta C^{\rm MB}_{\ell;ij} &= \ell(\ell+1)\frac{(\ell+2)!}{(\ell-2)!}\frac{1}{\ell^6}\int_0^\infty\frac{{\rm d}^2\boldsymbol{\ell'}}{(2\pi)^2}\cos(2\phi_{\ell'}-2\phi_\ell)\nonumber\\
&\times [(5s_i-2)B_{iij}^{\kappa\kappa\kappa}(\boldsymbol{\ell}, \boldsymbol{\ell'}, -\boldsymbol{\ell}-\boldsymbol{\ell'})\nonumber\\
&+ (5s_j-2)B_{ijj}^{\kappa\kappa\kappa}(\boldsymbol{\ell}, \boldsymbol{\ell'}, -\boldsymbol{\ell}-\boldsymbol{\ell'})].\end{aligned}$$ Note that the mathematical form of Eq. (\[eq:dClMB\]) is simply Eq. (\[eq:dCl\]) with factors of $(5s_i - 2)$ and $(5s_j - 2)$ applied to the corresponding bispectra. These additional prefactors are due to the magnification bias contribution from each bin depending on the slope of the luminosity function in that bin. Accordingly, we are able to compute both of these effects for the computational cost of one.
\[sec:23\]Intrinsic alignments
------------------------------
When galaxies form near each other, they do so in a similar tidal field. Such tidal process occurring during galaxy formation, together with other processes like spin correlations, can induce a preferred, intrinsically correlated, alignment of galaxy shapes [@IA1; @IA2; @IA3]. To first-order, this can be thought of as an additional contribution to the observed ellipticity of a galaxy, $\epsilon$: $$\label{eq:epsi1}
\epsilon = \gamma + \gamma^{\rm I} + \epsilon^s,$$ where $\gamma=\gamma^1+{\rm i}\gamma^2$ is the gravitational lensing shear, $\gamma^{\rm I}$ is the contribution to the observed shape resulting from IAs, and $\epsilon^s$ is the source ellipticity that the galaxy would have in the absence of the process causing the IA.
Using Eq. (\[eq:epsi1\]), we find that the theoretical two-point statistic (e.g. the two-point correlation function, or the power spectrum) consists of three types of terms: $\braket{\gamma\gamma},\braket{\gamma^{\rm I}\gamma}$, and $\braket{\gamma^{\rm I}\gamma^{\rm I}}$. The first of these terms leads to the standard lensing power spectra of Eq. (\[eq:Cl\]), while the other two terms lead to additional contributions to the observed power spectra, $C_{\ell;ij}^{\epsilon\epsilon}$, so that: $$\label{eq:ObsCl}
C_{\ell;ij}^{\epsilon\epsilon} = C_{\ell;ij}^{\gamma\gamma} + C_{\ell;ij}^{{\rm I}\gamma} + C_{\ell;ij}^{\rm II} + N_{\ell;ij}^\epsilon,$$ where $C_{\ell;ij}^{{\rm I}\gamma}$ represents the correlation between the background shear and the foreground IA, $C_{\ell;ij}^{\rm II}$ are the auto-correlation spectra of the IAs, and $N_{\ell;ij}^\epsilon$ is a shot noise term. The additional spectra can be described in a similar manner to the shear power spectra, by way of the ‘non-linear alignment’ (NLA) model [@IA_NLA]: $$\begin{aligned}
\label{eq:cllig}
C_{\ell;ij}^{{\rm I}\gamma} &= \frac{(\ell+2)!}{(\ell-2)!}\frac{1}{\ell^4}\int_0^{\chi_{\rm lim}}\frac{{\rm d}\chi}{d^{\,2}_{\rm A}(\chi)}[W_i(\chi)n_j(\chi)+n_i(\chi)W_j(\chi)] \nonumber\\
&\times P_{\delta {\rm I}}(k, \chi),\\
\label{eq:clli}
C_{\ell;ij}^{\rm II} &= \frac{(\ell+2)!}{(\ell-2)!}\frac{1}{\ell^4}\int_0^{\chi_{\rm lim}}\frac{{\rm d}\chi}{d^{\,2}_{\rm A}(\chi)}n_i(\chi)\,n_j(\chi)\,P_{\rm II}(k, \chi),\end{aligned}$$ where the intrinsic alignment power spectra, $P_{\delta {\rm I}}(k, \chi)$ and $P_{\rm II}(k, \chi)$, are expressed as functions of the matter power spectra: $$\begin{aligned}
\label{eq:pdi}
P_{\delta {\rm I}}(k, \chi) &= \frac{-\mathcal{A}_{\rm IA}\mathcal{C}_{\rm IA}\Omega_{\rm m}}{D(\chi)}P_{\delta\delta}(k,\chi),\\
\label{eq:pii}
P_{\rm II}(k, \chi) &= \bigg(\frac{-\mathcal{A}_{\rm IA}\mathcal{C}_{\rm IA}\Omega_{\rm m}}{D(\chi)}\bigg)^2P_{\delta\delta}(k,\chi),\end{aligned}$$ in which $\mathcal{A}_{\rm IA}$ and $\mathcal{C}_{\rm IA}$ are free model parameters to be determined by fitting to data or simulations, and $D(\chi)$ is the growth factor of density perturbations in the Universe, as a function of comoving distance.
\[sec:25\]IA-enhanced lensing bias
----------------------------------
The reduced shear approximation is also used when considering the impact of IAs, and magnification bias plays a role here too. We account for these by substituting the appropriate second-order expansions of the shear, Eq. (\[eq:gexpan\]) and Eq. (\[eq:MBshear\]), in place of $\gamma$ within Eq. (\[eq:epsi1\]). Neglecting source-lens clustering, the ellipticity now becomes: $$\label{eq:epsi2}
\epsilon \simeq \gamma + (1+5s-2)\gamma\kappa + \gamma^{\rm I} + \epsilon^s.$$ Constructing a theoretical expression for the two-point statistic from this revised expression for the ellipticity now gives us six types of terms: $\braket{\gamma\gamma}, \braket{\gamma^{\rm I}\gamma}, \braket{\gamma^{\rm I}\gamma^{\rm I}}, \braket{(\gamma\kappa)\gamma}, \braket{(\gamma\kappa)(\gamma\kappa)},$ and $\braket{(\gamma\kappa)\gamma^{\rm I}}$. The first three terms remain unchanged from the first-order case. The fourth term encompasses the basic reduced shear and magnification bias corrections, and results in the shear power spectrum corrections defined by Eq. (\[eq:dCl\]) and Eq. (\[eq:dClMB\]). The fifth of these terms can be neglected, as it is a fourth-order term. The final term creates an additional correction, $\delta C_{\ell;ij}^{\rm I}$, to the observed spectra that takes a form analogous to the basic reduced shear and magnification bias corrections: $$\begin{aligned}
\label{eq:dClIA}
\delta C_{\ell;ij}^{\rm I} &= \ell(\ell+1)\frac{(\ell+2)!}{(\ell-2)!}\frac{1}{\ell^6}\int_0^\infty\frac{{\rm d}^2\boldsymbol{\ell'}}{(2\pi)^2}\cos(2\phi_{\ell'}) \nonumber\\
&\times [(1+5s_i-2)B_{iij}^{\kappa\kappa {\rm I}}(\boldsymbol{\ell}, \boldsymbol{\ell'}, -\boldsymbol{\ell}-\boldsymbol{\ell'})\nonumber\\
&+ (1+5s_j-2)B_{jji}^{\kappa\kappa {\rm I}}(\boldsymbol{\ell}, \boldsymbol{\ell'}, -\boldsymbol{\ell}-\boldsymbol{\ell'})],\end{aligned}$$ where the convergence-IA bispectra, $B_{iij}^{\kappa\kappa {\rm I}}$ and $B_{jji}^{\kappa\kappa {\rm I}}$, are given by: $$\begin{aligned}
\label{eq:scIBi}
B_{iij}^{\kappa\kappa {\rm I}}(\boldsymbol{\ell_1}, \boldsymbol{\ell_2}, \boldsymbol{\ell_3}) &= \int_0^{\chi_{\rm lim}}\frac{{\rm d}\chi}{d^{\,4}_{\rm A}(\chi)}W^2_i(\chi)n_j(\chi)B_{\delta\delta {\rm I}}(\boldsymbol{k_1},\boldsymbol{k_2},\boldsymbol{k_3},\chi),\hspace{-0.1mm}\\
B_{jji}^{\kappa\kappa {\rm I}}(\boldsymbol{\ell_1}, \boldsymbol{\ell_2}, \boldsymbol{\ell_3}) &= \int_0^{\chi_{\rm lim}}\frac{{\rm d}\chi}{d^{\,4}_{\rm A}(\chi)}W^2_j(\chi)n_i(\chi)B_{\delta\delta {\rm I}}(\boldsymbol{k_1},\boldsymbol{k_2},\boldsymbol{k_3},\chi).\end{aligned}$$ The density perturbation-IA bispectrum, $B_{\delta\delta {\rm I}}(\boldsymbol{k_1},\boldsymbol{k_2},\boldsymbol{k_3},\chi)$, can be calculated in a similar way to the matter density perturbation bispectrum, using perturbation theory and the [@Scoccimarro01] fitting formula. Accordingly: $$\begin{aligned}
\label{eq:bispecEN}
B_{\delta\delta {\rm I}}(\boldsymbol{k_1},\boldsymbol{k_2},\boldsymbol{k_3},\chi) &= 2F_2^{\rm eff}(\boldsymbol{k_1},\boldsymbol{k_2}) P_{{\rm I}\delta}(k_1, \chi)P_{\delta\delta}(k_2, \chi) \nonumber\\
&+ 2F_2^{\rm eff}(\boldsymbol{k_2},\boldsymbol{k_3})P_{\delta\delta}(k_2, \chi)P_{\delta {\rm I}}(k_3, \chi) \nonumber\\
&+ 2F_2^{\rm eff}(\boldsymbol{k_1},\boldsymbol{k_3})P_{\delta {\rm I}}(k_1, \chi)P_{\delta\delta}(k_3, \chi),\end{aligned}$$ with $P_{{\rm I}\delta}(k_1, \chi)=P_{\delta {\rm I}}(k_1, \chi)$. This equation is an ansatz for how IAs behave in the non-linear regime, analogous to the NLA model. The described approach, and in particular the fitting functions, remain valid because, in the NLA model, we can treat IAs as a field proportional, by some redshift-dependence weighting, to the matter density contrast. Since the fitting functions, $F_2^{\rm eff}$, do not depend on the comoving distance, they remain unchanged. For the full derivation of this bispectrum term, and a generalisation for similar terms, see Appendix \[AGB\].
\[sec:24\]Shot noise
--------------------
The shot noise term in Eq. (\[eq:ObsCl\]) arises from the uncorrelated part of the unlensed source ellipticities; represented by $\epsilon^s$ in Eq. (\[eq:epsi1\]). This is zero for cross-correlation spectra, because the ellipticities of galaxies at different comoving distances should be uncorrelated. However, it is non-zero for auto-correlation spectra. It is written as: $$\label{eq:shotnoise}
N_{\ell;ij}^\epsilon = \frac{\sigma_\epsilon^2}{\bar{n}_{\rm g}/N_{\rm bin}}\delta_{ij}^{\rm K},$$ where $\sigma_\epsilon^2$ is the variance of the observed ellipticities in the galaxy sample, $\bar{n}_{\rm g}$ is the galaxy surface density of the survey, $N_{\rm bin}$ is the number of tomographic bins used, and $\delta_{ij}^{\rm K}$ is the Kronecker delta. Equation (\[eq:shotnoise\]) assumes the bins are equi-populated.
\[sec:26\]Fisher and bias formalism
-----------------------------------
To estimate the uncertainty on cosmological parameters that will be obtained from *Euclid*, we use the Fisher matrix approach [@Tegmark97; @ISTforecast]. In this formalism, the Fisher matrix is defined as the expectation of the Hessian of the likelihood: $$\label{eq:fishbasic}
F_{\tau\zeta} = \bigg\langle\frac{-\partial^2 \ln L}{\partial\theta_\tau\partial\theta_\zeta}\bigg\rangle,$$ where $L$ is the likelihood of the parameters given the data, and $\tau$ and $\zeta$ refer to parameters of interest, $\theta_\tau$ and $\theta_\zeta$. Assuming a Gaussian likelihood, the Fisher matrix can be rewritten in terms of only the covariance of the data, $\boldsymbol{C}$, and the mean of the data vector, $\boldsymbol{\mu}$: $$\label{eq:fishgauss}
F_{\tau\zeta} = \frac{1}{2}\:{\rm tr}\,\bigg[\frac{\partial\boldsymbol{C}}{\partial\theta_\tau}\boldsymbol{C}^{-1}\frac{\partial\boldsymbol{C}}{\partial\theta_\zeta}\boldsymbol{C}^{-1}\bigg] + \sum_{pq}\frac{\partial\mu_p}{\partial\theta_\tau}(\boldsymbol{C}^{-1})_{pq}\frac{\partial\mu_q}{\partial\theta_\zeta},$$ where the summations over $p$ and $q$ are summations over the variables in the data vector. In the case of cosmic shear, the mean of our data is zero, so the second term in Eq. (\[eq:fishgauss\]) vanishes.
In reality, the weak lensing likelihood is non-Gaussian (see e.g. [@nonGaussLik]). However, recent investigations indicate that the assumption of a Gaussian likelihood is unlikely to lead to significant biases in the cosmological parameters inferred from a Stage IV weak lensing experiment [@LSSTnonGauss]. Additionally, while this non-Gaussianity affects the shapes of the constraints on cosmological parameters, it will not affect the calculation of the reduced shear and magnification bias corrections, and accordingly will not significantly affect the corresponding relative biases. For these reasons, coupled with its simplicity, we proceed under the Gaussian assumption for this work.
The specific Fisher matrix used in this investigation can be expressed as: $$\label{eq:fishfin}
F_{\tau\zeta} = f_{\rm sky} \sum_{\ell=\ell_{\rm min}}^{\ell_{\rm max}} \Delta\ell\,\bigg(\ell+\frac{1}{2}\bigg)\:{\rm tr}\,\bigg[\frac{\partial\boldsymbol{C}_\ell^{\epsilon\epsilon}}{\partial\theta_\tau}({\boldsymbol{C}_\ell^{\epsilon\epsilon}})^{-1}\frac{\partial\boldsymbol{C}_\ell^{\epsilon\epsilon}}{\partial\theta_\zeta}({\boldsymbol{C}_\ell^{\epsilon\epsilon}})^{-1}\bigg],$$ where $\Delta\ell$ is the bandwidth of $\ell$-modes sampled and the sum is over the $\ell$-blocks, $f_{\rm sky}$ is the fraction of the sky surveyed, and $\boldsymbol{C}_\ell^{\epsilon\epsilon}$ is a matrix consisting of the values, at $\ell$, of the observed shear spectra for each tomographic bin combination. From this, we calculate the expected uncertainties on our parameters, $\sigma_\tau$, using the relation: $$\label{eq:sigma}
\sigma_\tau = \sqrt{({F^{-1}})_{\tau\tau}}.$$ The Fisher matrix can also be used to determine the projected confidence region ellipses of pairs of cosmological parameters [@ISTforecast].
In the presence of a systematic effect in the signal, the Fisher matrix formalism can be adapted to measure how biased the inferred cosmological parameter values will be if this systematic is not taken into consideration [@biaspap]. This bias is calculated as follows: $$\begin{aligned}
\label{eq:bias}
b(\theta_\tau) &= \sum_\zeta{(F^{-1})}_{\tau\zeta}\: f_{\rm sky}\sum_\ell \Delta\ell\,\bigg(\ell+\frac{1}{2}\bigg)\nonumber\\
&\times{\rm tr}\bigg[\delta\boldsymbol{C}_\ell\:{(\boldsymbol{C}_\ell^{\epsilon\epsilon})}^{-1}\,\frac{\partial\boldsymbol{C}_\ell^{\epsilon\epsilon}}{\partial\theta_\zeta}\,{(\boldsymbol{C}_\ell^{\epsilon\epsilon})}^{-1}\bigg],\end{aligned}$$ where $\delta\boldsymbol{C}_\ell$ is a matrix consisting of the values, at $\ell$, of the unaccounted for systematic effects, for each tomographic bin combination. In our case, these effects are the reduced shear and magnification bias corrections. Equation (\[eq:fishfin\]) and Eq. (\[eq:bias\]) are what the results in Appendix B of [@biaspap] reduce to when $\boldsymbol{C}^{\epsilon\epsilon}_\ell$ is block-diagonal in $\ell$.
\[sec:3\]Methodology
====================
In order to quantify the impact of the three corrections on *Euclid*, we adopt the forecasting specifications of [@ISTforecast]. Accordingly, we take there to be 10 equi-populated tomographic bins, with bin edges: {0.001, 0.418, 0.560, 0.678, 0.789, 0.900, 1.019, 1.155, 1.324, 1.576, 2.50}.
Cosmological Parameter Fiducial Value
------------------------ ----------------
$\Omega_{\rm m}$ 0.32
$\Omega_{\rm b}$ 0.05
$h$ 0.67
$n_{\rm s}$ 0.96
$\sigma_8$ 0.816
$\sum m_\nu$ (eV) 0.06
$\Omega_{\rm DE}$ 0.68
$w_0$ $-1$
$w_a$ 0
: Fiducial values of $w$CDM cosmological parameters for which the bias from reduced shear and magnification bias corrections is calculated. These values have been selected in accordance with *Euclid* Collaboration forecasting choices [@ISTforecast]; to facilitate consistent comparisons. Note that the value of the neutrino mass is kept fixed in the Fisher matrix calculations.[]{data-label="tab:cosmology"}
We investigate the ‘optimistic’ case for such a survey, in which $\ell$-modes of up to 5000 are probed. We consider the intrinsic variance of observed ellipticities to have two components, each with a value of 0.21, so that the intrinsic ellipticity root-mean-square value $\sigma_\epsilon = \sqrt{2}\times0.21 \approx 0.3$. For *Euclid*, we take the surface density of galaxies to be $\bar{n}_{\rm g}=30$ arcmin$^{-2}$, and the fraction of sky covered to be $f_{\rm sky}=0.36$.
Furthermore, we consider the $w$CDM model case in our calculations. This extension of the $\Lambda$CDM model accounts for a time-varying dark energy equation of state. In this model, we have the parameters: the present-day matter density parameter $\Omega_{\rm m}$, the present-day baryonic matter density parameter $\Omega_{\rm b}$, the Hubble parameter $h=H_0/100$kms$^{-1}$Mpc$^{-1}$, the spectral index $n_{\rm s}$, the RMS value of density fluctuations on 8 $h^{-1}$Mpc scales $\sigma_8$, the present-day dark energy density parameter $\Omega_{\rm DE}$, the present-day value of the dark energy equation of state $w_0$, and the high redshift value of the dark energy equation of state $w_a$. Additionally, we assume neutrinos to have masses. We denote the sum of neutrino masses by $\sum m_\nu\ne 0$. This quantity is kept fixed, and we do not generate confidence contours for it, in concordance with [@ISTforecast]. The fiducial values chosen for these parameters are given in Table \[tab:cosmology\]. These values are chosen to allow for a direct and consistent comparison of the two corrections with the forecasted precision of *Euclid*. The values provided in the forecasting specifications for the free parameters of the NLA model are also used in our work, in Eq. (\[eq:pdi\]) and Eq. (\[eq:pii\]). These are: $\mathcal{A}_{\rm IA}=1.72$ and $\mathcal{C}_{\rm IA}=0.0134$.
Model Parameter Fiducial Value
------------------ ----------------
$c_{\rm b}$ 1.0
$z_{\rm b}$ 0.0
$\sigma_{\rm b}$ 0.05
$c_{\rm o}$ 1.0
$z_{\rm o}$ 0.1
$\sigma_{\rm o}$ 0.05
$f_{\rm out}$ 0.1
: Choice of parameter values used to define the probability distribution function of the photometric redshift distribution of sources, in Eq. (\[eq:pphot\]). We do not consider how variation in the quality of photometric redshifts impacts the Fisher matrix predictions.
\[tab:phphotparams\]
As in [@ISTforecast], we choose to define the distributions of galaxies in our tomographic bins, for photometric redshift estimates, as: $$\label{eq:ncfht}
{\mathcal N}_i(z) = \frac{\int_{z_i^-}^{z_i^+}{\rm d}z_{\rm p}\,\mathfrak{n}(z)p_{\rm ph}(z_{\rm p}|z)}{\int_{z_{\rm min}}^{z_{\rm max}}{\rm d}z\int_{z_i^-}^{z_i^+}{\rm d}z_{\rm p}\,\mathfrak{n}(z)p_{\rm ph}(z_{\rm p}|z)},$$ where $z_{\rm p}$ is measured photometric redshift, $z_i^-$ and $z_i^+$ are edges of the $i$-th redshift bin, and $z_{\rm min}$ and $z_{\rm max}$ define the range of redshifts covered by the survey. Then, $n_i(\chi) = {\mathcal N}_i(z){\rm d}z/{\rm d}\chi$. In Eq. (\[eq:ncfht\]), $\mathfrak{n}(z)$ is the true distribution of galaxies with redshift, $z$; defined as in the *Euclid* Red Book [@EuclidRB]: $$\label{eq:ntrue}
\mathfrak{n}(z) \propto \bigg(\frac{z}{z_0}\bigg)^2\,{\rm exp}\bigg[-\bigg(\frac{z}{z_0}\bigg)^{3/2}\bigg],$$
where $z_0=z_{\rm m}/\sqrt{2}$, with $z_{\rm m}=0.9$ as the median redshift of the survey. Meanwhile, the function $p_{\rm ph}(z_{\rm p}|z)$ describes the probability that a galaxy at redshift $z$ is measured to have a redshift $z_{\rm p}$, and takes the parameterisation: $$\begin{aligned}
\label{eq:pphot}
p_{\rm ph}(z_{\rm p}|z) &= \frac{1-f_{\rm out}}{\sqrt{2\pi}\sigma_{\rm b}(1+z)}\,{\rm exp}\Bigg\{-\frac{1}{2}\bigg[\frac{z-c_{\rm b}z_{\rm p}-z_{\rm b}}{\sigma_{\rm b}(1+z)}\bigg]^2\Bigg\} \nonumber\\
&+ \frac{f_{\rm out}}{\sqrt{2\pi}\sigma_{\rm o}(1+z)}\:{\rm exp}\Bigg\{-\frac{1}{2}\bigg[\frac{z-c_{\rm o}z_{\rm p}-z_{\rm o}}{\sigma_{\rm o}(1+z)}\bigg]^2\Bigg\}.\end{aligned}$$ In this parameterisation, the first term describes the multiplicative and additive bias in redshift determination for the fraction of sources with a well measured redshift, whereas the second term accounts for the effect of a fraction of catastrophic outliers, $f_{\rm out}$. The values of these parameters, chosen to match the selection of [@ISTforecast], are stated in Table \[tab:phphotparams\]. By using this formalism, the impact of the photometric redshift uncertainties is also included in the derivatives, with respect to the cosmological parameters, of the shear power spectra.
The matter density power spectrum and growth factor used in our analyses are computed using the publicly available `CLASS`[^6] cosmology package [@Classpap]. Within the framework of `CLASS`, we include non-linear corrections to the matter density power spectrum, using the `Halofit` model [@Takahashi12]. Using these modelling specifics, we first calculate the basic reduced shear correction of Eq. (\[eq:dCl\]), and the resulting biases in the $w$CDM parameters. In doing so, we compute the derivatives of our tomographic matrices, at each sampled $\ell$-mode, using a simple finite-difference method. The correction for magnification bias, and the resulting biases in the cosmological parameters, are calculated in the same way.
Bin $i$ Central Redshift Slope $s_i$
--------- ------------------ -------------
1 0.2095 0.196
2 0.489 0.274
3 0.619 0.320
4 0.7335 0.365
5 0.8445 0.412
6 0.9595 0.464
7 1.087 0.525
8 1.2395 0.603
9 1.45 0.720
10 2.038 1.089
: Slope of the luminosity function for each redshift bin, calculated at the central redshifts of each bin. These are evaluated at the limiting magnitude 24.5 (AB in the *Euclid* VIS band [@VISpap]). The slopes are determined using finite difference methods with the fitting formula of [@JoachimiBridleMB], which is based on fitting to COMBO-17 and SDSS $r$-band results [@BlakeBridlenz].
\[tab:slopesbybin\]
The slope of the luminosity function, as defined in Eq. (\[eq:slope\]), is calculated for each redshift bin using the approach described in Appendix C of [@JoachimiBridleMB]. We apply a finite-difference method to the fitting formula for galaxy number density as a function of limiting magnitude stated here, in order to calculate the slope of the luminosity function at the limiting magnitude of *Euclid*, 24.5 [@EuclidRB]; or AB in the *Euclid* VIS band [@VISpap]. This technique produces slope values consistent with those generated from the Schechter function approach of [@LSSTMagBias]. The calculated slopes for each redshift bin are given in Table \[tab:slopesbybin\]. However, we emphasise that while this method allows the investigation of the impact of magnification bias at this stage, when the correction is computed for the true *Euclid* data, updated galaxy number counts determined directly from *Euclid* observations should be used to ensure accuracy.
We then combine the two corrections, and calculate the resulting biases, as well as the resulting confidence contours for parameter combinations. Next, the additional IA-lensing bias interaction term from Eq. (\[eq:dClIA\]) is included, and the biases are recomputed.
To validate the perturbative formalism based on a fitting formula for the matter bispectrum, we also compute the reduced shear correction using a forward model approach assuming the lognormal field approximation [@hilbert2011cosmic; @mancini20183d; @xavier2016improving]. This approximation was recently used to generate a covariance matrix in the Dark Energy Survey Year 1 analysis [@troxel2018dark]. Using the pipeline recently presented in [@2019arXiv190405364T] (which uses the public code [CAMB]{} [@lewis538efficient], [Halofit]{} [@Takahashi12], [Cosmosis]{} [@zuntz2015cosmosis] and the python wrapper of [HEALpix]{} [^7] [@gorski1999healpix; @2005ApJ...622..759G] – [HEALpy]{}) we compute the reduced shear correction by averaging over $100$ forward realizations. We compare our semi-analytic approach to the forward modelled approach, for the auto-correlation spectrum of a single tomographic bin spanning the entire redshift range of 0 – 2.5. To ensure a consistent comparison is made with the forward model approach, the correction computed from the perturbative formalism in this case uses the best-fitting photometric redshift galaxy distribution of the CFHTLenS catalogue [@cfhtnpap]: $$\label{eq:nfm}
\mathfrak{n}(z) = 1.5\,{\rm exp}\bigg[-\frac{(z-0.7)^2}{0.1024}\bigg] + 0.2\,{\rm exp}\bigg[-\frac{(z-1.2)^2}{0.2116}\bigg],$$ as this is used in [@2019arXiv190405364T]. In this comparison, we do not consider magnification bias, or the IA-enhanced lensing bias case.
\[sec:4\]Results and discussion
===============================
In this section, we report the impact of the various effects studied on *Euclid*. We first present the individual and combined impacts of the reduced shear and magnification bias corrections. The impact of IA-enhanced lensing bias is also discussed. Finally, we present a forward modelled approach for computing the reduced shear correction.
\[subsec:RSres\]The reduced shear correction
--------------------------------------------
The relative magnitude of the basic reduced shear correction described by Eq. (\[eq:dCl\]), to the observed shear auto-correlation spectra (excluding shot noise), at various redshifts, is shown in Fig. \[fig:dcl\_comp\_nonIA\]. The correction increases with $\ell$, and becomes particularly pronounced at scales above $\ell\sim 100$. This is expected, as small-scale modes grow faster in high-density regions, where the convergence tends to be larger, so there is more power in these regions. We can also see, from Fig. \[fig:dcl\_comp\_nonIA\], that the relative magnitude of the correction increases with redshift, as the reduced shear correction has an extra factor of the lensing kernel, $W_i(\chi)$, in comparison to the angular shear spectra. The lensing kernel increases with comoving distance and, accordingly, redshift. While only a selection of auto-correlation spectra are presented in Fig. \[fig:dcl\_comp\_nonIA\] for illustration purposes, the remaining auto and cross-correlation spectra exhibit the same trends. The uncertainties on the $w$CDM cosmological parameters that are predicted for *Euclid*, are stated in Table \[tab:unctab\]. Correspondingly, Table \[tab:nonIAbiastab\] shows the biases that are induced in the predicted cosmological parameters from neglecting the basic reduced shear correction.
Biases are typically considered acceptable when the biased and unbiased confidence regions have an overlap of at least 90$\%$; corresponding to the magnitude of the bias being $\leq$ $0.25\sigma$ [@BiasAceptPap]. The majority of the biases are not significant, with $\Omega_{\rm m}$, $\Omega_{\rm b}$, $h$, $n_{\rm s}$, and $\sigma_8$ remaining strongly consistent pre and post-correction. However, the three dark energy parameters, $\Omega_{\rm DE}$, $w_0$, and $w_a$, all exhibit significant biases of $0.31\sigma$, $-0.32\sigma$, and $0.40\sigma$, respectively. Since one of the chief goals of upcoming weak lensing surveys is the inference of dark energy parameters, these biases, which can be readily dealt with, indicate that the reduced shear correction must be included when constraining cosmological parameters from the surveys. Also shown in Table \[tab:unctab\] is the change in the uncertainty itself, when the reduced shear correction and its derivatives are included in the Fisher matrix used for prediction. In general, the change is negligible, because the reduced shear correction and its derivatives are relatively small in comparison to the shear spectra and derivatives.
\[subsec:MBres\]The magnification bias correction
-------------------------------------------------
\[fig:Comps\_dcl\]
Figure \[fig:dcl\_comp\_MB\] shows the magnitude of the basic magnification bias correction, relative to the shear auto-correlation spectra (again excluding shot noise) for the same redshift bins as in Fig. \[fig:dcl\_comp\_nonIA\]. In this case, the relative magnitude of the correction again increases with redshift. However, in the two lowest redshift bins shown, the correction is subtractive. This is the case for the five lowest redshift bins, of the ten that we consider. This is due to the dilution of galaxy density dominating over the magnification of individual galaxies, as there are fewer intrinsically fainter galaxies at lower redshifts. Conversely, at higher redshifts, there are more fainter sources which lie on the threshold of the survey’s magnitude cut, that are then magnified to be included in the sample.
The change in the uncertainty of the cosmological parameters if magnification bias is corrected for, and the bias in these parameters if magnification bias is neglected, are given in Table \[tab:unctab\] and Table \[tab:nonIAbiastab\], respectively. In the absence of any corrections, there are near-exact degeneracies which result in large uncertainties when the Fisher matrix is inverted. However, because we are dealing with near-zero eigenvalues in the Fisher matrix, even subtle changes to the models that encode information can significantly change the resulting parameter constraints. Accordingly, correcting for the magnification bias has a noticeable effect on the uncertainties of the parameters $\Omega_{\rm b}$, $h$, $n_{\rm s}$, $\Omega_{\rm DE}$, $w_0$, and $w_a$.
Since the magnification bias correction depends on the observed density of baryonic matter, including it improves the constraint on $\Omega_{\rm b}$. Also, the predicted uncertainties on $h$ are also reduced, as the correction term has an additional factor of the lensing kernel relative to the angular power spectrum; increasing sensitivity to $h$ by a power of two. The fitting formulae used to describe the matter bispectrum, as part of the correction term, also have a non-trivial dependence on $n_{\rm s}$. This means that the sensitivity to $n_{\rm s}$ is also increased, when the magnification bias correction is made.
{width="\textwidth"}
On the other hand, the uncertainty on $\Omega_{\rm DE}$ worsens upon correcting for magnification bias. This stems from the fact that the derivative of the correction term with respect to $\Omega_{\rm DE}$ is negative, as a higher dark energy density results in a Universe that has experienced a greater rate of expansion, and accordingly is more sparsely populated with matter. Then, the effect of magnification bias is lower, and the magnitude of the correction drops. Therefore, the magnitude of the magnification bias correction and the strength of the $\Omega_{\rm DE}$ signal are inversely correlated. This means that in the case where the magnification bias correction is made, $\Omega_{\rm DE}$ is less well constrained than in the case where there is no correction.
Conversely, increasing $w_0$ and $w_a$ decreases the rate of expansion of the Universe, and so sensitivity to $w_0$ and $w_a$ increases in the case when the correction is made. We note, however, that the changes in uncertainty induced by the inclusion of this correction will likely be dwarfed by those resulting from the combination of *Euclid* weak lensing data with other probes; both internal and external. For example, the combination of weak lensing with other *Euclid* probes alone, such as photometric and spectroscopic galaxy clustering, and the cross-correlation between weak lensing and photometric galaxy clustering, will significantly improve parameter constraints [@ISTforecast].
------------------- -------------------------- -------------------------- -------------------------- --------------------------
Cosmological W/o Correction With RS Correction With MB Correction With RS+MB Corrections
Parameter Uncertainty (1-$\sigma$) Uncertainty (1-$\sigma$) Uncertainty (1-$\sigma$) Uncertainty (1-$\sigma$)
$\Omega_{\rm m}$ 0.012 0.012 0.012 0.012
$\Omega_{\rm b}$ 0.021 0.021 0.018 0.018
$h$ 0.13 0.13 0.085 0.090
$n_{\rm s}$ 0.031 0.032 0.018 0.020
$\sigma_8$ 0.011 0.011 0.011 0.011
$\Omega_{\rm DE}$ 0.050 0.051 0.056 0.059
$w_0$ 0.16 0.16 0.15 0.15
$w_a$ 0.94 0.94 0.85 0.86
------------------- -------------------------- -------------------------- -------------------------- --------------------------
------------------- ------------------------- ------------------------- ------------------------- -------------------------
Cosmological Basic RS Correction Basic MB Correction Combined RS+MB IA-enhanced Correction
Parameter Cosmology Bias/$\sigma$ Cosmology Bias/$\sigma$ Cosmology Bias/$\sigma$ Cosmology Bias/$\sigma$
$\Omega_{\rm m}$ $-0.10$ $-0.41$ $-0.51$ $-0.58$
$\Omega_{\rm b}$ 0.023 $-0.19$ $-0.19$ $-0.22$
$h$ 0.072 $-0.013$ 0.059 0.019
$n_{\rm s}$ $-0.10$ $-0.26$ $-0.36$ $-0.30$
$\sigma_8$ 0.055 0.35 0.37 0.48
$\Omega_{\rm DE}$ 0.31 1.05 1.36 1.33
$w_0$ $-0.32$ $-0.34$ $-0.66$ $-0.64$
$w_a$ 0.40 0.81 1.21 1.14
------------------- ------------------------- ------------------------- ------------------------- -------------------------
If magnification bias is not corrected for, the values determined for the parameters $\Omega_{\rm m}$, $n_{\rm s}$, $\sigma_8$, $\Omega_{\rm DE}$, $w_0$, and $w_a$ are significantly biased at $-0.41\sigma$, $-0.26\sigma$, $0.35\sigma$, $1.05\sigma$, $-0.34\sigma$, and $0.81\sigma$, respectively. All of these biases are higher than the corresponding bias from making the reduced shear approximation. Given that all but two cosmological parameters are significantly biased if magnification bias is neglected, this correction is necessary for *Euclid*.
\[subsec:combres\]The combined correction
-----------------------------------------
The relative magnitude of the combined reduced shear and magnification bias correction is shown in Fig. \[fig:dcl\_comp\_MBRS\]. At the lowest redshifts considered, the subtractive magnification bias correction essentially cancels out the reduced shear correction. Then, at intermediate redshifts, the magnification bias is additive and comparable to the reduced shear correction. However, the dominant part of combined corrections is found at the highest redshifts, where the magnification bias correction is particularly strong. Therefore, the combined correction term is predominantly additive across the survey’s redshift bins.
The effects of the combined corrections, on the predicted cosmological parameter constraints, are stated in Table \[tab:unctab\] and shown in Fig. \[fig:bias\_nonIA\]. The constraints largely remain affected as they were by just the magnification bias correction. The constraints on $h$ and $n_{\rm s}$ worsen slightly when the two corrections are considered together, due to their differing behaviour at lower redshifts. The uncertainty on $\Omega_{\rm DE}$ also increases further. Additionally, Fig. \[fig:bias\_nonIA\] and Table \[tab:nonIAbiastab\] show the biases induced in the cosmological parameters if these corrections are neglected. As expected, the biases add together linearly, and their severity emphasises the need for these two corrections to be applied to the angular power spectra that will be obtained from *Euclid*. Furthermore, the combination of weak lensing with other probes will improve parameter constraints, whilst leaving the biases resulting from reduced shear and magnification bias unchanged; meaning that the relative biases in this scenario will be even larger. This further stresses the importance of these corrections.
\[subsec:IAEb\]The IA-enhanced lensing bias correction
------------------------------------------------------
When the IA-lensing bias interaction term, from Eq. (\[eq:dClIA\]), is also accounted for, the biases are minimally altered. These are displayed in Table \[tab:nonIAbiastab\]. From these, we see that the additional term, while non-trivial, does not induce significant biases in the cosmological parameters obtained at our current level of precision.
The nature of this additional correction, and its relatively minor impact, is explained by Fig. \[fig:dcl\_comp\_IA\]. This charts the change with $\ell$ and redshift, of the two components of the IA-enhanced lensing bias, $\delta C^{\rm RS+MB}_{\ell;ij}$ and $\delta C^{\rm I}_{\ell;ij}$. From this, we see that for the lowest redshift bins, the two already small terms cancel each other out and at higher redshifts, the latter term is evidently sub-dominant. Accordingly, while upcoming surveys must make the basic reduced shear and magnification bias corrections to extract accurate information, the IA-enhanced correction is not strictly necessary.
![The two components of the IA-enhanced lensing bias correction, Eq. (\[eq:dCl\]) and Eq. (\[eq:dClIA\]), for the cross-spectra of our first (0.001$\leq z \leq$0.418), and tenth bins (1.576$\leq z \leq$2.50). For the first bin, the basic correction is already sub-dominant, and the additional IA-enhanced terms cancels it out. In the higher redshift bin, the second term is sub-dominant. This trend persists across all bins.[]{data-label="fig:dcl_comp_IA"}](dcl_comp_IA){width="1.0\linewidth"}
\[subsec:FMres\]Forward modelling comparison
--------------------------------------------
Figure \[fig:compare\] compares the reduced shear corrections obtained from the perturbative and forward modelling approaches, for a singular tomographic bin spanning the entire probed redshift range of 0 – 2.5. There is remarkable agreement between the two within the range of $\ell$-modes that will be observed by *Euclid*, particularly at the highest and lowest $\ell$-modes. We see minor disagreements at intermediate $\ell$-modes, however, this is unsurprising given the various different approximations and assumptions made in the two techniques. We also note that at $\ell$-modes beyond the survey’s limit, the lognormal approach will under-predict the perturbative solution.
![The reduced shear correction using the bispectrum perturbative approach (see Sect. \[sec:22\]) and using the forward model in the lognormal field approximation presented in [@2019arXiv190405364T]. The data points are plotted at the geometric mean of the $\ell$-bin boundaries. There is mild disagreement at intermediate $\ell$-mode. This is to be expected given the approximations that go into the bispectrum fitting formula and the lognormal field approximation. Nevertheless the agreement at low-$\ell$ and in the highest $\ell$-bin are striking. Here $C_\ell^{gg}$ labels the reduced shear auto-correlation spectrum, while $C_\ell^{\gamma\gamma}$ denotes the shear-shear auto-correlation spectrum.[]{data-label="fig:compare"}](frwd_bispec_compare.pdf){width="\linewidth"}
Performing cosmological inference on full forward models of the data using density-estimation likelihood-free inference (DELFI) [@alsing2018generalized; @alsingndes] to compute the posteriors on cosmological parameters is emerging as a new paradigm in cosmic shear analyses [@2019arXiv190405364T; @alsing2018generalized]. It is shown in [@2019arXiv190405364T; @alsingndes] that $\mathcal{O}(1000)$ simulations are needed to perform inference on Stage IV data and in contrast to MCMC methods (see e.g. [@foreman2013emcee]) these can be run in parallel, at up to $100$ simulations at a time. In the future it may be easier to handle the reduced shear correction in this paradigm, rather than directly computing the lensing observable with a perturbative expansion.
The eventual aim for a DELFI pipeline [@2019arXiv190405364T] is to compute lensing observables from full N-body simulations (see e.g [@izard2017ice]). This would avoid the need to write a matter bispectrum emulator trained on simulations, although, the N-body simulations used for this purpose would need to accurately represent the physics of the bispectrum.
\[sec:5\]Conclusions
====================
In this work, we quantified the impact that making the reduced shear approximation and neglecting magnification bias will have on angular power spectra of upcoming weak lensing surveys. Specifically, we calculated the biases that would be expected in the cosmological parameters obtained from *Euclid*. By doing so, significant biases were found for $\Omega_{\rm m}$, $n_{\rm s}$, $\sigma_8$, $\Omega_{\rm DE}$, $w_0$, and $w_a$ of $-0.51\sigma$, $-0.36\sigma$, $0.37\sigma$, $1.36\sigma$, $-0.66\sigma$, and $1.21\sigma$, respectively. We also built the formalism for an IA-enhanced correction. This was discovered to be sub-dominant. Given the severity of our calculated biases, we conclude that it is necessary to make both the reduced shear and magnification bias corrections for Stage IV experiments.
However, there are important limitations to consider in the approach described here. In calculating these corrections, the Limber approximation is still made. This approximation is typically valid above $\ell\sim100$. But, for *Euclid* we expect to reach $\ell$-modes of $10$. Therefore, the impact of this simplification at the correction level must be evaluated. Given that the dominant contributions to the reduced shear and magnification bias corrections come from $\ell$-modes above $100$, we would not expect the Limber approximation to significantly affect the resulting cosmological biases. However, an explicit calculation is still warranted. Furthermore, the various correction terms depend on bispectra which are not well understood: they both involve making a plethora of assumptions, and using fitting formulae that have accuracies of only 30-50$\%$ on small scales.
In addition, this work does not consider the impact of baryonic feedback on the corrections. We would expect that baryonic feedback behaves in a similar way to lowering the fiducial value $\sigma_8$ (see Appendix \[A:VCosmo\]), i.e. they both suppress structure growth in high density regions. Accordingly, it is likely that the inclusion of baryonic feedback would have an effect on these corrections. If the matter power spectrum is suppressed by a greater fraction than the matter bispectrum, then the biases will increase. However, it is not currently clear to what degree the matter bispectrum is suppressed relative to the matter power spectrum, and existing numerical simulations propose seemingly inconsistent answers (see e.g. [@Illustrisbisup] in comparison to [@Owlsbisup]). For this reason, we cannot robustly quantify the impact of baryonic feedback on the biases. As knowledge of the impact of baryons on the bispectrum improves, the reduced shear and magnification bias corrections should be modified accordingly.
An additional hurdle is the large computational expense of these terms; arising from the multiple nested integrals needing numerical computation. Computing the reduced shear and magnification bias corrections for this work took on the order of $24$ hours when multiprocessing across 100 CPU threads. Including the IA-enhanced correction term increases this to $\sim48$ hours. This expense can be prohibitive if the correction is to be included in inference methods. Considering that forward modelling approaches, such as a DELFI pipeline, could both bypass the need for matter bispectrum fitting formulae, and reduce computation time, we recommend that forward modelling should be used to account for these corrections in the future. However, there is also merit in exploring whether the existing processes can be optimised.
We thank Paniez Paykari for her programming expertise. ACD and TDK are supported by the Royal Society. PLT acknowledges support for this work from a NASA Postdoctoral Program Fellowship and the UK Science and Technologies Facilities Council. Part of the research was carried out at the Jet Propulsion Laboratory, California Institute of Technology, under a contract with the National Aeronautics and Space Administration. The Euclid Consortium acknowledges the European Space Agency and the support of a number of agencies and institutes that have supported the development of *Euclid*, in particular the Academy of Finland, the Agenzia Spaziale Italiana, the Belgian Science Policy, the Canadian Euclid Consortium, the Centre National d’Etudes Spatiales, the Deutsches Zentrum für Luft- und Raumfahrt, the Danish Space Research Institute, the Fundação para a Ciência e a Tecnologia, the Ministerio de Economia y Competitividad, the National Aeronautics and Space Administration, the Netherlandse Onderzoekschool Voor Astronomie, the Norwegian Space Agency, the Romanian Space Agency, the State Secretariat for Education, Research and Innovation (SERI) at the Swiss Space Office (SSO), and the United Kingdom Space Agency. A detailed complete list is available on the *Euclid* web site (<http://www.euclid-ec.org>).
\[AGB\]Generalised lensing bispectra formulae
=============================================
We can extend the methodology used to describe the matter bispectrum, $B_{\delta\delta\delta}$, to describe the bispectrum of three related quantities, $B_{\mu\nu\eta}$. Here, the three fields $\mu$, $\nu$, and $\eta$ are proportional to the density contrast, $\delta$, by some redshift-dependent weightings. This means they behave as $\delta$ would, under a small change in the fiducial cosmology. In this way, the second-order perturbation theory approach of [@Fryperturb] remains valid. We also assume Gaussian random initial conditions. Accordingly, the bispectrum is defined by first and second-order terms: $$\begin{aligned}
\label{eq:bigen0}
B_{\mu\nu\eta}(\boldsymbol{k}_1, \boldsymbol{k}_2, \boldsymbol{k}_3) &= \langle[\widetilde{\mu}^{(1)}(\boldsymbol{k}_1)+\widetilde{\mu}^{(2)}(\boldsymbol{k}_1)] \nonumber\\
&\times[\widetilde{\nu}^{(1)}(\boldsymbol{k}_2)+\widetilde{\nu}^{(2)}(\boldsymbol{k}_2)] \nonumber\\
&\times[\widetilde{\eta}^{(1)}(\boldsymbol{k}_3)+\widetilde{\eta}^{(2)}(\boldsymbol{k}_3)]\rangle,\end{aligned}$$ where the superscripts (2) and (1) denote the second and first-order terms respectively. But because we take Gaussian random initial conditions, the value of the three-point correlation vanishes at the lowest-order. Additionally, we can neglect products of second-order terms, as these will be fourth-order terms. Equation (\[eq:bigen0\]) now becomes: $$\begin{aligned}
\label{eq:bigen1}
B_{\mu\nu\eta}(\boldsymbol{k}_1, \boldsymbol{k}_2, \boldsymbol{k}_3) &= \braket{\widetilde{\mu}^{(2)}(\boldsymbol{k}_1)\widetilde{\nu}^{(1)}(\boldsymbol{k}_2)\widetilde{\eta}^{(1)}(\boldsymbol{k}_3)} \nonumber\\
&+ \braket{\widetilde{\nu}^{(2)}(\boldsymbol{k}_2)\widetilde{\mu}^{(1)}(\boldsymbol{k}_1)\widetilde{\eta}^{(1)}(\boldsymbol{k}_3)} \nonumber\\
&+ \braket{\widetilde{\eta}^{(2)}(\boldsymbol{k}_3)\widetilde{\mu}^{(1)}(\boldsymbol{k}_1)\widetilde{\nu}^{(1)}(\boldsymbol{k}_2)}.\end{aligned}$$ The above assumption relating the three fields to $\delta$, also leads us to concluding that $\delta^{(1)}$ is related to $\delta^{(2)}$ in the same way that $\mu^{(1)}$, $\nu^{(1)}$, and $\eta^{(1)}$ are related to $\mu^{(2)}$, $\nu^{(2)}$, and $\eta^{(2)}$ respectively. In which case, we can directly adapt Eq. (40) of [@Fryperturb], to read: $$\begin{aligned}
\label{eq:bigen2}
B_{\mu\nu\eta}(\boldsymbol{k}_1, \boldsymbol{k}_2, \boldsymbol{k}_3) &= 2F_2(\boldsymbol{k}_2, \boldsymbol{k}_3)P_{\mu\nu}(\boldsymbol{k}_2)P_{\mu\eta}(\boldsymbol{k}_3) \nonumber\\
&+ 2F_2(\boldsymbol{k}_1, \boldsymbol{k}_3)P_{\nu\mu}(\boldsymbol{k}_1)P_{\nu\eta}(\boldsymbol{k}_3) \nonumber\\
&+ 2F_2(\boldsymbol{k}_1, \boldsymbol{k}_2)P_{\eta\mu}(\boldsymbol{k}_1)P_{\eta\nu}(\boldsymbol{k}_2),\end{aligned}$$ with: $$\label{eq:Fdef}
F_2(\boldsymbol{k}_1, \boldsymbol{k}_2) = \frac{5}{7} + \frac{1}{2}\frac{\boldsymbol{k_1}\cdot\boldsymbol{k_2}}{k_1k_2}\bigg(\frac{k_1}{k_2}+\frac{k_2}{k_1}\bigg)+\frac{2}{7}\bigg(\frac{\boldsymbol{k_1}\cdot\boldsymbol{k_2}}{k_1k_2}\bigg)^2.$$
As in [@Scoccimarro01], this can then be modified to include numerical fitting to N-body simulations by exchanging $F_2$ for $F_2^{\rm eff}$, as defined in Eq. (\[eq:Feff\]). The fitting formula determined in [@Scoccimarro01] still remains valid, because it does not have any redshift dependence and does not depend on the fiducial cosmology. The density perturbation-IA bispectrum, used in the IA-enhanced lensing bias correction, is then a specific case of this formula, where $\mu=\nu=\delta$, and $\eta=I$.
\[A:VCosmo\]The impact of varying the fiducial cosmology
========================================================
------------------- ---------------- ----------------- ----------------- ------------------
Cosmological Low $\sigma_8$ High $\sigma_8$ Low $n_{\rm s}$ High $n_{\rm s}$
Parameter 1-$\sigma$ 1-$\sigma$ 1-$\sigma$ 1-$\sigma$
$\Omega_{\rm m}$ 0.016 0.0085 0.014 0.012
$\Omega_{\rm b}$ 0.024 0.0043 0.020 0.023
$h$ 0.13 0.041 0.12 0.13
$n_{\rm s}$ 0.031 0.012 0.030 0.031
$\sigma_8$ 0.014 0.041 0.012 0.011
$\Omega_{\rm DE}$ 0.065 0.037 0.061 0.059
$w_0$ 0.21 0.13 0.17 0.16
$w_a$ 1.18 0.76 1.03 1.03
------------------- ---------------- ----------------- ----------------- ------------------
: Predicted 1-$\sigma$ uncertainties for the $w$CDM parameters that would be determined from *Euclid*, for fiducial cosmologies with lower and higher values of $\sigma_8$ and $n_{\rm s}$, (0.716, 0.916) and (0.86, 1.06) respectively, are shown.[]{data-label="tab:appunc"}
------------------- ---------------- ----------------- ----------------- ------------------
Cosmological Low $\sigma_8$ High $\sigma_8$ Low $n_{\rm s}$ High $n_{\rm s}$
Parameter Bias/$\sigma$ Bias/$\sigma$ Bias/$\sigma$ Bias/$\sigma$
$\Omega_{\rm m}$ $-0.33$ $-0.76$ $-0.70$ $-0.41$
$\Omega_{\rm b}$ $-0.097$ $-1.29$ $-0.22$ $-0.23$
$h$ 0.076 $-0.24$ 0.10 0.018
$n_{\rm s}$ $-0.29$ $-0.97$ $-0.44$ $-0.50$
$\sigma_8$ 0.28 0.41 0.54 0.20
$\Omega_{\rm DE}$ 0.89 2.07 1.31 1.43
$w_0$ $-0.41$ $-0.99$ $-0.67$ $-0.62$
$w_a$ 0.76 1.85 1.15 1.26
------------------- ---------------- ----------------- ----------------- ------------------
: Biases induced in the $w$CDM parameters from neglecting the two studied corrections, for lower and higher fiducial values of $\sigma_8$ and $n_{\rm s}$, (0.716, 0.916) and (0.86, 1.06) respectively, are shown.[]{data-label="tab:appbi"}
Owing to the fact that the reduced shear and magnification bias corrections are a projection of the matter bispectrum, while the shear auto and cross-spectra are projections of the matter power spectrum, the relative size of the correction in comparison to the shear spectra is strongly influenced by non-linearity [@Shapiro09]. The parameters $\sigma_8$ and $n_{\rm s}$ have the strongest effect on non-linearity, therefore we examine the effect of changing these parameters on the biases.
Table \[tab:appunc\] and Table \[tab:appbi\] show the recomputed uncertainties and biases, respectively, when the fiducial values of $\sigma_8$ are lowered to 0.716, and raised to 0.916. These biases are also visualised in Fig. \[fig:bias\_s8\]. As expected, lowering the fiducial value of $\sigma_8$ suppresses the biases, though they still remain significant, whilst raising this value aggravates the biases. Contributing to these changes is also the fact that the predicted uncertainties in the parameters generally decrease as $\sigma_8$ is increased, with the exception of $\sigma_8$ itself.
The effects on the uncertainties of varying $n_{\rm s}$, to 0.86 then 1.06, are shown in Table \[tab:appunc\]. Figure \[fig:bias\_ns\] and Table \[tab:appbi\] show the biases after this variation. The effect on the significances of the biases is less straightforward in this case. The parameters are affected relatively differently in comparison to the variation of $\sigma_8$. In general, the change in the ratio of the biases to the uncertainties is non-trivial, but relatively subtle. The exceptions to this being $\sigma_8$ and $\Omega_{\rm m}$. For these parameters, the biases reduce considerably in magnitude. Despite the changes, the biases in each of the previously affected parameters remain significant.
{width="\textwidth"}
{width="\textwidth"}
[^1]:
[^2]: This paper is published on behalf of the Euclid Consortium.
[^3]: <https://www.euclid-ec.org/>
[^4]: <https://www.nasa.gov/wfirst>
[^5]: <https://www.lsst.org/>
[^6]: <https://class-code.net/>
[^7]: <https://sourceforge.net/projects/healpix/>
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We have recently extended many-body perturbation theory and coupled-cluster theory performed on top of a Slater determinant breaking rotational symmetry to allow for the restoration of the angular momentum at any truncation order \[T. Duguet, J. Phys. G: Nucl. Part. Phys. 42 (2015) 025107\]. Following a similar route, we presently extend Bogoliubov many-body perturbation theory and Bogoliubov coupled cluster theory performed on top of a Bogoliubov reference state breaking global gauge symmetry to allow for the restoration of the particle number at any truncation order. Eventually, formalisms can be merged to handle $SU(2)$ and $U(1)$ symmetries at the same time. Several further extensions of the newly proposed many-body formalisms can be foreseen in the mid-term future. The long-term goal relates to the ab initio description of near-degenerate finite quantum systems with an open-shell character.'
author:
- 'T. Duguet'
- 'A. Signoracci'
bibliography:
- 'PNR-BCC.bib'
title: |
Symmetry broken and restored coupled-cluster theory\
II. Global gauge symmetry and particle number
---
Introduction
============
In Ref. [@Duguet:2014jja], hereafter referred to as Paper I, the motivations to tackle degenerate (or near-degenerate) finite quantum systems with an open-shell character via ab initio methods relying on the concept of symmetry breaking and restoration were explained at length. Dealing with singly-open shell atomic nuclei, i.e. nuclei displaying a good closed-shell character for either protons or neutrons, requires the breaking and the restoration of $U(1)$ global gauge symmetry associated with particle number conservation. Breaking the symmetry allows one to lift the degeneracy of the unperturbed reference state associated with the Cooper pair instability responsible for nuclear superfluidity. Extending the treatment to doubly open-shell nuclei demands to further break and restore $SU(2)$ symmetry associated with the conservation of angular momentum[^1].
Standard single-reference Rayleigh-Schroedinger many-body perturbation theory (MBPT) [@goldstone57a; @hugenholtz57a; @bloch58a; @kutzelnigg09a; @shavitt09a] and coupled cluster (CC) theory [@coester58a; @shavitt09a] expand the exact many-body ground-state energy around a reference state taking the form of a Slater determinant. Consequently, these methods do respect $U(1)$ global gauge symmetry all throughout. To allow for the breaking of $U(1)$ symmetry, one must expand the exact many-body state around a more general vacuum taking the form of a Bogoliubov product state [@ring80a]. As for perturbation theory, this leads to formulating single-reference Bogoliubov many-body perturbation theory (SR-BMBPT) or its variant based on a simpler Bardeen-Cooper-Schrieffer reference state [@mehta1; @balian62a; @henley]. In either form, such a particle-number-breaking many-body perturbation theory has been scarcely used in the physics literature. Going beyond perturbation theory, the single-reference Bogoliubov coupled-cluster (SR-BCC) theory was only recently formulated and applied [@StolarczykMonkhorst; @Signoracci:2014dia; @Henderson:2014vka].
While BMBPT and BCC ab initio methods can efficiently access open-shell systems, it remains necessary to restore $U(1)$ symmetry when dealing with a finite quantum system such as the atomic nucleus. Consequently, it is the goal of the present paper to generalize BMBPT and BCC formalisms to allow for the exact restoration of good neutron (proton) number at any truncation order. This will lead to the design of [*particle-number-restored Bogoliubov many-body perturbation theory*]{} (PNR-BMBPT)[^2] and [*particle-number-restored Bogoliubov coupled cluster*]{} (PNR-BCC) theory. This is achieved by adapting the work done for the $SU(2)$ group in Paper I to the $U(1)$ group, which effectively requires the entire reformulation of the formalism on the basis of a Bogoliubov reference state and making use of Bogoliubov algebra.
The paper is organized as follows. Section \[definitions\] provides the ingredients necessary to set up the formalism while Sec. \[normkernel\] elaborates on the general principles of the approach, independently of the actual many-body technique eventually employed to expand the exact solution of the Schroedinger equation. In Sec. \[sectionMBPT\], a generalized BMBPT is developed and acts as the foundation for the generalized BCC approach introduced in Sec. \[CCtheory\]. It is shown how generalized energy and norm kernels at play in the formalism can be computed from [*naturally terminating*]{} BCC expansions. The way to recover SR-BMBPT and SR-BCC theory on the one hand and particle-number-projected Hartree-Fock-Bogoliubov theory on the other hand is illustrated. Eventually, the algorithm to be implemented by the owner of a BCC code to incorporate the particle-number restoration is highlighted. The body of the paper is restricted to discussing the overall scheme, limiting technical details to the minimum. Complete analytic results are provided in an extended set of appendices.
Basic ingredients {#definitions}
=================
Let us introduce necessary ingredients to make the paper self-contained. Although pedestrian, this section displays definitions and identities that are crucial to the building of the formalism later on.
Hamiltonian
-----------
Let the Hamiltonian $H=T+V$ of the system be of the form[^3] $$H \equiv \frac{1}{(1!)^2} \sum _{pq} t_{pq} c^{\dagger}_{p} c_{q}+\frac{1}{(2!)^2} \sum _{pqrs} \bar{v}_{pqrs} c^{\dagger}_{p} c^{\dagger}_{q} c_{s} c_{r} \, , \label{e:ham}$$ where antisymmetric matrix elements of the two-body interaction are employed and where $\{c_{p};c^{\dagger}_{p}\}$ denote particle annihilation and creation operators associated with an arbitrary basis of the one-body Hilbert space ${\cal H}_1$.
Bogoliubov algebra
------------------
The unitary Bogoliubov transformation connects quasi-particle annihilation and creation operators $\{\beta_{k};\beta^{\dagger}_{k}\}$ to particle ones through [@ring80a]
\[e:p2qp\] $$\begin{aligned}
\beta_{k} &= \sum_{p} U^{*}_{pk} \, c_{p}
+ V^{*}_{pk} \, c^{\dagger}_{p} \, , \\
\beta_{k}^{\dagger} &= \sum_{p} U_{pk} \, c^{\dagger}_{p}
+ V_{pk} \, c_{p} \, .\end{aligned}$$
Both sets of operators obey anticommutation rules
\[anticom\] $$\begin{aligned}
\{c_{p},c_{q}\}=0 \,\,\,\,\,\,\,\, &; \,\,\, \{\beta_{k_1},\beta_{k_2}\}=0 \, , \\
\{c^{\dagger}_{p},c^{\dagger}_{q}\}=0 \,\,\,\,\,\,\,\, &; \,\,\, \{\beta^{\dagger}_{k_1},\beta^{\dagger}_{k_2}\}=0 \, , \\
\{c_{p},c^{\dagger}_{q}\} = \delta_{pq} \,\,\, &; \,\,\, \{\beta_{k_1},\beta^{\dagger}_{k_2}\} = \delta_{k_1k_2} \, .\end{aligned}$$
The Bogoliubov product state, which carries even number-parity as a quantum number, is defined as $$\label{e:bogvac}
| \Phi \rangle \equiv \mathcal{C} \displaystyle \prod_{k} \beta_{k} | 0 \rangle \, ,$$ and is the vacuum of the quasiparticle operators, i.e. $\beta_k | \Phi \rangle =0$ for all $k$. In Eq. , $\mathcal{C}$ is a complex normalization ensuring that $\langle \Phi | \Phi \rangle =1$. As quasiparticle operators mix particle creation and annihilation operators (see Eq. ), the Bogoliubov vacuum breaks $U(1)$ symmetry associated with particle number conservation, i.e. $| \Phi \rangle$ is not an eigenstate of the particle-number operator $A$, except in the limit where it reduces to a Slater determinant.
The Bogoliubov transformation can be written in matrix form $$\left(
\begin{array} {c}
\beta \\
\beta^{\dagger}
\end{array}
\right) = W^{\dagger} \left(
\begin{array} {c}
c \\
c^{\dagger}
\end{array}
\right) \, ,$$ where $$W \equiv \left(
\begin{array} {cc}
U & V^{\ast} \\
V & U^{\ast}
\end{array}
\right) \, . \label{bogomatrix}$$ One can further define the skew-symmetric matrix $$Z\equiv V^{\ast}[U^{\ast}]^{-1}$$ in terms of which $| \Phi \rangle$ can be expressed by virtue of Thouless’ theorem [@thouless60]. The anticommutation rules obeyed by the quasi-particle operators relate to the unitarity of $W$ that leads to four relations
\[unitarity\] $$\begin{aligned}
UU^{\dagger} + V^{\ast}V^{T} &=& 1 \, , \label{unitarityA} \\
VU^{\dagger} + U^{\ast}V^{T} &=& 0 \, , \label{unitarityB} \\
U^{\dagger}U + V^{\dagger}V &=& 1 \, , \label{unitarityC} \\
V^{T}U + U^{T}V &=& 0 \, , \label{unitarityD} \end{aligned}$$ originating from $W^{\dagger}W = 1$ and four relations $$\begin{aligned}
UV^{\dagger} + V^{\ast}U^{T} &=& 0 \, , \label{unitarityE} \\
VV^{\dagger} + U^{\ast}U^{T} &=& 1 \, , \label{unitarityF} \\
U^{\dagger}V^{\ast} + V^{\dagger}U^{\ast} &=& 0 \, , \label{unitarityG} \\
V^{T}V^{\ast} + U^{T}U^{\ast} &=& 1 \, , \label{unitarityH} \end{aligned}$$
originating from $WW^{\dagger} = 1$.
The Bogoliubov state $| \Phi \rangle$ is fully characterized by the generalized density matrix [@ring80a]
\[generalizeddensitymatrix\] $$\begin{aligned}
{\cal R} &\equiv&
\left(
\begin{array} {cc}
\frac{\langle \Phi | c^{\dagger}c^{\phantom{\dagger}} | \Phi \rangle}{\langle \Phi | \Phi \rangle} & \frac{\langle \Phi | c^{\phantom{\dagger}}c^{\phantom{\dagger}} | \Phi \rangle}{\langle \Phi | \Phi \rangle} \\
\frac{\langle \Phi | c^{\dagger}c^{\dagger} | \Phi \rangle}{\langle \Phi | \Phi \rangle} & \frac{\langle \Phi | c^{\phantom{\dagger}}c^{\dagger} | \Phi \rangle}{\langle \Phi | \Phi \rangle}
\end{array}
\right) \\
&\equiv&
\left(
\begin{array} {cc}
+\rho & +\kappa \\
-\bar{\kappa}^{\ast} & -\sigma^{\ast}
\end{array}
\right) \\
&=&
\left(
\begin{array} {cc}
V^{\ast}V^T & V^{\ast}U^T \\
U^{\ast}V^T & U^{\ast}U^T
\end{array}
\right)
\, ,\end{aligned}$$
where $\rho$ and $\kappa$ denote the normal one-body density matrix and the anomalous density matrix (or pairing tensor), respectively. Using anticommutation rules of particle creation and annihilation operators, one demonstrates that
\[densitymatrices\] $$\begin{aligned}
\rho_{qp} &=& +\rho^{\ast}_{pq} \, , \label{densitymatrices1} \\
\kappa_{qp} &=& - \kappa_{pq} \, , \label{densitymatrices2} \\
\sigma_{qp} &=& +\rho_{qp} - \delta_{qp} \, , \label{densitymatrices3} \\
\bar{\kappa}_{qp} &=& +\kappa_{qp} \, , \label{densitymatrices4} \end{aligned}$$
meaning that $\rho$ is hermitian (i.e. $\rho^{\dagger} = \rho$) while $\kappa$ is skew-symmetric (i.e. $\kappa^{T} = - \kappa$). Transforming the generalized density matrix to the quasi-particle basis via ${\bf R} \equiv W^{\dagger} {\cal R} W $ leads to
\[generalizeddensitymatrix2\] $$\begin{aligned}
{\bf R} &=&
\left(
\begin{array} {cc}
\frac{\langle \Phi | \beta^{\dagger}\beta^{\phantom{\dagger}}\, | \Phi \rangle}{\langle \Phi | \Phi \rangle} & \frac{\langle \Phi | \beta^{\phantom{\dagger}}\beta^{\phantom{\dagger}} | \Phi \rangle}{\langle \Phi | \Phi \rangle} \\
\frac{\langle \Phi | \beta^{\dagger}\beta^{\dagger} | \Phi \rangle}{\langle \Phi | \Phi \rangle} & \frac{\langle \Phi | \beta^{\phantom{\dagger}}\beta^{\dagger} | \Phi \rangle}{\langle \Phi | \Phi \rangle}
\end{array}
\right) \\
&\equiv&
\left(
\begin{array} {cc}
R^{+-} & R^{--} \\
R^{++} & R^{-+}
\end{array}
\right) \\
&=&
\left(
\begin{array} {cc}
0 & 0 \\
0 & 1
\end{array}
\right)
\, , \end{aligned}$$
where the result, trivially obtained by considering the action of quasi-particle operators on the vacuum, can also be recovered starting from Eq. \[generalizeddensitymatrix\] and making use of Eqs. \[e:p2qp\] and \[unitarity\].
Normal ordering
---------------
A Lagrange term is eventually required to constrain the particle number to the correct value on average, such that the grand potential $\Omega \equiv H - \lambda A$ is to be used in place of $H$, where the particle-number operator $A = \sum_{n=1}^{\text{A}} 1$ takes the second-quantized form $$A = \sum_{p} c^{\dagger}_{p} c_{p} \, .$$ The present formalism is best formulated in the quasiparticle basis introduced in Eq. by normal ordering all operators at play with respect to $| \Phi \rangle$ via Wick’s theorem [@wick50a]. Taking $\Omega$ as an example, and as extensively discussed in Ref. [@Signoracci:2014dia], its normal-ordered form expressed in terms of fully antisymmetric matrix elements[^4] reads as
\[e:h3qpas\] $$\begin{aligned}
\label{e:h3qpasa}
\Omega &\equiv \Omega^{[0]} + \Omega^{[2]} + \Omega^{[4]} \\
&\equiv \Omega^{00} + \big[\Omega^{20} + \Omega^{11} + \Omega^{02}\big]+\big[\Omega^{40} + \Omega^{31} + \Omega^{22} + \Omega^{13} +
\Omega^{04}\big] \\
&= \Omega^{00} \\
& \: \: \: \: \: \: \: \: \: \: \: \: \: + \frac{1}{1!}\displaystyle\sum_{k_1 k_2} \Omega^{11}_{k_1 k_2}\beta^{\dagger}_{k_1} \beta_{k_2} \\
& \: \: \: \: \: \: \: \: \: \: \: \: \: + \frac{1}{2!}\displaystyle\sum_{k_1 k_2} \Big \{\Omega^{20}_{k_1 k_2} \beta^{\dagger}_{k_1}
\beta^{\dagger}_{k_2} + \Omega^{02}_{k_1 k_2} \beta_{k_2} \beta_{k_1} \Big \} \\
\label{e:h3qpase}
& \: \: \: \: \: \: \: \: \: \: \: \: \: + \frac{1}{(2!)^{2}} \displaystyle\sum_{k_1 k_2 k_3 k_4} \Omega^{22}_{k_1 k_2 k_3 k_4}
\beta^{\dagger}_{k_1} \beta^{\dagger}_{k_2} \beta_{k_4}\beta_{k_3} \\
& \: \: \: \: \: \: \: \: \: \: \: \: \: + \frac{1}{3!}\displaystyle\sum_{k_1 k_2 k_3 k_4}\Big \{ \Omega^{31}_{k_1 k_2 k_3 k_4}
\beta^{\dagger}_{k_1}\beta^{\dagger}_{k_2}\beta^{\dagger}_{k_3}\beta_{k_4} +
\Omega^{13}_{k_1 k_2 k_3 k_4} \beta^{\dagger}_{k_1} \beta_{k_4} \beta_{k_3} \beta_{k_2} \Big \} \\
& \: \: \: \: \: \: \: \: \: \: \: \: \: + \frac{1}{4!} \displaystyle\sum_{k_1 k_2 k_3 k_4}\Big \{ \Omega^{40}_{k_1 k_2 k_3 k_4}
\beta^{\dagger}_{k_1}\beta^{\dagger}_{k_2}\beta^{\dagger}_{k_3}\beta^{\dagger}_{k_4} +
\Omega^{04}_{k_1 k_2 k_3 k_4} \beta_{k_4} \beta_{k_3} \beta_{k_2} \beta_{k_1} \Big \} \, ,
\end{aligned}$$
where
1. Each term $\Omega^{ij}$ is characterized by its number $i$ ($j$) of quasiparticle creation (annihilation) operators. Because $\Omega$ has been normal-ordered with respect to $| \Phi \rangle$, all quasiparticle creation operators (if any) are located to the left of all quasiparticle annihilation operators (if any). The class $\Omega^{[k]}$ groups all the terms $\Omega^{ij}$ with $i+j=k$. The first contribution $$\Omega^{[0]} = \Omega^{00} = \frac{\langle \Phi | \Omega | \Phi \rangle}{\langle \Phi | \Phi \rangle}$$ denotes the fully contracted part of $\Omega$ and is nothing but a (real) number.
2. The subscripts of the matrix elements $\Omega^{ij}_{k_1 \ldots k_i k_{i+1} \ldots k_{i+j}}$ are ordered sequentially, independently of the creation or annihilation character of the operators the indices refer to. While quasiparticle creation operators themselves also follow sequential order, quasiparticle annihilation operators follow inverse sequential order. In Eq. , for example, the two creation operators are ordered $\beta^{\dagger}_{k_1}\beta^{\dagger}_{k_2}$ while the two annihilation operators are ordered $\beta_{k_4}\beta_{k_3}$.
3. Matrix elements are fully antisymmetric, i.e. $$\begin{aligned}
\Omega^{ij}_{k_1 \ldots k_{i} k_{i+1} \ldots k_{i+j}} &=& (-1)^{\sigma(P)}
\Omega^{ij}_{P(k_1 \ldots k_i | k_{i+1} \ldots k_{i+j})} \end{aligned}$$ where $\sigma(P)$ refers to the signature of the permutation $P$. The notation $P(\ldots | \ldots)$ denotes a separation into the $i$ quasiparticle creation operators and the $j$ quasiparticle annihilation operators such that permutations are only considered between members of the same group.
4. As each $\Omega^{[k]}$ component is hermitian, matrix elements exhibit the following behavior under hermitian conjugation
\[e:me3sym\] $$\begin{aligned}
\label{e:me3syma}
\Omega^{11}_{k_1 k_2} &= \Omega^{11*}_{k_2 k_1} \, ,\\
\label{e:me3symc}
\Omega^{20}_{k_1 k_2} &= \Omega^{02*}_{k_1 k_2} \, , \\
\label{e:me3symb}
\Omega^{22}_{k_1 k_2 k_3 k_4} &= \Omega^{22*}_{k_3 k_4 k_1 k_2} \, , \\
\label{e:me3symf}
\Omega^{31}_{k_1 k_2 k_3 k_4} &= \Omega^{13*}_{k_4 k_1 k_2 k_3} \, , \\
\label{e:me3symj}
\Omega^{40}_{k_1 k_2 k_3 k_4} &= \Omega^{04*}_{k_1 k_2 k_3 k_4} \, .\end{aligned}$$
Similarly to $\Omega$, the normal-ordered form of the particle-number operator is obtained as[^5]
\[normalorderedA\] $$\begin{aligned}
A &\equiv& A^{[0]} + A^{[2]} \\
&\equiv& A^{00} + \big[A^{20} + A^{11} + A^{02}\big]\\
&\equiv& A^{00} \\
&& + \frac{1}{1!} \sum_{k_1 k_2} A^{11}_{k_1 k_2}\beta^{\dagger}_{k_1} \beta_{k_2} \\
&& + \frac{1}{2!}\sum_{k_1 k_2} \Big \{A^{20}_{k_1 k_2} \beta^{\dagger}_{k_1}
\beta^{\dagger}_{k_2} + A^{02}_{k_1 k_2} \beta_{k_2} \beta_{k_1} \Big \} \end{aligned}$$
and allows the extraction of the normal-ordered Hamiltonian itself whose various terms take the same form as those of $\Omega$ but with the modifications $H^{[0]} = \Omega^{[0]} + \lambda A^{[0]}$ and $H^{[2]} = \Omega^{[2]} + \lambda A^{[2]}$.
Diagrammatic representation of an operator {#diagramsforvertices}
------------------------------------------
Normal-ordered operators in the Schroedinger representation can be represented diagrammatically. Taking the grand potential and the particle-number operators as typical examples, canonical diagrams representing their normal-ordered contributions $\Omega^{ij}$ and $A^{ij}$ are displayed in Figs. \[variousvertices1\] and \[variousvertices2\], respectively. Focusing on $\Omega$ as an example, the various diagrams contributing to it must be understood in the following way.
1. One must associate the factor $\Omega^{ij}_{k_1 \ldots k_i k_{i+1} \ldots k_{i+j}}$ to the dot vertex, where $i$ denotes the number of lines traveling out of the vertex and representing quasiparticle creation operators while $j$ denotes the number of lines traveling into the vertex and representing quasiparticle annihilation operators.
2. A factor $1/[i!j!]$ must multiply $\Omega^{ij}_{k_1 \ldots k_i k_{i+1} \ldots k_{i+j}}$ given that the corresponding diagram contain $j$ equivalent ingoing lines and $i$ equivalent outgoing lines.
3. In the canonical representation used in Figs. \[variousvertices1\] and \[variousvertices2\], all oriented lines go up, i.e. lines representing quasiparticle creation (annihilation) operators appear above (below) the vertex. Accordingly, indices $k_1 \ldots k_i$ must be assigned consecutively from the leftmost to the rightmost line above the vertex, while $k_{i+1} \ldots k_{i+j}$ must be similarly assigned consecutively for lines below the vertex.
4. In the diagrammatic representation at play in the many-body formalism designed below, it is possible for a line to propagate downwards[^6]. This can be obtained unambiguously starting from the canonical representation given in Figs. \[variousvertices1\] and \[variousvertices2\] at the price of adding a specific rule. As illustrated in Fig. \[variousvertices3\] for the diagram representing $\Omega^{22}$, lines must only be rotated through the right of the diagram, i.e. going through the dashed line, while it is forbidden to rotate them through the full line. Additionally, a minus sign must be added to the amplitude $\Omega^{ij}_{k_1 \ldots k_i k_{i+1} \ldots k_{i+j}}$ associated with the canonical diagram each time two lines cross as illustrated in Fig. \[variousvertices3\].
![ \[variousvertices1\] Canonical diagrammatic representation of normal-ordered contributions to the grand potential $\Omega$ in the Schroedinger representation.](figures/omega){width="\columnwidth"}
\
\
$U(1)$ group
------------
We consider the abelian compact Lie group $U(1)\equiv \{S(\varphi), \varphi \in [0,2\pi]\}$ associated with the global rotation of an A-body fermion system in gauge space. As $U(1)$ is considered to be a symmetry group of $H$, commutation relations $$\begin{aligned}
\left[H,S(\varphi)\right]&=& \left[A,S(\varphi)\right] = \left[\Omega,S(\varphi)\right] =0 \, , \label{commutation}\end{aligned}$$ hold for any $\varphi \in [0,2\pi]$.
We utilize the unitary representation of $U(1)$ on Fock space ${\cal F}$ given by $$S(\varphi) = e^{iA\varphi} \, .$$ Matrix elements of the irreducible representations (IRREPs) of $U(1)$ are $$\langle \Psi^{\text{A}}_{\mu} | S(\varphi) |\Psi^{\text{A}'}_{\mu'} \rangle \equiv e^{i\text{A}\varphi} \, \delta_{\text{A}\text{A}'} \, \delta_{\mu\mu'} \, , \label{irreps}$$ where $| \Psi^{\text{A}}_{\mu} \rangle$ is an eigenstate of $A$ $$\begin{aligned}
A | \Psi^{\text{A}}_{\mu} \rangle &=& \text{A} | \Psi^{\text{A}}_{\mu} \rangle \, , \label{eigenequationA}\end{aligned}$$ and, by virtue of Eq. \[commutation\], of the Hamiltonian at the same time $$\begin{aligned}
H | \Psi^{\text{A}}_{\mu} \rangle &=& \text{E}^{\text{A}}_{\mu} \, | \Psi^{\text{A}}_{\mu} \rangle \,\,\, , \label{schroed}\end{aligned}$$ where $\text{E}^{\text{A}}_{\mu}$, with $\mu=0, 1, 2\ldots$, orders increasing eigenenergies for a fixed $\text{A}$. From the group theory point of view, $\text{A} \in \mathbb{Z}$ on the right-hand side of Eq. \[irreps\]. Since $\text{A}$ actually represents the number of fermions in the system, its value is constrained from the physics point of view to $\text{A} \in \mathbb{N}$. Equations \[eigenequationA\] and \[schroed\] trivially lead to $$\begin{aligned}
\Omega | \Psi^{\text{A}}_{\mu} \rangle &=& \Omega^{\text{A}}_{\mu} \, | \Psi^{\text{A}}_{\mu} \rangle \,\,\, , \label{grandpotschroed}\end{aligned}$$ where $\Omega^{\text{A}}_{\mu} \equiv \text{E}^{\text{A}}_{\mu} - \lambda \text{A}$. The volume of the group is $$\begin{aligned}
v_{U(1)} \equiv \int_{0}^{2\pi} \!d\varphi = 2\pi \,\,\, ,\end{aligned}$$ and the orthogonality of IRREPs reads as $$\frac{1}{2}\int_{0}^{2\pi} \!d\varphi \,e^{-i\text{A}\varphi}\,e^{+i\text{A}'\varphi} = \delta_{\text{A}\text{A}'} \, . \label{orthogonality}$$A tensor operator $O$ of rank[^7] $\text{A}$ and a state $| \Psi^{\text{A}}_{\mu} \rangle$ transform under global gauge rotation according to
\[eq:ten:def\] $$\begin{aligned}
S(\varphi) \, O \, S(\varphi)^{-1} &=& e^{i\text{A}\varphi} \, O \, \,\, ,\label{eq:ten:def1} \\
S(\varphi) \, | \Psi^{A}_{\mu} \rangle &=& e^{i\text{A}\varphi} \, | \Psi^{\text{A}}_{\mu} \rangle \, \,\, . \label{eq:ten:def2}\end{aligned}$$
A key feature for the following is that any integrable function $f(\varphi)$ defined on $[0,2\pi]$ can be expanded over the IRREPs of the $U(1)$ group. This constitutes nothing but the Fourier decomposition of the function $$f(\varphi) \equiv \sum_{\text{A} \in \mathbb{Z}} \, f^{\text{A}} \, e^{i\text{A}\varphi} \, , \label{decomposition_general}$$ which defines the set of expansion coefficients $\{f^{\text{A}}\}$. Last but not least, the IRREPs fulfill the first-order ordinary differential equation (ODE) $$-i \frac{d}{d \varphi} e^{i\text{A}\varphi} = \text{A} \, e^{i\text{A}\varphi} \label{ODE} \, .$$
Time-dependent state {#timedependentstate}
--------------------
The many-body formalism proposed in the present work is conveniently formulated within an imaginary-time framework. We thus introduce the evolution operator in imaginary time as[^8] $${\cal U}(\tau) \equiv e^{-\tau \Omega} \, , \label{evoloperator}$$ with $\tau$ real. A key quantity throughout the present study is the time-evolved many-body state defined as $$\begin{aligned}
| \Psi (\tau) \rangle &\equiv& {\cal U}(\tau) | \Phi \rangle \nonumber \\
&=& \sum_{\text{A} \in \mathbb{N}}\sum_{\mu} e^{-\tau \Omega^{\text{A}}_{\mu}} \, | \Psi^{\text{A}}_{\mu} \rangle \, \langle \Psi^{\text{A}}_{\mu} | \Phi \rangle \, , \label{evolstate}\end{aligned}$$ where we have inserted a completeness relationship on Fock space under the form $${\ensuremath{\mathbbm{1}}}= \sum_{\text{A} \in \mathbb{N}}\sum_{\mu} | \Psi^{\text{A}}_{\mu} \rangle \, \langle \Psi^{\text{A}}_{\mu} | \, . \label{completeness}$$ It is straightforward to demonstrate that $| \Psi (\tau) \rangle$ satisfies the time-dependent Schroedinger equation $$\Omega \, | \Psi (\tau) \rangle = -\partial_{\tau} | \Psi (\tau) \rangle \, . \label{schroedinger}$$
Large and infinite time limits {#timelimits}
------------------------------
Below, we will be interested in first looking at the [*large*]{} $\tau$ limit of various quantities before eventually taking their [*infinite*]{} time limit. Although we utilize the same mathematical symbol ($\lim\limits_{\tau \to \infty}$) in both cases for simplicity, the reader must not be confused by the fact that there remains a residual $\tau$ dependence in the first case, which typically disappears by considering ratios before actually promoting the time to infinity. The large $\tau$ limit is essentially defined as $\tau \gg \Delta E^{-1}$, where $\Delta E$ is the energy difference between the ground state and the first excited state of $\Omega$. Depending on the system, the latter can be the first excited state in the IRREP of the ground state or the lowest state of another IRREP.
Ground state {#groundstateSeC}
------------
Taking the large $\tau$ limit provides the ground state of $\Omega$ under the form[^9]
\[groundstate\] $$\begin{aligned}
| \Psi^{\text{A}_0}_{0} \rangle &\equiv& \lim\limits_{\tau \to \infty} | \Psi (\tau) \rangle \\
&=& e^{-\tau \Omega^{\text{A}_0}_0} | \Psi^{\text{A}_0}_{0} \rangle \, \langle \Psi^{\text{A}_0}_{0} | \Phi \rangle \, .\end{aligned}$$
As will become clear below, the many-body scheme developed in the present work relies on choosing the Bogoliubov product state $| \Phi \rangle$ as the ground state of an unperturbed grand potential $\Omega_0$ that breaks $U(1)$ symmetry. As such, $| \Phi \rangle$ mixes several IRREPS but is likely to contain a component belonging to the nucleus of interest given that it is typically chosen to have (close to) the number $\text{A}_0$ of particles in average. Eventually, one recovers from Eq. \[schroedinger\] that $$\begin{aligned}
\Omega | \Psi^{\text{A}_0}_{0} \rangle &=& \Omega^{\text{A}_0}_0 \, | \Psi^{\text{A}_0}_{0} \rangle \,\,\, , \label{schroedevolstate}\end{aligned}$$ in the large $\tau$ limit.
Off-diagonal kernels {#transitionkernels}
--------------------
We now introduce the off-diagonal, i.e. $\varphi$-dependent, time-dependent kernel of an operator[^10] $O$ through $$\begin{aligned}
O(\tau,\varphi) &\equiv& \langle \Psi (\tau) | O | \Phi(\varphi) \rangle \, ,\end{aligned}$$ where $| \Phi(\varphi) \rangle \equiv S(\varphi) | \Phi \rangle$ denotes the [*gauge-rotated*]{} Bogoliubov state. Doing so for the identity, the Hamiltonian, the particle number and the grand potential operators, we introduce the set of off-diagonal kernels
\[defkernels\] $$\begin{aligned}
N(\tau,\varphi) &\equiv& \langle \Psi (\tau) | {\ensuremath{\mathbbm{1}}}| \Phi(\varphi) \rangle \, , \label{defnormkernel} \\
H(\tau,\varphi) &\equiv& \langle \Psi (\tau) | H | \Phi(\varphi) \rangle \, , \label{defenergykernel} \\
A(\tau,\varphi) &\equiv& \langle \Psi (\tau) | A | \Phi(\varphi) \rangle \, , \label{defAkernel} \\
\Omega(\tau,\varphi) &\equiv& \langle \Psi (\tau) | \Omega | \Phi(\varphi) \rangle \, , \label{defomegakernel} \end{aligned}$$
where the first one denotes the off-diagonal norm kernel and where the latter three are related through $\Omega(\tau,\varphi) = H(\tau,\varphi) - \lambda A(\tau,\varphi)$. Focusing on the grand potential operator as an example, its kernel can be split into various contributions associated with its normal-ordered components, i.e.
\[defkernels1and2body\] $$\begin{aligned}
\Omega(\tau,\varphi) &\equiv& \Omega^{00}(\tau,\varphi) \\
&&+ \Omega^{20}(\tau,\varphi) + \Omega^{11}(\tau,\varphi) + \Omega^{02}(\tau,\varphi) \\
&&+ \Omega^{40}(\tau,\varphi) + \Omega^{31}(\tau,\varphi) + \Omega^{22}(\tau,\varphi) \\
&&+ \Omega^{13}(\tau,\varphi) + \Omega^{04}(\tau,\varphi) \, ,\end{aligned}$$
having trivially that $\Omega^{00}(\tau,\varphi) = \Omega^{00}\,N(\tau,\varphi)$. Similarly, the particle-number kernel can be split according to
\[defkernels1and2bodyB\] $$\begin{aligned}
A(\tau,\varphi) &\equiv& A^{00}(\tau,\varphi) \\
&&+ A^{20}(\tau,\varphi) + A^{11}(\tau,\varphi) + A^{02}(\tau,\varphi) \, ,\end{aligned}$$
with $A^{00}(\tau,\varphi) = A^{00}\,N(\tau,\varphi)$.
Finally, use will often be made of the [*reduced*]{} kernel of an operator $O$ defined through $${\cal O}(\tau,\varphi) \equiv \frac{O(\tau,\varphi)}{N(\tau,0)} \, , \label{reducedkernels}$$ which leads, for $O={\ensuremath{\mathbbm{1}}}$, to working with [*intermediate normalization*]{} at $\varphi=0$, i.e. to having a norm kernel that satisfies ${\cal N}(\tau,0) \equiv 1$ for all $\tau$.
Master equations {#normkernel}
================
This section presents a set of master equations providing the basis of the newly proposed many-body formalism, i.e. they constitute [*exact*]{} equations of reference on top of which actual many-body expansion schemes will be designed in the remainder of the paper.
Fourier expansion of the off-diagonal kernels {#kernelexpansion}
---------------------------------------------
Inserting twice Eq. \[completeness\] into Eqs. \[defkernels\] while making use of Eqs. \[eigenequationA\], \[schroed\] and \[eq:ten:def2\], one obtains the Fourier decomposition of the kernels
\[expandedkernels\] $$\begin{aligned}
N(\tau,\varphi) &=& \sum_{\text{A} \in \mathbb{N}}\sum_{\mu} \phantom{E^{\text{A}}_\mu \,} e^{-\tau \Omega^{\text{A}}_\mu} |\langle \Phi | \Psi^{\text{A}}_{\mu} \rangle |^2 \, e^{i\text{A}\varphi} \, , \label{expandedkernels1} \\
H(\tau,\varphi) &=& \sum_{\text{A} \in \mathbb{N}}\sum_{\mu} \text{E}^{\text{A}}_\mu \, e^{-\tau \Omega^{\text{A}}_\mu} |\langle \Phi | \Psi^{\text{A}}_{\mu} \rangle |^2 \, e^{i\text{A}\varphi} \, , \label{expandedkernels2} \\
A(\tau,\varphi) &=& \sum_{\text{A} \in \mathbb{N}}\sum_{\mu} \, \, \, \text{A} \, \, \, e^{-\tau \Omega^{\text{A}}_\mu} |\langle \Phi | \Psi^{\text{A}}_{\mu} \rangle |^2 \, e^{i\text{A}\varphi} \, , \label{expandedkernels3} \\
\Omega(\tau,\varphi) &=& \sum_{\text{A} \in \mathbb{N}}\sum_{\mu} \Omega^{\text{A}}_\mu \, e^{-\tau \Omega^{\text{A}}_\mu} |\langle \Phi | \Psi^{\text{A}}_{\mu} \rangle |^2 \, e^{i\text{A}\varphi} \, , \label{expandedkernels4} \end{aligned}$$
where one trivially notices that contributions associated with $\text{A} < 0$ are zero.
Ground-state energy {#energy}
-------------------
Defining the large $\tau$ limit of a kernel via $$\begin{aligned}
O(\varphi) &\equiv & \lim\limits_{\tau \to \infty} O(\tau,\varphi) \, , \label{limitoperator}\end{aligned}$$ one obtains
\[limitkernels\] $$\begin{aligned}
N(\varphi) &=& \phantom{E^{\text{A}_0}_0 \,} e^{-\tau \Omega^{\text{A}_0}_0} |\langle \Phi | \Psi^{\text{A}_0}_0 \rangle |^2 \, e^{i\text{A}_0\varphi} \, , \label{limitnorm} \\
H(\varphi) &=& \text{E}^{\text{A}_0}_0 \, e^{-\tau \Omega^{\text{A}_0}_0} |\langle \Phi | \Psi^{\text{A}_0}_0 \rangle |^2 \, e^{i\text{A}_0\varphi} \, , \label{limitenergy} \\
A(\varphi) &=& \, \, \, \text{A}_0 \, \, e^{-\tau \Omega^{\text{A}_0}_0} |\langle \Phi | \Psi^{\text{A}_0}_0 \rangle |^2 \, e^{i\text{A}_0\varphi} \, , \label{limitA} \\
\Omega(\varphi) &=& \Omega^{\text{A}_0}_0 \, e^{-\tau \Omega^{\text{A}_0}_0} |\langle \Phi | \Psi^{\text{A}_0}_0 \rangle |^2 \, e^{i\text{A}_0\varphi} \, , \label{limitgrandpotential} \end{aligned}$$
where the residual time dependence typically disappears by eventually employing reduced kernels as defined in Eq. \[reducedkernels\]. Expressions \[limitkernels\] relate in the large-time limit off-diagonal operator kernels of interest to the off-diagonal norm kernel through eigenvalue-like equations
\[kernelequation\] $$\begin{aligned}
H(\varphi) &=& \text{E}^{\text{A}_0}_0 \, N(\varphi) \,\, , \label{kernelequation1} \\
A(\varphi) &=& \text{A}_0 \, N(\varphi) \,\, , \label{kernelequation2} \\
\Omega(\varphi) &=& \Omega^{\text{A}_0}_0 \, N(\varphi) \,\, , \label{kernelequation3}\end{aligned}$$
and similarly for reduced kernels. In Eq. \[limitkernels\], the $\varphi$ dependence originally built into the time-dependent kernels reduces to that of the single IRREP $\text{A}_0$ of the target nucleus. Additionally, the expansion coefficient in the particle-number operator kernel equates the expected value $\text{A}_0$. These characteristic features, trivially valid for the exact kernels, testify that the selected eigenstate $| \Psi^{\text{A}_0}_{0} \rangle$ of $\Omega$ (and $H$) does carry good particle number $\text{A}_0$. Accordingly, Eq. \[kernelequation\] demonstrates that the straight ratio of the operator kernels to the norm kernel accesses, at any value of $\varphi$, the eigenvalues that are in one-to-one relationship with the physical IRREP.
Let us now consider the case of actual interest where the kernels are approximated in a way that breaks $U(1)$ symmetry. In this situation, reduced kernels in the infinite time limit display the typical structure
\[limitkernelsapprox\] $$\begin{aligned}
{\cal N}_{\text{app}}(\varphi) &\equiv& \sum_{\text{A} \in \mathbb{Z}} \textmd{N}^{\text{A}}_{\text{app}} \, e^{i\text{A}\varphi} \, , \label{limitnormapprox} \\
{\cal H}_{\text{app}}(\varphi) &\equiv& \sum_{\text{A} \in \mathbb{Z}} \textmd{E}^{\text{A}}_{\text{app}}\,\textmd{N}^{\text{A}}_{\text{app}} \, e^{i\text{A}\varphi} \, , \label{limitenergyapprox} \\
{\cal A}_{\text{app}}(\varphi) &\equiv& \sum_{\text{A} \in \mathbb{Z}} \textmd{A}^{\text{A}}_{\text{app}}\,\textmd{N}^{\text{A}}_{\text{app}} \, e^{i\text{A}\varphi} \, , \label{limitAapprox} \end{aligned}$$
where the condition ${\cal N}_{\text{app}}(0) = \sum_{\text{A} \in \mathbb{Z}} \textmd{N}^{\text{A}}_{\text{app}} = 1$ characterizes intermediate normalization at gauge angle $\varphi = 0$. In Eq. \[limitkernelsapprox\], the remaining sum over $\text{A}$ signals the breaking of the symmetry induced by the approximation. The Fourier expansion \[limitkernelsapprox\] of the approximate kernels defined on $[0,2\pi]$ exists by virtue of Eq. \[decomposition\_general\]. In the expansion, the sum over the IRREPS runs a priori through $\mathbb{Z}$. If the many-body approximation scheme is well behaved from the physics standpoint, coefficients corresponding to $\text{A} < 0$ must be zero, which acts as a check that the formalism is sensible [@Bender:2008rn; @Duguet:2013dga].
Except for going back to an exact computation of the kernels, such that all the expansion coefficients but the physical one are zero in Eq. \[limitkernelsapprox\], taking the straight ratio ${\cal H}_{\text{app}}(\varphi)/{\cal N}_{\text{app}}(\varphi)$ does not provide an approximate energy that is in one-to-one correspondence with the physical IRREP $\text{A}_0$. This materializes the contamination associated with the breaking of the symmetry. However, one can take advantage of the $\varphi$ dependence built into ${\cal N}_{\text{app}}(\varphi)$, ${\cal H}_{\text{app}}(\varphi)$ and ${\cal A}_{\text{app}}(\varphi)$ to extract the component associated with that physical IRREP. Indeed, by virtue of the orthogonality of the IRREPs (Eq. \[orthogonality\]), the approximation to $\text{E}^{\text{A}_0}_0$ can be extracted as $$\begin{aligned}
\text{E}^{\text{A}_0}_0 &\approx& \frac{\int_{0}^{2\pi} \!d\varphi \,e^{-i\text{A}_0\varphi} \, {\cal H}_{\text{app}}(\varphi)}{\int_{0}^{2\pi} \!d\varphi \,e^{-i\text{A}_0\varphi} \, \, {\cal N}_{\text{app}}(\varphi)} = \textmd{E}^{\text{A}_0}_{\text{app}} \, . \label{projected_energy}\end{aligned}$$ Following the same line for the particle-number operator kernel provides a case of particular interest. Indeed, the integral over the domain of the group not only allows one to extract the component in one-to-one relationship with the physical IRREP but should be such that the expansion coefficient $\textmd{A}^{\text{A}_0}_{\text{app}}$ thus obtained is actually equal to the correct result $\text{A}_0$, i.e. it should be such that $$\begin{aligned}
\textmd{A}^{\text{A}_0}_{\text{app}} &=& \frac{\int_{0}^{2\pi} \!d\varphi \,e^{-i\text{A}_0\varphi} \, {\cal A}_{\text{app}}(\varphi)}{\int_{0}^{2\pi} \!d\varphi \,e^{-i\text{A}_0\varphi} \, \, {\cal N}_{\text{app}}(\varphi)} \, \label{projected_A}\end{aligned}$$ is indeed equal to $\text{A}_0$. This is a necessary condition to claim that the particle-number symmetry is indeed restored at any truncation order in the many-body expansion. We will see in Sec. \[Secnormkernel\] how this key demand constrains the many-body expansion scheme in a very specific way.
Whenever $| \Phi \rangle$ is taken to be a Slater determinant rather than a Bogoliubov vacuum, the targeted IRREP $\text{A}_0$ is selected [*a priori*]{} at the level of Eq. \[expandedkernels\], i.e. the gauge-angle dependence of all the kernels reduces to the IRREP $e^{i\text{A}_0\varphi}$ at any finite time $\tau$. Correspondingly, the coefficients in the Fourier expansion of the approximate kernels in Eq. \[limitkernelsapprox\] directly provide $\textmd{E}^{\text{A}_0}_{\text{app}}$ and $\textmd{A}^{\text{A}_0}_{\text{app}}=\text{A}_0$ (after division by the coefficient in the norm kernel). The integration over the domain of the $U(1)$ group becomes obviously superfluous in this case as no symmetry was broken in the first place.
Comparison with standard approaches {#standardapproaches}
-----------------------------------
Applying particle-number breaking BMBPT [@mehta1; @balian62a; @henley] or BCC theory [@StolarczykMonkhorst; @Signoracci:2014dia] amounts to expanding the [*diagonal*]{} kernels $N(0)$, $H(0)$, $A(0)$ and $\Omega(0)$ around the Bogoliubov state $| \Phi \rangle$. These methods can efficiently tackle systems characterized by a near degeneracy of the unperturbed ground state associated with a Cooper pair infra-red instability. This is done at the price of breaking global gauge invariance, even though the symmetry is restored by definition in the limit of exact calculations, i.e. when summing all diagrams. In practice, approximate kernels obtained via a truncation of the expansion mix components associated with [*different*]{} IRREPs of $U(1)$ and thus contain spurious contaminations from the symmetry standpoint. The difficulty resides here in the fact that the kernels at play do not carry any $\varphi$ dependence, i.e. Eq. \[limitkernelsapprox\] reduces in this case to
\[limitkernelsapprox0\] $$\begin{aligned}
{\cal N}_{\text{app}}(0) &=& \sum_{\text{A} \in \mathbb{Z}} \textmd{N}^{\text{A}}_{\text{app}} \, , \label{limitnormapprox0} \\
{\cal H}_{\text{app}}(0) &=& \sum_{\text{A} \in \mathbb{Z}} \textmd{E}^{\text{A}}_{\text{app}}\,\textmd{N}^{\text{A}}_{\text{app}} \, , \label{limitenergyapprox0} \\
{\cal A}_{\text{app}}(0) &=& \sum_{\text{A} \in \mathbb{Z}} \textmd{A}^{\text{A}}_{\text{app}}\,\textmd{N}^{\text{A}}_{\text{app}} \, , \label{limitAapprox0}\end{aligned}$$
such that the coefficients associated with the physical IRREP $\text{A}_0$ cannot be extracted via the integral over the domain of the $U(1)$ group. Accordingly, the key feature of the generalized approach presently proposed is to utilize [*off-diagonal*]{} kernels incorporating, from the outset, the effect of the gauge rotation $S(\varphi)$. The associated $\varphi$ dependence leaves a fingerprint of the artificial symmetry breaking built into approximate kernels that can be exploited to extract the physical components of interest through Eqs. \[projected\_energy\] and \[projected\_A\], i.e. to remove the symmetry contaminants.
Accessing neighboring nuclei at once {#yrast}
------------------------------------
Now that the benefit of performing the integral over the domain of the $U(1)$ group has been highlighted for the lowest state of the target nucleus, let us step back to Eq. \[expandedkernels\] and slightly modify the procedure to access the lowest eigenenergy $\text{E}^{\text{A}}_0$ associated with [*each*]{} IRREP, i.e. to access within the same calculation the ground state of neighboring nuclei having a non-zero overlap with the reference state $| \Phi \rangle$. To do so, we invert the order in which the limit $\tau \to \infty$ and the integral over the domain of the group are performed. We first extract the component of the time-dependent kernels associated with the specific Hilbert space ${\cal H}_{\text{A}}$ of interest
\[integratedkernels\] $$\begin{aligned}
N^{\text{A}}(\tau) &\equiv & \frac{1}{2\pi} \int_{0}^{2\pi} \!d\varphi \, e^{-i\text{A}\varphi} \, \, {\cal N}(\tau,\varphi) \nonumber \\
&=& \sum_{\mu} e^{-\tau \Omega^{\text{A}}_{\mu}} \, |\langle \Phi | \Psi^{\text{A}}_{\mu} \rangle |^2 /N(\tau,0) \, , \label{integratednorm} \\
H^{\text{A}}(\tau) &\equiv & \frac{1}{2\pi} \int_{0}^{2\pi} \!d\varphi \, e^{-i\text{A}\varphi} \, {\cal H}(\tau,\varphi) \nonumber \\
&=& \sum_{\mu} e^{-\tau \Omega^{\text{A}}_{\mu}} E^{\text{A}}_{\mu} \, |\langle \Phi | \Psi^{\text{A}}_{\mu} \rangle |^2 /N(\tau,0) \, . \label{integratedenergy}\end{aligned}$$
and take the limit $\tau \to \infty$ to access the lowest eigenenergy $\text{E}^{\text{A}}_{0}$ through $$\begin{aligned}
\text{E}^{\text{A}}_{0} &=& \lim\limits_{\tau \to \infty} \frac{H^{\text{A}}(\tau)}{N^{\text{A}}(\tau)} . \label{yrast_projected_energy}\end{aligned}$$ The above analysis is based on exact kernels respecting the symmetries and requires the extraction of the IRREP of interest prior to taking the large time limit. As explained above, the large time limit of [*approximate*]{} kernels based on a symmetry breaking reference state still mixes the IRREPS of $U(1)$. This can be used as an advantage to actually extract the ground state associated with various $\text{A}$ from the infinite time kernels, i.e. Eq.\[yrast\_projected\_energy\] is eventually replaced by Eq. \[projected\_A\] applied to $\text{A}\neq \text{A}_0$. In practice, this procedure is limited to values of $\text{A}$ whose components in the infinite time kernels are larger than a given threshold.
Everything exposed so far is valid independently of the many-body method employed to expand and truncate the off-diagonal kernels. The remainder of the paper is devoted to the computation of $N(\tau,\varphi)$, $\Omega(\tau,\varphi)$ and $A(\tau,\varphi)$ via an extension of SR-BMBPT and SR-BCC theory.
Perturbation theory {#sectionMBPT}
===================
Single-reference BCC theory starts from the similarity-transformed grand potential [@Signoracci:2014dia] or could be formulated from an energy functional [@shavitt09a]. The Baker-Campbell-Hausdorff identity applied to the similarity-transformed grand potential on the basis of standard Wick’s theorem provides the naturally terminating expansion of the reduced diagonal grand-potential kernel. Unfortunately, this property cannot be obtained directly for the [*off-diagonal*]{} operator kernels presently at play. This is due to the fact that the off-diagonal Wick theorem [@balian69a] we will rely on to expand off-diagonal matrix elements of strings of quasi-particle operators does not grant a normal ordering of the operators themselves. This feature prevents us from straightforwardly recovering the connected structure of the kernels associated with an underlying exponentiated connected cluster operator. Doing so will require a detour via the perturbative expansion of the off-diagonal kernels. With off-diagonal BMBPT at hand, it will be possible to design the off-diagonal BCC scheme in Sec. \[CCtheory\].
Unperturbed system {#chap:slater}
------------------
The grand potential is split into an unperturbed part $\Omega_{0}$ and a residual part $\Omega_1$ $$\label{split1}
\Omega = \Omega_{0} + \Omega_{1} \, ,$$ such that
$$\begin{aligned}
\Omega_{0} &\equiv& \Omega^{00}+\bar{\Omega}^{11} \\
\Omega_{1} &\equiv& \Omega^{20} + \breve{\Omega}^{11} + \Omega^{02} \nonumber \\
&& + \Omega^{22} + \Omega^{31} + \Omega^{13} + \Omega^{40} + \Omega^{04} \label{perturbation}\end{aligned}$$
where $\breve{\Omega}^{11}\equiv\Omega^{11}- \bar{\Omega}^{11}$. The term $\bar{\Omega}^{11}$ has the same formal structure as $\Omega^{11}$ and remains to be specified.
For a given number of interacting fermions, the key is to choose $\Omega_0$ with a low-enough symmetry for its ground state $| \Phi \rangle$ to be non-degenerate with respect to elementary excitations. For open-shell superfluid nuclei, this leads to choosing an operator $\Omega_0$ that breaks particle number conservation, i.e. while $\Omega$ commutes with transformations of $U(1)$, we are interested in the case where $\Omega_0$, and thus $\Omega_1$, do [*not*]{} commute with $S(\varphi)$, i.e.
\[commutators\] $$\begin{aligned}
\left[\Omega_0,S(\varphi)\right]&\neq& 0 \, , \label{commutators2} \\
\left[\Omega_1,S(\varphi)\right]&\neq& 0 \, . \label{commutators3} \end{aligned}$$
In this context, the vacuum $| \Phi \rangle$ is a Bogoliubov state that is [*deformed*]{} in gauge space and that is thus not an eigenstate of $A$; i.e. it spans several IRREPs of $U(1)$.
The operator $\Omega_{0}$ can be written in diagonal form in terms of its one quasi-particle eigenstates $$\Omega_{0} \equiv \Omega^{00} + \sum_{k} E_k \beta^{\dagger}_k \beta_k \label{hzero} \, ,$$ with $E_k > 0$ for all $k$. Eventually, it remains to specify how the quasi-particle operators $\{\beta_k ; \beta^{\dagger}_k\}$ and energies $\{E_k\}$ are determined. This corresponds to fixing the Bogoliubov transformation $W$ (Eq. \[bogomatrix\]), and thus $| \Phi \rangle$, along with $\bar{\Omega}^{11}$. The traditional choice consists in requiring that $| \Phi \rangle$ minimizes $\Omega^{00}$, which amounts to solving so-called Hartree-Fock-Bogoliubov (HFB) equations [@ring80a] to fix both $W$ and the set of $E_k$. This corresponds to working within a Moller-Plesset scheme. While this choice conveniently leads to canceling $\Omega^{20}$ and $\Omega^{02}$ in $\Omega$ while diagonalizing $\bar{\Omega}^{11}=\Omega^{11}$, we do not impose such a choice in the present work in order to design the formalism in its general Rayleigh-Schroedinger form.
Introducing many-body states generated via an even number of quasi-particle excitations[^11] of the vacuum $$\begin{aligned}
| \Phi^{k_1 k_2\ldots} \rangle &\equiv& {\cal B}_{k_1 k_2\ldots} | \Phi \rangle \, , \end{aligned}$$ where $$\begin{aligned}
{\cal B}_{k_1 k_2\ldots} &\equiv& \beta^{\dagger}_{k_1} \, \beta^{\dagger}_{k_2} \, \ldots \, , \label{phexcitation}\end{aligned}$$ the unperturbed system is fully characterized by its complete set of eigenstates
$$\begin{aligned}
\Omega_{0}\, | \Phi \rangle &=& \Omega^{00} \, | \Phi \rangle \, , \\
\Omega_{0}\, | \Phi^{k_1 k_2\ldots} \rangle &=& \left[\Omega^{00} \!+\! E_{k_1} \!+\! E_{k_2}\!+\!\ldots\right] | \Phi^{k_1 k_2\ldots} \rangle \label{phi} \, . \; \;\end{aligned}$$
As mentioned above, the Bogoliubov vacuum $| \Phi \rangle$ necessarily possesses a [*closed-shell*]{} character with respect to elementary (quasi-particle) excitations. This means that there exists a finite energy gap between the vacuum state and the lowest two quasi-particle excitations, i.e. $E_{k_1}+E_{k_1}\geq 2\Delta_{\text{F}} > 0$ for all $(k_1,k_2)$, where $\Delta_{\text{F}}$ is traditionally characterized as the pairing gap.
Rotated reference state {#chap:rotatedbasis}
-----------------------
Particle creation and annihilation operators are tensor operators of rank $+1$ and $-1$, respectively. As a result, they transform under gauge rotation according to
$$\begin{aligned}
c_{\bar{p}} &\equiv& S(\varphi) \, c_{p} \, S^{-1}(\varphi) = e^{-i\varphi} \, c_{p} \, , \\
c_{\bar{p}}^{\dagger} &\equiv& S(\varphi) \, c^{\dagger}_{p} \, S^{-1}(\varphi) = e^{+i\varphi} \, c_{p}^{\dagger} \, ,\end{aligned}$$
where $S_{pq}(\varphi) \equiv \langle p | S(\varphi) | q \rangle = e^{i\varphi} \delta_{pq}$ is the unitary transformation matrix connecting the rotated particle basis to the unrotated one. This leads to rotated quasi-particle operators[^12]
$$\begin{aligned}
\beta_{\bar{k}} &\equiv& S(\varphi) \, \beta_{k} \, S^{-1}(\varphi) \nonumber \\
&=& \sum_{p} U^{*}_{pk} \, e^{-i\varphi} \, c_{p} + V^{*}_{pk} \, e^{+i\varphi} \, c^{\dagger}_{p} \, , \\
\beta_{\bar{k}}^{\dagger} &\equiv& S(\varphi) \, \beta_{k}^{\dagger} \, S^{-1}(\varphi) \nonumber \\
&=& \sum_{p} U_{pk} \, e^{+i\varphi} \, c^{\dagger}_{p} + V_{pk} \, e^{-i\varphi} \, c_{p} \, ,\end{aligned}$$
which are used to specify the rotated partner of $| \Phi \rangle$ under the form
$$\begin{aligned}
| \Phi (\varphi) \rangle &\equiv& S(\varphi) | \Phi \rangle \\
&=& \mathcal{C} \displaystyle \prod_{\bar{k}} \beta_{\bar{k}} | 0 \rangle \, .\end{aligned}$$
By virtue of Thouless’ theorem [@thouless60], the rotated state $| \Phi (\varphi) \rangle$ is itself a Bogoliubov state associated with the transformation
$$\begin{aligned}
W^{\varphi} &\equiv& \left(
\begin{array} {cc}
U^{\varphi} & V^{\varphi \ast} \\
V^{\varphi} & U^{\varphi \ast}
\end{array}
\right) \\
&=&
\left(
\begin{array} {cc}
e^{+i\varphi} U & e^{+i\varphi} V^{\ast} \\
e^{-i\varphi} V & e^{-i\varphi}U^{\ast}
\end{array}
\right)\, , \label{rotbogomatrix}\end{aligned}$$
that leads to defining the skew-symmetrix matrix
$$\begin{aligned}
Z^{\varphi} &\equiv& V^{\varphi \ast}[U^{\varphi \ast}]^{-1} \\
&=& e^{2i\varphi} Z\, .\end{aligned}$$
State $| \Phi(\varphi) \rangle$ is the ground-state of the rotated Hamiltonian $\Omega_0(\varphi)\equiv S(\varphi) \Omega_0 S^{-1}(\varphi)$ with the $\varphi$-independent eigenvalue $\Omega^{00}$. This feature characterizes the fact that, while the unperturbed ground-state is non-degenerate with respect to quasi-particle excitations, there exists a degeneracy, i.e. a zero mode, in the manifold of its gauge rotated partners. In other words, breaking $U(1)$ symmetry commutes the degeneracy of the unperturbed state with respect to individual excitations into a degeneracy with respect to collective rotations in gauge space. Lifting the latter degeneracy is eventually necessary for a finite quantum systems and it is the objective of the present work to do so within the frame of BMBPT and BCC theory.
Unperturbed off-diagonal norm kernel {#transitionover}
------------------------------------
As proven in Ref. [@Robledo:2009yd], the overlap between $| \Phi \rangle$ and $| \Phi (\varphi) \rangle$ is best expressed as a Pfaffian $$\langle \Phi | \Phi(\varphi) \rangle = |\mathcal{C}|^{2} (-1)^{N(N+1)/2} \textrm{pf}\left(\begin{array}{cc}
Z^{\varphi} & -1\\
1 & -Z^{\ast}
\end{array}\right) \, , \label{kernel}$$ with $N$ the (even) dimension of the (truncated) one-body Hilbert space ${\cal H}_1$ spanned by the basis $\{c_{p};c^{\dagger}_{p}\}$. In case both states $| \Phi \rangle$ and $| \Phi (\varphi) \rangle$ share a common discrete symmetry like simplex or time reversal[^13], the overlap can be reduced to a determinant [@Robledo:2009yd] without any loss of its sign.
In Sec. \[newunperturbedkernel\], a new alternative to Eq. \[kernel\] will be proposed.
Unperturbed off-diagonal density matrix {#transitiondens}
---------------------------------------
The transformation that links $| \Phi \rangle$ and $| \Phi (\varphi) \rangle$ is itself a Bogoliubov transformation built as the product $W^{\varphi \dagger}W$ such that
\[e:p2qpAA\] $$\begin{aligned}
\beta_{k_1} &= \sum_{k_2} A^{*}_{k_2k_1}(\varphi) \, \beta_{\bar{k}_2}
+ B^{*}_{k_2k_1}(\varphi) \, \beta^{\dagger}_{\bar{k}_2} \, , \\
\beta_{k_1}^{\dagger} &= \sum_{k_2} A_{k_2k_1}(\varphi) \, \beta^{\dagger}_{\bar{k}_2}
+ B_{k_2k_1}(\varphi) \, \beta_{\bar{k}_2} \, ,\end{aligned}$$
where
$$\begin{aligned}
A(\varphi) &\equiv& U^{\varphi \dagger}U + V^{\varphi \dagger}V \, , \\
B(\varphi) &\equiv& V^{\varphi T}U + U^{\varphi T}V \, .\end{aligned}$$
Having defined this transitional Bogoliubov transformation, one introduces the [*off-diagonal*]{} generalized density matrix expressed in the one-body basis as [@ring80a]
\[offdiaggeneralizeddensitymatrix\] $$\begin{aligned}
{\cal R}(\varphi) &\equiv&
\left(
\begin{array} {cc}
\frac{\langle \Phi | c^{\dagger}c^{\phantom{\dagger}} | \Phi(\varphi) \rangle}{\langle \Phi | \Phi(\varphi) \rangle} & \frac{\langle \Phi | c^{\phantom{\dagger}}c^{\phantom{\dagger}} | \Phi(\varphi) \rangle}{\langle \Phi | \Phi(\varphi) \rangle} \\
\frac{\langle \Phi | c^{\dagger}c^{\dagger} | \Phi(\varphi) \rangle}{\langle \Phi | \Phi(\varphi) \rangle} & \frac{\langle \Phi | c^{\phantom{\dagger}}c^{\dagger} | \Phi(\varphi) \rangle}{\langle \Phi | \Phi(\varphi) \rangle}
\end{array}
\right) \\
&\equiv&
\left(
\begin{array} {cc}
+\rho(\varphi) & +\kappa(\varphi) \\
-\bar{\kappa}^{\ast}(\varphi) & -\sigma^{\ast}(\varphi)
\end{array}
\right) \\
&=&
\left(
\begin{array} {cc}
V^{\varphi \ast}[A^{T}(\varphi)]^{-1}V^T & V^{\varphi \ast}[A^{T}(\varphi)]^{-1}U^T \\
U^{\varphi \ast}[A^{T}(\varphi)]^{-1}V^T & U^{\varphi \ast}[A^{T}(\varphi)]^{-1}U^T
\end{array}
\right)
\, , \end{aligned}$$
which, after transformation to the quasi-particle basis of $| \Phi \rangle$, becomes
\[offdiaggeneralizeddensitymatrix2\] $$\begin{aligned}
{\bf R}(\varphi) &=&
\left(
\begin{array} {cc}
\frac{\langle \Phi | \beta^{\dagger}\beta^{\phantom{\dagger}} | \Phi(\varphi) \rangle}{\langle \Phi | \Phi(\varphi) \rangle} & \frac{\langle \Phi | \beta^{\phantom{\dagger}}\beta^{\phantom{\dagger}} | \Phi(\varphi) \rangle}{\langle \Phi | \Phi(\varphi) \rangle} \\
\frac{\langle \Phi | \beta^{\dagger}\beta^{\dagger} | \Phi(\varphi) \rangle}{\langle \Phi | \Phi(\varphi) \rangle} & \frac{\langle \Phi | \beta^{\phantom{\dagger}}\beta^{\dagger} | \Phi(\varphi) \rangle}{\langle \Phi | \Phi(\varphi) \rangle}
\end{array}
\right) \\
&\equiv&
\left(
\begin{array} {cc}
R^{+-}(\varphi) & R^{--}(\varphi) \\
R^{++}(\varphi) & R^{-+}(\varphi)
\end{array}
\right) \\
&=&
\left(
\begin{array} {cc}
0 & B^{\dagger}(\varphi)[A^{T}(\varphi)]^{-1} \\
0 & 1
\end{array}
\right)
\, , \end{aligned}$$
with $$\begin{aligned}
R^{--}(\varphi)&=& B^{\dagger}(\varphi)[A^{T}(\varphi)]^{-1} \label{upperright} \\
&=& V^{\dagger} (1-e^{2i\varphi})(1-e^{2i\varphi}Z^{\ast}Z)^{-1}[U^{T}]^{-1}\, . \nonumber\end{aligned}$$ We note that ${\bf R}(0) = {\bf R}$ and that
\[propR\] $$\begin{aligned}
R^{--}(\varphi)&=& - R^{-- T}(\varphi) \, , \\
R^{--}(0) &=& 0 \, .\end{aligned}$$
Unperturbed off-diagonal propagator {#prop}
-----------------------------------
Quasi-particle creation and annihilation operators read in the interaction representation
\[aalphatau\] $$\begin{aligned}
\beta_{k} \left(\tau\right) &\equiv& e^{+\tau \Omega_{0}} \, \beta_{k} \, e^{-\tau \Omega_{0}}=e^{-\tau E_k} \, \beta_{k} \label{aalphatau2} \, , \\
\beta_{k}^{\dagger}\left(\tau\right) &\equiv& e^{+\tau \Omega_{0}} \, \beta_{k}^{\dagger} \, e^{-\tau
\Omega_{0}}=e^{+\tau E_{k}} \, \beta_{k}^{\dagger} \, . \label{aalphatau1}\end{aligned}$$
\
The generalized unperturbed off-diagonal one-body propagator is introduced as a $2\times 2$ matrix in Bogoliubov space $$\begin{aligned}
\bold{G}^{0}(\varphi)
&\equiv&
\left(
\begin{array} {cc}
G^{+- (0)}(\varphi) & G^{-- (0)}(\varphi) \\
G^{++ (0)}(\varphi) & G^{-+ (0)}(\varphi)
\end{array}
\right) \label{offdiaggeneralizedprog}\end{aligned}$$ whose four components are defined through their matrix elements in the quasi-particle basis $\{\beta_{k};\beta^{\dagger}_{k}\}$ according to
\[propagatorsA\] $$\begin{aligned}
G^{+- (0)}_{k_1k_2}(\tau_1, \tau_2 ; \varphi) &\equiv& \frac{\langle \Phi | \textmd{T}[\beta^{\dagger}_{k_1}(\tau_1) \beta_{k_2}(\tau_2)] | \Phi(\varphi) \rangle}{\langle \Phi | \Phi(\varphi) \rangle} \, , \label{propagatorsA1} \; \; \;\\
G^{-- (0)}_{k_1k_2}(\tau_1, \tau_2 ; \varphi) &\equiv& \frac{\langle \Phi | \textmd{T}[\beta_{k_1}(\tau_1) \beta_{k_2}(\tau_2)] | \Phi(\varphi) \rangle}{\langle \Phi | \Phi(\varphi) \rangle} \, , \label{propagatorsA2} \; \; \;\\
G^{++ (0)}_{k_1k_2}(\tau_1, \tau_2 ; \varphi) &\equiv& \frac{\langle \Phi | \textmd{T}[\beta^{\dagger}_{k_1}(\tau_1) \beta^{\dagger}_{k_2}(\tau_2)] | \Phi(\varphi) \rangle}{\langle \Phi | \Phi(\varphi) \rangle} \, , \label{propagatorsA3} \; \; \;\\
G^{-+ (0)}_{k_1k_2}(\tau_1, \tau_2 ; \varphi) &\equiv& \frac{\langle \Phi | \textmd{T}[\beta_{k_1}(\tau_1) \beta^{\dagger}_{k_2}(\tau_2)] | \Phi(\varphi) \rangle}{\langle \Phi | \Phi(\varphi) \rangle} \, , \label{propagatorsA4} \; \; \;\end{aligned}$$
where $\textmd{T}$ denotes the time ordering operator. The diagrammatic representation of the four elementary propagators $G^{gg' (0)}(\varphi)$, with $g\equiv \pm$ and $g'\equiv \pm$, is provided in Fig. \[prop1\]. The above definition of propagators implies the relations
\[relat\] $$\begin{aligned}
G^{+- (0)}_{k_1k_2}(\tau_1, \tau_2 ; \varphi) &=& -G^{+- (0)}_{k_2k_1}(\tau_2, \tau_1 ; \varphi) \, , \label{relat1} \; \; \;\\
G^{-- (0)}_{k_1k_2}(\tau_1, \tau_2 ; \varphi) &=& -G^{-- (0)}_{k_2k_1}(\tau_2, \tau_1 ; \varphi) \, , \label{relat2} \; \; \;\\
G^{++ (0)}_{k_1k_2}(\tau_1, \tau_2 ; \varphi) &=& -G^{++ (0)}_{k_2k_1}(\tau_2, \tau_1 ; \varphi). \; \; \;\end{aligned}$$
Combining Eqs. \[offdiaggeneralizeddensitymatrix2\], \[propR\] and \[aalphatau\], together with anticommutation rules of quasi-particle creation and annihilation operators, one obtains that
\[propagatorsB\] $$\begin{aligned}
G^{+- (0)}_{k_1k_2}(\tau_1, \tau_2 ; \varphi) &=& - e^{-(\tau_2-\tau_1)E_{k_1}} \theta(\tau_2-\tau_1) \delta_{k_1k_2} \, , \label{propagatorsB1}\; \; \; \; \; \; \\
G^{-- (0)}_{k_1k_2}(\tau_1, \tau_2 ; \varphi) &=& + e^{-\tau_1 E_{k_1}} e^{-\tau_2 E_{k_2}} R^{--}_{k_1k_2}(\varphi) \, , \label{propagatorsB2} \\
G^{++ (0)}_{k_1k_2}(\tau_1, \tau_2 ; \varphi) &=& 0 \, , \label{propagatorsB3} \\
G^{-+ (0)}_{k_1k_2}(\tau_1, \tau_2 ; \varphi) &=& + e^{-(\tau_1-\tau_2)E_{k_1}} \theta(\tau_1-\tau_2) \delta_{k_1k_2} \label{propagatorsB4} \, , \; \; \; \; \; \;\end{aligned}$$
where only $G^{-- (0)}(\varphi)$ actually depends on the gauge angle $\varphi$ and is such that $G^{-- (0)}(0)=0$.
The equal-time unperturbed propagator deserves special attention. Equal-time propagators will solely arise from contracting two quasi-particle operators belonging to the same normal-ordered operator displaying creation operators to the left of annihilation ones. In both $G^{+- (0)}_{k_1k_2}(\tau, \tau ; \varphi)$ and $G^{-+ (0)}_{k_1k_2}(\tau, \tau ; \varphi)$, this necessarily leads to selecting a contraction associated with $R^{+-}(\varphi)$ that is identically zero. As a result, a non-zero equal-time propagator is always of the anomalous type[^14], i.e.
\[propagatorsC\] $$\begin{aligned}
G^{+- (0)}_{k_1k_2}(\tau, \tau ; \varphi) &\equiv& 0 \, , \label{propagatorsC1} \\
G^{-- (0)}_{k_1k_2}(\tau, \tau ; \varphi) &\equiv& + e^{-\tau (E_{k_1}+E_{k_2})} R^{--}_{k_1k_2}(\varphi) \, , \label{propagatorsC2} \\
G^{++ (0)}_{k_1k_2}(\tau, \tau ; \varphi) &\equiv& 0 \, , \label{propagatorsC3} \\
G^{-+ (0)}_{k_1k_2}(\tau, \tau ; \varphi) &\equiv& 0 \label{propagatorsC4} \, ,\end{aligned}$$
such that no equal-time contraction, and thus no contraction of an interaction vertex onto itself, can occur in the diagonal case, i.e. for $\varphi=0$.
Expansion of the evolution operator
-----------------------------------
As recalled in App. \[perturbativeannexe\], the evolution operator can be expanded in powers of $\Omega_{1}$ under the form $$\begin{aligned}
{\cal U}(\tau) &=& e^{-\tau \Omega_{0}} \, \textmd{T}e^{-\int_{0}^{\tau}dt \Omega_{1}\left(t\right) \, ,
} \label{evol1}$$ where[^15] $$\Omega_{1}\left( \tau\right) \equiv e^{\tau \Omega_{0}}\Omega_{1}e^{-\tau \Omega_{0}} \, ,$$ defines the perturbation in the interaction representation.
Off-diagonal norm kernel {#normkernel2}
------------------------
### Off-diagonal BMBPT expansion
Expressing $\Omega_1$ in the eigenbasis of $\Omega_0$ and expanding the exponential in Eq. \[evol1\] in power series, one obtains the perturbative expansion of the off-diagonal norm kernel
$$\begin{aligned}
N(\tau,\varphi) &=& \langle \Phi | e^{-\tau \Omega_{0}} \, \textmd{T}e^{-\int_{0}^{\tau}dt \Omega_{1}\left(t\right)} | \Phi(\varphi) \rangle \label{expansionnormkernel} \\
&=& e^{-\tau\Omega^{00}} \langle \Phi |\Big\{ 1-\int_{0}^{\tau}d\tau_1 \Omega_{1}\left( \tau_1\right) +\frac{1}{2!}\int_{0}^{\tau}d\tau_{1}d\tau
_{2}\textmd{T}\left[ \Omega_{1}\left( \tau_{1}\right) \Omega_{1}\left( \tau_{2}\right) \right]
+... \Big\}| \Phi(\varphi) \rangle \nonumber\\
&=& e^{-\tau\Omega^{00}}\Big\{ \sum_{p=0}^{\infty}\frac{(-1)^{p}}{p!} \sum_{\substack{i_1+j_1=2,4 \\ \vdots \\i_p+j_p=2,4}} \int_{0}^{\tau}\!\!d\tau_{1}\ldots d\tau_{p} \hspace{-0.3cm}\sum_{\substack{k_1 \ldots k_{i_1} \\k_{i_1+1} \ldots k_{i_1+j_1} \\ \vdots \\ l_1 \ldots l_{i_p} \\l_{i_p+1} \ldots l_{i_p+j_p}}} \hspace{-0.3cm} \frac{\Omega^{i_1j_1}_{k_1 \ldots k_{i_1} k_{i_1+1} \ldots k_{i_1+j_1}}}{(i_1)!(j_1)!} \ldots \frac{\Omega^{i_pj_p}_{l_1 \ldots l_{i_p} l_{i_p+1} \ldots l_{i_p+j_p}}}{(i_p)!(j_p)!} \nonumber\\
&& \hspace{0.2cm} \times \langle \Phi | \textmd{T}\left[ \beta^{\dagger}_{k_1}\left( \tau_{1}\right)\ldots \beta^{\dagger}_{k_{i_1}}\left( \tau_{1}\right)\beta_{k_{i_1+j_1}}\left( \tau_{1}\right) \ldots \beta_{k_{i_1+1}}\left( \tau_{1}\right) \ldots \beta^{\dagger}_{l_1}\left( \tau_{p}\right)\ldots \beta^{\dagger}_{l_{i_p}}\left( \tau_{p}\right)\beta_{l_{i_p+j_p}}\left( \tau_{p}\right) \ldots \beta_{l_{i_p+1}}\left( \tau_{p}\right)
\right] | \Phi(\varphi) \rangle \Big\} . \nonumber \end{aligned}$$
where $\breve{\Omega}^{11}$ must be understood in place of $\Omega^{11}$ whenever necessary.
The off-diagonal matrix elements of products of time-dependent field operators appearing in Eq. \[expansionnormkernel\] can be expressed as the sum of all possible systems of products of elementary contractions $G^{gg' (0)}_{k_1k_2}(\tau_1, \tau_2 ; \varphi)$ (Eqs. \[propagatorsA\]-\[propagatorsC\]), eventually multiplied by the unperturbed norm kernel $\langle \Phi | \Phi(\varphi) \rangle$ (Eq. \[kernel\]). This derives from a generalized Wick theorem [@balian69a] applicable to matrix elements between [*different*]{} (non-orthogonal) left and right vacua, i.e. presently $\langle \Phi |$ and $| \Phi(\varphi) \rangle$, which constitutes a powerful way to deal exactly with the presence of the rotation operator $S(\varphi)$ in off-diagonal kernels. Eventually, this makes possible to represent $N(\tau,\varphi)$ diagrammatically following techniques [@blaizot86] usually applied to the diagonal norm kernel $N(\tau,0)$ [@bloch58a].
### Off-diagonal BMBPT diagrammatic rules {#diagrulenormMBPT}
Equation \[expansionnormkernel\] for $N(\tau,\varphi)$ can be translated into an infinite set of vacuum-to-vacuum diagrams. The rules to build and compute those diagrams are now detailed.
1. A vacuum-to-vacuum, i.e. closed, Feynman diagram of order $p$ consists of $p$ vertices $\Omega^{ij}(\tau_k)$ connected by fermionic quasi-particle lines, i.e. elementary propagators $G^{gg' (0)}(\varphi)$, forming a set of closed loops.
2. Each vertex is labeled by a time variable while each line is labeled by two quasi-particle indices and two time labels at its ends, the latter being associated with the two vertices the line is attached to. Each vertex contributes a factor $\Omega^{ij}_{k_1 \ldots k_i k_{i+1} \ldots k_{i+j}}$ with the sign convention detailed in Sec. \[diagramsforvertices\]. Each line contributes a factor $G^{gg' (0)}_{k_1k_2}(\tau_k, \tau_{k'} ; \varphi)$, where $g=\pm$ and $g'=\pm$ characterize the type of elementary propagator the line corresponds to[^16].
3. The contributions to $N(\tau,\varphi)$ of order $p$ are generated by drawing all possible vacuum-to-vacuum diagrams involving $p$ operators $\Omega_1(\tau_k)$. This is done by contracting the quasi-particle lines attached to the vertices in all possible ways, allowing both for normal and anomalous propagators. Eventually, the set of diagrams must be limited to [*topologically distinct*]{} diagrams, i.e. diagrams that cannot be obtained from one another via a mere displacement, i.e. translation, of the vertices.
4. All quasi-particle labels must be summed over while all time variables must be integrated over from $0$ to $\tau$.
5. A sign factor $(-1)^{p+n_c}$, where $p$ denotes the order of the diagram and $n_c$ denotes the number of crossing lines in the diagram, must be considered. The overall sign results from multiplying the latter factor with the sign associated with each factor $\Omega^{ij}_{k_1 \ldots k_i k_{i+1} \ldots k_{i+j}}$ as discussed above.
6. Each diagram comes with a numerical prefactor obtained from the following combination
- A factor $1/(n_e)!$ must be considered for [*each*]{} group of $n_e$ equivalent lines. Equivalent lines must all begin and end at the same vertices (or vertex, for anomalous propagators starting and ending at the same vertex), and must correspond to the same type of contractions, i.e. they must all correspond to propagators characterized by the same superscripts $g$ and $g'$ in addition to having identical time labels.
- Given the previous rule, an extra factor $1/2$ must be considered for [*each*]{} anomalous propagator that starts and ends at the same vertex. The proof of this unusual[^17] diagrammatic rule, already used in Ref. [@soma11a], is given in App. \[ruleanomalouscontraction\].
- A symmetry factor $1/n_s$ must be considered in connection with exchanging the time labels of the vertices in all possible ways. The factor $n_s$ corresponds to the number of ways exchanging the time labels provides a diagram that is topologically equivalent to the original one.
As each operator $\Omega_1(\tau)$ actually contains eight normal-ordered operators $\Omega^{ij}(\tau)$, with $i+j=2,4$, and given that four types of propagators must be considered, one may be worried about the proliferation of diagrams. Whereas the number of diagrams to be considered is indeed significantly larger than in standard, i.e. diagonal ($\varphi=0$), BMBPT, several “selection rules” can be identified by virtue of Eqs. \[propagatorsB\] and \[propagatorsC\] that limit drastically the number of non-zero diagrams. Let us detail those additional rules.
1. As $G^{++ (0)}(\varphi)$ is identically zero, non-zero anomalous off-diagonal contractions necessarily involve two quasi-particle [*annihilation*]{} operators, i.e. the diagram is identically zero anytime a contraction between two creation operators is considered. Whenever a string of operators contain more creation operators than annihilation operators, the result is thus necessarily zero, i.e. for an arbitrary matrix element $\langle \Phi | \Omega^{i_1j_1}(\tau_1) \Omega^{i_2j_2}(\tau_2) \ldots \Omega^{i_pj_p}(\tau_p) | \Phi(\varphi) \rangle$ to give non-zero contributions (diagrams), it is mandatory that $n_a = \sum_{k=1}^{p}(j_k-i_k) \ge 0$. Corresponding diagrams must contain exactly $n_a$ anomalous contractions to provide a non-zero result. Diagrams at play in diagonal ($\varphi=0$) BMBPT reduce to those characterized by $n_a=0$ as no anomalous contraction occurs in this case (i.e. $G^{-- (0)}(0)=0$).
2. Normal lines linking two given vertices $\Omega^{i_kj_k}(\tau_k)$ and $\Omega^{i_{k'}j_{k'}}(\tau_{k'})$ must propagate in the same direction. Equations \[propagatorsB1\] and \[propagatorsB4\] indeed indicate that two normal lines propagating in opposite directions induce a factor $\theta(\tau_k-\tau_{k'})\theta(\tau_{k'}-\tau_k)$ that makes the diagram to be zero.
3. As Eq. \[propagatorsC\] demonstrates, propagators starting and ending at the same vertex are necessarily of anomalous, i.e. $G^{-- (0)}(\varphi)$, type.
### Exponentiation of connected diagrams
Diagrams representing the off-diagonal norm kernel are vacuum-to-vacuum diagrams, i.e. diagrams with no incoming or outgoing external lines. In general, a diagram consists of disconnected parts which are joined neither by vertices nor by propagators. Consider a diagram $\Gamma(\tau,\varphi)$ contributing to Eq. \[expansionnormkernel\] and consisting of $n_{1}$ identical connected parts $\Gamma_{1}(\tau,\varphi)$, of $n_{2}$ identical connected parts $\Gamma_{2}(\tau,\varphi)$, and so on. Using for simplicity the same symbol to designate both the diagram and its contribution, the whole diagram gives $$\Gamma(\tau,\varphi) = \frac{\left[\Gamma_{1}(\tau,\varphi)\right]^{n_{1}}}{n_{1}!}\frac{\left[ \Gamma_{2}(\tau,\varphi)\right]^{n_{2}}}{n_{2}!}...$$ The factor $n_{i}!$ is the symmetry factor due to the exchange of time labels among the $n_{i}$ identical diagrams $\Gamma_{i}(\tau,\varphi)$. It follows that the sum of all vacuum-to-vacuum diagrams is equal to the exponential of the sum of [*connected*]{} vacuum-to-vacuum diagrams $$\begin{aligned}
\sum_{\Gamma} \Gamma(\tau,\varphi) &=& \sum_{n_{1}n_{2}...}\frac{\left[ \Gamma_{1}(\tau,\varphi)\right]^{n_{1}}}{n_{1}!}\frac{\left[\Gamma_{2}(\tau,\varphi)\right]^{n_{2}}}{n_{2}!}... \nonumber \\
&=& e^{\Gamma_{1}(\tau,\varphi)+\Gamma_{2}(\tau,\varphi)+...} \label{condiag} \, .\end{aligned}$$ Consequently, the norm can be written as $$N(\tau,\varphi) = e^{-\tau \Omega^{00} + n(\tau,\varphi)} \, \langle \Phi | \Phi(\varphi) \rangle \label{wexp3} \, ,$$ where $n(\tau,\varphi)\equiv\sum^{\infty}_{n=1}n^{(n)}(\tau,\varphi)$, with $n^{(n)}(\tau,\varphi)$ the sum of all $\varphi$-dependent connected vacuum-to-vacuum diagrams of order $n$. By virtue of Eqs. \[condiag\]-\[wexp3\], only connected diagrams have to be eventually considered in practice.
### Computing diagrams {#computingdiagramsNc}
![ \[diagramsNc\] First- and second-order connected Feynman diagrams contributing to $n(\tau,\varphi)$. Vertices referring to $\Omega^{11}$, i.e. the lower vertex in diagrams PN.4 and PN.15, must actually be understood as referring to $\breve{\Omega}^{11}$.](figures/pnrmbptN){width="1.03\columnwidth"}
The eighteen non-zero first- and second-order connected vacuum-to-vacuum diagrams contributing to $n(\tau,\varphi)$ are displayed in Fig. \[diagramsNc\], where they are classified according to the value of $n_a$, i.e. according to the number of anomalous lines they contain.
Choosing the reference state $|\Phi\rangle$ to be the solution of HFB equations amounts to setting $\breve{\Omega}^{11}=\Omega^{20}=\Omega^{02}=0$ such that diagrams PN.1, PN.3-PN.5 and PN.11-PN.15 are zero in the Moller-Plesset scheme, i.e. the set reduces from eighteen non-zero diagrams to nine non-zero diagrams at second order. Finally, $n(\tau,0)$ at play in diagonal BMBPT reduces to PN.3 and PN.6 ($n_a=0$) diagrams at second order (only PN.6 in the Moller-Plesset scheme).
![ \[diagramsNcex\] Example of a fully-labeled off-diagonal BMBPT diagram contributing to $n(\tau,\varphi)$, i.e. the second-order diagram labeled PN.8 in Fig. \[diagramsNc\].](figures/pnrmbptNex){width="0.35\columnwidth"}
While the full analytic expression of each of these diagrams is provided in App. \[diagramsN\], we presently detail the calculation of one of them for illustration. The second-order diagram labeled as PN.8 in Fig. \[diagramsNc\] is displayed in detail in Fig. \[diagramsNcex\]. It contains one $\Omega^{22}$ vertex and one $\Omega^{04}$ vertex. The diagram contains two anomalous lines ($n_a=2$), two vertices and no crossing lines ($(-1)^{p+n_c}=+1$), two equivalent lines of normal type propagating in the same direction along with two equivalent anomalous lines ($n_e=4$), and a symmetry factor $n_s=1$ as exchanging the time labels of the two vertices gives topologically distinct diagrams. Last but not least, the sign convention for the vertices requires to associate the factors $+\Omega^{22}_{k_1k_2k_3k_4}$ and $+\Omega^{04}_{k_5k_6k_7k_8}$ to the vertices as they appear on the diagram drawn in Fig. \[diagramsNcex\]. Eventually, diagram PN.8 reads as
$$\begin{aligned}
\text{PN}.8=& +\left(\frac{1}{2}\right)^2 \displaystyle\sum_{\substack{k_1 k_2 k_3 k_4 \\ k_5 k_6k_7k_8}} \hspace{-0.3cm} \Omega^{22}_{k_1k_2k_3k_4} \Omega^{04}_{k_5k_6k_7k_8} \int\limits_0^{\tau} \mathrm{d} \tau_1 \mathrm{d} \tau_2 G^{+-(0)}_{k_1k_5}(\tau_1, \tau_2; \varphi) G^{+-(0)}_{k_2k_6}(\tau_1, \tau_2; \varphi) G^{--(0)}_{k_3k_8}(\tau_1, \tau_2; \varphi) G^{--(0)}_{k_4k_7}(\tau_1, \tau_2; \varphi) \nonumber \\
=& +\frac{1}{4} \displaystyle\sum_{\substack{k_1 k_2 k_3 k_4 \\ k_5 k_6}} \hspace{-0.3cm}\Omega^{22}_{k_1k_2k_3k_4} \Omega^{04}_{k_1k_2k_5k_6} R^{--}_{k_3k_6}(\varphi) R^{--}_{k_4k_5}(\varphi) \int\limits_0^{\tau} \mathrm{d} \tau_1 \mathrm{d} \tau_2 \theta(\tau_2-\tau_1) e^{-\tau_2(E_{k_1}+E_{k_2}+E_{k_5}+E_{k_6}) -\tau_1(E_{k_3}+E_{k_4}-E_{k_1}-E_{k_2})} \nonumber \\
=& +\frac{1}{4} \displaystyle\sum_{\substack{k_1 k_2 k_3 k_4 \\ k_5 k_6}} \frac{\Omega^{22}_{k_1k_2k_3k_4} \Omega^{04}_{k_1k_2k_5k_6}}{E_{k_1}+E_{k_2}-E_{k_3}-E_{k_4}}\left[ \frac{1- e^{-\tau(E_{k_3}+E_{k_4}+E_{k_5}+E_{k_6})}}{E_{k_3}+E_{k_4}+E_{k_5}+E_{k_6}} \right.\nonumber \\
& \hspace{7cm} \left.- \frac{1- e^{-\tau(E_{k_1}+E_{k_2}+E_{k_5}+E_{k_6})}}{E_{k_1}+E_{k_2}+E_{k_5}+E_{k_6}}\right] R^{--}_{k_3k_6}(\varphi) R^{--}_{k_4k_5}(\varphi) \, ,
\label{PNex1}\end{aligned}$$
where use was made of the identities provided in App. \[usefulID\]. We note that the first line of Eq. \[PNex1\] was obtained by reading the diagram from bottom to top, i.e. in a ascendant fashion, in Fig. \[diagramsNcex\]. In the infinite $\tau$ limit, the result reduces to $$\begin{aligned}
\text{PN}.8=& +\frac{1}{4} \displaystyle\sum_{\substack{k_1 k_2 k_3 k_4 \\ k_5 k_6}} \frac{\Omega^{22}_{k_1k_2k_3k_4} \Omega^{04}_{k_1k_2k_5k_6}}{(E_{k_3}+E_{k_4}+E_{k_5}+E_{k_6})(E_{k_1}+E_{k_2}+E_{k_5}+E_{k_6})}R^{--}_{k_3k_6}(\varphi) R^{--}_{k_4k_5}(\varphi) \, .
\label{PNex2}\end{aligned}$$
This diagram is zero in diagonal BMBPT as $R^{--}(0)=0$.
### Dependence on $\tau$ and $\varphi$ {#structure1}
The set of connected diagrams contributing to $n\left(\tau,\varphi\right)$ can be split according to $$n\left(\tau,\varphi\right) \equiv n\left(\tau ; n_a =0\right) + n\left(\tau,\varphi ; n_a > 0 \right) \, , \label{split1a}$$ where $n\left(\tau ; n_a =0\right) \equiv n\left(\tau,0\right)$ is the sum of connected vacuum-to-vacuum diagrams containing no anomalous propagator and arising in standard, i.e. diagonal, BMBPT. The term $n\left(\tau,\varphi ; n_a > 0 \right)$ gathers all diagrams containing at least one anomalous propagator and carries the full gauge angle dependence of $n\left(\tau,\varphi\right)$. As a consequence of Eq. \[split1a\], Eq. \[wexp3\] becomes $$\begin{aligned}
N(\tau,\varphi) &\equiv& N(\tau,0) \, e^{n\left(\tau,\varphi ; n_a > 0 \right)} \, \langle \Phi | \Phi(\varphi) \rangle \, , \label{splittedN} \end{aligned}$$ with $$\begin{aligned}
N(\tau,0) &=& e^{-\tau\Omega^{00}+n\left(\tau, 0\right)} \label{betalimtruc1} \, , \end{aligned}$$ and where $n\left(\tau,0 ; n_a > 0 \right)=0$.
In view of Eq. \[limitnorm\], one is interested in the large $\tau$ limit
\[betalim\] $$\begin{aligned}
\underset{\tau\rightarrow\infty}{\lim} n\left(\tau, 0\right) &\equiv& -\tau \Delta \Omega^{A_0}_0 + \ln |\langle \Phi | \Psi^{A_0}_{0} \rangle|^2 \label{betalim1} \, , \; \; \; \; \\
\underset{\tau\rightarrow\infty}{\lim} n\left(\tau,\varphi ; n_a > 0 \right) &\equiv& n\left(\varphi ; n_a > 0 \right) \label{betalim2} \, .\end{aligned}$$
Equation \[betalim1\] relates to the known result applicable to the logarithm of the diagonal, i.e. $\varphi=0$, norm kernel whose part proportional to $\tau$ provides the correction to the unperturbed ground-state eigenvalue of $\Omega$ $$\begin{aligned}
\Delta \Omega^{A_0}_0 &\equiv& \Omega^{A_0}_0-\Omega^{00} \nonumber \\
&=& \langle \Phi | \Omega_{1} \sum_{k=1}^{\infty} \left(\frac{1}{\Omega^{00}-\Omega_0} \Omega_{1} \right)^{k-1} | \Phi \rangle_{c} \, , \label{frombloch}$$ given under the form of Goldstone’s formula [@goldstone57a], which is here computed relative to the superfluid (i.e. Bogoliubov) reference state $| \Phi \rangle$ breaking global gauge symmetry. This expansion of $\Delta \Omega^{A_0}_0$ based on diagonal BMBPT does not constitute the solution to the problem of present interest but is anyway recovered as a byproduct. Relation \[betalim1\] recalls that, in the large $\tau$ limit, the $\varphi$-independent part $n\left(\tau, 0\right)$ gathers a term independent of $\tau$ and a term linear in $\tau$. Contrarily, Eq. \[betalim2\] states that the $\varphi$-dependent counterpart $n\left(\tau,\varphi ; n_a > 0 \right)$ is independent of $\tau$ in that limit, i.e. it converges to a finite value when $\tau$ goes to infinity. These characteristic behaviors at large imaginary time can be proven for any arbitrary order by trivially adapting the proof given in App. B.7 of Paper I.
In Eq. \[betalim1\], the contribution that does not depend on $\tau$ provides the overlap between the unrotated unperturbed state and the correlated ground-state. This overlap is not equal to $1$, which underlines that the expansion of $N(\tau,\varphi)$ does not rely on intermediate normalization at $\varphi=0$. Equation \[betalim2\] only contains a term independent of $\tau$ because the presence of the operator $S(\varphi)$ in the off-diagonal norm kernel does not modify $\Delta \Omega^{A_0}_0$ but simply provides the overlap with the ground state selected in the large $\tau$ limit with a dependence on $\varphi$.
### Particle-number conserving case {#SD1}
If the reference state $| \Phi \rangle$ is chosen to be a Slater determinant, i.e. to be an eigenstate of $A$ with eigenvalue $\text{A}_0$, one trivially finds from Eq. \[offdiaggeneralizeddensitymatrix2\] that $R^{--}(\varphi)=0$ for all $\varphi$. This leads to the fact that $n\left(\tau,\varphi ; n_a > 0 \right)=0$ for all $\tau$ and $\varphi$. At the same time, the unperturbed off-diagonal norm kernel becomes $\langle \Phi | \Phi(\varphi) \rangle = e^{i\text{A}_0\varphi}$ such that the off-diagonal norm kernel reduces to $$\begin{aligned}
N(\tau,\varphi) &\equiv& N(\tau,0) \, e^{i\text{A}_0\varphi} \, . \label{splittedN} \end{aligned}$$ When particle-number symmetry is not broken by the reference state, the introduction of the rotation operator $S(\varphi)$ in the definition of the norm kernel simply leads to an overall phase and the particle-number-conserving MBPT of $N(\tau,0)$ is trivially recovered. As a matter of fact, Eq. \[splittedN\] complies with Eq. \[expandedkernels1\] as $| \Phi \rangle$ is orthogonal to all the eigenstates of $\Omega$ characterized by $\text{A}\neq \text{A}_0$ in this case.
Off-diagonal grand-potential kernel {#energykernel}
-----------------------------------
### Off-diagonal BMBPT expansion
Proceeding similarly to $N(\tau,\varphi)$, and taking the grand potential as a particular example, one obtains the perturbative expansion of an operator kernel according to
$$\begin{aligned}
\Omega(\tau,\varphi) &=& \langle \Phi | e^{-\tau \Omega_{0}} \, \textmd{T}e^{-\int_{0}^{\tau}dt \Omega_{1}\left(t\right)} \Omega | \Phi(\varphi) \rangle \label{expansionenergykernel} \\
&=& e^{-\tau\Omega^{00}} \langle \Phi |\Big\{ \Omega(0)-\int_{0}^{\tau}d\tau_1 \textmd{T}\left[ \Omega_{1}\left( \tau_1\right)\Omega(0)\right] +\frac{1}{2!}\int_{0}^{\tau}d\tau_{1}d\tau
_{2}\textmd{T}\left[ \Omega_{1}\left( \tau_{1}\right) \Omega_{1}\left( \tau_{2}\right) \Omega(0)\right]
+... \Big\}| \Phi(\varphi) \rangle , \nonumber \end{aligned}$$
where each term in the matrix element can be fully expanded in the way that was done for the norm kernel in Eq. \[expansionnormkernel\]. The one key difference with the norm kernel relates to the presence of the time-independent operator $\Omega$ to which a fixed time $t=0$ is attributed in order to insert it inside the time ordering at no cost.
As for $N(\tau,\varphi)$, $\Omega(\tau,\varphi)$ can be expressed diagrammatically according to $$\begin{aligned}
\Omega(\tau,\varphi) &\equiv& e^{-\tau\Omega^{00}} \!\!\! \sum_{i+j=0,2,4} \; \sum^{\infty}_{n=0} \Omega^{ij \, (n)}(\tau,\varphi) \langle \Phi | \Phi(\varphi) \rangle \, , \nonumber\end{aligned}$$ where $\Omega^{ij \, (n)}(\tau,\varphi)$ denotes the sum of all vacuum-to-vacuum diagrams of order $n$ including the operator $\Omega^{ij}$ at fixed time $t=0$. The convention is that the zero-order diagram $\Omega^{ij \, (0)}(\tau,\varphi)$ solely contains the fixed-time operator $\Omega^{ij}(0)$, i.e. the latter must not be considered when counting the order of the diagram to apply the diagrammatic rules listed in Sec. \[diagrulenormMBPT\].
### Exponentiation of disconnected diagrams
Any diagram $\Omega^{ij \, (n)}(\tau,\varphi)$ consists of a part that is [*linked*]{} to the operator $\Omega^{ij}(0)$, i.e. that results from contractions involving the creation and annihilation operators of $\Omega^{ij}(0)$, and parts that are disconnected. In the infinite series of diagrams obtained via the off-diagonal BMBPT expansion of $\Omega^{ij}(\tau,\varphi)$, each vacuum-to-vacuum diagram linked to $\Omega^{ij}(0)$ effectively multiplies the complete set of vacuum-to-vacuum diagrams making up $N(\tau,\varphi)$. Gathering those infinite sets of diagrams accordingly leads to the remarkable factorization $$\Omega^{ij}(\tau,\varphi) \equiv \omega^{ij}(\tau,\varphi) \, N(\tau,\varphi) \label{linked1} \, ,$$ where
\[linked2\] $$\begin{aligned}
\omega^{ij}(\tau,\varphi) &\equiv& \sum^{\infty}_{n=0} \omega^{ij \, (n)}(\tau,\varphi) \label{linked2a} \end{aligned}$$
sums all connected vacuum-to-vacuum diagrams of order $n$ [*linked*]{} to $\Omega^{ij}(0)$.
The fact that the (reduced) kernel $O(\tau,\varphi)$ (${\cal O}(\tau,\varphi)$) of any normal-ordered operator $O$ factorizes into its linked/connected part $o(\tau,\varphi)$ times the (reduced) norm kernel $N(\tau,\varphi)$ (${\cal N}(\tau,\varphi)$) similarly to Eq. \[linked1\] is a fundamental result that will be exploited extensively in the remainder of the paper.
### Computing diagrams {#computingdiagramsEL}
![ \[diagramsTL\] Zero- and first-order connected Feynman off-diagonal BMBPT diagrams contributing to $\omega(\tau,\varphi)$. All vertices referring to $\Omega^{11}$, i.e. the lower vertex in diagrams PE.5 and PE.16, must be understood as indeed referring to the full $\Omega^{11}=\bar{\Omega}^{11} +\breve{\Omega}^{11}$.](figures/pnrmbpt "fig:"){width="\columnwidth"}\
The twenty non-zero connected/linked zero- and first-order diagrams contributing to $\omega(\tau,\varphi)$ are displayed in Fig. \[diagramsTL\], where they are classified according to the value of $n_a$, i.e. according to the number of anomalous lines they contain. Given that all first-order diagrams involve $\Omega(\tau_1)$ and $\Omega(0)$ with the constraint that $\tau_1 > 0$, normal lines not only propagate in the same direction but are also limited to propagate upward.
Choosing the reference state $|\Phi\rangle$ to be the solution of HFB equations amounts to setting $\breve{\Omega}^{11}=\Omega^{20}=\Omega^{02}=0$ such that diagrams PE.2, PE.4-PE.6, PE.12-PE.15 and PE.17 are zero in the Moller-Plesset scheme, i.e. the set reduces from twenty non-zero diagrams to eleven non-zero diagrams at first order. Finally, $\omega(\tau,0)$ at play in diagonal BMBPT reduces to PE.1, PE.4 and PE.7 ($n_a=0$) diagrams at first order (only PE.1 and PE.7 in the Moller-Plesset scheme).
![ \[diagramsTLex\] Example of a fully-labeled Feynman off-diagonal BMBPT diagram contributing to $\omega(\tau,\varphi)$, i.e. the diagram labeled PE.13 in Fig. \[diagramsTL\].](figures/pnrmbptex "fig:"){width="0.35\columnwidth"}\
While the full analytic expression of the twenty diagrams is provided in App. \[diagramsE\], we presently detail the calculation of one of them for illustration. The first-order connected/linked diagram labeled as PE.13 in Fig. \[diagramsTL\] and displayed in details in Fig. \[diagramsTLex\] contains one $\Omega^{02}$ vertex at running time $\tau_1$ coming from the perturbative expansion of the evolution operator and one vertex $\Omega^{13}$ at fixed time 0, i.e. this diagram contributes to $\omega^{13 \, (1)}(\tau,\varphi)$. The diagram contains two anomalous lines ($n_a=2$), one vertex and no crossing line ($(-1)^{p+n_c}=-1$), one anomalous line beginning and ending at the $\Omega^{13}$ vertex, and a symmetry factor $n_s=1$ as only one vertex carries a running time and thus cannot be exchanged with any other. Last but not least, the sign convention requires to associate the factors $+\Omega^{02}_{k_5k_6}$ and $+\Omega^{13}_{k_1k_2k_3k_4}$ to the vertices as they appear in Fig \[diagramsTLex\]. Eventually, diagram PE.13 reads as
$$\begin{aligned}
\text{PE}.13=& -\frac{1}{2} \displaystyle\sum_{\substack{k_1 k_2 k_3 k_4 \\ k_5 k_6}} \hspace{-0.3cm} \Omega^{13}_{k_1k_2k_3k_4} \Omega^{02}_{k_5k_6} \int\limits_0^{\tau} \mathrm{d} \tau_1 G^{-+(0)}_{k_5k_1}(\tau_1, 0; \varphi) G^{--(0)}_{k_6k_4}(\tau_1, 0; \varphi) G^{--(0)}_{k_3k_2}(0, 0; \varphi) \nonumber \\
=& -\frac{1}{2} \displaystyle\sum_{k_1 k_2 k_3 k_4k_5} \hspace{-0.3cm} \Omega^{13}_{k_1k_2k_3k_4} \Omega^{02}_{k_1k_5} R^{--}_{k_5k_4}(\varphi) R^{--}_{k_3k_2}(\varphi) \int\limits_0^{\tau} \mathrm{d} \tau_1 e^{-\tau_1(E_{k_1}+E_{k_5})} \nonumber \\
=& -\frac{1}{2} \displaystyle\sum_{k_1 k_2 k_3 k_4k_5} \frac{\Omega^{13}_{k_1k_2k_3k_4} \Omega^{02}_{k_1k_5} }{E_{k_1}+E_{k_5}}\left[ 1- e^{-\tau(E_{k_1}+E_{k_5})}\right] R^{--}_{k_5k_4}(\varphi)R^{--}_{k_3k_2}(\varphi) \, ,
\label{PEex1B}\end{aligned}$$
where use was made of the identities provided in App. \[usefulID\]. We note that, at variance with the example worked out in Sec. \[computingdiagramsNc\], the first line of Eq. \[PEex1B\] was obtained by reading the diagram from top to bottom, i.e. in a descendant fashion, in Fig. \[diagramsTLex\]. In the infinite $\tau$ limit, this reduces to $$\begin{aligned}
\text{PE}.13=& -\frac{1}{2} \displaystyle\sum_{k_1 k_2 k_3 k_4k_5} \frac{\Omega^{13}_{k_1k_2k_3k_4} \Omega^{02}_{k_1k_5} }{E_{k_1}+E_{k_5}} R^{--}_{k_3k_2}(\varphi) R^{--}_{k_5k_4}(\varphi) \, .
\label{PEex2B}\end{aligned}$$
This diagram is zero in diagonal BMBPT as $R^{--}(0)=0$.
### Large $\tau$ limit and $\varphi$ dependence {#phidepatlargetau}
According to Eq. \[limitkernels\], $N(\tau,\varphi)$ and $\Omega(\tau,\varphi)$ carry the same dependence on $\varphi$ in the large $\tau$ limit, which leads to the remarkable result that the complete sum $\omega(\varphi)$ of [*all*]{} vacuum-to-vacuum diagrams linked to the fixed-time operator $\Omega(0)$ is actually independent of $\varphi$ in this limit. This corresponds to the fact that the expansion does fulfill the symmetry in the exact limit independently of whether the expansion is performed around a particle-number conserving Slater determinant or a particle-number breaking Bogoliubov state. In the latter case, however, each individual contribution $\omega^{(n)}(\varphi)$ or any partial sum of diagrams carries a dependence on $\varphi$ as a fingerprint of the particle-number breaking. To conclude, while the dependence of $n\left(\varphi ; n_a > 0 \right)$ on $\varphi$ is genuine, the dependence of $\omega(\varphi)$ is not and must be dealt with to restore the symmetry.
### Particle-number conserving case {#SD2}
If the reference state $| \Phi \rangle$ is chosen to be a Slater determinant, $\omega(\tau,\varphi)$ is independent of $\varphi$ for all $\tau$ at any truncation order. This relates to the fact that $R^{--}(\varphi)=0$ for all $\varphi$ in this case. It is thus a situation where $N(\tau,\varphi)$ and $\Omega(\tau,\varphi)$ carry the same dependence on $\varphi$ for all $\tau$ independently of the truncation employed. As discussed in Sec. \[SD1\], this dependence is trivial and reduces to the overall phase $e^{i\text{A}_0\varphi}$ in compliance with Eqs. \[expandedkernels1\] and \[expandedkernels4\]. The expansion of $\omega(\tau, \varphi)$ is, at any $\varphi$, nothing but the standard particle-number conserving MBPT in this case.
Coupled cluster theory {#CCtheory}
======================
Having the off-diagonal BMBPT expansion of $\Omega(\tau,\varphi)$ and $N(\tau,\varphi)$ at hand, we are now in position to design their off-diagonal BCC expansion.
Off-diagonal grand-potential kernel {#energykernelMBPTtoCC}
-----------------------------------
We first demonstrate that the perturbative expansion of the linked/connected kernel $\omega(\tau,\varphi)$ can be recast in terms of an exponentiated cluster operator whose expansion naturally terminates.
### From off-diagonal BMBPT to off-diagonal BCC
We introduce the $\tau$- and $\varphi$-dependent $n$-body Bogoliubov cluster operator through
$$\mathcal{T}^{\dagger}_n (\tau, \varphi) = \frac{1}{(2n)!} \displaystyle\sum_{k_1 \ldots k_{2n}} \int\limits_0^{\tau} \mathrm{d} \tau_1 \ldots
\mathrm{d} \tau_{2n} \mathcal{T}^{\dagger}_{k_1 \ldots k_{2n}}(\tau_1 \ldots \tau_{2n}; \varphi) T \Big[ \beta_{k_{2n}}(\tau_{2n})
\ldots \beta_{k_1}(\tau_1) \Big] , \label{clusteroperators}$$
where the Feynman amplitude $\mathcal{T}^{\dagger}_{k_1 \ldots k_{2n}}(\tau_1 \ldots \tau_{2n}; \varphi)$ is antisymmetric under the exchange of $(k_i,\tau_i)$ and $(k_j,\tau_j)$ for any $(i,j)\in\{1,\ldots 2n\}^2$. One- and two-body cluster amplitudes are represented diagrammatically in Fig. \[TD1B2BCA\]. For historical reasons, the operators introduced in Eq. \[clusteroperators\] reduce to the [*Hermitian conjugate*]{} of the traditional cluster operators appearing in diagonal BCC theory [@Signoracci:2014dia].
As discussed in Sec. \[energykernel\], $\omega(\tau,\varphi)$ represents the infinite set of connected off-diagonal BMBPT diagrams linked at time zero to the operator $\Omega$. By virtue of their linked character, diagrams entering $\omega(\tau,\varphi)$ necessarily possess the topology of one of the twenty diagrams represented in Fig. \[T1contribtokinetic\] and ordered according to the value of $n_a$. The restriction to these twenty topologies are dictated by the diagrammatic rules detailed in Sec. \[diagrulenormMBPT\] and by the fact that normal lines attached to an operator at fixed time $0$ necessarily propagate upward as already mentioned.
The first three diagrams displayed in Fig. \[T1contribtokinetic\] isolate the contributions with no lines propagating in time, i.e the zero-order contributions associated with the matrix element of $\Omega$ between the reference state $| \Phi \rangle$ and its rotated partner $| \Phi(\varphi) \rangle$. Non-zero contributions of this type are limited to contributions originating from $\Omega^{00}$, $\Omega^{02}$ and $\Omega^{04}$.
\
All diagrams entering $\omega(\tau,\varphi)$ beyond zero order are captured by the remaining seventeen topologies. This leads to defining the one-body cluster amplitude ${\cal T}^{\dagger}_{k_1k_2}(\tau_1 \tau_2 ; \varphi)$ as the [*complete*]{} sum of connected off-diagonal BMBPT diagrams with one line entering at an arbitrary time $\tau_1$ and another line entering at an arbitrary time $\tau_2$. In Fig. \[T1contribtokinetic\], these two lines contract with lines arising from the various components $\Omega^{ij}$ of $\Omega$ at time zero. Covering the remaining topologies requires the introduction of the two-body cluster amplitude ${\cal T}^{\dagger}_{k_1k_2k_3k_4}(\tau_1 \tau_2 \tau_3 \tau_4 ; \varphi)$ defined as the [*complete*]{} sum of connected off-diagonal BMBPT diagrams with four lines entering at arbitrary times $\tau_1$, $\tau_2$, $\tau_3$ and $ \tau_4$. In Fig. \[T1contribtokinetic\], these four lines contract with lines arising from the various components $\Omega^{ij}$ of $\Omega$ at time zero. This definition trivially extends to higher-body cluster operators. First-order expressions of ${\cal T}^{\dagger}_{k_1k_2}(\tau_1 \tau_2 ; \varphi)$ and ${\cal T}^{\dagger}_{k_1k_2k_3k_4}(\tau_1 \tau_2 \tau_3 \tau_4 ; \varphi)$ are provided in Sec. \[firstorderToperators\].
Thus, the introduction of cluster operators allows one to group the complete set of linked/connected vacuum-to-vacuum diagrams making up $\omega(\tau,\varphi)$ under the form
$$\label{CCenergyequation}
\omega(\tau, \varphi) = \langle \Phi | \big[1+\mathcal{T}^{\dagger}_1(\tau, \varphi) + \tfrac{1}{2}\mathcal{T}^{\dagger \; 2}_1(\tau, \varphi) + \mathcal{T}^{\dagger}_2(\tau,\varphi) \big] \Omega | \Phi (\varphi) \rangle _{\text{c}}
\langle \Phi | \Phi(\varphi) \rangle ^{-1} ,$$
which translates into the twenty different terms displayed in Fig. \[T1contribtokinetic\] when expanding $\Omega$ in terms of its normal-ordered components $\Omega^{ij}$ and only retaining the non-zero contributions
\[CCenergyequation2\] $$\begin{aligned}
\omega^{00}(\tau, \varphi) &=& \Omega^{00} \, , \label{CCenergyequation2A} \\
\omega^{20}(\tau, \varphi) &=& \langle \Phi | \mathcal{T}^{\dagger}_1(\tau, \varphi)\Omega^{20} | \Phi (\varphi) \rangle _{\text{c}}
\langle \Phi | \Phi(\varphi) \rangle ^{-1} \, , \label{CCenergyequation2B} \\
\omega^{11}(\tau, \varphi) &=& \langle \Phi | \mathcal{T}^{\dagger}_1(\tau, \varphi) \Omega^{11} | \Phi (\varphi) \rangle _{\text{c}}
\langle \Phi | \Phi(\varphi) \rangle ^{-1} \, , \label{CCenergyequation2C} \\
\omega^{02}(\tau, \varphi) &=& \langle \Phi | \big[1+\mathcal{T}^{\dagger}_1(\tau, \varphi) \big] \Omega^{02} | \Phi (\varphi) \rangle _{\text{c}}
\langle \Phi | \Phi(\varphi) \rangle ^{-1} \, , \label{CCenergyequation2D} \\
\omega^{40}(\tau, \varphi) &=& \langle \Phi | \big[\tfrac{1}{2}\mathcal{T}^{\dagger \; 2}_1(\tau, \varphi) + \mathcal{T}^{\dagger}_2(\tau,\varphi) \big] \Omega^{40} | \Phi (\varphi) \rangle _{\text{c}}
\langle \Phi | \Phi(\varphi) \rangle ^{-1} \, , \\
\omega^{31}(\tau, \varphi) &=& \langle \Phi | \big[\tfrac{1}{2}\mathcal{T}^{\dagger \; 2}_1(\tau, \varphi) + \mathcal{T}^{\dagger}_2(\tau,\varphi) \big] \Omega^{31} | \Phi (\varphi) \rangle _{\text{c}}
\langle \Phi | \Phi(\varphi) \rangle ^{-1} \, , \\
\omega^{22}(\tau, \varphi) &=& \langle \Phi | \big[\mathcal{T}^{\dagger}_1(\tau, \varphi) + \tfrac{1}{2}\mathcal{T}^{\dagger \; 2}_1(\tau, \varphi) + \mathcal{T}^{\dagger}_2(\tau,\varphi) \big] \Omega^{22} | \Phi (\varphi) \rangle _{\text{c}}
\langle \Phi | \Phi(\varphi) \rangle ^{-1} \, , \\
\omega^{13}(\tau, \varphi) &=& \langle \Phi | \big[\mathcal{T}^{\dagger}_1(\tau, \varphi) + \tfrac{1}{2}\mathcal{T}^{\dagger \; 2}_1(\tau, \varphi) + \mathcal{T}^{\dagger}_2(\tau,\varphi) \big] \Omega^{13} | \Phi (\varphi) \rangle _{\text{c}}
\langle \Phi | \Phi(\varphi) \rangle ^{-1} \, , \\
\omega^{04}(\tau, \varphi) &=& \langle \Phi | \big[1+\mathcal{T}^{\dagger}_1(\tau, \varphi) + \tfrac{1}{2}\mathcal{T}^{\dagger \; 2}_1(\tau, \varphi) + \mathcal{T}^{\dagger}_2(\tau,\varphi) \big] \Omega^{04} | \Phi (\varphi) \rangle _{\text{c}}
\langle \Phi | \Phi(\varphi) \rangle ^{-1} \, , \end{aligned}$$
In Eqs. \[CCenergyequation\] and \[CCenergyequation2\], the subscript $c$ means that (i) cluster operators must all be linked to $\Omega$ through strings of contractions and that (ii) no contraction can occur among cluster operators or within a given cluster operator. Contracting quasiparticle annihilation operators originating from different cluster operators (e.g. from ${\cal T}^{\dagger}_1$ and ${\cal T}^{\dagger}_2$) or within the same cluster operator generate diagrams that are already contained in a connected cluster of lower rank and would thus lead to double counting. As off-diagonal contractions within a given cluster operator are [*not*]{} zero a priori, the rule that those contractions are to be excluded when computing contributions to Eq. \[CCenergyequation\] must indeed be stated explicitly. The grand potential being of two-body character, the sum of terms in Eq. \[CCenergyequation\] does exhaust exactly the complete set of diagrams generated through perturbation theory.
The $1/2$ factor made explicit in the third term of Eq. \[CCenergyequation\] can be justified order by order by considering a contribution ${\cal T}^{\dagger(n)}_{1}(\tau,\varphi)$ extracted from an arbitrary diagram of order $n$ having the topology of, e.g., diagram E5 in Fig. \[T1contribtokinetic\]. The corresponding contribution to $\omega(\tau,\varphi)$ of order $2n$ associated with the third term of Eq. \[CCenergyequation2\] acquires a factor $1/2$ because exchanging at once all time labels entering the two identical ${\cal T}^{\dagger (n)}_{1}(\tau,\varphi)$ pieces provides an equivalent contribution. This, combined with the other diagrammatic rules, will eventually result into the rule detailed in Sec. \[1stCCdiagrammaticrules\] dealing with so-called equivalent ${\cal T}^{\dagger}_{m}(\tau,\varphi)$ vertices.
Eventually, one can rewrite Eq. \[CCenergyequation\] under the characteristic form
\[clusterexpansion\] $$\begin{aligned}
\omega(\tau,\varphi) &=& \frac{\langle \Phi | e^{{\cal T}^{\dagger}(\tau,\varphi)} \Omega |\Phi (\varphi) \rangle_{c}}{\langle \Phi | \Phi (\varphi) \rangle} \, , \label{clusterexpansiona} \\
{\cal T}^{\dagger}(\tau,\varphi) &\equiv& \sum_{n \in \mathbb{N}} {\cal T}^{\dagger}_{n}(\tau,\varphi) \, , \label{clusterexpansionb}\end{aligned}$$
given that no cluster operator beyond ${\cal T}^{\dagger}_{1}(\tau,\varphi)$ and ${\cal T}^{\dagger}_{2}(\tau,\varphi)$ can actually contribute to $\omega(\tau,\varphi)$ in view of its linked/connected character. The fact that Eq. \[clusterexpansiona\] does indeed reduce to Eq. \[CCenergyequation\] generalizes to the off-diagonal grand potential kernel the natural termination of the BCC expansion displayed by the diagonal one [@Signoracci:2014dia]. As mentioned earlier, the termination and the specific connected structure of the resulting terms are traditionally obtained for the diagonal kernel from the similarity transformed grand potential $\bar{\Omega}\equiv e^{-T}\Omega e^{T}$ on the basis of the Baker-Campbell-Hausdorff identity and of standard Wick’s theorem [@shavitt09a]. In the present case, the long detour via perturbation theory applied to off-diagonal kernels was necessary to obtain the same connected structure as in the diagonal case, including the necessity to omit contractions within a cluster operator or among two different cluster operators.
### Computation of diagrams {#1stCCdiagrammaticrules}
We now compute the algebraic BCC contributions to $\omega(\tau,\varphi)$. The diagrammatic rules to obtain them are essentially the same as those detailed in Sec. \[diagrulenormMBPT\] for BMBPT Feynman diagrams. The only modifications are that
1. One must attribute a factor $\mathcal{T}^{\dagger}_{k_1 \ldots k_{2n}}(\tau_1 \ldots \tau_{2n}; \varphi)$ to any vertex representing an $n$-body cluster operator (i.e. its hermitian conjugate). Indices $k_1 \ldots k_{2n}$ and $\tau_1 \ldots \tau_{2n}$ must be assigned consecutively from the leftmost to the rightmost line below the vertex.
2. All diagrams are [*connected*]{}, i.e. each contributing $\mathcal{T}^{\dagger}_{n}(\tau , \varphi)$ operator is contracted at least once with $\Omega$. No line may connect two cluster operators while lines belonging to a given cluster operator cannot be contracted together.
3. Following the above rule, construct all possible independent closed diagrams from the building blocks. Doing so typically limits which parts $\Omega^{ij}$ of $\Omega$ contribute to a given term.
4. The symmetry factor $1/n_s$ must be replaced by a factor $(\ell_m!)^{-1}$ for each set of $\ell_m$ equivalent ${\cal T}^{\dagger}_{m}(\tau,\varphi)$ vertices. Two ${\cal T}^{\dagger}_{m}(\tau,\varphi)$ vertices are equivalent if they have the same number of quasi-particle lines $n_l \, (n_l \leq 2m)$ connected to the interaction vertex via propagators of the [*same*]{} type.
5. The sign of the diagram is obtained by combining the factor $(-1)^{n_c}$ with the sign rule associated with each factor $\Omega^{ij}_{k_1 \ldots k_i k_{i+1} \ldots k_{i+j}}$.
The above rules result into the twenty non-zero BCC diagrams contributing to $\omega(\tau,\varphi)$ and displayed in Fig. \[T1contribtokinetic\] where they are classified according to the value of $n_a$, i.e. according to the number of anomalous lines they contain.
Choosing the reference state $|\Phi\rangle$ to be the solution of HFB equations amounts to setting $\Omega^{20}=\Omega^{02}=0$ such that diagrams E.2, E.4 and E.6 are zero in the Moller-Plesset scheme, i.e. the set reduces from twenty non-zero diagrams to seventeen non-zero diagrams in this case. Finally, $\omega(\tau,0)$ at play in diagonal BCC reduces to E.1, E.4, E.7 and E.12 ($n_a=0$) diagrams (only E.1, E.7 and E.12 in the Moller-Plesset scheme).
While the full analytic expression of each of the twenty diagrams is provided in App. \[CCenergycontrib\], we presently detail the calculation of one of them for illustration. We calculate diagram E.19 that is displayed in complete detail in Fig. \[pnrbccex\]. According to the diagrammatic rules, and reading the diagram in a descendant fashion, its contribution reads as
$$\begin{aligned}
\nonumber
\text{E}.19 &= + \frac{1}{2} \displaystyle\sum_{\substack{k_1 k_2 k_3 k_4 \\ k_5 k_6}} \Omega^{13}_{k_1k_2k_3 k_4} \int\limits_0^{\tau} \mathrm{d} \tau_1 \mathrm{d} \tau_2
\mathcal{T}^{\dagger}_{k_5 k_6}(\tau_1 \tau_2; \varphi) G^{-+(0)}_{k_5 k_1}(\tau_1,0; \varphi) G^{--(0)}_{k_6 k_4}(\tau_2,0;\varphi) G^{--(0)}_{k_3 k_2}(0,0;\varphi)\\
&= + \frac{1}{2} \displaystyle\sum_{k_1 k_2 k_3 k_4 k_5} \Omega^{13}_{k_1 k_2 k_3 k_4} \int\limits_0^{\tau} \mathrm{d} \tau_1 \mathrm{d} \tau_2
\mathcal{T}^{\dagger}_{k_1 k_5} (\tau_1 \tau_2; \varphi) e^{-E_{k_1} \tau_1}e^{-E_{k_5} \tau_2} R^{--}_{k_5k_4}(\varphi)R^{--}_{k_3k_2}(\varphi)\, .
\label{diag1}\end{aligned}$$
Defining a time-integrated one-body cluster amplitude $$\mathcal{T}^{\dagger}_{k_1 k_2}(\tau,\varphi) \equiv \int\limits_0^{\tau} \mathrm{d} \tau_1 \mathrm{d} \tau_2
\mathcal{T}^{\dagger}_{k_1 k_2} (\tau_1 \tau_2; \varphi) e^{-E_{k_1} \tau_1}e^{-E_{k_2} \tau_2} \, , \label{T1matrixelements}$$ such that $$\begin{aligned}
{\cal T}^{\dagger}_{1}(\tau,\varphi) &=& \frac{1}{2!} \sum_{k_1 k_2} \mathcal{T}^{\dagger}_{k_1 k_2}(\tau,\varphi) \, \beta_{k_2} \, \beta_{k_1} \, , \label{T1} \end{aligned}$$ one can finalize Eq. \[diag1\] under the form $$\text{E}.19 = +\frac{1}{2} \displaystyle\sum_{k_1 k_2 k_3 k_4 k_5} \mathcal{T}^{\dagger}_{k_1 k_5}(\tau,\varphi) \Omega^{13}_{k_1 k_2 k_3 k_4}R^{--}_{k_5k_4}(\varphi)R^{--}_{k_3k_2}(\varphi) \, .$$
\
Further introducing
$$\begin{aligned}
\mathcal{T}^{\dagger}_{k_1 k_2 k_3 k_4}(\tau,\varphi) &\equiv& \int\limits_0^{\tau} \mathrm{d} \tau_1 \mathrm{d} \tau_2
\mathrm{d} \tau_3 \mathrm{d} \tau_4 \mathcal{T}^{\dagger}_{k_1 k_2 k_3 k_4} (\tau_1 \tau_2 \tau_3 \tau_4; \varphi) e^{-E_{k_1} \tau_1}e^{-E_{k_2} \tau_2} e^{-E_{k_3} \tau_3}e^{-E_{k_4} \tau_4} \, \label{T2matrixelements} \\
{\cal T}^{\dagger}_{2}(\tau,\varphi) &=& \frac{1}{4!} \sum_{k_1 k_2 k_3 k_4} \mathcal{T}^{\dagger}_{k_1 k_2 k_3 k_4}(\tau,\varphi) \, \beta_{k_4} \, \beta_{k_3} \, \beta_{k_2} \, \beta_{k_1} \, , \label{T2} \end{aligned}$$
all twenty contributions to $\omega(\tau,\varphi)$ can be derived.
Determining off-diagonal BCC amplitudes {#dynamicalequations}
---------------------------------------
In order to effectively compute the various contributions to $\omega(\tau,\varphi)$, one must have the matrix elements of the $\varphi$-dependent cluster operators at hand.
### First-order in off-diagonal BMBPT {#firstorderToperators}
The first option consists of determining the cluster amplitudes via off-diagonal BMBPT. Feynman diagrams contributing to one- and two-body cluster amplitudes at first order in off-diagonal BMBPT are displayed in Fig. \[1storderTamplitudes\] and give
\[1stordercluster\] $$\begin{aligned}
\mathcal{T}^{\dagger (1)}_{k_1k_2}(\tau_1 \tau_{2}; \varphi) &=& - \Omega^{02}_{k_1k_2} \delta(\tau_1 - \tau_2) - \frac{1}{2}
\displaystyle\sum_{k_3k_4} \Omega^{04}_{k_1k_2k_3k_4} G^{--(0)}_{k_4k_3} (\tau_1,\tau_1; \varphi) \delta(\tau_1 - \tau_2) \, , \label{1storderclustera} \\
\mathcal{T}^{\dagger (1)}_{k_1k_2k_3k_4}(\tau_1 \tau_{2} \tau_3 \tau_4; \varphi) &=&
- \Omega^{04}_{k_1k_2k_3k_4} \delta(\tau_1 - \tau_2)\delta(\tau_2 - \tau_3)\delta(\tau_3 - \tau_4) \, . \label{1storderclusterb}\end{aligned}$$
Inserting these expressions into Eqs. \[T1matrixelements\] and \[T2matrixelements\] provides associated Goldstone amplitudes
\[1storderclusterB\] $$\begin{aligned}
\mathcal{T}^{\dagger (1)}_{k_1k_2}(\tau,\varphi) &=& - \frac{\Omega^{02}_{k_1k_2}}{E_{k_1}+E_{k_2}} \big[ 1- e^{- \tau(E_{k_1}+E_{k_2})} \big] \nonumber \\
&&-\frac{1}{2}
\displaystyle\sum_{k_3k_4} \frac{\Omega^{04}_{k_1k_2k_3k_4}}{E_{k_1}+E_{k_2}+E_{k_3}+E_{k_4}}
\big[ 1- e^{- \tau(E_{k_1}+E_{k_2}+E_{k_3}+E_{k_4})} \big] R^{--}_{k_4k_3} (\varphi) \, , \label{1storderclusterBa} \\
\mathcal{T}^{\dagger (1)}_{k_1k_2k_3k_4}(\tau,\varphi) &=& - \frac{\Omega^{04}_{k_1k_2k_3k_4}}{E_{k_1}+E_{k_2}+E_{k_3}+E_{k_4}}
\big[ 1- e^{- \tau(E_{k_1}+E_{k_2}+E_{k_3}+E_{k_4})} \big] \, \, , \label{1storderclusterBb}\end{aligned}$$
such that ${\cal T}^{\dagger \, (1)}_{2}(\tau,\varphi)$ does not depend on $\varphi$. One can check that ${\cal T}^{\dagger \, (1)}_{1}(0,\varphi)={\cal T}^{\dagger \, (1)}_{2}(0,\varphi)=0$ as it should be.
### Off-diagonal BCC amplitude equations {#amplitudeequations}
To work within a non-perturbative BCC framework, one must derive equations of motion for the $\varphi$-dependent cluster amplitudes. To do so, we introduce $n$-tuply excited off-diagonal norm, grand-potential and particle-number kernels through
\[amplitudekernels\] $$\begin{aligned}
N_{k_1k_2\ldots}(\tau,\varphi) &\equiv& \langle \Psi (\tau) | {\ensuremath{\mathbbm{1}}}| \Phi^{k_1k_2\ldots}(\varphi) \rangle \,\, , \label{amplitudekernels1} \\
\Omega_{k_1k_2\ldots}(\tau,\varphi) &\equiv& \langle \Psi (\tau) | \Omega | \Phi^{k_1k_2\ldots}(\varphi) \rangle \,\, , \label{amplitudekernels2} \\
A_{k_1k_2\ldots}(\tau,\varphi) &\equiv& \langle \Psi (\tau) | A | \Phi^{k_1k_2\ldots}(\varphi) \rangle \,\, , \label{amplitudekernels3}\end{aligned}$$
where $$| \Phi^{k_1k_2\ldots}(\varphi) \rangle \equiv {\cal B}_{k_1k_2\ldots} | \Phi(\varphi) \rangle \, ,$$ with the operator ${\cal B}_{k_1k_2\ldots}$ defined in Eq. \[phexcitation\]. From Eq. \[schroedinger\], one obtains that $$\Omega_{k_1k_2\ldots}(\tau,\varphi) = -\partial_{\tau} N_{k_1k_2\ldots}(\tau,\varphi) \, . \label{dynamicalkernels}$$ In App. \[amplitudeequations\], we demonstrate in detail how Eq. \[dynamicalkernels\] eventually provides the equations of motion satisfied by the $n$-body ($\tau$- and $\varphi$-dependent) cluster amplitudes under the form $$\omega_{k_1k_2\ldots}(\tau,\varphi) = -\partial_{\tau} {\cal T}^{\dagger}_{k_1k_2\ldots}(\tau,\varphi) \, \, , \label{reduceddynamicalkernels}$$ where the $n$-tuply excited [*connected*]{} grand potential kernel is defined through $$\omega_{k_1k_2\ldots}(\tau,\varphi) \equiv \frac{\langle \Phi | e^{{\cal T}^{\dagger}(\tau,\varphi)} \Omega | \Phi^{k_1k_2\ldots}(\varphi) \rangle_{c}}{\langle \Phi | \Phi(\varphi) \rangle} \, , \label{CCamplitudekernels}$$ and whose connected character denotes that (i) cluster operators are all connected to $\Omega$ and that (ii) no contraction is to be considered among cluster operators or within any given cluster operator such that the power series of the exponential naturally terminates.
As a result of this termination of the exponential, singly- and doubly-excited off-diagonal connected grand potential kernels read as
\[termination\] $$\begin{aligned}
\omega_{k_1k_2}(\tau,\varphi) &=& \langle \Phi | \big[1+\mathcal{T}^{\dagger}_1(\tau, \varphi) + \frac{1}{2}\mathcal{T}^{\dagger 2}_1(\tau, \varphi) +
\frac{1}{3!}\mathcal{T}^{\dagger 3}_1(\tau, \varphi) \nonumber \\
&& \,\,\,\,\,\,\,\,\,\, + \mathcal{T}^{\dagger}_2(\tau, \varphi) +
\mathcal{T}^{\dagger}_1(\tau, \varphi)\mathcal{T}^{\dagger}_2(\tau, \varphi) \big] \Omega | \Phi^{k_1 k_2} (\varphi) \rangle_{\text{c}}
\langle \Phi | \Phi(\varphi) \rangle^{-1} \, , \label{termination1} \\
\omega_{k_1k_2k_3k_4}(\tau,\varphi) &=& \langle \Phi | \big[1+\mathcal{T}^{\dagger}_1(\tau, \varphi) + \frac{1}{2}\mathcal{T}^{\dagger 2}_1(\tau, \varphi) +
\frac{1}{3!}\mathcal{T}^{\dagger 3}_1(\tau, \varphi) + \mathcal{T}^{\dagger}_2(\tau, \varphi) +
\mathcal{T}^{\dagger}_1(\tau, \varphi)\mathcal{T}^{\dagger}_2(\tau, \varphi) \nonumber \\
&& \,\,\,\,\,\,\,\,\,\, + \frac{1}{2} \mathcal{T}^{\dagger 2}_2(\tau, \varphi)+ \frac{1}{4!}\mathcal{T}^{\dagger 4}_1(\tau, \varphi) +
\frac{1}{2} \mathcal{T}^{\dagger 2}_1(\tau, \varphi) \mathcal{T}^{\dagger}_2 (\tau, \varphi) \big]
\Omega | \Phi^{k_1 k_2 k_3 k_4} (\varphi) \rangle_{\text{c}}
\langle \Phi | \Phi(\varphi) \rangle^{-1} \, , \label{termination2}\end{aligned}$$
respectively. Note that we refer to singly- and doubly-excited kernels to connect to the standard CC formalism as two-quasiparticle (four-quasiparticle) excitations reduce in the Slater determinant limit to one (two) particle-hole excitations.
### Computation of diagrams {#computation-of-diagrams}
To compute the algebraic contributions to $\omega_{k_1\ldots k_{2n}}(\tau,\varphi)$, a few diagrammatic rules beyond those stated in Sec. \[1stCCdiagrammaticrules\] to determine $\omega(\tau,\varphi)$ must be added
1. Diagrams making up $\omega_{k_1\ldots k_{2n}}(\tau,\varphi)$ are [*linked*]{} with $2n$ external quasi-particle lines, exiting from below. External lines must be labeled with quasi-particle indices $k_1\ldots k_{2n}$ coinciding with the left-right ordering of the indices observed in the ket defining the $n$-tuply excited kernel. Internal lines must be labeled with different quasi-particle indices.
2. Only internal quasi-particle line indices must be summed over.
3. All distinct permutations $P$ of labels of inequivalent external lines must be summed over, including a parity factor $(-1)^{\sigma(P)}$ from the signature of the permutation. External lines are equivalent if and only if they connect to the same vertex.
\
One example diagram contributing to the singly-excited off-diagonal connected grand potential kernel is displayed in Fig. \[pnrbccsingex\] with explicit labeling. Following the diagrammatic rules, its contribution reads as
$$\begin{aligned}
\nonumber
\text{S}.25 &= + \frac{1}{4} P(\alpha / \beta) \displaystyle\sum_{\substack{k_1 k_2 k_3 k_4 \\ k_5 k_6 k_7 k_8}} \Omega^{22}_{k_1k_2k_3 k_4} \int\limits_0^{\tau} \mathrm{d} \tau_1 \mathrm{d} \tau_2 \mathrm{d} \tau_3 \mathrm{d} \tau_4
\mathcal{T}^{\dagger}_{k_5 k_6}(\tau_1 \tau_2; \varphi) \mathcal{T}^{\dagger}_{k_7 k_8}(\tau_3 \tau_4; \varphi)G^{-+(0)}_{k_5 \alpha}(\tau_1,0;\varphi) \nonumber \\
& \hspace{7cm} \times G^{-+(0)}_{k_6 k_1}(\tau_2,0;\varphi) G^{--(0)}_{k_4 k_3}(0,0;\varphi)G^{-+(0)}_{k_7 k_2}(\tau_3,0;\varphi) G^{-+(0)}_{k_8 \beta}(\tau_4,0;\varphi) \nonumber \\
&= + \frac{1}{4} P(\alpha / \beta) \displaystyle\sum_{k_1 k_2 k_3 k_4} \Omega^{22}_{k_1 k_2 k_3 k_4} \int\limits_0^{\tau} \mathrm{d} \tau_1 \mathrm{d} \tau_2 \mathrm{d} \tau_3 \mathrm{d} \tau_4 \mathcal{T}^{\dagger}_{\alpha k_1} (\tau_1 \tau_2; \varphi) \mathcal{T}^{\dagger}_{k_2 \beta} (\tau_3 \tau_4; \varphi) e^{-E_{\alpha} \tau_1}e^{-E_{k_1} \tau_2}\nonumber \\
& \hspace{10cm} \times e^{-E_{k_2} \tau_3} e^{-E_{\beta} \tau_4}R^{--}_{k_4k_3}(\varphi) \nonumber \\
&= +\frac{1}{4}P(\alpha / \beta) \displaystyle\sum_{k_1 k_2 k_3 k_4} \mathcal{T}^{\dagger}_{\alpha k_1}(\tau,\varphi)
\mathcal{T}^{\dagger}_{k_2 \beta}(\tau,\varphi) \Omega^{22}_{k_1 k_2 k_3 k_4}R^{--}_{k_4k_3}(\varphi) \, ,
\label{diagsing1}\end{aligned}$$
where $P(\alpha / \beta) \equiv 1-P_{\alpha\beta}$, with $P_{\alpha\beta}$ the operator exchanging quasi-particle indices $\alpha$ and $\beta$ labelling the two external lines.
Although the operator form of $\omega_{k_1k_2}(\tau,\varphi)$ and $\omega_{k_1k_2k_3k_4}(\tau,\varphi)$ as displayed in Eq. \[termination\] is formally identical to diagonal ($\varphi=0$) BCC expressions [@Signoracci:2014dia], their expanded algebraic expressions are much lengthier. This translates into the fact that $\omega_{k_1k_2}(\tau,\varphi)$ ($\omega_{k_1k_2k_3k_4}(\tau,\varphi)$) is made out of fifty-seven (seventy-seven) diagrams at the BCCSD level that reduce to only ten (fourteen) diagrams in the diagonal ($\varphi =0$) limit. While the explicit algebraic expressions of the fifty-seven contributions to $\omega_{k_1k_2}(\tau,\varphi)$ are provided in App. \[CCsinglecontrib\], the contributions to $\omega_{k_1k_2k_3k_4}(\tau,\varphi)$ are too numerous and lengthy to be reported here. This is anyway unnecessary given that a compact form of these expressions containing just as many terms as in the diagonal ($\varphi=0$) limit will be identified in Sec. \[compact\] below. For this reason, we do not produce the full diagrammatic description of the off-diagonal amplitude equations at the BCCSD level as well.
In the end, one is only interested in the infinite imaginary-time limit. In this limit, the scheme becomes stationary such that the static amplitude equations are obtained by setting the right-hand side of Eq. \[reduceddynamicalkernels\] to zero, i.e. $$\omega_{k_1 k_2\ldots}(\varphi) = 0 \, , \label{staticamplitudeequations}$$ which naturally extend diagonal BCC amplitude equations. Coupled Eqs. \[staticamplitudeequations\] must be solved iteratively for each $\varphi \in [0,2\pi]$, typically employing first-order perturbation theory expressions as an initial guess (see Eq. \[1storderclusterB\] for single and double amplitudes). Once $\varphi$-dependent cluster amplitudes have been obtained, the off-diagonal connected grand-potential kernel $\omega(\tau,\varphi)$ can be computed on the basis of the expressions provided in App. \[CCenergycontrib\].
Compact formulation {#compact}
-------------------
As alluded to above, the algebraic expressions of $\omega(\tau,\varphi)$, $\omega_{k_1k_2}(\tau,\varphi)$ and $\omega_{k_1k_2k_3k_4}(\tau,\varphi)$ (see Apps. \[CCenergycontrib\] and \[CCsinglecontrib\] for the first two) are lengthy and translate into a large number of diagrams. This is not optimal, both for bookkeeping and from the numerical implementation viewpoint. It happens that the generalized BCC expansion of those off-diagonal kernels can eventually be reformulated in terms of a transformed, $\varphi$-dependent, grand potential operator such that their algebraic expressions are not only made much more compact but formally identical to their diagonal counterpart.
### Transformed grand-potential operator {#transformedoperator}
We introduce a [*non-unitary*]{} Bogoliubov transformation that transforms quasi-particle operators defining the vacuum $| \Phi \rangle$ into a new set of quasi-particle operators according to $$\begin{aligned}
\left(
\begin{array} {c}
\tilde{\beta} \\
\tilde{\beta}^{\dagger}
\end{array}
\right)(\varphi) &\equiv& M(\varphi) \left(
\begin{array} {c}
\beta \\
\beta^{\dagger}
\end{array}
\right) M^{-1}(\varphi) \nonumber \\
&=& {\cal M}^{\dagger}(\varphi) \left(
\begin{array} {c}
\beta \\
\beta^{\dagger}
\end{array}
\right) \, ,\end{aligned}$$ where $${\cal M}(\varphi) \equiv \left(
\begin{array} {cc}
1 & 0 \\
R^{-- \ast}(\varphi) & 1
\end{array}
\right) \, ,$$ such that
\[newbogo\] $$\begin{aligned}
\tilde{\beta}_{k_1}(\varphi) &= \beta_{k_1} + \sum_{k_2} R^{--}_{k_2k_1}(\varphi) \beta^{\dagger}_{k_2} \, , \\
\tilde{\beta}^{\dagger}_{k_1}(\varphi) &= \beta^{\dagger}_{k_1} \, .\end{aligned}$$
Next, we introduce the non-hermitian transformed grand potential operator $\tilde{\Omega}(\varphi) \equiv M(\varphi) \Omega M^{-1}(\varphi)$. Starting from the normal-ordered form of $\Omega$, performing the non-unitary Bogoliubov transformation, normal-ordering the resulting $\tilde{\Omega}(\varphi)$ with respect to $| \Phi \rangle$ and gathering appropriately the terms thus generated allows one to write $\tilde{\Omega}(\varphi)$ under the typical form
\[e:h3qpastransformed\] $$\begin{aligned}
\tilde{\Omega}(\varphi) &\equiv \tilde{\Omega}^{[0]}(\varphi) + \tilde{\Omega}^{[2]}(\varphi) + \tilde{\Omega}^{[4]}(\varphi) \\
&\equiv \tilde{\Omega}^{00}(\varphi) + \big[\tilde{\Omega}^{20}(\varphi) + \tilde{\Omega}^{11}(\varphi) + \tilde{\Omega}^{02}(\varphi)\big] + \big[\tilde{\Omega}^{40}(\varphi) + \tilde{\Omega}^{31}(\varphi) + \tilde{\Omega}^{22}(\varphi) + \tilde{\Omega}^{13}(\varphi) + \tilde{\Omega}^{04}(\varphi)\big] \\
&= \tilde{\Omega}^{00}(\varphi) \\*
& \: \: \: \: \: \: \: \: \: \: \: \: \: + \frac{1}{1!}\displaystyle\sum_{k_1 k_2} \tilde{\Omega}^{11}_{k_1 k_2}(\varphi)\beta^{\dagger}_{k_1} \beta_{k_2} \\*
& \: \: \: \: \: \: \: \: \: \: \: \: \: + \frac{1}{2!}\displaystyle\sum_{k_1 k_2} \Big \{\tilde{\Omega}^{20}_{k_1 k_2}(\varphi) \beta^{\dagger}_{k_1}
\beta^{\dagger}_{k_2} + \tilde{\Omega}^{02}_{k_1 k_2}(\varphi) \beta_{k_2} \beta_{k_1} \Big \} \\*
& \: \: \: \: \: \: \: \: \: \: \: \: \: + \frac{1}{(2!)^{2}} \displaystyle\sum_{k_1 k_2 k_3 k_4} \tilde{\Omega}^{22}_{k_1 k_2 k_3 k_4}(\varphi)
\beta^{\dagger}_{k_1} \beta^{\dagger}_{k_2} \beta_{k_4}\beta_{k_3} \\*
& \: \: \: \: \: \: \: \: \: \: \: \: \: + \frac{1}{3!}\displaystyle\sum_{k_1 k_2 k_3 k_4}\Big \{ \tilde{\Omega}^{31}_{k_1 k_2 k_3 k_4}(\varphi)
\beta^{\dagger}_{k_1}\beta^{\dagger}_{k_2}\beta^{\dagger}_{k_3}\beta_{k_4} +
\tilde{\Omega}^{13}_{k_1 k_2 k_3 k_4}(\varphi) \beta^{\dagger}_{k_1} \beta_{k_4} \beta_{k_3} \beta_{k_2} \Big \} \\*
& \: \: \: \: \: \: \: \: \: \: \: \: \: + \frac{1}{4!} \displaystyle\sum_{k_1 k_2 k_3 k_4}\Big \{ \tilde{\Omega}^{40}_{k_1 k_2 k_3 k_4}(\varphi)
\beta^{\dagger}_{k_1}\beta^{\dagger}_{k_2}\beta^{\dagger}_{k_3}\beta^{\dagger}_{k_4} +
\tilde{\Omega}^{04}_{k_1 k_2 k_3 k_4}(\varphi) \beta_{k_4} \beta_{k_3} \beta_{k_2} \beta_{k_1} \Big \} \, .\end{aligned}$$
The expressions of the transformed matrix elements in terms of the original ones are provided in App. \[transformME\]. As a testimony of the non-hermitian character of $\tilde{\Omega}(\varphi)$, itself the consequence of the non-unitary character of $M(\varphi)$, matrix elements $\tilde{\Omega}^{ij}_{k_1 \ldots k_{i} k_{i+1}}(\varphi)$ do [*not*]{} display relationships characterized by Eq. \[e:me3sym\]. One also notices that transformation \[newbogo\] reduces to the identity for $\varphi=0$, i.e. $\tilde{\Omega}(0)=\Omega$.
### Compact algebraic expressions {#compactexpressions}
It is tedious but straightforward to demonstrate that the algebraic expressions of $\omega(\tau,\varphi)$, $\omega_{k_1k_2}(\tau,\varphi)$ and $\omega_{k_1k_2k_3k_4}(\tau,\varphi)$ obtained through the application of the off-diagonal Wick theorem are formally identical to diagonal ($\varphi=0$) BCC formulae [@shavitt09a], as long as one uses the [*transformed*]{} grand potential $\tilde{\Omega}(\varphi)$ in place of the original one $\Omega$ on the basis of standard Wick’s theorem, i.e.
\[compactification\] $$\begin{aligned}
\omega_{k_1k_2\ldots}(\tau,\varphi) &=& \frac{\langle \Phi | e^{{\cal T}^{\dagger}(\tau,\varphi)} \Omega | \Phi^{k_1k_2\ldots}(\varphi) \rangle_{c}}{\langle \Phi | \Phi(\varphi) \rangle} \label{CCamplitudekernelsbis0} \\
&=& \langle \Phi | e^{{\cal T}^{\dagger}(\tau,\varphi)} \tilde{\Omega}(\varphi) | \Phi_{k_1k_2\ldots} \rangle_{c} \, . \label{CCamplitudekernelsbis}\end{aligned}$$
To illustrate the severe shortening of the algebraic expressions accomplished by employing Eq. \[CCamplitudekernelsbis\] instead of Eq. \[CCamplitudekernelsbis0\], let us focus on the energy kernel $\omega(\tau,\varphi)$ and refer to Ref. [@Signoracci:2014dia] for single- and double-amplitude equations. While the lengthy expression associated with Eq. \[CCamplitudekernelsbis0\] (App. \[CCenergycontrib\]) has been derived from twenty different diagrams, the compact form associated with Eq. \[CCamplitudekernelsbis\] reads as
\[rewritingomega\] $$\begin{aligned}
\omega(\tau,\varphi) &=& \tilde{\Omega}^{00}(\varphi) \label{rewritingomega1} \hspace{6.5cm} \\
&+& \frac{1}{2} \displaystyle\sum_{k_1 k_2} \mathcal{T}^{\dagger}_{k_1 k_2}(\tau,\varphi) \, \tilde{\Omega}^{20}_{k_1 k_2}(\varphi) \label{rewritingomega2} \\
&+& \frac{1}{8} \displaystyle\sum_{k_1 k_2 k_3 k_4} \mathcal{T}^{\dagger}_{k_1 k_2}(\tau,\varphi)\mathcal{T}^{\dagger}_{k_3 k_4}(\tau,\varphi) \, \tilde{\Omega}^{40}_{k_1 k_2 k_3 k_4}(\varphi) \label{rewritingomega3} \\
& +& \frac{1}{4!} \displaystyle\sum_{k_1 k_2 k_3 k_4} \mathcal{T}^{\dagger}_{k_1 k_2 k_3 k_4}(\tau,\varphi) \, \tilde{\Omega}^{40}_{k_1 k_2 k_3 k_4}(\varphi) \, , \label{rewritingomega4}\end{aligned}$$
and relates to the four Goldstone, i.e. time-independent, diagrams displayed in Fig. \[diagramst\] and involving vertices $\tilde{\Omega}^{ij}(\varphi)$ of the transformed grand potential. Those four Goldstone diagrams, along with the associated expression \[rewritingomega\], are indeed formally identical to the four $n_a=0$ diagrams at play in diagonal BCC theory [@Signoracci:2014dia]. As a matter of fact, one can now entirely rephrase the generalized BCC formalism applicable to off-diagonal kernels in terms of a time-independent diagrammatic technique that parallels exactly BCC theory, i.e. diagrammatic rules, diagrams and algebraic expressions are identical except that transformed interaction vertices $\tilde{\Omega}^{ij}_{k_1 \ldots k_{i} k_{i+1} \ldots k_{i+j}}(\varphi)$ must be used in place of $\Omega^{ij}_{k_1 \ldots k_{i} k_{i+1} \ldots k_{i+j}}$, such that cluster amplitudes acquire an explicit $\varphi$ dependence.
\
The result obtained in Eq. \[compactification\] is remarkable and constitutes a drastic simplification both from a formal and a practical standpoint. Regarding the latter, it means that a previously built single-reference BCC code can be employed almost straightforwardly. The additional cost, which is not negligible, consists of building and storing matrix elements of $\tilde{\Omega}(\varphi)$ for each $\varphi \in [0,2\pi]$ according to App. \[transformME\]. With those transformed matrix elements in input, the BCC code can be used essentially as it is.
Norm kernel {#Secnormkernel}
-----------
In Sec. \[normkernel2\], we obtained the perturbative expansion of $N(\tau,\varphi)$ that eventually led to the connected expansion of $\ln N(\tau,\varphi)$. In the present section, we wish to identify a method to compute $N(\tau,\varphi)$ non-perturbatively in a way that is consistent with the BCC expansion of the connected grand potential kernel $\omega(\tau,\varphi)$. More specifically, we wish to design a naturally terminating expansion of $\ln N(\tau,\varphi)$. It happens that a naturally terminating expansion does not trivially emerge from the perturbative expansion of $\ln N(\tau,\varphi)$ as the corresponding diagrams are not linked to an operator at a fixed time. We now explain how this apparent difficulty can be overcome by following an alternative route from the outset.
### Key property
In the case of [*exact*]{} kernels, Eqs. \[expandedkernels1\] and \[expandedkernels3\] trivially lead, for any $\tau$, to $$\begin{aligned}
\frac{\int_{0}^{2\pi} \!d\varphi \,e^{-i\text{A}\varphi} \, {\cal A}(\tau,\varphi)}{\int_{0}^{2\pi} \!d\varphi \,e^{-i\text{A}\varphi} \, \, {\cal N}(\tau,\varphi)} &=& \text{A} \, , \label{restoreLieoperators1} \end{aligned}$$ which testifies that the implicit many-body state at play is indeed an eigenstate of the particle-number operator with eigenvalue A. Equation \[restoreLieoperators1\] stresses the fact that we [*know*]{} a priori the value that must be obtained through the integral over the domain of the $U(1)$ group for the coefficient associated with the physical IRREP in the Fourier expansion of the particle number operator kernel (once it is divided by the corresponding expansion coefficient in the norm kernel). This differs from $H$ or $\Omega$ for which we can only require to extract the expansion coefficient that is in one-to-one correspondence with the physical IRREP of interest without knowing the [*value*]{} this coefficient should take (once it is divided by the corresponding expansion coefficient in the norm kernel).
Consequently, the key question is: what happens to Eq. \[restoreLieoperators1\] when ${\cal A}(\tau,\varphi)$ and ${\cal N}(\tau,\varphi)$ are approximated? Or rephrasing the question more appropriately: what constraint(s) does restoring the symmetry, i.e. fulfilling Eqs. \[restoreLieoperators1\], impose on the truncation scheme used to approximate the kernels? Addressing this question below delivers the proper approach to the reduced norm kernel.
### Differential equation
We derive a first-order ordinary differential equation (ODE) fulfilled, at each imaginary time $\tau$, by ${\cal N}(\tau,\varphi)$. To do so, we employ Eq. \[ODE\] to relate ${\cal N}(\tau,\varphi)$ (Eqs. \[expandedkernels1\]) and ${\cal A}(\tau,\varphi)$ (Eqs. \[expandedkernels3\]). Exploiting that the reduced kernel of the operator $A$ can be factorized according to ${\cal A}(\tau,\varphi)=a(\tau,\varphi) {\cal N}(\tau,\varphi)$, where $a(\tau,\varphi)$ denotes the corresponding linked/connected kernel, we arrive at the first-order ODE $$\frac{d}{d \varphi} \, {\cal N}(\tau,\varphi) - i \, a(\tau,\varphi) \, {\cal N}(\tau,\varphi) = 0 \, ,\label{NormkernelODE1}$$ with the initial condition ${\cal N}(\tau,0)=1$ associated with intermediate normalization at $\varphi = 0$. Equation \[NormkernelODE1\] possesses a closed-form solution $${\cal N}(\tau,\varphi) = e^{i \int_{0}^{\varphi} \!d\phi \, a(\tau,\phi)} \, , \label{solNormkernelODE}$$ which demonstrates that the logarithm of the off-diagonal norm kernel can be related to the linked/connected kernel of $A$ via an integral over the gauge angle. The linked/connected kernel of $A$ possesses a naturally terminating BCC expansion $$a(\tau,\varphi) = a^{00}(\tau,\varphi) + a^{20}(\tau,\varphi) + a^{11}(\tau,\varphi) + a^{02}(\tau,\varphi) \, ,\nonumber$$ which is obtained by substituting $\Omega^{ij}$ with $A^{ij}$ in Eqs. \[CCenergyequation2A\]-\[CCenergyequation2D\]. Employing the transformed particle-number operator, the algebraic expression takes the compact form $$a(\tau,\varphi) = \tilde{A}^{00}(\varphi) + \frac{1}{2} \displaystyle\sum_{k_1 k_2} \mathcal{T}^{\dagger}_{k_1 k_2}(\tau,\varphi) \, \tilde{A}^{20}_{k_1 k_2}(\varphi)\, ,\label{linkedkernelA2}$$ where the formulae for $\tilde{A}^{00}(\varphi)$ and $\tilde{A}^{20}(\varphi)$ are provided in App \[transformME\]. The diagrams corresponding to Eq. \[linkedkernelA2\] are displayed in Fig. \[diagramsa\] for illustration.
In addition to authorizing the computation of ${\cal N}(\tau,\varphi)$ from a kernel displaying a naturally terminating BCC expansion, the scheme proposed above ensures that the particle number is indeed restored at any truncation order in the proposed many-body method. Indeed, $a(\tau,\varphi)$ and ${\cal N}(\tau,\varphi)$ being related through Eq. \[NormkernelODE1\], one has that $$\begin{aligned}
\int_{0}^{2\pi} \!d\varphi \,e^{-i\text{A}\varphi} \, {\cal A}(\tau,\varphi) &=& -i\int_{0}^{2\pi} \!d\varphi \,e^{-i\text{A}\varphi} \, \frac{d}{d \varphi} \, {\cal N}(\tau,\varphi) \nonumber \\
&=& +i \int_{0}^{2\pi} \!d\varphi \,\frac{d}{d \varphi} \, e^{-i\text{A}\varphi} \, {\cal N}(\tau,\varphi) \nonumber \\
&=& \text{A} \int_{0}^{2\pi} \!d\varphi \, e^{-i\text{A}\varphi} \, {\cal N}(\tau,\varphi) \, , \nonumber\end{aligned}$$ where an integration by part was performed to go from the first to the second line. This demonstrates that, if the reduced norm kernel satisfies Eq. \[NormkernelODE1\], then Eq. \[restoreLieoperators1\] is fulfilled [*independently of the approximation made on*]{} $a(\tau,\varphi)$, i.e. independently of the order at which the BCC expansion of linked/connected operator kernels is truncated. Eventually, the fact that ${\cal N}(\tau,\varphi)$ is determined from the structure of the $U(1)$ group, i.e. from the kernel of $A$, is very natural in the present context. Once extracted through Eq. \[solNormkernelODE\] at a given BCC order, the off-diagonal reduced norm kernel can be consistently used in the computation of the energy as is discussed in Sec. \[symrestE\] below.
### Particle-number conserving case {#normnobreaking}
In case the expansion is performed around a Slater determinant, one must recover the trivial behavior displayed by the norm kernel when global gauge symmetry is conserved. It is indeed easy to demonstrate that
$$\begin{aligned}
\tilde{A}^{00}(\varphi) &=& A^{00}=\text{A}_0 \, , \\
\tilde{A}^{20}_{k_1 k_2}(\varphi) &=& A^{20}_{k_1 k_2}=0 \, ,\end{aligned}$$
in this case, such that $a(\tau,\varphi) =\text{A}_0$. As a result, Eq. \[solNormkernelODE\] does provide the expected result for the reduced norm kernel $${\cal N}(\tau,\varphi) = e^{i \text{A}_0 \varphi} \, , \label{normSD}$$ independently of the truncation employed in the many-body expansion.
### Lowest order {#newunperturbedkernel}
Reducing the calculation to lowest order, i.e. at order $n=0$ in off-diagonal BMBPT or or equivalently taking ${\cal T}^{\dagger}_{n}(\tau,\varphi)=0$ for all $n$ in the off-diagonal BCC scheme, one obtains $$\begin{aligned}
a^{(0)}(\tau,\varphi) &=& \frac{\langle \Phi | A | \Phi(\varphi) \rangle}{\langle \Phi | \Phi(\varphi) \rangle} = -i \frac{d}{d \varphi} \ln \langle \Phi | \Phi(\varphi) \rangle \, , \end{aligned}$$ and thus recovers from Eq. \[solNormkernelODE\] that $$\begin{aligned}
{\cal N}^{(0)}(\tau,\varphi) &=& e^{i \int_{0}^{\varphi} \!d\phi \, a^{(0)}(\tau,\phi)} \nonumber \\
&=& e^{[\ln \langle \Phi | \Phi(\phi) \rangle]^{\varphi}_{0}} \nonumber \\
&=& \langle \Phi | \Phi(\varphi) \rangle \, . \label{solNormkernelODE0order} \end{aligned}$$ Given that at lowest order one has
$$\begin{aligned}
a^{(0)}(\tau,\varphi) &=& \tilde{A}^{00}(\varphi) \\
&=& A^{00} + \frac{1}{2} \displaystyle\sum_{k_1 k_2} A^{02}_{k_1 k_2} R^{--}_{k_2 k_1}(\varphi) \, , \end{aligned}$$
Eq. \[solNormkernelODE0order\] provides a way to compute the overlap between a HFB state and its gauge rotated partner according to $$\begin{aligned}
\langle \Phi | \Phi(\varphi) \rangle &=& e^{i \langle \Phi | A | \Phi \rangle \varphi} \, e^{\frac{i}{2} \sum_{k_1 k_2} A^{02}_{k_1 k_2} \int_{0}^{\varphi} \!d\phi \,R^{--}_{k_2 k_1}(\phi)} \, , \nonumber\end{aligned}$$ as $A^{00}=\langle \Phi | A | \Phi \rangle$. The above expression constitutes an interesting alternative to the Pfaffian formula [@Robledo:2009yd] provided in Eq. \[kernel\] or to even older approaches to the unperturbed norm kernel.
Energy {#symrestE}
======
Particle-number-restored energy
-------------------------------
The particle-number-restored energy is computed according to Eq. \[yrast\_projected\_energy2nd\], now written as $$\begin{aligned}
\text{E}^{\text{A}}_{0} &=& \frac{\int_{0}^{2\pi} \!d\varphi \,e^{-i\text{A}\varphi} \, \, h(\varphi) \,\, {\cal N}(\varphi)}{\int_{0}^{2\pi} \!d\varphi \,e^{-i\text{A}\varphi} \, \, {\cal N}(\varphi)} , \label{projected_energy_MBPT}\end{aligned}$$ where $h(\varphi) \equiv \omega(\varphi) + \lambda a(\varphi)$. Expressing the energy in terms of the [*reduced*]{} norm kernel in Eq. \[projected\_energy\_MBPT\] is essential. Indeed, the fact that ${\cal N}(\tau,\varphi)$ goes to a finite number in the large $\tau$ limit, contrarily to $N(\tau,\varphi)$ that goes exponentially to zero, is mandatory to make the ratio in Eq. \[projected\_energy\_MBPT\] well defined and numerically controllable. The connected/linked kernels $\omega(\varphi)$ and $a(\varphi)$ are to be truncated consistently, i.e. at a given order $n$ in off-diagonal BMBPT or at a given order in off-diagonal BCC amplitudes (including singles, doubles, triples…). The approximate kernel $a(\varphi)$ is also employed to access ${\cal N}(\varphi)$ consistently via Eq. \[solNormkernelODE\] .
If one were to sum all diagrams in the computation of $h(\varphi)$ and ${\cal N}(\varphi)$ or were to expand them around a Slater determinant, the symmetry restoration would become dispensable by definition. This relates to the fact that $h(\varphi)$ becomes independent of $\varphi$ while ${\cal N}(\varphi)$ becomes trivially proportional to the targeted IRREP $e^{i \text{A}_0 \varphi}$ in these two cases. As a result, Eq. \[projected\_energy\_MBPT\] reduces to $$\begin{aligned}
\text{E}^{\text{A}}_0 &=& h(0) \, , \label{standard_MBPT}\end{aligned}$$ for $\text{A}=\text{A}_0$ and zero otherwise.
The benefit of the method arises when the many-body expansion is performed around a Bogoliubov state and is eventually truncated. Indeed, $h(\varphi)$ acquires a dependence on $\varphi$ that signals the breaking of the symmetry generated by that truncation. The method authorizes the summation of standard sets of diagrams (i.e. dealing with so-called [*dynamical*]{} correlations) while leaving the non-perturbative symmetry-restoration process (i.e. dealing with so-called [*static*]{} correlations) to be achieved at each truncation order through the integration over the domain of the group. As a matter of fact, one can rewrite $$\begin{aligned}
\text{E}^{\text{A}}_0 &=& h(0) + \frac{\int_{0}^{2\pi} \!d\varphi \,e^{-i\text{A}\varphi} \,\, \left[h(\varphi)-h(0) \right] \,\, {\cal N}(\varphi)}{\int_{0}^{2\pi} \!d\varphi \,e^{-i\text{A}\varphi} \,\, {\cal N}(\varphi)} \, , \label{rewritingresult}\end{aligned}$$ such that the effect of the particle-number restoration itself can be viewed, at any truncation order, as a [*correction*]{} to the particle-number-breaking BMBPT or BCC results provided by $h(0)$.
Particle-number projected HFB theory {#PHFsection}
------------------------------------
Reducing the calculation to lowest order, i.e. at order $n=0$ in off-diagonal BMBPT or or equivalently taking ${\cal T}^{\dagger}_{n}(\tau,\varphi)=0$ for all $n$ in the off-diagonal BCC scheme, one recovers particle-number projected Hartree-Fock-Bogoliubov (PNP-HFB) theory [@ring80a; @blaizot86] (assuming that the reference state $| \Phi \rangle$ is obtained from a HFB calculation). The associated kernels read as
\[PHFkernels\] $$\begin{aligned}
h^{(0)}(\tau,\varphi) &=& \frac{\langle \Phi | H | \Phi(\varphi) \rangle}{\langle \Phi | \Phi(\varphi) \rangle} \, , \label{PHFkernelsH} \\
{\cal N}^{(0)}(\tau,\varphi) &=& \langle \Phi | \Phi(\varphi) \rangle \label{PHFkernelsN} \, ,\end{aligned}$$
for all $\tau$, such that the symmetry-restored energy becomes
\[projected\_HF\] $$\begin{aligned}
\text{E}^{\text{A} (0)}_{0} &=& \frac{\int_{0}^{2\pi} \!d\varphi \,e^{-i\text{A}\varphi} \, \, \langle \Phi | H | \Phi(\varphi) \rangle}{\int_{0}^{2\pi} \!d\varphi \,e^{-i\text{A}\varphi} \,\, \langle \Phi | \Phi(\varphi) \rangle} \label{projected_HF1} \\
&=& \frac{\langle \Phi | H | \Theta^{\text{A}} \rangle}{\langle \Phi | \Theta^{\text{A}} \rangle} \label{projected_HF2} \\
&=& \frac{\langle \Theta^{\text{A}} | H | \Theta^{\text{A}} \rangle}{\langle \Theta^{\text{A}} | \Theta^{\text{A}} \rangle} \, ,\end{aligned}$$
where the (un-normalized) PNP-HFB wave-function $| \Theta^{\text{A}} \rangle \equiv P^{\text{A}} | \Phi \rangle$ is defined from the projection operator $$P^{\text{A}} \equiv \frac{1}{2\pi}\int_{0}^{2\pi} \!d\varphi \,e^{-i\text{A}\varphi} \, S(\varphi) \, ,$$ satisfying
\[projector\] $$\begin{aligned}
P^{\text{A}}\,P^{\text{A}} &=& P^{\text{A}} \, , \label{projector1} \\
P^{\text{A} \dagger} &=& P^{\text{A}} \, , \label{projector2} \\
\left[H,P^{\text{A}}\right] &=& \left[A,P^{\text{A}}\right] = \left[\Omega,P^{\text{A}}\right] = 0 \, .\end{aligned}$$
Discussion
==========
Let us provide one last set of comments
- In the end, off-diagonal BMBPT and BCC schemes only need to be applied at $\tau=+\infty$, i.e. the imaginary-time formulation becomes superfluous and one is left with the static version of the many-body formalisms.
- PNR-BMBPT and PNR-BCC formalisms are of multi-reference character but reduce in practice to a set of $N_{\text{sym}}$ single-reference-like off-diagonal BMBPT and BCC calculations, where $N_{\text{sym}}\sim 10$. The factor of $10$ is an estimation based on the discretization of the integral over the gauge angle in Eq. \[projected\_energy\_MBPT\] typically needed to achieve convergence in the ground-state computation of even-even nuclei at the PNP-HFB level. Beyond PNP-HFB, the appropriate value of $N_{\text{sym}}$ will have to be validated through numerical tests.
- It could be of interest to design an approximation of the presently proposed many-body formalisms based on a Lipkin expansion method [@lipkin60a]. This would permit the extension of this well-performing approximation to particle-number restoration before variation beyond PNP-HFB [@Wang:2014nza].
- The chemical potential $\lambda$ entering the definition of $\Omega$ must be specified in the calculation. In principle, it must be such that the exact ground state of the $\text{A}_0$-body system of interest is the eigenstate of $\Omega$ with the lowest eigenergy. In practice, $\lambda$ is to be fixed at the single reference level, e.g. when solving the BCC equations ($\varphi =0$) at the chosen level of many-body truncation, e.g. BCCSD. As a further simplification, one can envision to fix $\lambda$ to the value obtained by solving the simpler HFB equations at $\varphi =0$. It remains to be seen how much the choice made to fix $\lambda$ impacts PNR-BMBPT and PNR-BCC results at various levels of truncation.
- The angular-momentum-restored MBPT and CC theory [@Duguet:2014jja] based on the three-parameter non-Abelian $SU(2)$ Lie group has recently undergone its first numerical implementation [@binder15a]. Doing so, it was identified that the exact fulfillment of the second-order ODE necessary to extract the norm kernel associated with an exact restoration of $J^2$ is compromised when going beyond lowest, i.e. HF, order. The remedy to this issue requires an extension of the angular-momentum-restored MBPT and CC theory formulated in Ref. [@Duguet:2014jja] in order to compute $J_z$ and $J^2$ kernels via a “symmetric”, e.g. expectation value, rather than a “projective” formula. This is in fact necessary to match single-reference CC theory at zero Euler angles where properties as $J_z$ and $J^2$ are known to be best approximated from a “symmetric” formula, which is typically achieved via the $\Lambda$CC method [@shavitt09a]. It is thus our goal to formulate such an extension of angular-momentum-restored MBPT and angular-momentum-restored MBPT CC theory in the near future. As for the one-parameter Abelian $U(1)$ group of present interest, while a similar extension will most probably be beneficial, it is not as critical as for $SU(2)$ given that the first-order ODE (Eq. \[NormkernelODE1\]) possesses a closed-form, i.e. exact, solution (Eq. \[solNormkernelODE\]) independently on the approximation used to compute the “projective” particle-number kernel $a(\varphi)$. Thus, one has confidence that PNR-BMBPT and PNR-BCC calculations can be safely performed on the basis of the present formulation while awaiting for their $\Lambda$-like extension in which the particle-number kernel will be evaluated via a symmetric rather than a projective formula.
Implementation algorithm
========================
Let us eventually synthesize the steps the owner of a single-reference BCC code must follow to implement the particle-number restoration procedure[^18].
1. Solve, e.g., Hartree-Fock-Bogoliubov equations in the single-particle basis of interest to obtain the reference state $| \Phi \rangle$ (Eq. \[e:bogvac\]), i.e. to determine the $(U,V)$ matrices.
2. Build matrix elements $\Omega^{ij}_{k_1 \ldots k_{i} k_{i+1}}$ according to App. A of Ref. [@Signoracci:2014dia].
3. Discretize the interval of integration over the gauge angle $\varphi \in [0,2\pi]$.
4. For each angle $\varphi$
1. Compute matrix $R^{--}(\varphi)$ (Eqs. \[offdiaggeneralizeddensitymatrix2\]-\[upperright\]).
2. Build transformed matrix elements $\tilde{\Omega}^{ij}_{k_1 \ldots k_{i} k_{i+1}}(\varphi)$ (App. \[transformME\]).
3. Initialize the cluster amplitudes through first-order perturbation theory; e.g. at the singles and doubles level, apply Eq. \[1storderclusterB\] for $\tau\rightarrow+\infty$ to obtain ${\cal T}^{\dagger (1)}_{k_1k_2}(\varphi)$ and ${\cal T}^{\dagger (1)}_{k_1k_2k_3k_4}(\varphi)$.
4. Run the single-reference BCC code using matrix elements $\tilde{\Omega}^{ij}_{k_1 \ldots k_{i} k_{i+1}}(\varphi)$ and the initial cluster amplitudes ${\cal T}^{\dagger (1)}_{k_1k_2}(\varphi)$ and ${\cal T}^{\dagger (1)}_{k_1k_2k_3k_4}(\varphi)$ as inputs.
5. Use the converged amplitudes, e.g. ${\cal T}^{\dagger}_{1}(\varphi)$ and ${\cal T}^{\dagger}_{2}(\varphi)$, to compute and store the linked/connected kernels $\omega(\varphi)$ and $a(\varphi)$.
5. Using the values of $a(\varphi)$ stored for discretized values of the gauge angle, extract the reduced norm kernel ${\cal N}(\varphi)$ via Eq. \[solNormkernelODE\][^19].
6. Calculate the particle-number restored energy according to Eq. \[projected\_energy\_MBPT\].
Two remarks can be added
- Step 4 can be carried out independently for each value of the gauge angle and is thus amenable to a trivial parallelization. Eventually one solely needs to retrieve and store $\omega(\varphi)$ and $a(\varphi)$.
- In practice, the domain of integration in step 5 can be reduced thanks to symmetries of the reference state $| \Phi \rangle$. For instance, the domain can be limited to $\varphi \in [0,\pi]$ whenever using a reference state $| \Phi \rangle$ with good number parity.
- If limiting oneself to a perturbative approach, one simply needs to replace steps 4.c-4.e by the evaluation of $\omega(\varphi)$ and $a(\varphi)$ at order $n$ in the off-diagonal BMBPT scheme.
Conclusions
===========
Both in Paper I (Ref. [@Duguet:2014jja]) and in the present work, we have addressed a long-term challenge of ab initio many-body theory, i.e. extend symmetry-unrestricted Rayleigh-Schroedinger many-body perturbation theory and coupled-cluster theory in such a way that a broken symmetry is [*exactly*]{} restored at any truncation order. The newly proposed symmetry-restored MBPT and CC formalisms authorize the computation of connected diagrams associated with dynamical correlations while consistently incorporating static correlations through the non-perturbative restoration of the broken symmetry. These approaches are meant to be valid for any symmetry that can be (spontaneously) broken by the reference state and to be applicable to any system independently of its closed-shell, near degenerate or open-shell character.
In Paper I, we focused on the breaking and the restoration of $SU(2)$ rotational symmetry associated with angular momentum conservation. The proposed scheme provides access to the yrast spectroscopy, i.e. the lowest energy for each value of the angular momentum. Standard symmetry-restricted and symmetry-unrestricted MBPT and CC theories, along with angular-momentum-projected Hartree-Fock theory, were recovered as particular cases of the newly developed many-body formalism.
In the present paper, we extend the work to $U(1)$ global gauge symmetry associated with particle-number conservation. In this case, the symmetry-unrestricted single-reference many-body methods upon which the extended formalism builds are Bogoliubov MBPT [@mehta1; @balian62a; @henley] and Bogoliubov CC theory [@StolarczykMonkhorst; @Signoracci:2014dia; @Henderson:2014vka] that can tackle open-shell nuclei displaying a superfluid character at the price of breaking particle-number conservation. The presently proposed formalism overcomes the latter limitation by allowing the restoration of global gauge invariance associated with particle-number conservation, which is mandatory in a finite quantum system such as the atomic nucleus. Technically speaking, the present work follows the same steps as for $SU(2)$ but requires the use of Bogoliubov algebra and thus more general diagrammatic techniques. It leads to designing particle-number-restored Bogoliubov many-body perturbation theory and particle-number-restored Bogoliubov coupled cluster theory.
As the goal is to resolve the near-degenerate nature of the ground state, the proposed many-body formalisms are necessarily of multi-reference character. However, the multi-reference nature is different from any of the multi-reference MBPT or CC methods developed in quantum chemistry [@bartlett07a], i.e. reference states are not obtained from one another via elementary (i.e. dynamical) excitations but via highly non-perturbative symmetry transformations, e.g. rotation in gauge space. Most importantly, the presently proposed method corresponds to performing a set of $N_{\text{sym}}$ single-reference-like Bogoliubov MBPT or Bogoliubov CC calculations, where $N_{\text{sym}}$ corresponds to the discretization of the integral over the gauge angle that is typically of the order of $10$.
The present work offers a wealth of potential applications and further extensions appropriate to the ab initio description of open-shell atomic nuclei. For example, mid-mass singly open-shell nuclei are also being addressed through particle-number-breaking Green’s function calculations under the form of self-consistent Gorkov Green’s function theory [@soma11a; @Soma:2012zd; @Barbieri:2012rd; @Soma:2013vca; @Soma:2013xha]. It is thus of interest to develop the equivalent to the particle-number-restored BCC formalism within the framework of self-consistent Gorkov-Green’s function techniques.
Last but not least, symmetry-restored MBPT and CC theories provide well-founded, formally exact, references for the so-far empirical multi-reference nuclear energy density functional (EDF) method. Multi-reference EDF calculations are known to be compromised with serious pathologies when the off-diagonal EDF kernel is not strictly computed as the matrix element of an [*effective*]{} Hamilton operator between a product state and its rotated partner [@dobaczewski07; @Lacroix:2008rj; @Bender:2008rn; @Duguet:2008rr; @Duguet:2010cv; @Duguet:2013dga], i.e. when it does not take the strict form of an (effective) projected Hartree-Fock-Bogoliubov theory. Starting from the newly proposed many-body formalisms, one can contemplate the derivation of [*safe*]{} parametrizations of off-diagonal EDF kernel that go beyond projected Hartree-Fock-Bogoliubov, most probably under the form of orbital- and symmetry-angle-dependent energy functionals [@Duguet:2015nna]. This remains to be investigated in the future.
T.D. wishes to thank G. Ripka very deeply for enlightening discussions that were instrumental in making this work possible. The authors thank B. Bally and V. Somà for their careful proofreading of the manuscript. This work was supported in part by the U.S. Department of Energy (Oak Ridge National Laboratory), under Grant Nos. DEFG02-96ER40963 (University of Tennessee), DE-SC0008499 (NUCLEI Sci-DAC collaboration), and the Field Work Proposal ERKBP57.
Perturbative expansion of ${\cal U}(\tau)$ {#perturbativeannexe}
==========================================
The imaginary-time evolution operator ${\cal U}(\tau)$ can be expanded in powers of $\Omega_{1}$. Taking $\tau$ real, one writes $${\cal U}(\tau) \equiv e^{-\tau \Omega_{0}} \, {\cal U}_1(\tau) \label{evoloperatorApp1}$$ where $$\begin{aligned}
{\cal U}_1(\tau) &=& e^{\tau \Omega_{0}} \, e^{-\tau(\Omega_{0}+\Omega_{1})} \, , \end{aligned}$$ and thus $$\begin{aligned}
\partial_{\tau} {\cal U}_1(\tau) &=&-e^{\tau \Omega_{0}} \, \Omega_{1} \, e^{-\tau \Omega_{0}} \, {\cal U}_1(\tau) \, \, .\end{aligned}$$ The formal solution to the latter equation reads $$\begin{aligned}
{\cal U}_1(\tau) &=& \textmd{T}e^{-\int_{0}^{\tau}dt \Omega_{1}(t)} \, , \label{reducedevolutionop1} \end{aligned}$$ where $\textmd{T}$ is a time-ordering operator and where $\Omega_{1}(\tau)$ defines the perturbation in the interaction representation $$\begin{aligned}
\Omega_{1}(\tau) &\equiv& e^{\tau \Omega_{0}} \, \Omega_{1} \, e^{-\tau \Omega_{0}} \, . \label{reducedevolutionop2}\end{aligned}$$ Eventually, the full solution reads $${\cal U}(\tau)=e^{-\tau \Omega_{0}} \, \textmd{T}e^{-\int_{0}^{\tau}d\tau \Omega_{1}(\tau)
} \, . \label{exp1}$$
Weight of diagrams and anomalous contractions {#ruleanomalouscontraction}
=============================================
We exemplify the diagrammatic rule, i.e. a factor $1/2$, associated with anomalous lines contracted onto a given vertex. This rule comes in combination with the standard rule associated with equivalent lines. We presently consider the simplest case of a string of operators $$\beta^{\dagger}_1 (\tau_1)\beta^{\dagger}_2 (\tau_1) \ldots \beta^{\dagger}_i (\tau_1) \beta_{i+1} (\tau_1) \beta_{i+2} (\tau_1) \ldots \beta_{i+j} (\tau_1) \nonumber$$ consisting of $i$ quasiparticle creation operators and $j$ quasiparticle annihilation operators[^20] belonging to a vertex at time $\tau_1$. One can then extend the proof to more general cases, i.e. cases where part of the operators belong to one vertex at time $\tau_1$ while the others belong to another vertex at time $\tau_2$…Those more general cases are key as they demonstrate how each anomalous line contracted onto a given vertex requires, once the rule for equivalent lines has been properly taken into account, an extra factor $1/2$ whereas anomalous lines linking two different vertices do not. The proof for more general cases can be straightforwardly generalized from the present one and are thus not reproduced here.
For the present scenario, we assume that $j > i$, but the derivation from $j < i$ proceeds in a similar way to the same result. One must keep track of two aspects of performing contractions: (i) the number of ways a given set of contractions can be obtained and (ii) the rule associated with equivalent lines, either anomalous or normal.
Minimal number of anomalous contractions
----------------------------------------
We begin by evaluating the result when all $i$ quasi-particle creation operators are members of a normal contraction with a quasi-particle annihilation operator, which leaves $(j-i)$ quasi-particle annihilation operators for anomalous contractions[^21]. Thus, the number of ways to pair the $i+j$ operators with minimal anomalous contractions is $$j(j-1)(j-2)\ldots(j-i+1) (j-i-1)!! = \frac{j!}{(j-i)!} (j-i-1)!! \, ,$$ since the first quasi-particle creation operator can be paired with any of the $j$ quasi-particle annihilation operators, while the second can be paired with any of the remaining $j-1$ operators, etc., until the $i^{th}$ can be paired with any of the remaining $j-i+1$. The remaining $j-i$ quasi-particle annihilation operators must contract with each other. Starting from one, it can be paired with any of the remaining $j-i-1$. At this point, $j-i-2$ remain to contract with each other. Selecting one, $j-i-3$ options are available, and once a contraction has occurred, $j-i-4$ operators remain. This results in the double factorial $(j-i-1)!!$.
Applying the diagrammatic rules, one must associate a factor to equivalent lines, which here is $$\left[i! \left(\frac{j-i}{2}\right)!\right]^{-1} \, ,$$ given that the $i$ normal contractions are all equivalent while the $(j-i)/2$ anomalous contractions are also all equivalent in the present example. In further considering the factor $[i!j!]^{-1}$ from the definition of the operator giving rise to the string of creation and annihilation operators, one is left with a factor $$\frac{j!}{(j-i)!} (j-i-1)!! \left[i! \left(\frac{j-i}{2}\right)!\right] \left[i! j! \right]^{-1}=\frac{[(j-i)/2]!}{(j-i)!!} \label{factor1}$$ to account for. Given that $$\begin{aligned}
(2n)!! &=& 2n(2n-2)(2n-4)(2n-6)....2 \nonumber \\
&=& 2n*2(n-1)*2(n-2)*2(n-3)....*2(1) \nonumber \\
&=& 2^n * n! \, ,\end{aligned}$$ the right-hand side of Eq. \[factor1\] can be rewritten as $(1/2)^{(j-i)/2}$, which corresponds to a factor $1/2$ to be attributed to each of the $(j-i)/2$ anomalous lines starting and ending at the operator.
Two additional anomalous contractions
-------------------------------------
When one anomalous contraction occurs between two quasiparticle creation operators, the permutations possible from the pairing of operators is
$$\frac{i(i-1)}{2} j(j-1)(j-2)\ldots(j-i+3) (j-i+1)!! = \frac{j!}{(j-i)!}(j-i-1)!! \frac{i(i-1)}{2} \frac{1}{j-i+2} \; ,$$
where on the left side of the equation, any two of the $i$ quasiparticle creation operators can be combined in the anomalous contraction (i.e., pick one from $i$, then one from remaining $i-1$, and divide by 2 since the order of selection is irrelevant), the remaining $i-2$ quasiparticle creation operators are contracted with quasi-particle annihilation operators, and the remaining $j-i+2$ quasi-particle annihilation operators are contracted amongst themselves. The factors are obtained analogously to the prior subsection. The right-hand side of the equation above is found by separating $(j-1+1)!!$ into $(j-i+1)(j-i-1)!!$ and multiplying by $(j-i+2)/(j-i+2)$ to put the expression in a similar form to the prior subsection.
The factor accounted for by the rule on equivalent lines is in this example $$(i-2)! \left(\frac{j-i+2}{2}\right)! = \frac{i!}{i(i-1)} \left( \frac{j-i+2}{2} \right)\left(\frac{j-i}{2}\right)! \; .$$
Taking the ratio of the counting factor to the one obtained in the previous case gives
$$\frac{ \frac{j!}{(j-i)!}(j-i-1)!! \frac{i(i-1)}{2} \frac{1}{j-i+2} \frac{i!}{i(i-1)} ( \frac{j-i+2}{2} )(\frac{j-i}{2})!}{\frac{j!}{(j-i)!} (j-i-1)!! \, i! (\frac{j-i}{2})!} =
\frac{i(i-1)}{2} \frac{1}{j-i+2} ( \frac{j-i+2}{2} )\frac{1}{i(i-1)} = \frac{1}{4} \: ,$$
which corresponds to a factor $1/2$ for each additional anomalous contraction, since the anomalous contraction among quasi-particle creation operators has induced an additional anomalous contraction among quasi-particle annihilation operators. This again justifies the rule that can be further extended up to the point where all quasi-particle creation operators are contracted amongst themselves.
$n(\tau,\varphi)$ at BMBPT(2) {#diagramsN}
=============================
Applying the diagrammatic rules explicated in Sec. \[diagrulenormMBPT\] to the 18 first- and second-order off-diagonal BMBPT diagrams contributing to $n(\tau,\varphi)$ and displayed in Fig. \[diagramsNc\] gives
$$\begin{aligned}
\text{PN}.1=& + \frac{1}{2} \displaystyle\sum_{k_1 k_2} \frac{\Omega^{02}_{k_1 k_2}}{E_{k_1}+E_{k_2}} \left[ 1-
e^{-\tau(E_{k_1}+E_{k_2})} \right] R^{--}_{k_1k_2}(\varphi) \\
\text{PN}.2=& - \frac{1}{8} \displaystyle\sum_{k_1 k_2 k_3 k_4} \frac{\Omega^{04}_{k_1k_2k_3k_4}}{E_{k_1}+E_{k_2}+E_{k_3}+E_{k_4}} \left[ 1 - e^{-\tau(E_{k_1}+E_{k_2}+E_{k_3}+E_{k_4})} \right] R^{--}_{k_1k_2}(\varphi) R^{--}_{k_3k_4}(\varphi)\\
\text{PN}.3=& + \frac{1}{2} \displaystyle\sum_{k_1 k_2} \frac{\Omega^{02}_{k_1k_2}\Omega^{20}_{k_1k_2}}{E_{k_1}+E_{k_2}}
\left[ \tau - \frac{1-e^{-\tau(E_{k_1}+E_{k_2})}}{E_{k_1}+E_{k_2}} \right]\\
\text{PN}.4=& +\frac{1}{2} \displaystyle\sum_{k_1 k_2 k_3} \frac{\Omega^{02}_{k_1k_3}\Omega^{11}_{k_1k_2}}{E_{k_1}-E_{k_2}} \left[
\frac{1-e^{-\tau(E_{k_1}+E_{k_3})}}{E_{k_1}+E_{k_3}} - \frac{1-e^{-\tau(E_{k_2}+E_{k_3})}}{E_{k_2}+E_{k_3}} \right] R^{--}_{k_2 k_3}(\varphi)\\
\text{PN}.5=& +\frac{1}{2}\displaystyle\sum_{k_1k_2k_3k_4}\frac{\Omega^{02}_{k_1k_2}\Omega^{02}_{k_3k_4}}{(E_{k_1}+E_{k_2})(E_{k_3}+E_{k_4})}
\left[ (1-e^{-\tau(E_{k_1}+E_{k_2})})(1-e^{-\tau(E_{k_3}+E_{k_4})}) \right] R^{--}_{k_1 k_4}(\varphi) R^{--}_{k_2 k_3}(\varphi) \\
\text{PN}.6=& +\frac{1}{4!} \displaystyle\sum_{k_1 k_2 k_3 k_4} \frac{\Omega^{04}_{k_1k_2k_3k_4}\Omega^{40}_{k_1k_2k_3k_4}}{E_{k_1}+E_{k_2}+E_{k_3}+E_{k_4}} \left[ \tau - \frac{1-e^{-\tau(E_{k_1}+E_{k_2}+E_{k_3}+E_{k_4})}}{E_{k_1}+E_{k_2}+E_{k_3}+E_{k_4}} \right]\\
\text{PN}.7=& +\frac{1}{3!} \displaystyle\sum_{\substack{k_1 k_2 k_3 k_4 \\ k_5}} \frac{\Omega^{31}_{k_1k_2k_3k_4} \Omega^{04}_{k_5k_1k_2k_3}}{E_{k_1}+E_{k_2}+E_{k_3}-E_{k_4}} \left[ \frac{1- e^{-\tau(E_{k_4}+E_{k_5})}}{E_{k_4}+E_{k_5}} - \frac{1- e^{-\tau(E_{k_1}+E_{k_2}+E_{k_3}+E_{k_5})}}{E_{k_1}+E_{k_2}+E_{k_3}+E_{k_5}}\right] R^{--}_{k_4k_5}(\varphi) \\
\text{PN}.8=& + \frac{1}{4} \!\!\!\displaystyle\sum_{\substack{k_1 k_2 k_3 k_4 \\ k_5 k_6}} \frac{\Omega^{22}_{k_1k_2k_3k_4} \Omega^{04}_{k_5k_6k_1k_2}}{E_{k_1}+E_{k_2}-E_{k_3}-E_{k_4}} \left[ \frac{1- e^{-\tau(E_{k_3}+E_{k_4}+E_{k_5}+E_{k_6})}}{E_{k_3}+E_{k_4}+E_{k_5}+E_{k_6}} \right.\nonumber \\
& \hspace{7cm} \left.- \frac{1- e^{-\tau(E_{k_1}+E_{k_2}+E_{k_5}+E_{k_6})}}{E_{k_1}+E_{k_2}+E_{k_5}+E_{k_6}}\right] R^{--}_{k_3k_6}(\varphi) R^{--}_{k_4k_5}(\varphi) \\
\text{PN}.9=& +\frac{1}{3!} \displaystyle\sum_{\substack{k_1 k_2 k_3 k_4 \\ k_5 k_6 k_7}} \frac{\Omega^{13}_{k_1k_2k_3k_4} \Omega^{04}_{k_1k_5k_6k_7}}{E_{k_1}-E_{k_2}-E_{k_3}-E_{k_4}} \left[ \frac{1- e^{-\tau(E_{k_2}+E_{k_3}+E_{k_4}+E_{k_5}+E_{k_6}+E_{k_7})}}{E_{k_2}+E_{k_3}+E_{k_4}+E_{k_5}+E_{k_6}+E_{k_7}} \right.\nonumber \\
& \hspace{6cm} \left.- \frac{1- e^{-\tau(E_{k_1}+E_{k_5}+E_{k_6}+E_{k_7})}}{E_{k_1}+E_{k_5}+E_{k_6}+E_{k_7}}\right] R^{--}_{k_2k_7}(\varphi) R^{--}_{k_3k_6}(\varphi) R^{--}_{k_4k_5}(\varphi) \\
\text{PN}.10=& +\frac{1}{4!} \displaystyle\sum_{\substack{k_1 k_2 k_3 k_4 \\ k_5 k_6 k_7 k_8}} \frac{\Omega^{04}_{k_1k_2k_3k_4} \Omega^{04}_{k_5k_6k_7k_8}}{(E_{k_1}+E_{k_2}+E_{k_3}+E_{k_4})(E_{k_5}+E_{k_6}+E_{k_7}+E_{k_8})} \left[1- e^{-\tau(E_{k_1}+E_{k_2}+E_{k_3}+E_{k_4})} \right] \nonumber \\
& \hspace{4cm} \times \left[1- e^{-\tau(E_{k_5}+E_{k_6}+E_{k_7}+E_{k_8})} \right] R^{--}_{k_1k_8}(\varphi) R^{--}_{k_2k_7}(\varphi) R^{--}_{k_3k_6}(\varphi) R^{--}_{k_4k_5}(\varphi) \\
\text{PN}.11=& +\frac{1}{4} \displaystyle\sum_{k_1 k_2 k_3 k_4} \frac{\Omega^{22}_{k_1k_2k_3k_4} \Omega^{02}_{k_1k_2}}{E_{k_1}+E_{k_2}-E_{k_3}-E_{k_4}} \left[ \frac{1- e^{-\tau(E_{k_3}+E_{k_4})}}{E_{k_3}+E_{k_4}} - \frac{1- e^{-\tau(E_{k_1}+E_{k_2})}}{E_{k_1}+E_{k_2}}\right] R^{--}_{k_4k_3}(\varphi) \\
\text{PN}.12=& -\frac{1}{2} \displaystyle\sum_{\substack{k_1 k_2 k_3 k_4 \\ k_5}} \frac{\Omega^{13}_{k_1 k_2 k_3 k_4} \Omega^{02}_{k_1 k_5}}{E_{k_1}-E_{k_2}-E_{k_3}-E_{k_4}} \left[ \frac{1- e^{-\tau(E_{k_1}+E_{k_5})}}{E_{k_1}+E_{k_5}} - \frac{1- e^{-\tau(E_{k_2}+E_{k_3}+E_{k_4}+E_{k_5})}}{E_{k_2}+E_{k_3}+E_{k_4}+E_{k_5}}\right] R^{--}_{k_4k_5}(\varphi) R^{--}_{k_3k_2}(\varphi) \\
\text{PN}.13=& +\frac{1}{4} \displaystyle\sum_{\substack{k_1 k_2 k_3 k_4 \\ k_5 k_6}} \frac{\Omega^{02}_{k_1 k_2} \Omega^{04}_{k_3 k_4 k_5 k_6}}{E_{k_1}+E_{k_2}} \left[ \frac{1- e^{-\tau(E_{k_3}+E_{k_4}+E_{k_5}+E_{k_6})}}{E_{k_3}+E_{k_4}+E_{k_5}+E_{k_6}}\right.\nonumber \\
& \hspace{5cm} \left. - \frac{1- e^{-\tau(E_{k_1}+E_{k_2}+E_{k_3}+E_{k_4}+E_{k_5}+E_{k_6})}}{E_{k_1}+E_{k_2}+E_{k_3}+E_{k_4}+E_{k_5}+E_{k_6}}\right]
R^{--}_{k_2 k_3}(\varphi) R^{--}_{k_1 k_4}(\varphi) R^{--}_{k_6 k_5}(\varphi) \\
\text{PN}.14=& +\frac{1}{4} \displaystyle\sum_{k_1 k_2 k_3 k_4} \frac{\Omega^{20}_{k_1k_2} \Omega^{04}_{k_3 k_4 k_1 k_2}}{E_{k_1}+E_{k_2}} \left[ \frac{1- e^{-\tau(E_{k_3}+E_{k_4})}}{E_{k_3}+E_{k_4}} - \frac{1- e^{-\tau(E_{k_1}+E_{k_2}+E_{k_3}+E_{k_4})}}{E_{k_1}+E_{k_2}+E_{k_3}+E_{k_4}}\right] R^{--}_{k_4k_3}(\varphi) \\
\text{PN}.15=& +\frac{1}{2} \displaystyle\sum_{\substack{k_1 k_2 k_3 k_4 \\ k_5}} \frac{\Omega^{11}_{k_1 k_2} \Omega^{04}_{k_3 k_4 k_1 k_5}}{E_{k_1}-E_{k_2}} \left[ \frac{1- e^{-\tau(E_{k_2}+E_{k_3}+E_{k_4}+E_{k_5})}}{E_{k_2}+E_{k_3}+E_{k_4}+E_{k_5}} \right.\nonumber \\
& \hspace{6cm} \left.- \frac{1- e^{-\tau(E_{k_1}+E_{k_3}+E_{k_4}+E_{k_5})}}{E_{k_1}+E_{k_3}+E_{k_4}+E_{k_5}}\right] R^{--}_{k_2 k_5}(\varphi) R^{--}_{k_4 k_3}(\varphi) \\
\text{PN}.16=& + \frac{1}{8} \!\!\!\displaystyle\sum_{\substack{k_1 k_2 k_3 k_4 \\ k_5 k_6}} \frac{\Omega^{22}_{k_1k_2k_3k_4} \Omega^{04}_{k_5k_6k_1k_2}}{E_{k_1}+E_{k_2}-E_{k_3}-E_{k_4}} \left[ \frac{1- e^{-\tau(E_{k_3}+E_{k_4}+E_{k_5}+E_{k_6})}}{E_{k_3}+E_{k_4}+E_{k_5}+E_{k_6}} \right.\nonumber \\
& \hspace{7cm} \left.- \frac{1- e^{-\tau(E_{k_1}+E_{k_2}+E_{k_5}+E_{k_6})}}{E_{k_1}+E_{k_2}+E_{k_5}+E_{k_6}}\right] R^{--}_{k_3k_4}(\varphi) R^{--}_{k_5k_6}(\varphi) \\
\text{PN}.17=& +\frac{1}{4} \displaystyle\sum_{\substack{k_1 k_2 k_3 k_4 \\ k_5 k_6 k_7}} \frac{\Omega^{13}_{k_1k_2k_3k_4} \Omega^{04}_{k_1k_5k_6k_7}}{E_{k_1}-E_{k_2}-E_{k_3}-E_{k_4}} \left[ \frac{1- e^{-\tau(E_{k_2}+E_{k_3}+E_{k_4}+E_{k_5}+E_{k_6}+E_{k_7})}}{E_{k_2}+E_{k_3}+E_{k_4}+E_{k_5}+E_{k_6}+E_{k_7}} \right.\nonumber \\
& \hspace{6cm} \left.- \frac{1- e^{-\tau(E_{k_1}+E_{k_5}+E_{k_6}+E_{k_7})}}{E_{k_1}+E_{k_5}+E_{k_6}+E_{k_7}}\right] R^{--}_{k_2k_7}(\varphi) R^{--}_{k_3k_4}(\varphi) R^{--}_{k_5k_6}(\varphi) \\
\text{PN}.18=& +\frac{1}{8} \displaystyle\sum_{\substack{k_1 k_2 k_3 k_4 \\ k_5 k_6 k_7 k_8}} \frac{\Omega^{04}_{k_1k_2k_3k_4} \Omega^{04}_{k_5k_6k_7k_8}}{(E_{k_1}+E_{k_2}+E_{k_3}+E_{k_4})(E_{k_5}+E_{k_6}+E_{k_7}+E_{k_8})} \left[1- e^{-\tau(E_{k_1}+E_{k_2}+E_{k_3}+E_{k_4})} \right] \nonumber \\
& \hspace{4cm} \times \left[1- e^{-\tau(E_{k_5}+E_{k_6}+E_{k_7}+E_{k_8})} \right] R^{--}_{k_1k_8}(\varphi) R^{--}_{k_2k_7}(\varphi) R^{--}_{k_3k_4}(\varphi) R^{--}_{k_5k_6}(\varphi) \, ,
$$
which reduce in the infinite $\tau$ limit to
$$\begin{aligned}
\text{PN}.1=& +\frac{1}{2} \displaystyle\sum_{k_1 k_2} \frac{\Omega^{02}_{k_1 k_2}}{E_{k_1}+E_{k_2}} R^{--}_{k_1k_2}(\varphi) \\
\text{PN}.2=& -\frac{1}{8} \displaystyle\sum_{k_1 k_2 k_3 k_4} \frac{\Omega^{04}_{k_1k_2k_3k_4}}{E_{k_1}+E_{k_2}+E_{k_3}+E_{k_4}} R^{--}_{k_1k_2}(\varphi) R^{--}_{k_3k_4}(\varphi) \\
\text{PN}.3=& +\frac{1}{2} \displaystyle\sum_{k_1 k_2} \frac{\Omega^{02}_{k_1k_2}\Omega^{20}_{k_1k_2}}{E_{k_1}+E_{k_2}} \left[ \tau - \frac{1}{E_{k_1}+E_{k_2}} \right] \\
\text{PN}.4=& -\frac{1}{2} \displaystyle\sum_{k_1 k_2 k_3} \frac{\Omega^{02}_{k_1k_3}\Omega^{11}_{k_1k_2}}{(E_{k_1}+E_{k_2})(E_{k_2}+E_{k_3})} R^{--}_{k_2 k_3}(\varphi)\\
\text{PN}.5=& +\frac{1}{2}\displaystyle\sum_{k_1k_2k_3k_4}\frac{\Omega^{02}_{k_1k_2}\Omega^{02}_{k_3k_4}}{(E_{k_1}+E_{k_2})(E_{k_3}+E_{k_4})}
R^{--}_{k_1 k_4}(\varphi) R^{--}_{k_2 k_3}(\varphi) \\
\text{PN}.6=& +\frac{1}{4!} \displaystyle\sum_{k_1 k_2 k_3 k_4} \frac{\Omega^{04}_{k_1k_2k_3k_4}\Omega^{40}_{k_1k_2k_3k_4}}{E_{k_1}+E_{k_2}+E_{k_3}+E_{k_4}} \left[ \tau - \frac{1}{E_{k_1}+E_{k_2}+E_{k_3}+E_{k_4}} \right] \\
\text{PN}.7=& +\frac{1}{3!} \displaystyle\sum_{\substack{k_1 k_2 k_3 k_4 \\ k_5}} \frac{\Omega^{31}_{k_1k_2k_3k_4} \Omega^{04}_{k_5k_1k_2k_3}}{(E_{k_4}+E_{k_5})(E_{k_1}+E_{k_2}+E_{k_3}+E_{k_5})}R^{--}_{k_4k_5}(\varphi) \\
\text{PN}.8=& + \frac{1}{4} \displaystyle\sum_{\substack{k_1 k_2 k_3 k_4 \\ k_5 k_6}} \frac{\Omega^{22}_{k_1k_2k_3k_4} \Omega^{04}_{k_5k_6k_1k_2}}{(E_{k_3}+E_{k_4}+E_{k_5}+E_{k_6})(E_{k_1}+E_{k_2}+E_{k_5}+E_{k_6})} R^{--}_{k_3k_6}(\varphi) R^{--}_{k_4k_5}(\varphi) \\
\text{PN}.9=& +\frac{1}{3!} \displaystyle\sum_{\substack{k_1 k_2 k_3 k_4 \\ k_5 k_6 k_7}} \frac{\Omega^{13}_{k_1k_2k_3k_4} \Omega^{04}_{k_1k_5k_6k_7}}{(E_{k_2}+E_{k_3}+E_{k_4}+E_{k_5}+E_{k_6}+E_{k_7})(E_{k_1}+E_{k_5}+E_{k_6}+E_{k_7})} R^{--}_{k_2k_7}(\varphi) R^{--}_{k_3k_6}(\varphi) R^{--}_{k_4k_5}(\varphi) \\
\text{PN}.10=& +\frac{1}{4!} \displaystyle\sum_{\substack{k_1 k_2 k_3 k_4 \\ k_5 k_6 k_7 k_8}} \frac{\Omega^{04}_{k_1k_2k_3k_4} \Omega^{04}_{k_5k_6k_7k_8}}{(E_{k_1}+E_{k_2}+E_{k_3}+E_{k_4})(E_{k_5}+E_{k_6}+E_{k_7}+E_{k_8})} R^{--}_{k_1k_8}(\varphi) R^{--}_{k_2k_7}(\varphi) R^{--}_{k_3k_6}(\varphi) R^{--}_{k_4k_5}(\varphi) \\
\text{PN}.11=& +\frac{1}{4} \displaystyle\sum_{k_1 k_2 k_3 k_4} \frac{\Omega^{22}_{k_1k_2k_3k_4} \Omega^{02}_{k_1k_2}}
{(E_{k_3}+E_{k_4})(E_{k_1}+E_{k_2})} R^{--}_{k_4k_3}(\varphi) \\
\text{PN}.12=& +\frac{1}{2} \displaystyle\sum_{\substack{k_1 k_2 k_3 k_4 \\ k_5}} \frac{\Omega^{13}_{k_1 k_2 k_3 k_4} \Omega^{02}_{k_1 k_5}}{(E_{k_1}+E_{k_5})(E_{k_2}+E_{k_3}+E_{k_4}+E_{k_5})} R^{--}_{k_4k_5}(\varphi) R^{--}_{k_3k_2}(\varphi) \\
\text{PN}.13=& +\frac{1}{4} \displaystyle\sum_{\substack{k_1 k_2 k_3 k_4 \\ k_5 k_6}} \frac{\Omega^{02}_{k_1 k_2} \Omega^{04}_{k_3 k_4 k_5 k_6}}{(E_{k_3}+E_{k_4}+E_{k_5}+E_{k_6})(E_{k_1}+E_{k_2}+E_{k_3}+E_{k_4}+E_{k_5}+E_{k_6})} R^{--}_{k_2 k_3}(\varphi) R^{--}_{k_1 k_4}(\varphi) R^{--}_{k_6 k_5}(\varphi) \\
\text{PN}.14=& +\frac{1}{4} \displaystyle\sum_{k_1 k_2 k_3 k_4} \frac{\Omega^{20}_{k_1k_2} \Omega^{04}_{k_3 k_4 k_1 k_2}}{(E_{k_3}+E_{k_4})(E_{k_1}+E_{k_2}+E_{k_3}+E_{k_4})} R^{--}_{k_4k_3}(\varphi) \\
\text{PN}.15=& +\frac{1}{2} \displaystyle\sum_{\substack{k_1 k_2 k_3 k_4 \\ k_5}} \frac{\Omega^{11}_{k_1 k_2} \Omega^{04}_{k_3 k_4 k_1 k_5}}{(E_{k_2}+E_{k_3}+E_{k_4}+E_{k_5})(E_{k_1}+E_{k_3}+E_{k_4}+E_{k_5})} R^{--}_{k_2 k_5}(\varphi) R^{--}_{k_4 k_3}(\varphi) \\
\text{PN}.16=& + \frac{1}{8} \displaystyle\sum_{\substack{k_1 k_2 k_3 k_4 \\ k_5 k_6}} \frac{\Omega^{22}_{k_1k_2k_3k_4} \Omega^{04}_{k_5k_6k_1k_2}}{(E_{k_3}+E_{k_4}+E_{k_5}+E_{k_6})(E_{k_1}+E_{k_2}+E_{k_5}+E_{k_6})} R^{--}_{k_3k_4}(\varphi) R^{--}_{k_5k_6}(\varphi) \\
\text{PN}.17=& +\frac{1}{4} \displaystyle\sum_{\substack{k_1 k_2 k_3 k_4 \\ k_5 k_6 k_7}} \frac{\Omega^{13}_{k_1k_2k_3k_4} \Omega^{04}_{k_1k_5k_6k_7}}{(E_{k_2}+E_{k_3}+E_{k_4}+E_{k_5}+E_{k_6}+E_{k_7})(E_{k_1}+E_{k_5}+E_{k_6}+E_{k_7})} R^{--}_{k_2k_7}(\varphi) R^{--}_{k_3k_4}(\varphi) R^{--}_{k_5k_6}(\varphi) \\
\text{PN}.18=& +\frac{1}{8} \displaystyle\sum_{\substack{k_1 k_2 k_3 k_4 \\ k_5 k_6 k_7 k_8}} \frac{\Omega^{04}_{k_1k_2k_3k_4} \Omega^{04}_{k_5k_6k_7k_8}}{(E_{k_1}+E_{k_2}+E_{k_3}+E_{k_4})(E_{k_5}+E_{k_6}+E_{k_7}+E_{k_8})} R^{--}_{k_1k_8}(\varphi) R^{--}_{k_2k_7}(\varphi) R^{--}_{k_3k_4}(\varphi) R^{--}_{k_5k_6}(\varphi) \, .
$$
$\omega(\tau,\varphi)$ at BMBPT(1) {#diagramsE}
==================================
Applying diagrammatic rules to the 20 zero- and first-order off-diagonal BMBPT connected/linked diagrams contributing to $\omega(\tau,\varphi)$ and displayed in Fig. \[diagramsTL\] gives
$$\begin{aligned}
\text{PE}.1=& + \Omega^{00} \, , \\
\text{PE}.2=& + \frac{1}{2} \displaystyle\sum_{k_1 k_2} \Omega^{02}_{k_1 k_2} R^{--}_{k_2k_1}(\varphi) \, , \\
\text{PE}.3=& + \frac{1}{8} \displaystyle\sum_{k_1 k_2 k_3 k_4} \Omega^{04}_{k_1 k_2 k_3 k_4} R^{--}_{k_2k_1}(\varphi) R^{--}_{k_4k_3}(\varphi) \, , \\
\text{PE}.4=& - \frac{1}{2} \displaystyle\sum_{k_1 k_2} \frac{\Omega^{02}_{k_1 k_2} \Omega^{20}_{k_1 k_2}}{E_{k_1}+E_{k_2}} \left[ 1- e^{-\tau(E_{k_1}+E_{k_2})}\right] \, , \\
\text{PE}.5=& - \displaystyle\sum_{k_1 k_2 k_3} \frac{\Omega^{02}_{k_1 k_2} \Omega^{11}_{k_1 k_3}}{E_{k_1}+E_{k_2}} \left[ 1- e^{-\tau(E_{k_1}+E_{k_2})}\right] R^{--}_{k_2k_3}(\varphi) \, , \\
\text{PE}.6=& - \frac{1}{2} \displaystyle\sum_{k_1 k_2 k_3 k_4} \frac{\Omega^{02}_{k_1 k_2} \Omega^{02}_{k_3 k_4}}{E_{k_1}+E_{k_2}} \left[ 1- e^{-\tau(E_{k_1}+E_{k_2})}\right] R^{--}_{k_1k_4}(\varphi) R^{--}_{k_2k_3}(\varphi) \, , \\
\text{PE}.7=& -\frac{1}{4!} \displaystyle\sum_{k_1 k_2 k_3 k_4} \frac{\Omega^{04}_{k_1 k_2 k_3 k_4}\Omega^{40}_{k_1 k_2 k_3 k_4}}{E_{k_1}+E_{k_2}+E_{k_3}+E_{k_4}} \left[ 1- e^{-\tau(E_{k_1}+E_{k_2}+E_{k_3}+E_{k_4})}\right] \, , \\
\text{PE}.8=& -\frac{1}{3!} \displaystyle\sum_{\substack{k_1 k_2 k_3 k_4 \\ k_5}} \frac{\Omega^{04}_{k_1k_2k_3k_4} \Omega^{31}_{k_1k_2k_3k_5} }{E_{k_1}+E_{k_2}+E_{k_3}+E_{k_4}} \left[ 1- e^{-\tau(E_{k_1}+E_{k_2}+E_{k_3}+E_{k_4})}\right] R^{--}_{k_4k_5}(\varphi) \, , \\
\text{PE}.9=& -\frac{1}{4} \displaystyle\sum_{\substack{k_1 k_2 k_3 k_4 \\ k_5 k_6}} \frac{\Omega^{04}_{k_1 k_2 k_3 k_4}\Omega^{22}_{k_1 k_2 k_5 k_6}}{E_{k_1}+E_{k_2}+E_{k_3}+E_{k_4}} \left[ 1- e^{-\tau(E_{k_1}+E_{k_2}+E_{k_3}+E_{k_4})}\right] R^{--}_{k_3k_6}(\varphi) R^{--}_{k_4k_5}(\varphi) \, , \\
\text{PE}.10=& -\frac{1}{3!} \displaystyle\sum_{\substack{k_1 k_2 k_3 k_4 \\ k_5 k_6 k_7}} \frac{\Omega^{04}_{k_1 k_2 k_3 k_4}\Omega^{13}_{k_1 k_5 k_6 k_7}}{E_{k_1}+E_{k_2}+E_{k_3}+E_{k_4}} \left[ 1- e^{-\tau(E_{k_1}+E_{k_2}+E_{k_3}+E_{k_4})}\right] R^{--}_{k_2k_8}(\varphi) R^{--}_{k_3k_6}(\varphi) R^{--}_{k_4k_5}(\varphi) \, , \\
\text{PE}.11=& -\frac{1}{4!} \displaystyle\sum_{\substack{k_1 k_2 k_3 k_4 \\ k_5 k_6 k_7 k_8}} \frac{\Omega^{04}_{k_1 k_2 k_3 k_4}\Omega^{04}_{k_5 k_6 k_7 k_8}}{E_{k_1}+E_{k_2}+E_{k_3}+E_{k_4}} \left[ 1- e^{-\tau(E_{k_1}+E_{k_2}+E_{k_3}+E_{k_4})}\right] R^{--}_{k_1k_8}(\varphi) R^{--}_{k_2k_7}(\varphi) R^{--}_{k_3k_6}(\varphi) R^{--}_{k_4k_5}(\varphi) \, , \\
\text{PE}.12=& - \frac{1}{4} \displaystyle\sum_{k_1 k_2 k_3 k_4} \frac{\Omega^{02}_{k_1 k_2} \Omega^{22}_{k_1 k_2 k_3 k_4}}{E_{k_1}+E_{k_2}} \left[ 1- e^{-\tau(E_{k_1}+E_{k_2})}\right] R^{--}_{k_4k_3}(\varphi) \, , \\
\text{PE}.13=& - \frac{1}{2} \displaystyle\sum_{\substack{k_1 k_2 k_3 k_4 \\ k_5}} \frac{\Omega^{02}_{k_1 k_2} \Omega^{13}_{k_1 k_3 k_4 k_5}}{E_{k_1}+E_{k_2}} \left[ 1- e^{-\tau(E_{k_1}+E_{k_2})}\right] R^{--}_{k_2k_5}(\varphi) R^{--}_{k_4k_3}(\varphi) \, , \\
\text{PE}.14=& - \frac{1}{4} \displaystyle\sum_{\substack{k_1 k_2 k_3 k_4 \\ k_5 k_6}} \frac{\Omega^{02}_{k_1 k_2} \Omega^{04}_{k_3 k_4 k_5 k_6}}{E_{k_1}+E_{k_2}} \left[ 1- e^{-\tau(E_{k_1}+E_{k_2})}\right] R^{--}_{k_1k_6}(\varphi) R^{--}_{k_2k_5}(\varphi) R^{--}_{k_4k_3}(\varphi) \, , \\
\text{PE}.15=& - \frac{1}{4} \displaystyle\sum_{k_1 k_2 k_3 k_4} \frac{\Omega^{04}_{k_1 k_2 k_3 k_4} \Omega^{20}_{k_1 k_2}}{E_{k_1}+E_{k_2}+E_{k_3}+E_{k_4}} \left[ 1- e^{-\tau(E_{k_1}+E_{k_2}+E_{k_3}+E_{k_4})}\right] R^{--}_{k_4k_3}(\varphi) \, , \\
\text{PE}.16=& -\frac{1}{2} \displaystyle\sum_{\substack{k_1 k_2 k_3 k_4 \\ k_5}} \frac{\Omega^{04}_{k_1k_2k_3k_4} \Omega^{11}_{k_1k_5} }{E_{k_1}+E_{k_2}+E_{k_3}+E_{k_4}} \left[ 1- e^{-\tau(E_{k_1}+E_{k_2}+E_{k_3}+E_{k_4})}\right] R^{--}_{k_2k_5}(\varphi) R^{--}_{k_4k_3}(\varphi) \, , \\
\text{PE}.17=& -\frac{1}{4} \displaystyle\sum_{\substack{k_1 k_2 k_3 k_4 \\ k_5 k_6}} \frac{\Omega^{04}_{k_1 k_2 k_3 k_4}\Omega^{02}_{k_5 k_6}}{E_{k_1}+E_{k_2}+E_{k_3}+E_{k_4}} \left[ 1- e^{-\tau(E_{k_1}+E_{k_2}+E_{k_3}+E_{k_4})}\right] R^{--}_{k_1k_6}(\varphi) R^{--}_{k_2k_5}(\varphi) R^{--}_{k_4k_3}(\varphi) \, , \\
\text{PE}.18=& -\frac{1}{8} \displaystyle\sum_{\substack{k_1 k_2 k_3 k_4 \\ k_5 k_6}} \frac{\Omega^{04}_{k_1 k_2 k_3 k_4}\Omega^{22}_{k_1 k_2 k_5 k_6}}{E_{k_1}+E_{k_2}+E_{k_3}+E_{k_4}} \left[ 1- e^{-\tau(E_{k_1}+E_{k_2}+E_{k_3}+E_{k_4})}\right] R^{--}_{k_3k_4}(\varphi) R^{--}_{k_5k_6}(\varphi) \, , \\
\text{PE}.19=& -\frac{1}{4} \displaystyle\sum_{\substack{k_1 k_2 k_3 k_4 \\ k_5 k_6 k_7}} \frac{\Omega^{04}_{k_1 k_2 k_3 k_4}\Omega^{13}_{k_1 k_5 k_6 k_7}}{E_{k_1}+E_{k_2}+E_{k_3}+E_{k_4}} \left[ 1- e^{-\tau(E_{k_1}+E_{k_2}+E_{k_3}+E_{k_4})}\right] R^{--}_{k_2k_8}(\varphi) R^{--}_{k_3k_4}(\varphi) R^{--}_{k_5k_6}(\varphi) \, , \\
\text{PE}.20=& -\frac{1}{8} \displaystyle\sum_{\substack{k_1 k_2 k_3 k_4 \\ k_5 k_6 k_7 k_8}} \frac{\Omega^{04}_{k_1 k_2 k_3 k_4}\Omega^{04}_{k_5 k_6 k_7 k_8}}{E_{k_1}+E_{k_2}+E_{k_3}+E_{k_4}} \left[ 1- e^{-\tau(E_{k_1}+E_{k_2}+E_{k_3}+E_{k_4})}\right] R^{--}_{k_1k_8}(\varphi) R^{--}_{k_2k_7}(\varphi) R^{--}_{k_3k_4}(\varphi) R^{--}_{k_5k_6}(\varphi) \, ,
\label{PE1}\end{aligned}$$
which reduce in the infinite $\tau$ limit to
$$\begin{aligned}
\text{PE}.1=& + \Omega^{00} \, , \\
\text{PE}.2=& + \frac{1}{2} \displaystyle\sum_{k_1 k_2} \Omega^{02}_{k_1 k_2} R^{--}_{k_2k_1}(\varphi) \, , \\
\text{PE}.3=& + \frac{1}{8} \displaystyle\sum_{k_1 k_2 k_3 k_4} \Omega^{04}_{k_1 k_2 k_3 k_4} R^{--}_{k_2k_1}(\varphi) R^{--}_{k_4k_3}(\varphi) \, , \\
\text{PE}.4=& - \frac{1}{2} \displaystyle\sum_{k_1 k_2} \frac{\Omega^{02}_{k_1 k_2} \Omega^{20}_{k_1 k_2}}{E_{k_1}+E_{k_2}} \, , \\
\text{PE}.5=& - \displaystyle\sum_{k_1 k_2 k_3} \frac{\Omega^{02}_{k_1 k_2} \Omega^{11}_{k_1 k_3}}{E_{k_1}+E_{k_2}} R^{--}_{k_2k_3}(\varphi) \, , \\
\text{PE}.6=& - \frac{1}{2} \displaystyle\sum_{k_1 k_2 k_3 k_4} \frac{\Omega^{02}_{k_1 k_2} \Omega^{02}_{k_3 k_4}}{E_{k_1}+E_{k_2}} R^{--}_{k_1k_4}(\varphi) R^{--}_{k_2k_3}(\varphi) \, , \\
\text{PE}.7=& -\frac{1}{4!} \displaystyle\sum_{k_1 k_2 k_3 k_4} \frac{\Omega^{04}_{k_1 k_2 k_3 k_4}\Omega^{40}_{k_1 k_2 k_3 k_4}}{E_{k_1}+E_{k_2}+E_{k_3}+E_{k_4}} \, , \\
\text{PE}.8=& -\frac{1}{3!} \displaystyle\sum_{\substack{k_1 k_2 k_3 k_4 \\ k_5}} \frac{\Omega^{04}_{k_1k_2k_3k_4} \Omega^{31}_{k_1k_2k_3k_5} }{E_{k_1}+E_{k_2}+E_{k_3}+E_{k_4}} R^{--}_{k_4k_5}(\varphi) \, , \\
\text{PE}.9=& -\frac{1}{4} \displaystyle\sum_{\substack{k_1 k_2 k_3 k_4 \\ k_5 k_6}} \frac{\Omega^{04}_{k_1 k_2 k_3 k_4}\Omega^{22}_{k_1 k_2 k_5 k_6}}{E_{k_1}+E_{k_2}+E_{k_3}+E_{k_4}} R^{--}_{k_3k_6}(\varphi) R^{--}_{k_4k_5}(\varphi) \, , \\
\text{PE}.10=& -\frac{1}{3!} \displaystyle\sum_{\substack{k_1 k_2 k_3 k_4 \\ k_5 k_6 k_7}} \frac{\Omega^{04}_{k_1 k_2 k_3 k_4}\Omega^{13}_{k_1 k_5 k_6 k_7}}{E_{k_1}+E_{k_2}+E_{k_3}+E_{k_4}} R^{--}_{k_2k_8}(\varphi) R^{--}_{k_3k_6}(\varphi) R^{--}_{k_4k_5}(\varphi) \, , \\
\text{PE}.11=& -\frac{1}{4!} \displaystyle\sum_{\substack{k_1 k_2 k_3 k_4 \\ k_5 k_6 k_7 k_8}} \frac{\Omega^{04}_{k_1 k_2 k_3 k_4}\Omega^{04}_{k_5 k_6 k_7 k_8}}{E_{k_1}+E_{k_2}+E_{k_3}+E_{k_4}} R^{--}_{k_1k_8}(\varphi) R^{--}_{k_2k_7}(\varphi) R^{--}_{k_3k_6}(\varphi) R^{--}_{k_4k_5}(\varphi) \, , \\
\text{PE}.12=& - \frac{1}{4} \displaystyle\sum_{k_1 k_2 k_3 k_4} \frac{\Omega^{02}_{k_1 k_2} \Omega^{22}_{k_1 k_2 k_3 k_4}}{E_{k_1}+E_{k_2}} R^{--}_{k_4k_3}(\varphi) \, , \\
\text{PE}.13=& - \frac{1}{2} \displaystyle\sum_{\substack{k_1 k_2 k_3 k_4 \\ k_5}} \frac{\Omega^{02}_{k_1 k_2} \Omega^{13}_{k_1 k_3 k_4 k_5}}{E_{k_1}+E_{k_2}} R^{--}_{k_2k_5}(\varphi) R^{--}_{k_4k_3}(\varphi) \, , \\
\text{PE}.14=& - \frac{1}{4} \displaystyle\sum_{\substack{k_1 k_2 k_3 k_4 \\ k_5 k_6}} \frac{\Omega^{02}_{k_1 k_2} \Omega^{04}_{k_3 k_4 k_5 k_6}}{E_{k_1}+E_{k_2}} R^{--}_{k_1k_6}(\varphi) R^{--}_{k_2k_5}(\varphi) R^{--}_{k_4k_3}(\varphi) \, , \\
\text{PE}.15=& - \frac{1}{4} \displaystyle\sum_{k_1 k_2 k_3 k_4} \frac{\Omega^{04}_{k_1 k_2 k_3 k_4} \Omega^{20}_{k_1 k_2}}{E_{k_1}+E_{k_2}+E_{k_3}+E_{k_4}} R^{--}_{k_4k_3}(\varphi) \, , \\
\text{PE}.16=& -\frac{1}{2} \displaystyle\sum_{\substack{k_1 k_2 k_3 k_4 \\ k_5}} \frac{\Omega^{04}_{k_1k_2k_3k_4} \Omega^{11}_{k_1k_5} }{E_{k_1}+E_{k_2}+E_{k_3}+E_{k_4}} R^{--}_{k_2k_5}(\varphi) R^{--}_{k_4k_3}(\varphi) \, , \\
\text{PE}.17=& -\frac{1}{4} \displaystyle\sum_{\substack{k_1 k_2 k_3 k_4 \\ k_5 k_6}} \frac{\Omega^{04}_{k_1 k_2 k_3 k_4}\Omega^{02}_{k_5 k_6}}{E_{k_1}+E_{k_2}+E_{k_3}+E_{k_4}} R^{--}_{k_1k_6}(\varphi) R^{--}_{k_2k_5}(\varphi) R^{--}_{k_4k_3}(\varphi) \, , \\
\text{PE}.18=& -\frac{1}{8} \displaystyle\sum_{\substack{k_1 k_2 k_3 k_4 \\ k_5 k_6}} \frac{\Omega^{04}_{k_1 k_2 k_3 k_4}\Omega^{22}_{k_1 k_2 k_5 k_6}}{E_{k_1}+E_{k_2}+E_{k_3}+E_{k_4}} R^{--}_{k_3k_4}(\varphi) R^{--}_{k_5k_6}(\varphi) \, , \\
\text{PE}.19=& -\frac{1}{4} \displaystyle\sum_{\substack{k_1 k_2 k_3 k_4 \\ k_5 k_6 k_7}} \frac{\Omega^{04}_{k_1 k_2 k_3 k_4}\Omega^{13}_{k_1 k_5 k_6 k_7}}{E_{k_1}+E_{k_2}+E_{k_3}+E_{k_4}} R^{--}_{k_2k_8}(\varphi) R^{--}_{k_3k_4}(\varphi) R^{--}_{k_5k_6}(\varphi) \, , \\
\text{PE}.20=& -\frac{1}{8} \displaystyle\sum_{\substack{k_1 k_2 k_3 k_4 \\ k_5 k_6 k_7 k_8}} \frac{\Omega^{04}_{k_1 k_2 k_3 k_4}\Omega^{04}_{k_5 k_6 k_7 k_8}}{E_{k_1}+E_{k_2}+E_{k_3}+E_{k_4}} R^{--}_{k_1k_8}(\varphi) R^{--}_{k_2k_7}(\varphi) R^{--}_{k_3k_4}(\varphi) R^{--}_{k_5k_6}(\varphi) \, .
$$
$\omega(\tau,\varphi)$ from BCC {#CCenergycontrib}
===============================
The algebraic expressions of the twenty off-diagonal BCC diagrams contributing to $\omega(\tau, \varphi)$ and displayed in Fig. \[T1contribtokinetic\] are
\[variousEcontrib\] $$\begin{aligned}
\text{E}.1=& +\Omega^{00} \, , \\
\text{E}.2=& + \frac{1}{2} \displaystyle\sum_{k_1 k_2} \Omega^{02}_{k_1 k_2} R^{--}_{k_2 k_1}(\varphi) \, , \\
\text{E}.3=& +\frac{1}{8} \displaystyle\sum_{k_1 k_2 k_3 k_4} \Omega^{04}_{k_1 k_2 k_3 k_4} R^{--}_{k_2 k_1}(\varphi)R^{--}_{k_4 k_3}(\varphi) \, , \\
\text{E}.4 =& + \frac{1}{2} \displaystyle\sum_{k_1 k_2} \mathcal{T}^{\dagger}_{k_1 k_2}(\tau,\varphi) \, \Omega^{20}_{k_1 k_2} \, , \\
\text{E}.5 =& +\displaystyle\sum_{k_1 k_2 k_3} \mathcal{T}^{\dagger}_{k_1 k_2}(\tau,\varphi) \, \Omega^{11}_{k_1 k_3} R^{--}_{k_2 k_3} (\varphi) \, , \\
\text{E}.6 =& +\frac{1}{2} \displaystyle\sum_{k_1 k_2 k_3 k_4} \mathcal{T}^{\dagger}_{k_1 k_2}(\tau,\varphi) \, \Omega^{02}_{k_3 k_4} R^{--}_{k_1 k_4} (\varphi) R^{--}_{k_2 k_3} (\varphi) \, , \\
\text{E}.7 =& + \frac{1}{4!} \displaystyle\sum_{k_1 k_2 k_3 k_4} \mathcal{T}^{\dagger}_{k_1 k_2 k_3 k_4}(\tau,\varphi) \, \Omega^{40}_{k_1 k_2 k_3 k_4} \, , \\
\text{E}.8 =& +\frac{1}{3!} \displaystyle\sum_{\substack{k_1 k_2 k_3 k_4 \\ k_5}} \mathcal{T}^{\dagger}_{k_1 k_2 k_3 k_4}(\tau,\varphi) \, \Omega^{31}_{k_1 k_2 k_3 k_5}
R^{--}_{k_4 k_5}(\varphi) \, , \\
\text{E}.9 =& + \frac{1}{4} \displaystyle\sum_{\substack{k_1 k_2 k_3 k_4 \\ k_5 k_6}} \mathcal{T}^{\dagger}_{k_1 k_2 k_3 k_4}(\tau,\varphi) \, \Omega^{22}_{k_1 k_2 k_5 k_6}
R^{--}_{k_4 k_5}(\varphi) R^{--}_{k_3 k_6}(\varphi) \, , \\
\text{E}.10 =& +\frac{1}{3!} \displaystyle\sum_{\substack{k_1 k_2 k_3 k_4\\ k_5 k_6 k_7}} \mathcal{T}^{\dagger}_{k_1 k_2 k_3 k_4}(\tau,\varphi) \, \Omega^{13}_{k_1 k_5 k_6 k_7} R^{--}_{k_4 k_5}(\varphi) R^{--}_{k_3 k_6}(\varphi) R^{--}_{k_2 k_7}(\varphi) \, , \\
\text{E}.11 =& +\frac{1}{4!} \displaystyle\sum_{\substack{k_1 k_2 k_3 k_4\\k_5 k_6 k_7 k_8}} \mathcal{T}^{\dagger}_{k_1 k_2 k_3 k_4}(\tau,\varphi) \, \Omega^{04}_{k_5 k_6 k_7 k_8} R^{--}_{k_4 k_5}(\varphi) R^{--}_{k_3 k_6}(\varphi)R^{--}_{k_2 k_7}(\varphi) R^{--}_{k_1 k_8}(\varphi) \, , \\
\text{E}.12 =& +\frac{1}{8} \displaystyle\sum_{k_1 k_2 k_3 k_4} \mathcal{T}^{\dagger}_{k_1 k_2}(\tau,\varphi) \mathcal{T}^{\dagger}_{k_3 k_4}(\tau,\varphi) \, \Omega^{40}_{k_1 k_2 k_3 k_4} \, , \\
\text{E}.13 =& +\frac{1}{2} \displaystyle\sum_{\substack{k_1 k_2 k_3 k_4 \\ k_5}} \mathcal{T}^{\dagger}_{k_1 k_2}(\tau,\varphi) \mathcal{T}^{\dagger}_{k_3 k_4}(\tau,\varphi) \, \Omega^{31}_{k_1 k_2 k_3 k_5} R^{--}_{k_4 k_5}(\varphi) \, , \\
\text{E}.14 =& +\frac{1}{4} \displaystyle\sum_{\substack{k_1 k_2 k_3 k_4 \\ k_5 k_6}} \mathcal{T}^{\dagger}_{k_1 k_2}(\tau,\varphi) \mathcal{T}^{\dagger}_{k_3 k_4}(\tau,\varphi) \, \Omega^{22}_{k_1 k_2 k_5 k_6} R^{--}_{k_4 k_5}(\varphi) R^{--}_{k_3 k_6}(\varphi) \, , \\
\text{E}.15 =& +\frac{1}{2} \displaystyle\sum_{\substack{k_1 k_2 k_3 k_4 \\ k_5 k_6}} \mathcal{T}^{\dagger}_{k_1 k_2}(\tau,\varphi) \mathcal{T}^{\dagger}_{k_3 k_4}(\tau,\varphi) \, \Omega^{22}_{k_1 k_3 k_5 k_6} R^{--}_{k_2 k_5}(\varphi) R^{--}_{k_4 k_6}(\varphi) \, , \\
\text{E}.16 =& +\frac{1}{2} \displaystyle\sum_{\substack{k_1 k_2 k_3 k_4\\ k_5 k_6 k_7}} \mathcal{T}^{\dagger}_{k_1 k_2}(\tau,\varphi) \mathcal{T}^{\dagger}_{k_3 k_4}(\tau,\varphi) \, \Omega^{13}_{k_1 k_5 k_6 k_7} R^{--}_{k_4 k_5}(\varphi) R^{--}_{k_3 k_6}(\varphi) R^{--}_{k_2 k_7}(\varphi) \, , \\
\text{E}.17 =& +\frac{1}{8} \displaystyle\sum_{\substack{k_1 k_2 k_3 k_4\\k_5 k_6 k_7 k_8}} \mathcal{T}^{\dagger}_{k_1 k_2}(\tau,\varphi) \mathcal{T}^{\dagger}_{k_3 k_4}(\tau,\varphi) \, \Omega^{04}_{k_5 k_6 k_7 k_8} R^{--}_{k_4 k_5}(\varphi) R^{--}_{k_3 k_6}(\varphi)R^{--}_{k_2 k_7}(\varphi) R^{--}_{k_1 k_8}(\varphi) \\
\text{E}.18 =& + \frac{1}{4} \displaystyle\sum_{k_1 k_2 k_3 k_4} \mathcal{T}^{\dagger}_{k_1 k_2}(\tau,\varphi) \, \Omega^{22}_{k_1 k_2 k_3 k_4}
R^{--}_{k_4 k_3} (\varphi) \, , \\
\text{E}.19 =& +\frac{1}{2} \displaystyle\sum_{\substack{k_1 k_2 k_3 k_4 \\ k_5}} \mathcal{T}^{\dagger}_{k_1 k_2}(\tau,\varphi) \, \Omega^{13}_{k_1 k_3 k_4 k_5}
R^{--}_{k_2 k_5} (\varphi) R^{--}_{k_4 k_3} (\varphi) \, , \\
\text{E}.20 =& +\frac{1}{4} \displaystyle\sum_{\substack{k_1 k_2 k_3 k_4 \\ k_5 k_6}} \mathcal{T}^{\dagger}_{k_1 k_2}(\tau,\varphi) \, \Omega^{04}_{k_3 k_4 k_5 k_6}
R^{--}_{k_1 k_6} (\varphi) R^{--}_{k_2 k_5} (\varphi) R^{--}_{k_4 k_3} (\varphi) \, .\end{aligned}$$
The infinite time limit of these expressions is simply obtained by replacing $\mathcal{T}^{\dagger}_{n}(\tau,\varphi)$ everywhere by $\mathcal{T}^{\dagger}_{n}(\varphi)$.
Off-diagonal BCC amplitude equations {#amplitudeequations}
====================================
Starting from Eq. \[dynamicalkernels\], we derive the equations of motion (Eqs. \[reduceddynamicalkernels\] and \[CCamplitudekernels\]) satisfied by matrix elements of the off-diagonal cluster operators ${\cal T}^{\dagger}_{n}(\tau,\varphi)$. The derivations below are obtained by adapting to $n$-tuply excited grand-potential and norm kernels the steps taken in Secs. \[energykernel\] and \[energykernelMBPTtoCC\] for the non-excited grand potential kernel. Being formally similar, those steps are not detailed here.
Grand potential equation
------------------------
With ${\cal B}_{k_1 k_2 \ldots}={\ensuremath{\mathbbm{1}}}$, Eq. \[dynamicalkernels\] provides $$\Omega(\tau,\varphi) = -\partial_{\tau} N(\tau,\varphi) \, , \label{nonconnectedenergy}$$ which expresses the off-diagonal grand-potential kernel as the (imaginary-)time derivative of the off-diagonal norm kernel.
Single amplitude equation {#singles}
-------------------------
Considering the operator ${\cal B}_{k_1 k_2}=\beta^{\dagger}_{k_1} \beta^{\dagger}_{k_2}$ that creates a two-quasiparticle (i.e. single) excitation, Eq. \[dynamicalkernels\] $$\Omega_{k_1 k_2}(\tau,\varphi) = -\partial_{\tau} N_{k_1 k_2}(\tau,\varphi) \, . \label{nonconnectedsingleamplitude}$$
Let us start with $N_{k_1 k_2}(\tau,\varphi)$. Rewriting $| \Psi (\tau) \rangle$ in terms of ${\cal U}(\tau)$, expanding the latter through perturbation theory and applying off-diagonal Wick’s theorem [@balian69a], one obtains the factorization of the singly-excited norm kernel as $$N_{k_1 k_2}(\tau,\varphi) = n_{k_1 k_2}(\tau,\varphi) \, N(\tau,\varphi) \, , \label{factorizesinglenormkernel}$$ where $$n_{k_1 k_2}(\tau,\varphi) \equiv \frac{\langle \Phi | {\cal U}(\tau) {\cal B}_{k_1 k_2} | \Phi(\varphi) \rangle_{c}}{\langle \Phi | \Phi(\varphi) \rangle} \, \label{factorizesinglenormkernela}$$ contains the complete set of BMBPT connected vacuum-to-vacuum diagrams [*linked*]{} to ${\cal B}_{k_1 k_2}$. In the next step, this complete set of diagrams can be rewritten as
\[singleconnectednormkernel\] $$\begin{aligned}
n_{k_1 k_2}(\tau,\varphi) &=& \frac{\langle \Phi | {\cal T}^{\dagger}_{1}(\tau,\varphi) {\cal B}_{k_1 k_2} | \Phi(\varphi) \rangle_{c}}{\langle \Phi | \Phi(\varphi) \rangle} \label{singleconnectednormkernel2} \\
&=& {\cal T}^{\dagger}_{k_1 k_2}(\tau,\varphi) \label{singleconnectednormkernel1} \\
&=& \frac{\langle \Phi | e^{{\cal T}^{\dagger}(\tau,\varphi)} {\cal B}_{k_1 k_2} | \Phi(\varphi) \rangle_{c}}{\langle \Phi | \Phi(\varphi) \rangle} \label{singleconnectednormkernel3} \, ,\end{aligned}$$
where the rule is that no contraction is to be considered among cluster operators or within a cluster operator when expanding the exponential. Off-diagonal contractions within the operator ${\cal B}_{k_1 k_2 \ldots}$ are zero (Eq. \[propagatorsB3\]). Expression \[singleconnectednormkernel3\] can be equated at no cost to Eq. \[singleconnectednormkernel2\] by virtue of the linked/connected character of the kernel.
Let us now come to $\Omega_{k_1 k_2}(\tau,\varphi)$. Because of the presence of two fixed-time operators ${\cal B}_{k_1 k_2}$ and $\Omega$ in the matrix elements, perturbation theory leads to the typical structure $$\Omega_{k_1 k_2}(\tau,\varphi) = \omega_{k_1 k_2}(\tau,\varphi) \, N(\tau,\varphi) + n_{k_1 k_2}(\tau,\varphi) \, \Omega(\tau,\varphi) \, . \label{factorizesingleenergykernel}$$ In Eq. \[factorizesingleenergykernel\] was introduced the kernel
\[singleconnectedenergykernel\] $$\begin{aligned}
\omega_{k_1 k_2}(\tau,\varphi) &\equiv& \frac{\langle \Phi | {\cal U}(\tau) \Omega {\cal B}_{k_1 k_2} | \Phi(\varphi) \rangle_{c}}{\langle \Phi | \Phi(\varphi) \rangle} \label{singleconnectedenergykernel1} \\
&=& \frac{\langle \Phi | e^{{\cal T}^{\dagger}(\tau,\varphi)} \Omega {\cal B}_{k_1 k_2} | \Phi(\varphi) \rangle_{c}}{\langle \Phi | \Phi(\varphi) \rangle} \label{singleconnectedenergykernel2} \, ,\end{aligned}$$
where operators in the matrix element are all connected together by strings of contractions and where no contraction is to be considered among cluster operators or within a cluster operator. A crucial remark is here in order. In Eq. \[singleconnectedenergykernel2\], the only cluster operator that could contract exclusively with the operators entering ${\cal B}_{k_1 k_2}$ is ${\cal T}^{\dagger}_{1}(\tau,\varphi)$. However, if it were to happen, the product ${\cal T}^{\dagger}_{1}(\tau,\varphi){\cal B}_{k_1 k_2}$ would be disconnected from $\Omega$ and from the other allowed ${\cal T}^{\dagger}_{n}(\tau,\varphi)$, which would contradict the fact that the matrix elements are connected, i.e. such contractions actually contribute to the second term on the right-hand side of Eq. \[factorizesingleenergykernel\]. Consequently, all allowed cluster operators are only partially contracted with ${\cal B}_{k_1 k_2}$ and are thus necessarily contracted with $\Omega$. Eventually, this is the actual meaning carried by the label $c$ in Eq. \[singleconnectedenergykernel2\]. This result allows us to recover the natural termination of the expanded exponential at play in standard BCC theory.
Inserting Eqs. \[factorizesinglenormkernel\] and \[factorizesingleenergykernel\] into Eq. \[nonconnectedsingleamplitude\], utilizing Eq. \[singleconnectednormkernel1\] and combining the result with Eq. \[nonconnectedenergy\] eventually leads to the single amplitude equation under the practical form of Eq. \[reduceddynamicalkernels\], i.e. $$\omega_{k_1 k_2}(\tau,\varphi) = -\partial_{\tau} {\cal T}^{\dagger}_{k_1 k_2}(\tau,\varphi) \, . \label{reducedsingleamplitudeequation}$$
Double amplitude equation {#doubles}
-------------------------
Considering the operator ${\cal B}_{k_1 k_2 k_3 k_4}=\beta^{\dagger}_{k_1} \beta^{\dagger}_{k_2} \beta^{\dagger}_{k_3} \beta^{\dagger}_{k_4}$ that creates a four-quasiparticle (i.e. double) excitation, Eq. \[dynamicalkernels\] provides $$\Omega_{k_1 k_2 k_3 k_4}(\tau,\varphi) = -\partial_{\tau} N_{k_1 k_2 k_3 k_4}(\tau,\varphi) \, . \label{nonconnecteddoubleamplitude}$$ Following the same steps as before, one first obtains $$N_{k_1 k_2 k_3 k_4}(\tau,\varphi) = n_{k_1 k_2 k_3 k_4}(\tau,\varphi) \, N(\tau,\varphi) \, , \label{factorizedoublenormkernel}$$ along with
\[doubleconnectednormkernel\] $$\begin{aligned}
n_{k_1 k_2 k_3 k_4}(\tau,\varphi) &=& \frac{\langle \Phi | \Big[{\cal T}^{\dagger}_{2}(\tau,\varphi) + \frac{1}{2} {\cal T}^{\dagger\, 2}_{1}(\tau,\varphi)\Big] {\cal B}_{k_1 k_2 k_3 k_4} | \Phi(\varphi) \rangle_{c}}{\langle \Phi | \Phi(\varphi) \rangle} \label{doubleconnectednormkernel1} \\
&=& {\cal T}^{\dagger}_{k_1 k_2 k_3 k_4}(\tau,\varphi) + {\cal T}^{\dagger}_{k_1 k_2}(\tau,\varphi) \, {\cal T}^{\dagger}_{k_3 k_4}(\tau,\varphi) - {\cal T}^{\dagger}_{k_1 k_3}(\tau,\varphi) \, {\cal T}^{\dagger}_{k_2 k_4}(\tau,\varphi) + {\cal T}^{\dagger}_{k_1 k_4}(\tau,\varphi) \, {\cal T}^{\dagger}_{k_2 k_3}(\tau,\varphi) \label{doubleconnectednormkernel2} \\
&=& \frac{\langle \Phi | e^{{\cal T}^{\dagger}(\tau,\varphi)} {\cal B}_{k_1 k_2 k_3 k_4} | \Phi(\varphi) \rangle_{c}}{\langle \Phi | \Phi(\varphi) \rangle} \label{doubleconnectednormkernel3} \, ,\end{aligned}$$
where the same rules and explanations as before apply.
Coming to $\Omega_{k_1 k_2 k_3 k_4}(\tau,\varphi)$, perturbation theory leads once again to the typical structure $$\begin{aligned}
\Omega_{k_1 k_2 k_3 k_4}(\tau,\varphi) &=& \frac{\langle \Phi | {\cal U}(\tau) \Omega {\cal B}_{k_1 k_2 k_3 k_4} | \Phi(\varphi) \rangle_{c}}{\langle \Phi | \Phi(\varphi) \rangle} \, N(\tau,\varphi) \nonumber \\
&& + n_{k_1 k_2 k_3 k_4}(\tau,\varphi) \, \Omega(\tau,\varphi) \, , \label{factorizedoubleenergykernel}\end{aligned}$$ where operators in the first kernel on the right-hand side are all connected together by strings of contractions. Following the same steps as before, one proceeds to the identification of the clusters, which leads to $$\begin{aligned}
\frac{\langle \Phi | {\cal U}(\tau) \Omega {\cal B}_{k_1 k_2 k_3 k_4} | \Phi(\varphi) \rangle_{c}}{\langle \Phi | \Phi(\varphi) \rangle} &=& \frac{\langle \Phi | e^{{\cal T}^{\dagger}(\tau,\varphi)} \Omega {\cal B}_{k_1 k_2 k_3 k_4} | \Phi(\varphi) \rangle_{c}}{\langle \Phi | \Phi(\varphi) \rangle} \, , \nonumber \\
&& \label{doubleconnectedenergykernelintermediate}\end{aligned}$$ where no contraction is to be considered among cluster operators or within a cluster operator.
Again, it is essential to detail the connected structure of this kernel. At this point, it can only be stated that the operators at play on the righthand side of Eq. \[doubleconnectedenergykernelintermediate\] are all connected together through strings of contractions by virtue of the connected character of the associated diagrams. A priori, this leaves the possibility that a cluster operator is solely, and thus entirely, connected to ${\cal B}_{k_1 k_2 k_3 k_4}$, i.e. that it is not connected to $\Omega$. In the present case, it can at most happen for ${\cal T}^{\dagger}_{1}(\tau,\varphi)$ or ${\cal T}^{\dagger}_{2}(\tau,\varphi)$. Contracting fully ${\cal T}^{\dagger}_{2}(\tau,\varphi)$ with ${\cal B}_{k_1 k_2 k_3 k_4}$ leaves no possibility for the latter to further connect to $\Omega$ and contradicts the fact that all the operators are connected together through strings of contractions, i.e. such a contribution is already included in the second term on the right-hand side of Eq. \[factorizedoubleenergykernel\]. As for ${\cal T}^{\dagger}_{1}(\tau,\varphi)$, the situation is more subtle. Let us thus consider contributions to Eq. \[doubleconnectedenergykernelintermediate\] where ${\cal T}^{\dagger}_{1}(\tau,\varphi)$ is fully contracted with ${\cal B}_{k_1 k_2 k_3 k_4}$. This leaves two quasi-particle creation operators originating from ${\cal B}_{k_1 k_2 k_3 k_4}$, i.e. a single-excitation operator ${\cal B}_{j_1 j_2}$ with $j_1, j_2$ to be chosen among the four $k_i$ indices to operate further contractions with $\Omega$. For each term with $p\geq 1$ powers of ${\cal T}^{\dagger}_{1}(\tau,\varphi)$ in the exponential, i.e. terms proportional to ${\cal T}^{\dagger \, p}_{1}(\tau,\varphi)/p!$, there are $p$ possibilities to fully contract a ${\cal T}^{\dagger}_{1}(\tau,\varphi)$ operator with ${\cal B}_{k_1 k_2 k_3 k_4}$, which leaves ${\cal T}^{\dagger \, p\!-\!1}_{1}(\tau,\varphi)/(p\!-\!1)!$ for further contractions. Summing over all terms stemming from the exponential, one can eventually re-factorize each time the full contribution of ${\cal T}^{\dagger}_{1}(\tau,\varphi)$ to the exponential. Performing the algebraic manipulations in details, one eventually arrives at $$\begin{aligned}
\frac{\langle \Phi | e^{{\cal T}^{\dagger}(\tau,\varphi)} \Omega {\cal B}_{k_1 k_2 k_3 k_4} | \Phi(\varphi) \rangle_{c}}{\langle \Phi | \Phi(\varphi) \rangle} &=& \omega_{k_1 k_2 k_3 k_4}(\tau,\varphi) \label{doubleconnectedenergykernel} \\
&& + \omega_{k_1 k_2}(\tau,\varphi) \, {\cal T}^{\dagger}_{k_3 k_4}(\tau,\varphi) \nonumber \\
&& - \omega_{k_1 k_3}(\tau,\varphi) \, {\cal T}^{\dagger}_{k_2 k_4}(\tau,\varphi) \nonumber \\
&& + \omega_{k_1 k_4}(\tau,\varphi) \, {\cal T}^{\dagger}_{k_2 k_3}(\tau,\varphi) \nonumber \\
&& + \omega_{k_2 k_3}(\tau,\varphi) \, {\cal T}^{\dagger}_{k_1 k_4}(\tau,\varphi) \nonumber \\
&& - \omega_{k_2 k_4}(\tau,\varphi) \, {\cal T}^{\dagger}_{k_1 k_3}(\tau,\varphi) \nonumber \\
&& + \omega_{k_3 k_4}(\tau,\varphi) \, {\cal T}^{\dagger}_{k_1 k_2}(\tau,\varphi) \nonumber \, , \end{aligned}$$ where $\omega_{k_1 k_2 k_3 k_4}(\tau,\varphi)$ denotes the contributions to the matrix elements where all cluster operators are [*necessarily*]{} contracted with $\Omega$, which ultimately leads to the usual termination of the exponential. The last six terms in Eq. \[doubleconnectedenergykernel\] gather all the contributions where a ${\cal T}^{\dagger}_{1}(\tau,\varphi)$ was fully contracted to ${\cal B}_{k_1 k_2 k_3 k_4}$.
To eventually obtain the practical form of the double amplitude equation (Eq. \[reduceddynamicalkernels\]), one needs not only to insert Eqs. \[doubleconnectednormkernel2\] and \[doubleconnectedenergykernel\] into Eq. \[nonconnecteddoubleamplitude\], but one must also invoke the single amplitude equation (Eq. \[reducedsingleamplitudeequation\]) along with the norm equation (Eq. \[nonconnectedenergy\]). In doing so, one finally arrives at the equation of motion for double amplitudes under the desired form $$\omega_{k_1 k_2 k_3 k_4}(\tau,\varphi) = -\partial_{\tau} {\cal T}^{\dagger}_{k_1 k_2 k_3 k_4}(\tau,\varphi) \, . \label{reduceddoubleamplitudeequation}$$
$n$-tuple amplitude equation {#ntuple}
----------------------------
As for single and double amplitude equations, the derivation of the $n$-tuple amplitude equation invokes all the amplitude equations of lower rank. Reasoning by recurrence, one can prove Eq. \[reduceddynamicalkernels\] for any $n$-tuply excited amplitude.
BCCSD contributions to $\omega_{k_1k_2}(\tau, \varphi)$ {#CCsinglecontrib}
=======================================================
The fifty-six off-diagonal BCCSD diagrams contributing to $\omega_{k_1k_2}(\tau, \varphi)$ are given by
$$\begin{aligned}
\text{S}.1&=& + \Omega^{02}_{\alpha \beta} \, , \\
\text{S}.2&=& + \frac{1}{2} \sum_{k_1 k_2} \Omega^{04}_{\alpha \beta k_1 k_2} R^{--}_{k_2 k_1}(\varphi) \, , \\
\text{S}.3&=& + \frac{1}{2} \displaystyle\sum_{k_1 k_2} \mathcal{T}^{\dagger}_{\alpha \beta k_1 k_2}(\tau,\varphi) \Omega^{20}_{k_1 k_2} \, , \\
\text{S}.4 &=& + \frac{1}{2} P(\alpha / \beta) \displaystyle\sum_{k_1 k_2} \mathcal{T}^{\dagger}_{\alpha k_1}(\tau,\varphi)
\mathcal{T}^{\dagger}_{k_2 \beta}(\tau,\varphi) \Omega^{20}_{k_1 k_2} \, , \\
\text{S}.5&=& + P(\alpha / \beta) \displaystyle\sum_{k_1} \mathcal{T}^{\dagger}_{\alpha k_1}(\tau,\varphi) \Omega^{11}_{k_1 \beta} \, , \\
\text{S}.6&=& +\displaystyle\sum_{k_1 k_2 k_3} \mathcal{T}^{\dagger}_{\alpha \beta k_1 k_2}(\tau,\varphi) \Omega^{11}_{k_1 k_3} R^{--}_{k_2 k_3}(\varphi) \, , \\
\text{S}.7&=& - P(\alpha / \beta) \displaystyle\sum_{k_1 k_2 k_3} \mathcal{T}^{\dagger}_{\alpha k_1}(\tau,\varphi)
\mathcal{T}^{\dagger}_{k_3 \beta}(\tau,\varphi) \Omega^{11}_{k_3 k_2} R^{--}_{k_1 k_2}(\varphi) \, , \\
\text{S}.8&=& +P(\alpha / \beta) \displaystyle\sum_{k_1 k_2} \mathcal{T}^{\dagger}_{\alpha k_1}(\tau,\varphi)
\Omega^{02}_{\beta k_2} R^{--}_{k_1 k_2}(\varphi) \, , \\
\text{S}.9&=& +\frac{1}{2} \displaystyle\sum_{k_1 k_2 k_3 k_4} \mathcal{T}^{\dagger}_{\alpha \beta k_1 k_2}(\tau,\varphi) \Omega^{02}_{k_4 k_3} R^{--}_{k_1 k_3}(\varphi) R^{--}_{k_2 k_4}(\varphi) \, , \\
\text{S}.10&=& +\frac{1}{2} P(\alpha / \beta) \displaystyle\sum_{k_1 k_2 k_3 k_4} \mathcal{T}^{\dagger}_{\alpha k_1}(\tau,\varphi)
\mathcal{T}^{\dagger}_{k_2 \beta}(\tau,\varphi) \Omega^{02}_{k_4 k_3} R^{--}_{k_1 k_3}(\varphi) R^{--}_{k_2 k_4}(\varphi) \, , \\
\text{S}.11&=& +\frac{1}{4} \displaystyle\sum_{k_1 k_2 k_3 k_4} \mathcal{T}^{\dagger}_{\alpha \beta k_1 k_2}(\tau,\varphi)
\mathcal{T}^{\dagger}_{k_3 k_4}(\tau,\varphi) \Omega^{40}_{k_1 k_2 k_3 k_4} \, , \\
\text{S}.12&=& +\frac{1}{3!} P(\alpha / \beta) \displaystyle\sum_{k_1 k_2 k_3 k_4} \mathcal{T}^{\dagger}_{\alpha k_1 k_2 k_3}(\tau,\varphi)
\mathcal{T}^{\dagger}_{k_4 \beta}(\tau,\varphi) \Omega^{40}_{k_1 k_2 k_3 k_4} \, , \\
\text{S}.13&=& +\frac{1}{4} P(\alpha / \beta) \displaystyle\sum_{k_1 k_2 k_3 k_4} \mathcal{T}^{\dagger}_{\alpha k_1}(\tau,\varphi)
\mathcal{T}^{\dagger}_{k_2 k_3}(\tau,\varphi) \mathcal{T}^{\dagger}_{k_4 \beta}(\tau,\varphi) \Omega^{40}_{k_1 k_2 k_3 k_4} \, , \\
\text{S}.14&=& +\frac{1}{6} P(\alpha / \beta) \displaystyle\sum_{k_1 k_2 k_3} \mathcal{T}^{\dagger}_{\alpha k_1 k_2 k_3}(\tau,\varphi)
\Omega^{31}_{k_1 k_2 k_3 \beta} \, , \\
\text{S}.15&=& +\frac{1}{2} P(\alpha / \beta) \displaystyle\sum_{k_1 k_2 k_3} \mathcal{T}^{\dagger}_{\alpha k_1}(\tau,\varphi)
\mathcal{T}^{\dagger}_{k_2 k_3}(\tau,\varphi) \Omega^{31}_{k_1 k_2 k_3 \beta} \, , \\
\text{S}.16&=& +\frac{1}{2} \displaystyle\sum_{k_1 k_2 k_3 k_4 k_5} \mathcal{T}^{\dagger}_{\alpha \beta k_1 k_2}(\tau,\varphi)
\mathcal{T}^{\dagger}_{k_3 k_4}(\tau,\varphi) \Omega^{31}_{k_1 k_2 k_3 k_5} R^{--}_{k_4 k_5}(\varphi) \, , \\
\text{S}.17&=& +\frac{1}{2} \displaystyle\sum_{k_1 k_2 k_3 k_4 k_5} \mathcal{T}^{\dagger}_{\alpha \beta k_1 k_2}(\tau,\varphi)
\mathcal{T}^{\dagger}_{k_3 k_4}(\tau,\varphi) \Omega^{31}_{k_1 k_3 k_4 k_5} R^{--}_{k_2 k_5}(\varphi) \, , \\
\text{S}.18&=& +\frac{1}{3!} P(\alpha / \beta) \displaystyle\sum_{k_1 k_2 k_3 k_4 k_5} \mathcal{T}^{\dagger}_{\alpha k_1 k_2 k_3}(\tau,\varphi)
\mathcal{T}^{\dagger}_{k_4 \beta}(\tau,\varphi) \Omega^{31}_{k_1 k_2 k_3 k_5} R^{--}_{k_4 k_5}(\varphi) \, , \\
\text{S}.19&=& +\frac{1}{2} P(\alpha / \beta) \displaystyle\sum_{k_1 k_2 k_3 k_4 k_5} \mathcal{T}^{\dagger}_{\alpha k_1 k_2 k_3}(\tau,\varphi)
\mathcal{T}^{\dagger}_{k_4 \beta}(\tau,\varphi) \Omega^{31}_{k_1 k_3 k_4 k_5} R^{--}_{k_2 k_5}(\varphi) \, , \\
\text{S}.20&=& +\frac{1}{2} P(\alpha / \beta) \displaystyle\sum_{k_1 k_2 k_3 k_4 k_5} \mathcal{T}^{\dagger}_{\alpha k_1}(\tau, \varphi)
\mathcal{T}^{\dagger}_{k_2 k_3}(\tau,\varphi) \mathcal{T}^{\dagger}_{k_4 \beta}(\tau,\varphi) \Omega^{31}_{k_1 k_2 k_3 k_5} R^{--}_{k_4 k_5}(\varphi) \, , \\
\text{S}.21&=& +\frac{1}{2} P(\alpha / \beta) \displaystyle\sum_{k_1 k_2 k_3 k_4 k_5} \mathcal{T}^{\dagger}_{\alpha k_1}(\tau, \varphi)
\mathcal{T}^{\dagger}_{k_2 k_3}(\tau,\varphi) \mathcal{T}^{\dagger}_{k_4 \beta}(\tau,\varphi) \Omega^{31}_{k_1 k_3 k_4 k_5} R^{--}_{k_2 k_5}(\varphi) \, , \\
\text{S}.22&=& +\frac{1}{2} \displaystyle\sum_{k_1 k_2} \mathcal{T}^{\dagger}_{k_1 k_2}(\tau,\varphi) \Omega^{22}_{k_1 k_2 \alpha \beta} \, , \\
\text{S}.23&=& +\frac{1}{4} \displaystyle\sum_{k_1 k_2 k_3 k_4} \mathcal{T}^{\dagger}_{\alpha \beta k_1 k_2}(\tau,\varphi)
\Omega^{22}_{k_1 k_2 k_3 k_4} R^{--}_{k_4 k_3} (\varphi) \, , \\
\text{S}.24&=& +\frac{1}{2} P(\alpha / \beta) \displaystyle\sum_{k_1 k_2 k_3 k_4} \mathcal{T}^{\dagger}_{\alpha k_1 k_2 k_3}(\tau,\varphi)
\Omega^{22}_{k_1 k_2 \beta k_4}R^{--}_{k_3 k_4}(\varphi) \, , \\
\text{S}.25&=& +\frac{1}{4} P(\alpha / \beta) \displaystyle\sum_{k_1 k_2 k_3 k_4} \mathcal{T}^{\dagger}_{\alpha k_1}(\tau,\varphi)
\mathcal{T}^{\dagger}_{k_2 \beta}(\tau,\varphi) \Omega^{22}_{k_1 k_2 k_3 k_4} R^{--}_{k_4 k_3}(\varphi) \, , \\
\text{S}.26&=& +P(\alpha / \beta) \displaystyle\sum_{k_1 k_2 k_3 k_4} \mathcal{T}^{\dagger}_{\alpha k_1}(\tau,\varphi)
\mathcal{T}^{\dagger}_{k_2 k_3}(\tau,\varphi) \Omega^{22}_{k_1 k_2 \beta k_4} R^{--}_{k_3 k_4}(\varphi) \, , \\
\text{S}.27&=& +\frac{1}{2} P(\alpha / \beta) \displaystyle\sum_{k_1 k_2 k_3 k_4} \mathcal{T}^{\dagger}_{\alpha k_1}(\tau,\varphi)
\mathcal{T}^{\dagger}_{k_2 k_3}(\tau,\varphi) \Omega^{22}_{k_2 k_3 \beta k_4} R^{--}_{k_1 k_4}(\varphi) \, , \\
\text{S}.28&=& +\frac{1}{4} \displaystyle\sum_{\substack{k_1 k_2 k_3 k_4 \\ k_5 k_6}} \mathcal{T}^{\dagger}_{\alpha \beta k_1 k_2}(\tau,\varphi)
\mathcal{T}^{\dagger}_{k_3 k_4}(\tau,\varphi) \Omega^{22}_{k_1 k_2 k_5 k_6} R^{--}_{k_3 k_6}(\varphi) R^{--}_{k_4 k_5}(\varphi) \, , \\
\text{S}.29&=& +\frac{1}{4} \displaystyle\sum_{\substack{k_1 k_2 k_3 k_4 \\ k_5 k_6}} \mathcal{T}^{\dagger}_{\alpha \beta k_1 k_2}(\tau,\varphi)
\mathcal{T}^{\dagger}_{k_3 k_4}(\tau,\varphi) \Omega^{22}_{k_3 k_4 k_5 k_6} R^{--}_{k_1 k_6}(\varphi) R^{--}_{k_2 k_5}(\varphi) \, , \\
\text{S}.30&=& +\displaystyle\sum_{\substack{k_1 k_2 k_3 k_4 \\ k_5 k_6}} \mathcal{T}^{\dagger}_{\alpha \beta k_1 k_2}(\tau,\varphi)
\mathcal{T}^{\dagger}_{k_3 k_4}(\tau,\varphi) \Omega^{22}_{k_1 k_4 k_5 k_6} R^{--}_{k_2 k_6}(\varphi) R^{--}_{k_3 k_5}(\varphi) \, , \\
\text{S}.31&=& +\frac{1}{2} P(\alpha / \beta) \displaystyle\sum_{\substack{k_1 k_2 k_3 k_4 \\ k_5 k_6}} \mathcal{T}^{\dagger}_{\alpha k_1 k_2 k_3}(\tau,\varphi)
\mathcal{T}^{\dagger}_{k_4 \beta}(\tau,\varphi) \Omega^{22}_{k_1 k_2 k_5 k_6} R^{--}_{k_3 k_6}(\varphi) R^{--}_{k_4 k_5}(\varphi) \, , \\
\text{S}.32&=& +\frac{1}{2} P(\alpha / \beta) \displaystyle\sum_{\substack{k_1 k_2 k_3 k_4 \\ k_5 k_6}} \mathcal{T}^{\dagger}_{\alpha k_1 k_2 k_3}(\tau,\varphi)
\mathcal{T}^{\dagger}_{k_4 \beta}(\tau,\varphi) \Omega^{22}_{k_1 k_4 k_5 k_6} R^{--}_{k_2 k_6}(\varphi) R^{--}_{k_3 k_5}(\varphi) \, , \\
\text{S}.33&=& +P(\alpha / \beta) \displaystyle\sum_{\substack{k_1 k_2 k_3 k_4 \\ k_5 k_6}} \mathcal{T}^{\dagger}_{\alpha k_1}(\tau, \varphi)
\mathcal{T}^{\dagger}_{k_2 k_3}(\tau,\varphi) \mathcal{T}^{\dagger}_{k_4 \beta}(\tau,\varphi) \Omega^{22}_{k_1 k_2 k_5 k_6}
R^{--}_{k_3 k_6}(\varphi) R^{--}_{k_4 k_5}(\varphi) \, , \\
\text{S}.34&=& +\frac{1}{4} P(\alpha / \beta) \displaystyle\sum_{\substack{k_1 k_2 k_3 k_4 \\ k_5 k_6}} \mathcal{T}^{\dagger}_{\alpha k_1}(\tau, \varphi)
\mathcal{T}^{\dagger}_{k_2 k_3}(\tau,\varphi) \mathcal{T}^{\dagger}_{k_4 \beta}(\tau,\varphi) \Omega^{22}_{k_1 k_4 k_5 k_6}
R^{--}_{k_2 k_6}(\varphi) R^{--}_{k_3 k_5}(\varphi) \, , \\
\text{S}.35&=& +\frac{1}{4} P(\alpha / \beta) \displaystyle\sum_{\substack{k_1 k_2 k_3 k_4 \\ k_5 k_6}} \mathcal{T}^{\dagger}_{\alpha k_1}(\tau, \varphi)
\mathcal{T}^{\dagger}_{k_2 k_3}(\tau,\varphi) \mathcal{T}^{\dagger}_{k_4 \beta}(\tau,\varphi) \Omega^{22}_{k_2 k_3 k_5 k_6}
R^{--}_{k_1 k_6}(\varphi) R^{--}_{k_4 k_5}(\varphi) \, , \\
\text{S}.36&=& +\frac{1}{2} P(\alpha / \beta) \displaystyle\sum_{k_1 k_2 k_3} \mathcal{T}^{\dagger}_{\alpha k_1}(\tau,\varphi)
\Omega^{13}_{k_1 \beta k_2 k_3} R^{--}_{k_3 k_2}(\varphi) \, , \\
\text{S}.37&=& +\displaystyle\sum_{k_1 k_2 k_3} \mathcal{T}^{\dagger}_{k_1 k_2}(\tau,\varphi) \Omega^{13}_{k_1 \alpha \beta k_3} R^{--}_{k_2 k_3}(\varphi) \, , \\
\text{S}.38&=& +\frac{1}{2} \displaystyle\sum_{k_1 k_2 k_3 k_4 k_5} \mathcal{T}^{\dagger}_{\alpha \beta k_1 k_2}(\tau,\varphi)
\Omega^{13}_{k_1 k_3 k_4 k_5} R^{--}_{k_2 k_5}(\varphi) R^{--}_{k_4 k_3}(\varphi) \, , \\
\text{S}.39&=& +\frac{1}{2} P(\alpha / \beta) \displaystyle\sum_{k_1 k_2 k_3 k_4 k_5} \mathcal{T}^{\dagger}_{\alpha k_1 k_2 k_3}(\tau,\varphi)
\Omega^{13}_{k_1 \beta k_4 k_5} R^{--}_{k_2 k_5}(\varphi) R^{--}_{k_3 k_4}(\varphi) \, , \\
\text{S}.40&=& +\frac{1}{2} P(\alpha / \beta) \displaystyle\sum_{k_1 k_2 k_3 k_4 k_5} \mathcal{T}^{\dagger}_{\alpha k_1}(\tau,\varphi)
\mathcal{T}^{\dagger}_{k_2 \beta}(\tau,\varphi) \Omega^{13}_{k_1 k_3 k_4 k_5} R^{--}_{k_2 k_5}(\varphi) R^{--}_{k_4 k_3}(\varphi) \, , \\
\text{S}.41&=& +\frac{1}{2} P(\alpha / \beta) \displaystyle\sum_{k_1 k_2 k_3 k_4 k_5} \mathcal{T}^{\dagger}_{\alpha k_1}(\tau,\varphi)
\mathcal{T}^{\dagger}_{k_2 k_3}(\tau,\varphi) \Omega^{13}_{k_1 \beta k_4 k_5} R^{--}_{k_2 k_5}(\varphi) R^{--}_{k_3 k_4}(\varphi) \, , \\
\text{S}.42&=& - P(\alpha / \beta) \displaystyle\sum_{k_1 k_2 k_3 k_4 k_5} \mathcal{T}^{\dagger}_{\alpha k_1}(\tau,\varphi)
\mathcal{T}^{\dagger}_{k_2 k_3}(\tau,\varphi) \Omega^{13}_{k_2 \beta k_4 k_5} R^{--}_{k_1 k_5}(\varphi) R^{--}_{k_3 k_4}(\varphi) \, , \\
\text{S}.43&=& +\frac{1}{2} \displaystyle\sum_{\substack{k_1 k_2 k_3 k_4 \\ k_5 k_6 k_7}} \mathcal{T}^{\dagger}_{\alpha \beta k_1 k_2}(\tau,\varphi)
\mathcal{T}^{\dagger}_{k_3 k_4}(\tau,\varphi) \Omega^{13}_{k_1 k_5 k_6 k_7} R^{--}_{k_2 k_7}(\varphi) R^{--}_{k_3 k_6}(\varphi) R^{--}_{k_4 k_5}(\varphi) \, , \\
\text{S}.44&=& +\frac{1}{2} \displaystyle\sum_{\substack{k_1 k_2 k_3 k_4 \\ k_5 k_6 k_7}} \mathcal{T}^{\dagger}_{\alpha \beta k_1 k_2}(\tau,\varphi)
\mathcal{T}^{\dagger}_{k_3 k_4}(\tau,\varphi) \Omega^{13}_{k_3 k_5 k_6 k_7} R^{--}_{k_1 k_7}(\varphi) R^{--}_{k_2 k_6}(\varphi)R^{--}_{k_4 k_5}(\varphi) \, , \\
\text{S}.45&=& +\frac{1}{2} P(\alpha / \beta) \displaystyle\sum_{\substack{k_1 k_2 k_3 k_4 \\ k_5 k_6 k_7}}
\mathcal{T}^{\dagger}_{\alpha k_1 k_2 k_3}(\tau,\varphi) \mathcal{T}^{\dagger}_{k_4 \beta}(\tau,\varphi)
\Omega^{13}_{k_1 k_5 k_6 k_7} R^{--}_{k_2 k_7}(\varphi) R^{--}_{k_3 k_6}(\varphi) R^{--}_{k_4 k_5}(\varphi) \, , \\
\text{S}.46&=& - \frac{1}{3!} P(\alpha / \beta) \displaystyle\sum_{\substack{k_1 k_2 k_3 k_4 \\ k_5 k_6 k_7}}
\mathcal{T}^{\dagger}_{\alpha k_1 k_2 k_3}(\tau,\varphi) \mathcal{T}^{\dagger}_{k_4 \beta}(\tau,\varphi)
\Omega^{13}_{k_4 k_5 k_6 k_7} R^{--}_{k_2 k_7}(\varphi) R^{--}_{k_3 k_6}(\varphi) R^{--}_{k_4 k_5}(\varphi) \, , \\
\text{S}.47&=& +\frac{1}{2} P(\alpha / \beta) \displaystyle\sum_{\substack{k_1 k_2 k_3 k_4 \\ k_5 k_6 k_7}}
\mathcal{T}^{\dagger}_{\alpha k_1}(\tau, \varphi) \mathcal{T}^{\dagger}_{k_2 k_3}(\tau,\varphi) \mathcal{T}^{\dagger}_{k_4 \beta}(\tau,\varphi)
\Omega^{13}_{k_1 k_5 k_6 k_7} R^{--}_{k_2 k_7}(\varphi) R^{--}_{k_3 k_6}(\varphi) R^{--}_{k_4 k_5}(\varphi) \, , \\
\text{S}.48&=& +\frac{1}{2} P(\alpha / \beta) \displaystyle\sum_{\substack{k_1 k_2 k_3 k_4 \\ k_5 k_6 k_7}}
\mathcal{T}^{\dagger}_{\alpha k_1}(\tau, \varphi) \mathcal{T}^{\dagger}_{k_2 k_3}(\tau,\varphi) \mathcal{T}^{\dagger}_{k_4 \beta}(\tau,\varphi)
\Omega^{13}_{k_3 k_5 k_6 k_7} R^{--}_{k_1 k_7}(\varphi) R^{--}_{k_2 k_6}(\varphi) R^{--}_{k_4 k_5}(\varphi) \, , \\
\text{S}.49&=& +\frac{1}{2} P(\alpha / \beta) \displaystyle\sum_{k_1 k_2 k_3 k_4} \mathcal{T}^{\dagger}_{\alpha k_1}(\tau,\varphi)
\Omega^{04}_{\beta k_2 k_3 k_4} R^{--}_{k_1 k_4}(\varphi) R^{--}_{k_3 k_2}(\varphi) \, , \\
\text{S}.50&=& +\frac{1}{2} \displaystyle\sum_{k_1 k_2 k_3 k_4} \mathcal{T}^{\dagger}_{k_1 k_2}(\tau,\varphi)
\Omega^{04}_{\alpha \beta k_3 k_4} R^{--}_{k_1 k_4}(\varphi) R^{--}_{k_2 k_3}(\varphi) \, , \\
\text{S}.51&=& +\frac{1}{4} \displaystyle\sum_{\substack{k_1 k_2 k_3 k_4 \\ k_5 k_6}} \mathcal{T}^{\dagger}_{\alpha \beta k_1 k_2}(\tau,\varphi)
\Omega^{04}_{k_3 k_4 k_5 k_6} R^{--}_{k_1 k_6}(\varphi) R^{--}_{k_2 k_5}(\varphi) R^{--}_{k_4 k_3}(\varphi) \, , \\
\text{S}.52&=& +\frac{1}{3!} P(\alpha / \beta) \displaystyle\sum_{\substack{k_1 k_2 k_3 k_4 \\ k_5 k_6}}
\mathcal{T}^{\dagger}_{\alpha k_1 k_2 k_3}(\tau,\varphi) \Omega^{04}_{\beta k_4 k_5 k_6}
R^{--}_{k_1 k_6} R^{--}_{k_2 k_5}(\varphi) R^{--}_{k_3 k_4}(\varphi) \, , \\
\text{S}.53&=& +\frac{1}{4} P(\alpha / \beta) \displaystyle\sum_{\substack{k_1 k_2 k_3 k_4 \\ k_5 k_6}} \mathcal{T}^{\dagger}_{\alpha k_1}(\tau,\varphi)
\mathcal{T}^{\dagger}_{k_2 \beta}(\tau,\varphi) \Omega^{04}_{k_3 k_4 k_5 k_6} R^{--}_{k_1 k_6}(\varphi) R^{--}_{k_2 k_5}(\varphi) R^{--}_{k_4 k_3}(\varphi) \, , \\
\text{S}.54&=& +\frac{1}{2} P(\alpha / \beta) \displaystyle\sum_{\substack{k_1 k_2 k_3 k_4 \\ k_5 k_6}} \mathcal{T}^{\dagger}_{\alpha k_1}(\tau,\varphi)
\mathcal{T}^{\dagger}_{k_2 k_3}(\tau,\varphi) \Omega^{04}_{\beta k_4 k_5 k_6} R^{--}_{k_1 k_6}(\varphi) R^{--}_{k_2 k_5}(\varphi) R^{--}_{k_3 k_4}(\varphi) \, , \\
\text{S}.55&=& +\frac{1}{4} \displaystyle\sum_{\substack{k_1 k_2 k_3 k_4 \\ k_5 k_6 k_7 k_8}} \mathcal{T}^{\dagger}_{\alpha \beta k_1 k_2}(\tau,\varphi)
\mathcal{T}^{\dagger}_{k_3 k_4}(\tau,\varphi) \Omega^{04}_{k_5 k_6 k_7 k_8} R^{--}_{k_1 k_8}(\varphi) R^{--}_{k_2 k_7}(\varphi) R^{--}_{k_3 k_6}(\varphi) R^{--}_{k_4 k_5}(\varphi) \, , \\
\text{S}.56&=& +\frac{1}{3!} P(\alpha / \beta) \displaystyle\sum_{\substack{k_1 k_2 k_3 k_4 \\ k_5 k_6 k_7 k_8}}
\mathcal{T}^{\dagger}_{\alpha k_1 k_2 k_3}(\tau,\varphi) \mathcal{T}^{\dagger}_{k_4 \beta}(\tau,\varphi)
\Omega^{04}_{k_5 k_6 k_7 k_8} R^{--}_{k_1 k_8}(\varphi) R^{--}_{k_2 k_7}(\varphi) R^{--}_{k_3 k_6}(\varphi) R^{--}_{k_4 k_5}(\varphi) \, , \\
\text{S}.57&=& +\frac{1}{4} P(\alpha / \beta) \displaystyle\sum_{\substack{k_1 k_2 k_3 k_4 \\ k_5 k_6 k_7 k_8}}
\mathcal{T}^{\dagger}_{\alpha k_1}(\tau, \varphi) \mathcal{T}^{\dagger}_{k_2 k_3}(\tau,\varphi) \mathcal{T}^{\dagger}_{k_4 \beta}(\tau,\varphi)
\Omega^{04}_{k_5 k_6 k_7 k_8} R^{--}_{k_1 k_8}(\varphi) R^{--}_{k_2 k_7}(\varphi) R^{--}_{k_3 k_6}(\varphi) R^{--}_{k_4 k_5}(\varphi) \, .
$$
The infinite time limit of these expressions is simply obtained by replacing $\mathcal{T}^{\dagger}_{n}(\tau,\varphi)$ everywhere by $\mathcal{T}^{\dagger}_{n}(\varphi)$.
Matrix elements of $\tilde{\Omega}(\varphi)$ and $\tilde{A}(\varphi)$ {#transformME}
=====================================================================
The matrix elements of the various normal-ordered contributions to the transformed grand potential $\tilde{\Omega}(\varphi)$ are expressed in terms of those of $\Omega$ through
\[e:me3defas\] $$\begin{aligned}
\tilde{\Omega}^{00}(\varphi) & \equiv \tilde{\Omega}^{00(00)}(\varphi) + \tilde{\Omega}^{00(02)}(\varphi) + \tilde{\Omega}^{00(04)} (\varphi) \\
& = \Omega^{00} + \frac{1}{2} \displaystyle\sum_{k_1 k_2} \Omega^{02}_{k_1 k_2} R^{--}_{k_2 k_1}(\varphi)
+ \frac{1}{8} \displaystyle\sum_{k_1 k_2 k_3 k_4} \Omega^{04}_{k_1 k_2 k_3 k_4} R^{--}_{k_4 k_3}(\varphi) R^{--}_{k_2 k_1}(\varphi) \, , \nonumber \\
\tilde{\Omega}^{11}_{k_1 k_2}(\varphi) & \equiv \tilde{\Omega}^{11(11)}_{k_1 k_2}(\varphi) + \tilde{\Omega}^{11(02)}_{k_1 k_2}(\varphi) +
\tilde{\Omega}^{11(13)}_{k_1 k_2}(\varphi) + \tilde{\Omega}^{11(04)}_{k_1 k_2}(\varphi) \\
& = \Omega^{11}_{k_1 k_2} + \displaystyle\sum_{k_3} \Omega^{02}_{k_2 k_3} R^{--}_{k_1 k_3}(\varphi) + \frac{1}{2} \displaystyle\sum_{k_3 k_4}
\Omega^{13}_{k_1 k_2 k_3 k_4} R^{--}_{k_4 k_3}(\varphi) + \frac{1}{2} \displaystyle\sum_{k_3 k_4 k_5} \Omega^{04}_{k_2 k_3 k_4 k_5}
R^{--}_{k_1 k_5}(\varphi) R^{--}_{k_4 k_3}(\varphi) \, , \nonumber \\
\tilde{\Omega}^{20}_{k_1 k_2}(\varphi) & \equiv \tilde{\Omega}^{20(20)}_{k_1 k_2}(\varphi) + \tilde{\Omega}^{20(11)}_{k_1 k_2}(\varphi) +
\tilde{\Omega}^{20(02)}_{k_1 k_2}(\varphi) + \tilde{\Omega}^{20(22)}_{k_1 k_2}(\varphi) +
\tilde{\Omega}^{20(13)}_{k_1 k_2}(\varphi) + \tilde{\Omega}^{20(04)}_{k_1 k_2}(\varphi) \\
\nonumber
& = \Omega^{20}_{k_1 k_2} + \displaystyle\sum_{k_3} \big[ \Omega^{11}_{k_1 k_3}R^{--}_{k_2 k_3}(\varphi) -
\Omega^{11}_{k_2 k_3}R^{--}_{k_1 k_3}(\varphi) \big] + \displaystyle\sum_{k_3 k_4} \Omega^{02}_{k_4 k_3} R^{--}_{k_1 k_3}(\varphi)R^{--}_{k_2 k_4} (\varphi) \\
\nonumber
& + \frac{1}{2} \displaystyle\sum_{k_3 k_4} \Omega^{22}_{k_1 k_2 k_3 k_4} R^{--}_{k_4 k_3}(\varphi)
+ \frac{1}{2} \displaystyle\sum_{k_3 k_4 k_5} \big[\Omega^{13}_{k_1 k_3 k_4 k_5} R^{--}_{k_2 k_5}(\varphi) R^{--}_{k_4 k_3}(\varphi) -
\Omega^{13}_{k_2 k_3 k_4 k_5}R^{--}_{k_1 k_5}(\varphi) R^{--}_{k_4 k_3}(\varphi) \big] \\
& + \frac{1}{2} \displaystyle\sum_{k_3 k_4 k_5 k_6}
\Omega^{04}_{k_3 k_4 k_5 k_6} R^{--}_{k_1 k_6}(\varphi)R^{--}_{k_2 k_5}(\varphi) R^{--}_{k_4 k_3}(\varphi) \, , \nonumber \\
\tilde{\Omega}^{02}_{k_1 k_2}(\varphi) & \equiv \tilde{\Omega}^{02(02)}_{k_1 k_2}(\varphi) + \tilde{\Omega}^{02(04)}_{k_1 k_2}(\varphi) \\
& = \Omega^{02}_{k_1 k_2} + \frac{1}{2} \displaystyle\sum_{k_3 k_4} \Omega^{04}_{k_1 k_2 k_3 k_4} R^{--}_{k_4 k_3} (\varphi) \, , \nonumber \\
\tilde{\Omega}^{22}_{k_1 k_2 k_3 k_4}(\varphi) &\equiv \tilde{\Omega}^{22(22)}_{k_1 k_2 k_3 k_4}(\varphi) + \tilde{\Omega}^{22(13)}_{k_1 k_2 k_3 k_4}(\varphi) + \tilde{\Omega}^{22(04)}_{k_1 k_2 k_3 k_4}(\varphi) \\
& = \Omega^{22}_{k_1 k_2 k_3 k_4} + \displaystyle\sum_{k_5} \big[ \Omega^{13}_{k_1 k_3 k_4 k_5} R^{--}_{k_2 k_5}(\varphi) - \Omega^{13}_{k_2 k_3 k_4 k_5} R^{--}_{k_1 k_5}(\varphi) \big] + \displaystyle\sum_{k_5 k_6} \Omega^{04}_{k_3 k_4 k_5 k_6} R^{--}_{k_1 k_6}(\varphi)R^{--}_{k_2 k_5}(\varphi) \, , \nonumber \\
\tilde{\Omega}^{31}_{k_1 k_2 k_3 k_4}(\varphi) &\equiv \tilde{\Omega}^{31(31)}_{k_1 k_2 k_3 k_4}(\varphi) + \tilde{\Omega}^{31(22)}_{k_1 k_2 k_3 k_4}(\varphi) +
\tilde{\Omega}^{31(13)}_{k_1 k_2 k_3 k_4}(\varphi) + \tilde{\Omega}^{31(04)}_{k_1 k_2 k_3 k_4}(\varphi)\\
\nonumber
& = \Omega^{31}_{k_1 k_2 k_3 k_4} + \displaystyle\sum_{k_5} \big[ \Omega^{22}_{k_2 k_3 k_4 k_5} R^{--}_{k_1 k_5} (\varphi) + \Omega^{22}_{k_1 k_2 k_4 k_5} R^{--}_{k_3 k_5}(\varphi) - \Omega^{22}_{k_1 k_3 k_4 k_5} R^{--}_{k_2 k_5}(\varphi) \big] \\
\nonumber
& + \displaystyle\sum_{k_5 k_6} \big[ \Omega^{13}_{k_1 k_4 k_5 k_6} R^{--}_{k_2 k_6} (\varphi) R^{--}_{k_3 k_5} (\varphi) + \Omega^{13}_{k_3 k_4 k_5 k_6} R^{--}_{k_1 k_6} (\varphi) R^{--}_{k_2 k_5}(\varphi) - \Omega^{13}_{k_2 k_4 k_5 k_6} R^{--}_{k_1 k_6} (\varphi) R^{--}_{k_3 k_5}(\varphi) \big] \\
& + \displaystyle\sum_{k_5 k_6 k_7} \Omega^{04}_{k_4 k_5 k_6 k_7} R^{--}_{k_1 k_7}(\varphi) R^{--}_{k_2 k_6} (\varphi) R^{--}_{k_3 k_5}(\varphi) \, , \nonumber \\
\tilde{\Omega}^{13}_{k_1 k_2 k_3 k_4}(\varphi) &\equiv \tilde{\Omega}^{13(13)}_{k_1 k_2 k_3 k_4}(\varphi) + \tilde{\Omega}^{13(04)}_{k_1 k_2 k_3 k_4}(\varphi) \\
& = \Omega^{13}_{k_1 k_2 k_3 k_4} + \displaystyle\sum_{k_5} \Omega^{04}_{k_2 k_3 k_4 k_5} R^{--}_{k_1 k_5} (\varphi) \, , \nonumber \\
\tilde{\Omega}^{40}_{k_1 k_2 k_3 k_4}(\varphi) &\equiv \tilde{\Omega}^{40(40)}_{k_1 k_2 k_3 k_4}(\varphi) + \tilde{\Omega}^{40(31)}_{k_1 k_2 k_3 k_4}(\varphi) + \tilde{\Omega}^{40(22)}_{k_1 k_2 k_3 k_4}(\varphi) + \tilde{\Omega}^{40(13)}_{k_1 k_2 k_3 k_4}(\varphi) + \tilde{\Omega}^{40(04)}_{k_1 k_2 k_3 k_4}(\varphi) \\
\nonumber
& = \Omega^{40}_{k_1 k_2 k_3 k_4} +
\displaystyle\sum_{k_5} \big[ \Omega^{31}_{k_1 k_2 k_3 k_5} R^{--}_{k_4 k_5} (\varphi) - \Omega^{31}_{k_2 k_3 k_4 k_5} R^{--}_{k_1 k_5}(\varphi) - \Omega^{31}_{k_1 k_2 k_4 k_5} R^{--}_{k_3 k_5}(\varphi) + \Omega^{31}_{k_1 k_3 k_4 k_5} R^{--}_{k_2 k_5}(\varphi) \big] \\
\nonumber
& + \displaystyle\sum_{k_5 k_6} \big[ \Omega^{22}_{k_1 k_4 k_5 k_6} R^{--}_{k_2 k_6} (\varphi) R^{--}_{k_3 k_5} (\varphi) + \Omega^{22}_{k_4 k_3 k_5 k_6} R^{--}_{k_2 k_6} (\varphi) R^{--}_{k_1 k_5}(\varphi) - \Omega^{22}_{k_4 k_2 k_5 k_6} R^{--}_{k_3 k_6} (\varphi) R^{--}_{k_1 k_5}(\varphi) \\
\nonumber
&+ \Omega^{22}_{k_1 k_2 k_5 k_6} R^{--}_{k_3 k_6} (\varphi) R^{--}_{k_4 k_5} (\varphi) + \Omega^{22}_{k_2 k_3 k_5 k_6} R^{--}_{k_1 k_6} (\varphi) R^{--}_{k_4 k_5}(\varphi) - \Omega^{22}_{k_1 k_3 k_5 k_6} R^{--}_{k_2 k_6} (\varphi) R^{--}_{k_4 k_5}(\varphi) \big] \\
\nonumber
&+ \displaystyle\sum_{k_5 k_6 k_7} \big[ \Omega^{13}_{k_3 k_5 k_6 k_7} R^{--}_{k_1 k_7} (\varphi) R^{--}_{k_2 k_6} (\varphi) R^{--}_{k_4 k_5} (\varphi) - \Omega^{13}_{k_2 k_5 k_6 k_7} R^{--}_{k_1 k_7} (\varphi) R^{--}_{k_3 k_6}(\varphi) R^{--}_{k_4 k_5} (\varphi) \\
\nonumber
& + \Omega^{13}_{k_1 k_5 k_6 k_7} R^{--}_{k_2 k_7} (\varphi) R^{--}_{k_3 k_6} (\varphi) R^{--}_{k_4 k_5} (\varphi) - \Omega^{13}_{k_4 k_5 k_6 k_7} R^{--}_{k_2 k_7} (\varphi) R^{--}_{k_3 k_6}(\varphi) R^{--}_{k_1 k_5} (\varphi) \big] \\
& + \displaystyle\sum_{k_5 k_6 k_7 k_8} \Omega^{04}_{k_5 k_6 k_7 k_8} R^{--}_{k_1 k_8}(\varphi) R^{--}_{k_2 k_7} (\varphi) R^{--}_{k_3 k_6}(\varphi) R^{--}_{k_4 k_5} (\varphi) \, , \nonumber \\
\tilde{\Omega}^{04}_{k_1 k_2 k_3 k_4}(\varphi) &\equiv \tilde{\Omega}^{04(04)}_{k_1 k_2 k_3 k_4}(\varphi) \\
&= \Omega^{04}_{k_1 k_2 k_3 k_4} \, . \nonumber \end{aligned}$$
Similarly for $A$
\[AA\] $$\begin{aligned}
\tilde{A}^{00}(\varphi) & \equiv \tilde{A}^{00(00)}(\varphi) + \tilde{A}^{00(02)}(\varphi) \\
& = A^{00} + \frac{1}{2} \displaystyle\sum_{k_1 k_2} A^{02}_{k_1 k_2} R^{--}_{k_2 k_1}(\varphi) \, , \nonumber \\
\tilde{A}^{11}_{k_1 k_2}(\varphi) & \equiv \tilde{A}^{11(11)}_{k_1 k_2}(\varphi) + \tilde{A}^{11(02)}_{k_1 k_2}(\varphi) \\
& = A^{11}_{k_1 k_2} + \displaystyle\sum_{k_3} A^{02}_{k_2 k_3} R^{--}_{k_1 k_3}(\varphi) \, , \nonumber \\
\tilde{A}^{20}_{k_1 k_2}(\varphi) & \equiv \tilde{A}^{20(20)}_{k_1 k_2}(\varphi) + \tilde{A}^{20(11)}_{k_1 k_2}(\varphi) +
\tilde{A}^{20(02)}_{k_1 k_2}(\varphi) \\
\nonumber
& = A^{20}_{k_1 k_2} + \displaystyle\sum_{k_3} \big[ A^{11}_{k_1 k_3}R^{--}_{k_2 k_3}(\varphi) -
A^{11}_{k_2 k_3}R^{--}_{k_1 k_3}(\varphi) \big] + \displaystyle\sum_{k_3 k_4} A^{02}_{k_4 k_3} R^{--}_{k_1 k_3}(\varphi)R^{--}_{k_2 k_4} (\varphi) \, , \nonumber \\
\tilde{A}^{02}_{k_1 k_2}(\varphi) & \equiv \tilde{A}^{02(02)}_{k_1 k_2}(\varphi) \\
& = A^{02}_{k_1 k_2} \, . \nonumber \end{aligned}$$
Useful identities {#usefulID}
=================
\[integral\] $$\begin{aligned}
\int_{0}^{\tau}d\tau_1 \, e^{a\tau_1} &=& \frac{1}{a}\Big(e^{\tau a}-1\Big) \, , \label{integral1} \\
\int_{0}^{\tau}d\tau_{1}d\tau_{2}\,\theta\left( \tau_{1}-\tau_{2}\right)
e^{a\left( \tau_{1}-\tau_{2}\right) } &=&\int_{0}^{\tau}d\tau_{1} \, e^{a\tau_{1}}\int_{0}^{\tau_{1}}d\tau
_{2} \, e^{-a\tau_{2}} = -\frac{\tau}{a}+\frac{1}{a^{2}}\Big( e^{\tau a}-1\Big) \, , \label{integral2} \\
\int_{0}^{\tau}d\tau_{1}d\tau_{2}\,\theta\left( \tau_{1}-\tau_{2}\right)
e^{a\tau_{1}+b\tau_{2}} &=& \int_{0}^{\tau}d\tau_{1} \, e^{a\tau_{1}}\int_{0}^{\tau_{1}}d\tau
_{2} \, e^{b\tau_{2}} = \frac{1}{b\left( a+b\right) } \Big(
e^{\tau\left( a+b\right) }-1\Big) -\frac{1}{ab}\Big( e^{\tau a}-1\Big) \, . \label{integral3} \end{aligned}$$
Given that such integrals only appear in the theory with $a<0$ and $a+b<0$, one obtains
\[limitintegral\] $$\begin{aligned}
\lim\limits_{\tau \to \infty} \int_{0}^{\tau}d\tau \, e^{a\tau} &=& -\frac{1}{a} \, , \label{limitintegral1} \\
\lim\limits_{\tau \to \infty} \int_{0}^{\tau}d\tau_{1}d\tau_{2}\,\theta\left( \tau_{1}-\tau_{2}\right)
e^{a\left( \tau_{1}-\tau_{2}\right) } &=& -\frac{\tau}{a}-\frac{1}{a^{2}} \, , \label{limitintegral2} \\
\lim\limits_{\tau \to \infty} \int_{0}^{\tau}d\tau_{1}d\tau_{2}\,\theta\left( \tau_{1}-\tau_{2}\right)
e^{a\tau_{1}+b\tau_{2}} &=& \frac{1}{a(a+b)} \, , \label{limitintegral3} \end{aligned}$$
where the first and third result are necessarily positive.
[^1]: Although it can be beneficial to indeed break and restore both symmetries, it is possible to limit oneself to breaking and restoring $SU(2)$ symmetry in this case. If breaking $U(1)$ symmetry, one must do it both for neutrons and protons.
[^2]: We will employ a general Rayleigh-Schroedinger scheme that can be eventually reduced to a Moller-Plesset scheme by using the Bogoliubov state solution of Hartree-Fock-Bogoliubov equations [@ring80a] as a reference state.
[^3]: The formalism can be extended to a Hamiltonian containing three- and higher-body forces without running into any fundamental problem. Also, one subtracts the center of mass kinetic energy to the Hamiltonian in actual calculations of finite nuclei. As far as the present work is concerned, this simply leads to a redefinition of one- and two-body matrix elements $t_{pq}$ and $\bar{v}_{pqrs}$ in the Hamiltonian without changing any aspect of the many-body formalism that follows.
[^4]: Explicit expressions of $\Omega^{ij}_{k_1 \ldots k_i k_{i+1} \ldots k_{i+j}}$ in terms of matrix elements $t_{pq}$ and $\bar{v}_{pqrs}$ and of ($U,V$) matrix elements are provided in Ref. [@Signoracci:2014dia].
[^5]: Explicit expressions of $A^{00}$, $A^{11}_{k_1 k_2}$, $A^{20}_{k_1 k_2}$ and $A^{02}_{k_1 k_2}$ are provided in Ref. [@Signoracci:2014dia].
[^6]: As explained in Ref. [@Signoracci:2014dia], downwards quasi-particle lines do not occur in SR-BCC theory. This will be recovered in Section \[CCtheory\] as a particular case of the present diagrammatics.
[^7]: A tensor operator of rank $\text{A}$ with respect to the $U(1)$ group is an operator that associates a state of the $(\text{N}\!+\!\text{A})$-body Hilbert space ${\cal H}_{\text{N}+\text{A}}$ to a state of the $\text{N}$-body Hilbert space ${\cal H}_{\text{N}}$, i.e. that changes the number of particles by $\text{A}$ units.
[^8]: The time is given in units of MeV$^{-1}$.
[^9]: The chemical potential $\lambda$ is fixed such that $\Omega^{\text{A}_0}_0$ for the targeted particle number $\text{A}_0$ is the lowest value of all $\Omega^{\text{A}}_\mu$ over Fock space, i.e. it penalizes systems with larger number of particles such that $\Omega^{\text{A}_0}_0 < \Omega^{\text{A}}_\mu$ for all $\text{A}>\text{A}_0$ while maintaining at the same time that $\Omega^{\text{A}_0}_0 < \Omega^{\text{A}}_\mu$ for all $\text{A}<\text{A}_0$. This is practically achievable only if $E^{\text{A}}_0$ is strictly convex in the neighborhood of $\text{A}_0$, which is generally but not always true for atomic nuclei.
[^10]: We are currently interested in operators that commute with $\Omega$ and that are scalars under transformations of the $U(1)$ group, i.e. that are of rank $\text{A}=0$. Dealing with operators of rank $\text{A}\neq 0$ and with amplitudes between different many-body eigenstates of $\Omega$ requires an extension of the presently developed formalism.
[^11]: The present many-body formalism only requires to consider Bogoliubov states with a given, i.e. odd or even, number parity. As such, Bogoliubov states involved necessarily differ from one another by an [*even*]{} number of quasi-particle excitations.
[^12]: Mixing particle creation and annihilation operators, quasi-particle operators do not correspond to tensor operator of specific rank.
[^13]: This will typically be the case when applying the present approach to even-even nuclei.
[^14]: As a general wording, [*normal*]{} contractions involve one (quasi-)particle creation operator and one (quasi-)particle annihilation operator whereas [*anomalous*]{} contractions involve two (quasi-)particle operators of the same type.
[^15]: A time-dependent [*operator*]{} $O(\tau)$ should not be confused with the gauge-angle dependent [*kernel*]{} $O(\varphi)$ (or its time-dependent partner $O(\tau,\varphi)$). Later on (see Sec. \[transformedoperator\]), the notion of transformed, gauge-angle dependent, [*operator*]{} $\tilde{O}(\varphi)$ will be introduced and distinguished by the “tilde”. It should be confused neither with the kernel $O(\varphi)$ nor with the operator $O(\tau)$.
[^16]: A normal line can be interpreted as $G^{-+(0)}(\varphi)$ or $G^{+-(0)}(\varphi)$ depending on the ascendant or descendant reading of the diagram. Similarly, the ordering of quasi-particle and time labels of a propagator depends on the ascendant or descendant reading of the diagram. While both ways are allowed, one must consistently interpret [*all*]{} the lines involved in a given diagram in the [*same*]{} way, i.e. sticking to an ascendant or descendant way of reading the diagram all throughout.
[^17]: This rule is “unusual” only because many-body methods based on diagrammatic techniques invoking anomalous contractions are scarce in the physics literature.
[^18]: In order to actually fit with an existing SR-BCC code, one must proceed to the hermitian conjugation of the quantities and equations that are referred to.
[^19]: As one is solely interested in the $\tau\rightarrow+\infty$ limit, the time argument can be simply ignored in Eq. \[solNormkernelODE\].
[^20]: The order in which the operators are written is important to know whether contractions arising from the string of operators are eventually non zero but is irrelevant as far as dealing with the numbering, which is what concerns us in the present proof.
[^21]: It is easiest to assume $i$ and $j$ are both even, although this is not necessary in principle.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In Hybrid Monte Carlo(HMC) simulations for full QCD, the gauge fields evolve smoothly as a function of Molecular Dynamics (MD) time. Thus we investigate improved methods of estimating the trial solutions to the Dirac propagator as superpositions of the solutions in the recent past. So far our best extrapolation method reduces the number of Conjugate Gradient iterations per unit MD time by about a factor of 4. Further improvements should be forthcoming as we further exploit the information of past trajectories.'
address:
- 'Department of Physics, Boston University, 590 Commonwealth Ave., Boston, MA 02215, USA'
- 'Department of Physics, Brown University, Providence, RI 02912, USA'
author:
- 'R. C. Brower, A.R. Levi and K. Orginos'
title: 'Extrapolation Methods for the Dirac Inverter in Hybrid Monte Carlo[^1]'
---
INTRODUCTION
============
The inclusion of internal fermion loops in the vacuum of QCD is a major challenge. The present state of the art for generating full QCD configurations is the so called Hybrid Monte Carlo algorithm which uses Molecular Dynamic evolution in a “fifth time” coordinate t. The Hamiltonian for this evolution is $$S = {{\textstyle{{1}\over{2}}}} Tr~P^2 + S_g(U) +
\varphi^\dagger [ M^\dagger M ]^{-1}\varphi,$$ where $P_\mu(x)$ are the angular momenta conjugate to the gauge fields $U_\mu(x)$, $S_g(U)$ is the pure gauge action, $M(U)$ is the Dirac matrix and $\varphi$ is the pseudofermion field. In our discussion the precise form of the gauge action is not important. What is relevant is the need to accurately integrate the equations of motion, calculating the force on $U$ due to the pseudofermions at each time step. This requires solving, over and over again, the linear equation, $$A(t)~ \chi(t) ~=~ \varphi,
\label{eq:linear}$$ where $A(t)\equiv M(U)^\dagger M(U)$ and $\chi(t)$ is the solution of the inverted Dirac operator. Technically this is achieved by starting with a trial value $\chi_{trial}$ and iteratively solving Eq. \[eq:linear\] for $\chi(t)$. On the order of 100 MD steps are taken holding the pseudofermion source fix, so that the operator A(t) changes smoothly as a function of the MD time t, as new values of U are generated. These iterations, usually done by the conjugate gradient (CG) method, are the most computational expensive part of Hybrid Monte Carlo algorithms.
This raises an obvious question. As we move in MD time, why are we not able to “learn” from the recent past enough about the space of likely solutions to vastly improve our iterative scheme?
One should be able to give a very good estimate, $\chi_{trial}$ before starting the conjugate gradient routine. A crucial ingredient in this approach, is the fact that detail balance is in principle preserved independent of the starting trial configuration so long as one converges accurately to the solution. Therefore, we propose to estimate carefully the starting configuration[@blk]. To accomplish this some information on the configurations in the previous MD steps have to be stored. Although these algorithms will use more memory, memory is often not a severe constraint in modern super computer simulations.
ANALYTIC EXTRAPOLATIONS
=======================
To motivate our extrapolation methods, consider the function $\chi(t)$ as an analytic function of t. For simplicity of notation suppose we want the value at t = 0, given past values at $t_1, t_2, t_3, \cdots$. In practice this is usually a regular series of values $t_i = - i \; \epsilon$ with a integrations step of size $dt = \epsilon$. Then a trial value for the the solution, $\chi(0)$ of the new Dirac matrix $A(0)$, might be considered as a linear superposition of old solutions, $$\chi_{trial} = c_1 \; \chi(t_1) + c_2 \; \chi(t_2)
+ \cdots + c_N \; \chi(t_N).$$ If $N \epsilon$ is sufficiently small, we may Taylor expand each term around t = 0 and determine the coefficients by canceling all terms for $\epsilon^k$ to $O(\epsilon^N)$. $$\sum_{i=1}^{N} (t_i)^{n-1} \; c_i = \delta_{1,n}
\label{equ:taylor}$$ As we will demonstrate this procedure is equivalent to the familiar N-1 order polynomial fit to N points.
Polynomial Extrapolation
------------------------
Good result can be obtained even with a si mple polynomial fit of degree $N-1$. To estimate the configuration it is necessary to store in memory the previous $(N+1)$ configurations. However the polynomial extrapolation does not require significant computational effort, it is just a local sum on each lattice point with fixed coefficient, that is less than a single CG step. The $\chi_{trial}$ is expressed as a polynomial, $\chi_{trial}~=~y_1 + t \; y_2 +\cdots + \;t^{N-1}\; y_N$, whose coefficient satisfy the constraint, $$\sum_{n=1}^{N} (t_i)^{n-1} \; y_n = \chi(t_i) \; .
\label{equ:poly}$$ One can easily prove that Eq. \[equ:taylor\] and Eq. \[equ:poly\] define identical extrapolation ($ y_1 = \sum_i c_i \chi(t_i) =\chi_{trial}$. For equally spaced time steps $t_i = - i \; \epsilon$, the coefficients given by $$c_i~=~ (-1)^{i-1}~ {N! \over i! (N-i)!} \; .$$ Table \[tab:tavola1\] shows the results of simulations for polynomial extrapolation. The number of CG steps needed to reach convergence is plotted in function of the degree of extrapolation $N$, and the MD time step $\epsilon$. The case $N=1$ corresponds to starting with the old solution.
‘?=?
[llllllll]{} N=1 (previous data) & 0.98 & 0.93 & 0.93 & 0.88 & 0.85 & 0.77 & 0.63\
N=2 ($1^{st}$ order extrap.) & 0.92 & 1.07 & 0.86 & 0.74 & 0.72 & 0.62 & 0.27\
N=3 ($2^{nd}$ order extrap.) & 0.91 & 0.82 & 0.74 & 0.61 & 0.59 & 0.47 & 0.21\
N=4 ($3^{th}$ order extrap.) & 0.96 & 0.77 & 0.56 & 0.44 & 0.41 & 0.31 & 0.21\
N=5 ($4^{th}$ order extrap.) & 0.86 & 0.77 & 0.54 & 0.45 & 0.42 & 0.35 & 0.26\
N=6 ($5^{th}$ order extrap.) & 0.70 & 0.48 & 0.39 & 0.38 & 0.40 & 0.40 & 0.41\
N=7 ($6^{th}$ order extrap.) & 0.70 & 0.59 & 0.40 & 0.42 & 0.42 & 0.39 & 0.46\
$\delta t=10^{-3}*$ & 15 & 10 & 9 & 8 & 7 & 5 & 2\
‘?=?
[llllllll]{} N=1 & 0.99 & 0.93 & 0.92 & 0.87 & 0.84 & 0.81 & 0.64\
N=2 & 0.87 & 0.72 & 0.71 & 0.69 & 0.63 & 0.49 & 0.25\
N=3 & 0.85 & 0.64 & 0.57 & 0.52 & 0.49 & 0.32 & 0.06\
N=4 & 0.72 & 0.49 & 0.41 & 0.34 & 0.28 & 0.16 & 0.10\
N=5 & 0.72 & 0.37 & 0.31 & 0.23 & 0.19 & 0.12 & 0.07\
N=6 & 0.51 & 0.28 & 0.26 & 0.21 & 0.16 & 0.13 & 0.06\
N=7 & 0.49 & 0.25 & 0.24 & 0.21 & 0.19 & 0.12 & 0.06\
N=8 & 0.40 & 0.26 & 0.23 & 0.18 & 0.15 & 0.09 & 0.06\
N=9 & 0.38 & 0.23 & 0.19 & 0.16 & 0.13 & 0.09 & 0.06\
N=10 & 0.39 & 0.22 & 0.19 & 0.14 & 0.12 & 0.10 & 0.04\
N=11 & 0.38 & 0.18 & 0.17 & 0.14 & 0.11 & 0.08 & 0.04\
$\delta t=10^{-3}*$ & 15 & 10 & 9 & 8 & 7 & 5 & 2\
Minimum Residual Extrapolation
-------------------------------
An alternate, perhaps more appealing, approach is to consider the past history of solutions, $\chi(t_i)$ as defining an important linear subspace for seeking an optimal trial solution. Since the Conjugate Gradient method is in fact just a minimal residual technique confined to the Krylov subspace spanned by vectors $A^{j-1} \chi_{trial}$, why not start by looking at a “smarter” subspace based on past success for nearby times?
In this spirit, we suggest minimizing the norm of the residual, $$r^\dagger r = \chi^\dagger M^\dagger M \chi -
\varphi^\dagger \chi - \chi^\dagger \varphi
+ b^\dagger b \; ,$$ in the subspace spanned by $\chi_i \equiv \chi(t_i)$, where $r = b - M \chi$ and $b \equiv M^\dagger \varphi$. The minimization condition reduces to $$\sum_{j=1}^{N}~{ \chi_i }^\dagger M^\dagger M \chi_j \; c_j =
{ \chi_i }^{\dagger} \varphi \; .$$ The only problem is that this system can be ill conditioned because the past solutions $ \chi_i $ differ from each other by order $ \epsilon $. Nevertheless if we only want to get the minimum of $r^\dagger r$ in [*span*]{}($\chi_i$) using a Gram Schmitt orthonormalization, we can solve the system avoiding the singularities.
This method requires $(N^2+5N)/2$ dots product and $(N)$ $M\chi$ matrix-vector applications, and the storage of $(N)$ past configuration. It is interesting to note that this method gives coefficients that for the first few order reproduce very close to the polynomial extrapolation. Table 2 show the number of CG steps using this method.
CONCLUSIONS
===========
To compare efficiencies, the CG iterations should be divided by $\epsilon$, so we compare the total number of CG steps needed to evolve the system for a fixed distance in configuration space. Note that if $\epsilon$ is too large, the overall performance is good, but the acceptance will drop drastically. If $\epsilon$ is too small, the extrapolation is excellent, but the system will evolve too slowly in phase space. It is not trivial that we find a window in $\epsilon$ where both the acceptance is good and the extrapolation works well. Moreover this window has parameters close to those used in actual simulations.
Our present approach is clearly not the only one worth considering. In fact we have emphasized the analytic properties of $\chi(t)$ because it suggests ways to understand and further improve our approach. For example the failure at fixed $\epsilon$ to improve the polynomial extrapolation by increasing indefinitely the number of terms is probably a signal of nearby singularities in t. Our success so far is probably due to low eigenvalues of the Dirac operator changing more slowly with time.
Many other ways of using the past to avoid needless repetition can be imagined[@sant]. For example the CG routine itself generates vectors, that may be useful than solutions in the more distant past. A Karman filter can be used top introduce exponential decreasing information from the past without increasing the storage requirement. The subspace of past vectors might be used not just as a way to arrive at an initial guess, but also as a way to precondition the iterative process itself. For now, we are pleased for now that even a few vectors from the past can be combined with simple extrapolation ideas to give a very useful acceleration method.
[9]{}
S. Gottlieb, W. Liu, D. Toussaint, R.L. Renken and R.L. Sugar, Phys Rev. D 35 (1987) 2531.
R. Brower, A.R. Levi and K. Orginos, in preparation.
, George Santayana, The Life of Reason or Phases of Human Progress, New York, 1917.
[^1]: Talk presented by A.R. Levi
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In this paper, we demonstrate the fragility of decentralized load-side frequency algorithms proposed in [@Low_Zhao_control] against stochastic parametric uncertainty in power network model. The stochastic parametric uncertainty is motivated through the presence of renewable energy resources in power system model. We show that relatively small variance value of the parametric uncertainty affecting the system bus voltages cause the decentralized load-side frequency regulation algorithm to become stochastically unstable. The critical variance value of the stochastic bus voltages above which the decentralized control algorithm become mean square unstable is computed using an analytical framework developed in [@Sai_arxiv; @pushpak2015stability]. Furthermore, the critical variance value is shown to decrease with the increase in the cost of the controllable loads and with the increase in penetration of renewable energy resources. Finally, simulation results on IEEE 68 bus system are presented to verify the main findings of the paper.'
author:
- 'Sai Pushpak and Umesh Vaidya[^1]'
bibliography:
- 'references.bib'
title: 'Fragility of Decentralized Load-Side Frequency Control in Stochastic Environment'
---
Introduction
============
In this paper, we show that the decentralized load-side frequency control algorithm proposed in [@Low_Zhao_control] is very fragile to parametric uncertainty in power network model. One of the essential components of the smart grid vision is the active participation of loads for improved operation and performance of network power system at different time scales [@doe_demandresponse]. The technology is maturing to the point where the smart grid vision can be realized for actively controlling the loads to absorb not only the long-term variability or uncertainty from renewable power generation but also short term fluctuations. The application of load-side control for frequency regulation falls into the latter category. With the potential benefits of active load control, there are increased research efforts towards the development of systematic analytical methods and optimization-based tools for distributed load control. Most of the literature on this topic primarily focus on stability properties of control algorithms developed for load-side frequency regulation [@trudnowski2006power; @short2007stabilization; @molina2011decentralized; @andreasson2013distributed; @Namerikawa_distributed; @Low_Zhao_distributed]. In particular, [@Low_Zhao_control] proves the asymptotic stability of primal-dual gradient system leading to the decentralized algorithm for load-side frequency control. However, the important issue related to the performance of these algorithms is not addressed primarily in the presence of parametric uncertainty in power network model.
For the successful implementation of the various developed algorithms, it is important to analyze the performance of these algorithms against both parametric and additive sources of uncertainties. In this paper, we extended the primal-dual gradient system model of power swing equation with controllable loads developed in [@Low_Zhao_control] to incorporate the stochastic parametric uncertainty. There are various sources of parametric uncertainty in power system model. In this paper, we argue that renewable energy resources in the form of wind and solar are the potential source of parametric uncertainty in the network power system, where stochasticity in the availability of renewable energy resources will lead to stochastic bus voltages. We develop stochastic power network model with the additive as well as the multiplicative source of uncertainty. The stochastic notion of mean square stability is used to analyze the stability of network power system using stochastic stability framework for the continuous-time system developed in [@Sai_arxiv; @pushpak2015stability; @sai_acc2016]. The continuous-time stochastic stability framework discussed in this paper is motivated from stochastic stability analysis and control results developed for linear and nonlinear systems in [@diwadkar2015control; @diwadkar2014stabilization; @vaidya2010limitations; @elia2005remote]. The developed framework is used to determine the critical variance, $\sigma_*^2$, of parametric stochastic uncertainty above which, the system is mean square unstable. We show that decentralized load-side frequency regulation algorithm developed in [@Low_Zhao_control] is extremely fragile to stochastic fluctuations in bus voltages. In particular, we show that, with an increase in the cost of the controllable loads, the value of critical variance, $\sigma^2_*$, above which the system is unstable decreases. Furthermore, $\sigma_*^2$ value also decreases with the increase in the penetration of the renewable energy resources in the power network.
The organization of the paper is as follows. In section \[sec\_stochastic\_model\], we provide a framework for incorporating stochastic parametric uncertainty in dynamic power system model employed for solving the load-side frequency regulation problem. Results on analyzing the mean square stability of stochastic power network model are discussed in section \[sec\_stochastic\_stability\]. Simulation results on IEEE 68 bus system are presented in section \[sec\_68bus\] followed by conclusion in section \[sec\_conclusion\].
Load-side frequency control model with stochasticity {#sec_stochastic_model}
====================================================
In this section, we develop an integrated load-side frequency control model with stochastic uncertainty. There are various sources of stochastic uncertainty in a power network and renewables energy sources such as solar and wind energy forms the major contributors to the uncertainty. We first discuss briefly the deterministic load-side frequency control model as developed in [@Low_Zhao_control]. We refer the readers to [@Low_Zhao_control] for more detailed discussion on this model. The basic idea behind the load-side frequency control is to control the loads, so that, the system-wide frequency can be regulated following a small change in power injection at one of the system bus. The load control comes at a cost and is measured by the aggregate disutility of the loads. The objective is to regulate the system frequency while minimizing the aggregate disutility of the loads.
To set up the problem, consider a power network as a graph network with generator and load buses as nodes and transmission lines as edges. Let the sets $\mathcal{G}, \mathcal{L}$, and $\mathcal{E}$ respectively, denote the set of generator nodes, set of load nodes, and the set containing network edges. The cardinalities of the sets $\mathcal{G}, \mathcal{L}$, and $\mathcal{E}$ are denoted by $n_g, n_l$, and $p$ respectively. Let $\mathcal{N}$ denote the set of generator and load nodes and its cardinality is given by $n$, where $n = n_g + n_l$. The dynamic network model for regulating system frequency using load can be written as follows (we refer the readers to [@Low_Zhao_control] for various assumption leading to this model). $$\begin{aligned}
\dot \omega_j&=&-\frac{1}{M_j} (\hat d_j +d_j-P_j^m+P_j^{out}-P_j^{in}),\;\forall j\in {\cal G}\nonumber\\
0&=&\hat d_j +d_j-P_j^m+P_j^{out}-P_j^{in},\;\;\forall j\in {\cal L}\nonumber\\
\dot P_{ij}&=&W_{ij}(\omega_i-\omega_j),\;\;\forall (i,j)\in {\cal E}\label{system}\end{aligned}$$ where $\omega_j$ is the frequency deviation at $j^{th}$ bus, $W_{ij} := 3 \frac{\vert V_i \vert \vert V_j \vert}{X_{ij}} \cos(\theta_i^0 - \theta_j^0),$ where $V_i, \theta_i^0$ are the voltage and nominal phase angle at bus $i$, and $X_{ij}$ is the reactance between the buses $i$ and $j$. Three types of loads are distinguished in the above model namely, frequency-sensitive, frequency-insensitive but controllable, and uncontrollable loads. The quantity, $\hat d_j$ models the frequency-sensitive load and is assumed to be of the form $\hat d_j=D_j \omega_j$, i.e., it responds linearly to frequency deviation. Further, $P_j^m$ incorporates the part of load which is frequency-insensitive and uncontrollable and $d_j$ models the load which is frequency-insensitive but controllable.
The objective is to design a feedback controller $d_j(\omega(t),P(t))$ for the controllable loads, so that, frequency can be regulated following disturbance, i.e., the system (\[system\]) is globally asymptotically stable. In [@Low_Zhao_control], an alternate optimization-based approach is proposed for adjusting the controllable load, $d_j$. The design of feedback controller, $d_j(\omega(t),P(t))$, is posed as an optimal load control (OLC) problem and the feedback controller is derived as a distributed algorithm to solve the OLC. The optimization problem for OLC is formulated as follows. $$\begin{aligned}
\min_{\underline{d}\leq d \leq \overline{d},\hat{d}} \quad & {\sum_{j \in \mathcal{N}}} \left(c_j(d_j) + \frac{1}{2D_j} \hat{d}_j^2\right) \label{lfc_optm_cost}\\
\rm{subject\; to}\quad & \sum_{j \in \mathcal{N}} (d_j + \hat{d_j}) = \sum_{j \in \mathcal{N}} P_j^m \label{lfc_optm_constraints}\end{aligned}$$ where $c_j(d_j)$ is the cost on the controllable load at bus $j$, when it is changed by $d_j$ and $\hat{d}_j := \hat D_j \omega_j$ denotes the frequency deviation, $\omega_j$ of the frequency sensitive load at bus $j$. The change in either generator or load bus $j$ is denoted by $P_j^m$ and the loads always satisfy $-\infty < \underline{d}_j \leq d_j\leq \overline{d}_j < \infty$. Furthermore, the cost function $c_j$ at every bus $j$ is assumed to be strictly convex and twice continuously differentiable on $[\underline{d}_j,\; \overline{d}_j]$. A dual to the OLC problem (\[lfc\_optm\_cost\])-(\[lfc\_optm\_constraints\]) which can be solved or implemented with a distributed architecture is written as $$\begin{aligned}
\max_{\nu} \quad & \textstyle{\sum_{j \in \mathcal{N}}} \Phi_j(\nu_j)\label{dual_cost} \\
\rm{subject\; to}\quad & \nu_i = \nu_j, \quad \forall (i,j) \in \mathcal{E},\label{dual_constraints}\end{aligned}$$ where $\Phi_j(\nu_j) = c_j(d_j(\nu_j)) - \nu_j d_j(\nu_j) - \frac{1}{2} \hat D_j \nu_j^2 + \nu_j P_j^m$, $d_j(\nu_j) = \left[c_j^{'-1}(\nu_j) \right]^{\overline{d}_j}_{\underline{d}_j}$, and $c_j^{'}(\nu_j)$ is the derivative of the cost function. Note that $\Phi_j$ is only a function of $\nu_j$, the dual variable, and all $\nu_j$ are constrainted to be equal at optimality. After some change of variables, it can be shown that the primal-dual gradient system corresponding to optimization problem (\[dual\_cost\])-(\[dual\_constraints\]) takes the form as given below [@Low_Zhao_control]. $$\begin{aligned}
\dot{\omega}_j = & \frac{-1}{M_j} (\hat d_j + d_j - P_j^m + P_j^{out} - P_j^{in}), \forall j \in \mathcal{G}, \label{eq_gen1} \\
0 = & \hat d_j + d_j - P_j^m + P_j^{out} - P_j^{in}, \forall j \in \mathcal{L}, \label{eq_load} \\
\dot{P}_{ij} = & W_{ij} ( \omega_i - \omega_j), \forall (i,j) \in \mathcal{E}, \label{eq_power} \\
\hat d_j = & \hat D_j \omega_j, \forall j \in \mathcal{N}, \label{eq_freq_sensitive_load} \\
d_j = & \left[c_j^{'-1} ( \omega_j) \right]^{\overline{{d}}_j}_{\underline{d}_j}, \; \forall j \in \mathcal{N}. \label{control_law}\end{aligned}$$
Notice that, the primal-dual gradient system is same as the dynamic network model for load frequency control, i.e., Eq. (\[system\]) except for the fact that the decentralized feedback control law for the controllable load is obtained as the solution of this optimization problem. This implies that the dynamic power network model is essentially implementing the primal-dual gradient algorithm, where the feedback control law (\[control\_law\]) needs to be implemented at each controllable load for decentralized load-side frequency regulation of power network. In [@Low_Zhao_control], the authors prove the existence of unique equilibrium point for primal-dual gradient system which is globally asymptotically stable. The objective of this paper is to show that, the asymptotically stable property of this equilibrium point is very fragile to the presence of multiplicative stochastic uncertainty in the power network. In the following section, we motivate the stochastic dynamic network model using uncertain and intermittent nature of wind energy.
Stochastic Renewables and Parametric Uncertainty
------------------------------------------------
In this subsection, we show how the parametric uncertainty enters in the power system dynamics. We motivate the parametric uncertainty in the power system dynamics through the presence of renewables where the intermittent nature of wind and solar energy resources are modeled as stochastic random variables. The network power system with parametric uncertainty can be modeled as a set of differential algebraic equations (DAEs) written as follows: $$\begin{aligned}
\dot x&=&f(x,y,\xi)\label{dynamic}\\
0&=&g(x,y,\xi)\label{algebra}\end{aligned}$$ where, $x$ are the dynamic states corresponding to generator angular velocities, generator excitation voltages, power across the transmission lines, etc., and $y$ are the network as well as generator algebraic states corresponding to the bus voltages, bus angles, currents, etc. and $\xi$ denotes the stochastic parametric uncertainty. In the following discussion we will identify the precise source of parametric uncertainty in both the differential equation and algebraic equation. The network power system model with renewable wind energy resources will consists of conventional synchronous generators as well as doubly fed induction generators (DFIG). We refer the readers to [@pulgar2011wind] for more detailed discussion on the deterministic modeling of network power system with renewable wind generation. A zero-axis model [@pulgar2011wind] for DFIG obtained by neglecting the dynamics of the stator and rotor flux linkages is given by, $$\begin{aligned}
\frac{d\omega_r}{dt} = & \frac{\omega_s}{2H_D} \left(T_{mD} - \bar X_m I_{qs} I_{dr} +\bar X_m I_{ds} I_{qr} \right) \label{dfigd}\label{diff_eq1} \end{aligned}$$ where $\omega_r$ is the electrical rotor speed of DFIG, $I_{qr},I_{qs},I_{dr}, I_{ds}$ are the algebraic states of DFIG, and $T_{mD} = \bar B \omega_b C_p(\lambda, \theta) \frac{v_{wind}^3}{\omega_r}$. The wind speed $v_{wind}$ is intermittent in nature and hence can be modeled as stochastic random variable as follows: $$\begin{aligned}
v_{wind}^3 = & v_{{wind}_0}^3 + \xi \label{winds}\end{aligned}$$ where $v_{{wind}_0}$ is the nominal wind speed and $\xi$ is the stochastic uncertainty. Now substituting (\[winds\]) in the expression of $T_{mD}$ and after substituting $T_{mD}$ in (\[dfigd\]), we notice that the random variable, $\xi$, multiply system state $\frac{1}{\omega_r}$ and hence enter parametrically in the dynamic state equation of power system. The presence of parametric uncertainty in the algebraic equation of the power system can be explained as follows. The algebraic equation corresponding to DFIG affected by the wind speed is written as follows [@pulgar2011wind] $$0 = -V_{qr} +K_{P2}[K_{P1}(P_{ref} - P_{gen})+z_1 - I_{qr}]+z_2,$$ where $V_{qr}, I_{qr}$ are the algebraic states of DFIG and $z_1,z_2$ are the dynamic states of speed controller of DFIG. The power reference input is written as $$P_{ref} = \omega_{ref} T_{mD}.$$ Again substituting $T_{mD} = \bar B \omega_b C_p(\lambda, \theta) \frac{1}{\omega_r} (v_{{wind}_0}^3 + \xi)$, we notice that the stochastic parametric uncertainty enters the DAE equations of DFIG.
In the absence of stochastic uncertainty, i.e., when $\xi = 0$ and under the assumption that $\frac{\partial g}{\partial y}\neq 0$, implicit function theorem can be applied to network algebraic equation (\[algebra\]) to eliminate the algebraic state $y$ by expressing $y=h(x)$. In the presence of stochastic uncertainty, an argument involving center manifold based reduction for stochastic system and singular perturbation theory for stochastic system [@arnold2013random; @berglund2003geometric] the algebraic states, $y$ can be expressed as a stochastic function of states $x$ i.e., $y=h(x,\xi)$. Using this in Eq. , we obtain, $$\dot x = f(x,h(x,\xi),\xi)$$ The above system can be linearized at a nominal operating point to obtain a linear system, where the stochastic uncertainty enters the linearized system parametrically. In the following, we show how the stochastic uncertainty in the algebraic states propagates into the network power system. One of the algebraic states that is of particular interest to us is the bus voltages. It is clear that uncertainty in renewables will cause the voltage to behave randomly. Apart from voltages, there are other network parameters that one can assume to be uncertain and hence modeled as a stochastic random variable. For example, the frequency-sensitive loads can be assumed to be uncertain, i.e., $\hat d_j=(D_o+d\xi D_1)\omega_j$, where $d\xi$ is stochastic process. Note that the uncertainty is assumed to be parametric, where the damping coefficient is changing over time. The loads are constantly turned on and off in the grid and thereby changing the effective damping coefficient of the frequency-sensitive loads. Similarly, the frequency-insensitive uncontrollable loads can also be uncertain. However, in this paper, we mainly focus on the bus voltages being uncertain and analyze the impact of stochastic voltage fluctuations on the system stability. Suppose, $p_1 < n_g$ generators in the power network are now replaced with a renewable energy source. As we are modeling the voltages at renewable buses to be stochastic, the voltages at buses connecting the renewables are also stochastic. Let ${\cal S}$ be the set of a pair of buses whose voltages are stochastic and its cardinality be denoted by $s < p$, where $p$ is the number of total links in the network. Under the assumption that the nominal voltages are $1$ p.u., stochastic voltage fluctuations are modeled as follows: $$\begin{aligned}
|V_i||V_j|= 1+\sigma d\xi_{k} \;\;\forall (i,j)\in{\cal S}\label{stoc_voltage}\end{aligned}$$ where $d\xi_{k}$ is the standard Wiener process and $\sigma$ is the standard deviation assumed to be same for all links. We use unique index $k$ to identify and denote the edge pair $(i,j)\in {\cal S}$ and hence $k=1,\ldots, s$.
Furthermore $d\xi_{k}$ is assumed to be independent of $d\xi_\ell$ for $k\neq \ell$. For the simplicity of presentation, we assume that all the links in the networks have the same variance, $\sigma^2$.
Notice that instead of assuming individual bus voltages to be random, we are assuming product of voltages to be random. This is a modeling assumption and is made to avoid technical difficulty that arises while multiplying two stochastic processes.
We now make following assumption on the cost function $c_j$.
\[cost\_assumption\]We assume that the cost function $c_j$ is quadratic and hence of the form $c_j(d_j)=\frac{d_j^2}{2\alpha_j}$ for all $j\in {\cal N}$. Furthermore, we neglect the saturation constraints on the cost function and hence optimal decentralized control law for the controllable load is of the form $d_j=\alpha_j\omega_j$.
Stochastic stability of power network {#sec_stochastic_stability}
=====================================
The deterministic dynamic network model (\[eq\_gen1\])-(\[control\_law\]) can be combined with the stochastic voltage fluctuation model (\[stoc\_voltage\]) to write a power system model with multiplicative stochastic uncertainty. First, we write the deterministic power network model (\[eq\_gen1\])-(\[control\_law\]) using Assumption \[cost\_assumption\] in compact form after eliminating the algebraic equation (\[eq\_load\]) as follows: $$\begin{aligned}
\dot{\omega}_G = & - M_G^{-1} (D_G \omega_G + E_G P - P_G^m) \\
\dot P = & W (E_G^{\top} \omega_G + E_L^{\top} D_L^{-1}(P_L^m-E_L P))
\end{aligned}
\label{det_model}$$ where $\omega_G \in \mathbb{R}^{n_g}$, $\omega_L \in \mathbb{R}^{n_l}$ and $P \in \mathbb{R}^{p}$. Observe that $M_G, D_G, D_L$ are diagonal matrices. The weight matrix, $W \in \mathbb{R}^{p \times p}$ is defined as a diagonal matrix with entries $W_{ij}$ for all $(i,j) \in \mathcal{E}$. The matrices $E_G$ and $E_L$ are the incidence matrices corresponding to generator and load buses respectively. Next, we incorporate the uncertainty in the above deterministic model. Using Eq. (\[stoc\_voltage\]), we can write the stochastic link weight, $W_{ij}$ as follows $W_{ij}=3\frac{(1+\sigma d\xi_k)}{X_{ij}}\cos (\theta_i^0-\theta_j^0).$ Define $W_{ij}^0:=3\frac{1}{X_{ij}}\cos (\theta_i^0-\theta_j^0)$, and hence, we have $$\begin{aligned}
W_{ij}=W_{ij}^0+\sigma W_{ij}^0 d\xi_k\label{stoc_weight}.\end{aligned}$$ Substituting (\[stoc\_weight\]) in (\[det\_model\]), we obtain following stochastic power network model with some abuse of notation. [$$\begin{aligned}
\dot{\omega}_G &= & - M_G^{-1} (D_G \omega_G + E_G P - P_G^m) \label{stoc_model}\\
\dot P &= & (W^0+\sigma W^0\circ d\xi) (E_G^{\top} \omega_G + E_L^{\top} D_L^{-1}(P_L^m-E_L P)) \nonumber\end{aligned}$$ ]{} where $W^0={\rm diag} (W_{ij}^0)$ for $(i,j)\in {\cal E}$, $d\xi$ is a diagonal matrix with zeros and $d\xi_1,\ldots,d\xi_s$. The nonzero entries of $d\xi$ correspond to the links given in set ${\cal S}$. The symbol, $\circ$ denotes element-wise matrix multiplication. To represent the system (\[stoc\_model\]) in standard robust control form (refer to Fig. \[fig\_cl\_sys\_unc\]), we rewrite the system equation in slightly different form. We first define $u:=[\omega_G^\top\;\ P^\top]^\top$ and $$\begin{aligned}
du = & A u dt + b dt + \textstyle{\sum_{k=1}^{s}} \sigma \bar B_k ( \bar C_k u + \bar G_k) d\xi_k\end{aligned}$$ where,
$$\begin{aligned}
& A = \begin{bmatrix}
-M_G^{-1} D_G & -M_G^{-1} E_G \\ W^0 E_G^{\top} & -W^0E_L^{\top}D_L^{-1} E_L \end{bmatrix},
b = \begin{bmatrix}
M_G^{-1}P_G^m \\
W^0 E_L^{\top} D_L^{-1} P_L^m
\end{bmatrix}, \\
& \bar B_k = \begin{bmatrix}
0 \\
e_k
\end{bmatrix},
\bar C_k = \begin{bmatrix}
(W^0 E_G^{\top})_k & -(W^0 E_L^{\top} D_L^{-1} E_L)_k
\end{bmatrix}, \\
& G_k = \begin{bmatrix}
( W^0 E_L^{\top} D_L^{-1} P_L^m)_k
\end{bmatrix},\end{aligned}$$
where $e_k\in \mathbb{R}^p$ is a vector of all zeros except for $1$ in the $k^{th}$ location. Chose $u^*$, such that, $A u^* + {b} = 0$ and define $ v = u - u^*$ to shift the equilibrium of the deterministic system to origin. Then, we have [$$\begin{aligned}
d v = & A v dt + \sum_{k=1}^{s} \sigma \bar B_k \bar C_k v d\xi_k + \sum_{k=1}^{s} \sigma \bar B_k ( \bar C_k u^* + \bar G_k) d\xi_k \label{eq_full_stoch_sys}\end{aligned}$$ ]{} The matrix $A$ is singular and consists of non-zero null space. In the following, we perform a change of coordinates to separate the dynamics on and off the null space of $A$ matrix.
Let ${\cal N}_s(A)$ and ${\cal R}_s(A)$ denotes the set of vectors which span the null space and range space of $A$. Then, define the transformation matrix, $V = \begin{bmatrix}
{\cal N}_s(A) & {\cal R}_s(A)
\end{bmatrix}$. Using the transformation matrix, $V$, we define $[dx^\top\;\ dy^\top]^\top:=V^\top dv$. It can be shown, after the transformation, various transformed matrices has the following structure.
$$\begin{aligned}
\begin{bmatrix}
0 & A_{yx} \\ 0 & A_{xx}
\end{bmatrix} := & V^\top AV \\
\begin{bmatrix}
0 & (V^{\top} \bar B_k \bar C_k V)_{yx} \\
0 & (V^{\top} \bar B_k \bar C_k V)_{xx}
\end{bmatrix} := & V^{\top} \bar B_k \bar C_k V \\
\begin{bmatrix}
(V^{\top} \bar B_k (\bar{C}_k u^*+\bar{G}_k))_{yx} \\
(V^{\top} \bar B_k (\bar{C}_k u^*+\bar{G}_k))_{xx}
\end{bmatrix}:= & V^{\top} \bar B_k (\bar {C}_k u^*+\bar {G}_k)\end{aligned}$$
where $A_{xx}\in \mathbb{R}^{n\times n}$, $A_{yx}\in\mathbb{R}^{n_l\times n} $. Using the fact that $\bar B_k$ and $\bar C_k$ are column vector and row vector respectively, the matrix $\bar B_k \bar C_k$ is rank deficient. It is easy to show that $(V^{\top} \bar B_k \bar C_k V)_{xx}$ is also rank deficient and hence we write $(V^{\top} \bar B_k \bar C_k V)_{xx}\in \mathbb{R}^{n\times n}$ also as a product of column vector and row vector. In particular, we write, $B_kC_k:=(V^{\top} \bar B_k \bar C_k V)_{xx},$ where $B_k$ and $C_k$ are $n$ dimensional column vector and row vector respectively. Note that, the decomposition of the matrix as a product of two rank one matrices is not unique, but the final stability results are independent of the decomposition. Defining $G_k:=(V^{\top} \bar B_k (\bar{C}_k u^*+\bar{G}_k))_{xx} \in \mathbb{R}^n$ and ${\cal A}:=A_{xx}$ we write Eq. (\[eq\_full\_stoch\_sys\]) in the transformed coordinates as follows $$\begin{aligned}
dy = & A_{yx} x dt + \textstyle{\sum_{k=1}^{s}} \sigma(V^{\top} \bar B_k \bar C_kV)_{yx} x d\xi_k \nonumber \\
& + \textstyle{\sum_{k=1}^{s}} \sigma(V^{\top} \bar B_k (\bar {C}_ku^*+\bar {G}_k))_{yx} d\xi_k, \label{eq_synchro} \\
dx = & {\cal A} x dt + \textstyle{\sum_{k=1}^{s}} \sigma B_k C_k x d\xi_k + \textstyle{\sum_{k=1}^{s}} \sigma G_k d\xi_k, \label{eq_nonsynchro}\end{aligned}$$ where $x \in \mathbb{R}^n$ and $y \in \mathbb{R}^{{n_g} - n}$. We notice that the $y$ dynamics is completely driven by $x$ dynamics and noise processes whereas, $x$ dynamics is not influenced by $y$ dynamics. Hence, the necessary condition for the stability of the above system of equations (\[eq\_synchro\])-(\[eq\_nonsynchro\]) is that, $x$ dynamics is stable. We define following notion of second moment bounded stability for system (\[eq\_nonsynchro\]).
\[def\_smb\]\[Second Moment Bounded\] System (\[eq\_nonsynchro\]) is said to be second moment bounded if there exists a positive constant $\bar K$, such that $\lim_{t\to \infty} E[x(t)^\top x(t)]\leq \bar K$ for all $x(0)\in \mathbb{R}^n$.
Now, consider the stochastic power network without the additive noise term as follows. $$\begin{aligned}
dx = & {\cal A} x dt + \textstyle{\sum_{i=1}^{s}} \sigma B_k C_k x d\xi_k. \label{eq_nonsynchro_noadd}\end{aligned}$$ Following notion of mean square exponential stability can be introduced for (\[eq\_nonsynchro\_noadd\]).
\[def\_mse\]\[Mean Square Exponentially Stable\] System (\[eq\_nonsynchro\_noadd\]) is mean square exponentially stable, if there exist positive constants $K$ and $\beta$, such that $$E[ x(t)^\top x(t) ]\leq K\exp^{-\beta t}E[{x(0)}^\top x(0)]\;\; \forall\; x(0)\in \mathbb{R}^n.$$
The connection between the stability of systems given in Eqs. and is established in the following Lemma.
The system (\[eq\_nonsynchro\_noadd\]) is mean square exponentially stable if and only if system (\[eq\_nonsynchro\]) is second moment bounded. \[lemma\_equivalence\]
Refer to Appendix for the proof.
Using Lemma \[lemma\_equivalence\], it suffices to analyze the mean square exponential stability of system without additive noise given in Eq. .
In order to apply the results of mean square stability from [@Sai_arxiv; @pushpak2015stability], system should be rewritten in the standard robust control form with a deterministic system in feedback with stochastic uncertainty (refer to Fig. \[fig\_cl\_sys\_unc\]). In doing so, we rewrite the stochastic power network model as feedback interconnection of the deterministic system and stochastic uncertainty.
[l]{}[4.1 cm]{}
[[{height="1.4"}]{}]{}
The deterministic part of the system is given by $$\mathcal{G}: \left\{
\begin{aligned}
\dot{x} = & \mathcal{A} x + \mathcal{B} w \\
z = & \mathcal{C} x
\end{aligned}
\right.
\label{eq_mean_sys}$$ where the control and disturbance variables are respectively $z \in \mathbb{R}^{s}$ and $w \in \mathbb{R}^{s}$. The matrix $\mathcal{B}$ is formed by stacking $B_k$’s in the columns and $\mathcal{C}$ matrix is formed by stacking $C_k$’s in the rows. This deterministic system, $\mathcal{G}$ now interacts with the stochastic uncertainty, $\Delta$ and the interconnected system is denoted by $\mathcal{F}(\mathcal{G},\Delta)$. The stochastic power network model is written as feedback interconnection of deterministic and stochastic uncertainty as follows: $$\mathcal{F}(\mathcal{G},\Delta): \left\{
\begin{aligned}
\dot{x} = & \mathcal{A} x + \mathcal{B} w \\
z = & \mathcal{C} x \\
w = & \Delta z
\end{aligned} \right.
\label{eq_full_sys}$$ where the matrix, $\Delta$, is a diagonal matrix whose entries are $\sigma \rm{diag}( \frac{d\xi_1}{dt},\dots,\frac{d\xi_{s}}{dt})$. The stochastic uncertainty is interacting with the deterministic system through the control and disturbance variables. Clearly, for the stochastic system to be mean square stable, we require that the deterministic system is stable, i.e., ${\cal A}$ is Hurwitz. Note that, in the power system model, the matrix $\cal A$ is Hurwitz. Under the assumption that the system matrix $\cal A$ is Hurwitz, we have following necessary and sufficient condition for mean square stability.
[@Sai_arxiv] \[theorem\_stability\] The feedback interconnection ${\cal F}(\mathcal{G},\Delta)$ is mean square exponentially stable if and only if $\sigma \rho ( \hat{G} )<1,$ where $\rho$ stands for the spectral radius of a matrix and $$\begin{aligned}
{\hat G} := \begin{pmatrix}\parallel \mathcal{G}_{11}\parallel_2^2&\ldots &\parallel \mathcal{G}_{1s}\parallel_2^2\\\vdots&\ddots&\vdots\\\parallel \mathcal{G}_{s1}\parallel_2^2&\ldots&\parallel \mathcal{G}_{ss}\parallel_2^2 \end{pmatrix}\label{G_fornorm}.\end{aligned}$$ The notation, $\parallel \mathcal{G}_{ij}\parallel_2$ is the ${\cal H}_2$ norm of the system $\mathcal{G}$ from disturbance input, $j$ and controlled output, $i$.
Refer to [@Sai_arxiv] for the proof. The result of the above theorem can be used to compute critical value of $\sigma_*$ above which, the system is mean square unstable. In particular, the critical value, $\sigma_*$ is given by, $\sigma_*=\frac{1}{\rho(\hat G)}$.
IEEE 68 bus system {#sec_68bus}
==================
In this section, we consider the IEEE $68$ bus network to analyze the load-side primary frequency control with stochastic load voltages. IEEE $68$ bus New England/New York interconnection test system consists of $16$ generator buses and $52$ load buses. The single-line diagram of the $68$ bus test system is shown in Fig. \[fig\_68bus\]. This system contains induction motor loads, constant power loads, and controllable loads. The relevant data for this system is obtained from the data files of power system toolbox [@Chow_toolbox]. As discussed in Section \[sec\_stochastic\_model\], we include the renewables in the power network, and the uncertain, intermittent nature of renewables are modeled into the power network by considering the voltages to be stochastic. Changing the controllable loads involves a cost measured in the form of aggregate disutility of loads, and it has to be minimized.
In this $68$ bus system, there are $29$ induction motor loads which are sensitive to frequency, $35$ controllable loads and the remaining loads are uncontrollable frequency insensitive loads. Now, we replace few of the classical generators with renewable energy sources such as either solar or wind power. As there are no large moving parts at these renewables, the inertia values at these buses are relatively smaller, when compared to the classical generators [@Gautam_Vittal_wind_impact]. Further, integration of renewables into the power network increases the damping slightly [@Slootweg_wind_impact]. Therefore, we assume a relatively smaller value for inertia at renewables location and relatively bigger value for damping at those places. For the simulation purpose, we consider the renewable energy source at buses $54,55,56,60,63,64$ and $65$ replacing the generator buses. Now, the buses connecting the renewable buses are $6, 10, 19, 25, 32, 36$ and $52$ and hence, $s = 7$.
In the given data, the inertia values at generator buses lie between $1-5$, whereas, at the buses with renewable energy, we have considered it as $0.5$. Similarly, the damping values at generator buses are in the range of $0-5$, and we consider the damping at renewable energy buses to be $6$. The simulations results discussed below are consistent with the range of inertia values between $0.5-1$ and damping values between $5-6$.
Next, we analyze the effect of cost on controllable loads on the load-side primary frequency control with stochastic renewables. If the cost on controllable loads is high, then it is difficult to vary the controllable loads. Using the analytical framework discussed in section \[sec\_stochastic\_stability\], we identify the critical variance that can be tolerated in the voltages while maintaining the mean square exponential stability of power network with a decentralized controller. The critical variance value, $\sigma_*^2$, is observed to be very small in the order of $10^{-3}$ with the maximum variance value of $1.9\times 10^{-3}$ which is obtained when the cost coefficient on the controllable load is equal to $\alpha=0.5$ (Refer to Assumption \[cost\_assumption\]). In Fig. \[fig\_cost\_vs\_variance\], observe that, if the cost of controllable loads is further increased, the critical variance that can be tolerated by the stochastic power network reduces. It is important to notice that, for most of the cost values on controllable loads, the critical variance is very small. In generating Fig. \[fig\_cost\_vs\_variance\], all the damping values at generator and load buses are kept constant.
Observe that, if the variance that can be tolerated by the system is small, then the system is on the verge of stability. This nature of the system can be seen, when we consider the stochastic voltages with a variance, $\sigma^2 > \sigma^2_*$, the frequencies grow out of bounds, and the power network becomes mean square unstable. This phenomenon is seen in Figs. \[fig\_voltage\_vs\_time\] and \[fig\_frequency\_vs\_time\_all\_buses\]. The stochastic voltage variation with respect to time is shown in Fig. \[fig\_voltage\_vs\_time\]. In Fig. \[fig\_voltage\_vs\_time\], for the chosen $\sigma^2 > \sigma^2_*$, it is important to emphasize that although the voltages values lie within the safe operating limits of $0.95$ pu to $1.05$ pu, but the frequencies violate the operating limits as seen in Fig. \[fig\_frequency\_vs\_time\_all\_buses\]. This shows the fragility of the decentralized controller in the presence of renewables, as it is inadequate to regulate the frequency by means of controllable loads.
Consider a step change in the power network. For this step change in power, the decentralized frequency controller is ineffective in controlling the controllable loads to regulate the frequency. In Fig. \[fig\_frequency\_with\_disturbance\], initially the system was in stable operating condition with frequencies within the operating limits. We created a step change in power after $10$ seconds, and then the frequencies oscillate and go out of the operating range and continue to oscillate. This phenomenon is not desirable, as it has the impact to damage the power system equipment. Hence, to counteract the fragility of this decentralized controller, a modified robust distributive controller must be designed to regulate the frequency that can tolerate uncertainty in the renewables.
The higher penetration of renewables in the power network will make the decentralized frequency control algorithm more fragile. In particular, with the increase in the number of renewable energy resources, more bus voltages will become uncertain, and this has an adverse impact on the frequency regulation. Fig. \[fig\_variance\_vs\_num\_renewables\] shows the effect of increasing the penetration of renewables in the power network. We notice, with the increase in penetration (i.e., with an increase in the value of $s$) the critical variance that can be tolerated by the system decreases. Note that, this figure will change based upon which locations in the network are chosen for renewables. However, the trend of decrease in the value of critical variance with the increase in the number of renewables will continue to hold true.
Conclusion {#sec_conclusion}
==========
We showed that the decentralized load-side frequency regulation algorithm is fragile to stochastic parametric uncertainty in a power network. The presence of stochastic uncertainty is motivated through uncertainty in renewables. We show that the decentralized algorithm becomes more fragile with the increase in the cost of the controllable loads and also with the increased degree of penetration of renewable resources. System theoretic-based analysis and synthesis framework developed for the stochastic networked system in [@Sai_arxiv; @sai_acc2016] is used to prove the main results. Our future research efforts will be focused on the design of distributed frequency regulation algorithm robust to stochastic uncertainty in power network using the synthesis framework developed in [@Sai_arxiv].
appendix {#appendix .unnumbered}
========
Here, first we recall from [@Sai_arxiv], the covariance propagation equation for the systems given in and . Let $Q(t)$ and $\bar Q(t)$ be the covariance matrices corresponding to and . Then, they satisfy the following matrix differential equations (MDE’s).
$$\begin{aligned}
\dot{{Q}}(t)=& {Q}(t){\cal A}^\top + {\cal A} {Q}(t) + \sum_{k=1}^p \sigma_k^2 B_k {Q}(t)B_k^\top, \label{cov_mde} \\
\dot {\bar Q}(t)=& \bar Q(t) {\cal A}^\top+{\cal A} \bar Q(t)+\sum_{k=1}^p \sigma_k^2 B_k \bar Q(t) B_k^\top + \sum_{k=1}^p G_kG_k^\top\nonumber \\
& + \sum_{k=1}^p \sigma_k G_k \mu(t)^{\top} B_k^{\top} + \sum_{k=1}^p \sigma_k B_k \mu(t) G_k^{\top}. \label{cov_mde_add} \end{aligned}$$
The following equation shows the mean propagation equation for the system with additive noise given in . $$\begin{aligned}
\dot \mu(t) = & {\cal A} \mu(t)
\label{eq_mean_propagation}\end{aligned}$$ Next, we present the proof of Lemma \[lemma\_equivalence\].
Using the operator $\phi$, that transforms a matrix into a vector as defined in [@Costa_book]\[Chapter 2\], the MDE’s given in Eq. and Eq. are written as linear differential equations as given below. $$\begin{aligned}
\dot{q} = & \mathscr{A} q,
\label{cov_lde} \\
\dot{\bar q} = & \mathscr{A} \bar q + \mathscr{B},
\label{cov_lde_add}\end{aligned}$$ where $q = \phi( Q), \bar q = \phi(\bar Q)$, $\mathscr{B} = \sum_{k=1}^p ((G_k \otimes G_k) + ((\sigma_k G_k \mu^{\top}) \otimes B_k) + (B_k \otimes (\sigma_k \mu^{\top} G_k)))\phi(I) \in \mathbb{R}^{n^2} \text{and}$ $\mathscr{A} = {\cal A} \oplus {\cal A} + \sum_{k=1}^p \sigma_k^2 (B_k \otimes B_k)\in\mathbb{R}^{n^2\times n^2},$ where $I$ is the identity matrix of size $n \times n$ and $\otimes$ denotes the Kronecker product, $\oplus$ is the Kronecker sum.
*Necessity*: The mean square exponential stability of system yields stability of system , that is, $\mathscr{A}$ is Hurwitz. Since $\mathscr{A}$ is Hurwitz, the steady state value of $\bar q$ is given by $\lim_{t \to \infty} \bar q(t) = \lim_{t \to \infty} \phi(\bar Q(t)) = -\mathscr{A}^{-1} (\sum_{k=1}^p G_k \otimes G_k) \phi(I).$ Now, taking the inverse $ \phi$ operator, we obtain, $ \lim_{t \to \infty} E[x(t)x(t)^{\top}] = - \phi^{-1}(\mathscr{A}^{-1} (\sum_{k=1}^p G_k \otimes G_k) \phi(I)),$ which is finite. Further, the necessary condition for $\mathscr{A}$ to be Hurwitz is ${\cal A}$ being Hurwitz. This implies that the mean propagation system of , shown in Eq. has a stable evolution. Therefore, system is second moment bounded.
*Sufficiency:* If system is second moment stable, then $ \lim_{t \to \infty} \bar Q(t) $ is a finite value and the mean system has a stable evolution. Taking the operator, it can be alternately written as, $\lim_{t \to \infty} \phi(\bar Q(t)) =\lim_{t\to \infty} e^{\mathscr{A}t} \phi(\bar Q(0)) - \mathscr{A}^{-1} (1-e^{\mathscr{A}t}) (\sum_{k=1}^p G_k \otimes G_k) \phi(I) + e^{\mathscr{A}t} (\sum_{k=1}^p (\sigma_k G_k \mu^{\top}) \otimes B_k) \phi(I) + e^{\mathscr{A}t} (\sum_{k=1}^p B_k \otimes (\sigma_k \mu^{\top} G_k)) \phi(I).$ The limit on the right-hand side is finite, if and only if $\mathscr A$ is Hurwitz. If $\mathscr{A}$ is Hurwitz, then the system (\[eq\_nonsynchro\_noadd\]) is mean square exponentially stable.
[^1]: Financial support from the National Science Foundation grant CNS-1329915 and ECCS-1150405 is gratefully acknowledged. U. Vaidya is with the Department of Electrical & Computer Engineering, Iowa State University, Ames, IA 50011.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Let $p\not =\ell$ be primes. We study the etale cohomology $H^{*}_{et}(BGL_n({{\mathbb F}}_{p^s});{{\mathbb Z}}/{\ell})$ over the algebraically closed field $\bar {{\mathbb F}}_p$ by using the stratification methods from Molina-Vistoli. To compute this cohomology, we use the Delinge-Lusztig variety.'
address: |
Department of mathematics, Faculty of Science, Ryukyu University, Okinawa, Japan\
Department of Mathematics, Faculty of Education, Ibaraki University, Mito, Ibaraki, Japan
author:
- 'M.Tezuka and N.Yagita'
title: The etale cohomology of the general linear group over a finite field and the Deligne and Lusztig variety
---
Introduction
============
Let $p$ and $\ell$ be primes with $p \not =\ell$. Let $G_n=GL_n({{\mathbb F}}_q)$ the general linear group over a finite field ${{\mathbb F}}_q$ with $q=p^s$. Then Quillen computed the cohomology of this group in the famous paper \[Qu\].
(Quillen \[Qu\]) Let $r$ be the smallest number such that $q^r-1=0\ mod(\ell)$. Then we have an isomorphism $$H^*(BG_n;{{\mathbb Z}}/{\ell})\cong {{\mathbb Z}}/\ell[c_r,...,c_{r[n/r]}]
\otimes \Delta(e_r,...,e_{r[n/r]})\quad (1.1)$$ where $|c_{rj}|=2rj,$ $|e_{rj}|=2rj-1$.
To prove this theorem, Quillen used the topological arguments, for example, the Eilenberg-Moore spectral sequences, and spaces of the kernel of the map $\psi^q-1$ defined by the Adams operation. In this paper, we give an elementary algebraic proof for this theorem, in the sense without using the above topological arguments.
By induction on $n$ and the equivariant cohomology theory (stratified methods) from Molina and Vistoli \[Mo-Vi\], \[Vi\], we can compute the etale cohomology over $k=\bar {{\mathbb F}}_p$, i.e., $H_{et}^*(BG_n;{{\mathbb Z}}/{\ell})\cong (1.1)$. Then the base change theorem implies the Quillen theorem.
The Molina and Vistoli stratified methods also work for the motivic cohomology. Let $H^{*,*'}(-;{{\mathbb Z}}/{\ell})$ be the motivic cohomology over the field $\bar {{\mathbb F}}_p$ and $0\not =\tau\in H^{0,1}(Spec(\bar{{\mathbb F}}_p);{{\mathbb Z}}/{\ell})$.
We have an isomorphism $H^{*,*'}(G_n;{{\mathbb Z}}/{\ell})\cong
{{\mathbb Z}}/{\ell}[\tau]\otimes(1.1)$ with degree $deg(c_{rj})=(2rj,rj)$ and $deg(e_{rj})=(2rj-1,rj)$.
To compute the equivariant cohomology, we consider the $G_n$-variety $$Q'=Spec(k[x_1,...,x_n]/((-1)^{n-1}det(x_i^{q^{j-1}})^{q-1}=1)),$$ and prove $Q'/G_n\cong {{\mathbb A}}^{n-1}$. This implies the equivariant cohomology $$H_{G_n}^*(Q'\times _{\mu_{q^n-1}}{{\mathbb G}}_m;{{\mathbb Z}}/p)\cong \Delta(f),\quad |f|=1.$$ The computation of the above isomorphism is the crucial point to compute $H_{G_n}^*(pt.;{{\mathbb Z}}/\ell)\cong H^*(BG_n;{{\mathbb Z}}/\ell)$.
Let $G$ be a connected reductive algebraic group defined over a finite field ${{{\mathbb F}}}_q$, $q=p^r$, let $F\colon G\rightarrow G$ be the Frobenius and let $G^F$ be the (finite) group of fixed points of $F$ in $G$, e.g., $GL_n^F=G_n$ in our notation. In the paper \[De-Lu\], Deligne and Lusztig studied the representation theory of $G^F$ over fields of characteristic 0. The main idea is to construct such representations in the $\ell$-adic cohomology spaces $ H_c{}^*(\tilde X(\dot w),{{\mathbb Q}}_{\ell})$ of certain algebraic varieties $\tilde X(\dot w)$ over ${{{\mathbb F}}}_q$, on which $G^F$ acts. (see $\S 6$ for the definition of $\tilde X(\dot w)$.)
For the $G=GL_n$ and$w=(1,\cdots,n)$, we see that $Q'\cong \tilde X(\dot w)$. One of our theorems is to show $\tilde X(\dot w)/G_n\cong {{\mathbb A}}^{n-1}$ for the above case by completely different arguments. The authors thank to Masaharu Kaneda and Shuichi Tsukuda for their useful suggestions.
Dickson Invariants
==================
At first, we recall the Dickson algebra. Let us write $G_n=GL_n({{\mathbb F}}_q)$. The Dickson algebra is the invariant ring of a polynomial of $n$ variables under the usual $G_n$-action, namely, $${{\mathbb F}}_q[x_1,..,x_n]^{G_n}={{\mathbb F}}_q[c_{n,0},c_{n,1},...,c_{n,n-1}]$$ where each $c_{n,i}$ is defined by $$\sum c_{n,i}X^{q^i}=\prod_{x\in {{\mathbb F}}_q\{x_1,...,x_n\}}(X+x)=
\prod_{(\lambda_1,...,\lambda_n)\in{{\mathbb F}}_q^{\times n}}(X+\lambda_1 x_1+...\lambda_n x_n)$$ Hence the degree $|c_{n,i}|=q^n-q^i$ letting $|x_i|=1$. Let us write $e_n=c_{n,0}^{1/(q-1)}$, namely, $$e_n=(\prod_{0\not =x\in {{\mathbb F}}_q\{x_1,...,x_n\}}(x))^{1/(q-1)}
=\left|\begin{array}{cccc}
x_1&x_1^q&\ldots&x_1^{q^{n-1}}\\
x_2&x_2^q&\ldots&x_2^{q^{n-1}}\\
\multicolumn{4}{c}{\dotfill}\\
x_n&x_n^q&\ldots&x_n^{q^{n-1}}
\end{array}\right|.$$ Then each $c_{n,i}$ is written as $$c_{n,s}=\left| \begin{array}{ccccc}
x_1 &\ldots& \hat x_1^{q^s}& \ldots & x_1^{q^{n}}\\
x_2 & \ldots&\hat x_2^{q ^s}& \ldots & x_2^{q^{n}}\\
\multicolumn{5}{c}{\dotfill} \\
x_n & \ldots &\hat x_n^{q^s} & \ldots & x_n^{q^{n}}
\end{array} \right|/e(x).$$ Note that the Dickson algebra for $SG_n=SL_n({{\mathbb F}}_q)$ is given as $${{\mathbb F}}_q[x_1,..,x_n]^{SG_n}={{\mathbb F}}_q[e_n,c_{n,1},...,c_{n,n-1}].$$
Let us write $k=\bar {{\mathbb F}}_p$. We consider the algebraic variety $$F=Spec(k[x_1,...,x_n]/(e_n)).$$ We want to study the $G_n$-space structure of $X=X(n)={{\mathbb A}}^n-\{0\}$ and $X(1)=X-F$. For this, we consider the following variety (the Deligne-Lusztig variety for $w=(1,...,n)$, see $\S 6$ for details) $$Q=Spec(k[x_1,...,x_n]/(e_n-1)).$$
[**Example.**]{} When $q=p$ and $n=2$, we see $$Q=\{(x,y)|x^py-xy^p=1\}\subset {{\mathbb A}}^2,$$ $$F=\{(x,y)|x^py-xy^p=0\}=\cup _{i\in {{\mathbb F}}_p\cup\{\infty\}}F_{i}$$ where $F_i=\{(x,ix)|x\in k\}$ and $F_{\infty}=\{(0,x)|x\in k\}$.
The corresponding projective variety $\bar Q$ is written $$\bar Q=Proj(k[x_0,...,x_n]/(e_n=x_0^{1+q+..+q^{n-1}})).$$
Let us write $q(n)=1+q+...+q^{n-1}=(q^n-1)/(q-1)$. Then we have an isomorphism $Q\times_{\mu_{q(n)}}{{\mathbb G}}_m\cong X(1)$ of varieties.
We consider the map $$p:Q\times {{\mathbb G}}_m\to X(1)\quad by\ (x,t)\mapsto tx.$$ We see $$e_n(p(x,t))=e_n(tx_1,...,tx_n)=t^{1+q+...+q^{n-1}}e_n(x_1,...,x_n).$$ It is easily seen that this map is onto. Moreover if $x\in Q$ and $t\in \mu_{q(n)}$, then $p(x,t)=tx\in Q$. In fact $\mu_{q(n)}$ acts on $Q$. Since $p(x,t)=p(tx,1)$, we have the isomorphism in this lemma.
Remark 2.1 .It is immediate that the left $SG_{n}$-action and the right $\mu_{q(n)}$-action on Q is compatitive.i.e$(gx)\mu =g(x\mu) for g\in SG_
{n} and \mu \in \mu_{q(n)}.$
We have $Q({{\mathbb F}}_q)=\emptyset$.
Let $(x_1,...,x_n)$ be a ${{\mathbb F}}_q$-rational ponits. Then $x_i^q=x_i$. Hence we see $$e_n=\left|\begin{array}{cccc}
x_1&x_2&\ldots&x_n\\
x_1^q&x_2^q&\ldots&x_n^q\\
\multicolumn{4}{c}{\dotfill}\\
x_1^{q^{n-1}}&x_2^{q^{n-1}}&\ldots&x_n^{q^{n-1}}
\end{array}\right| =\left|\begin{array}{cccc}
x_1&x_2&\ldots&x_n\\
x_1&x_2&\ldots&x_n\\
\multicolumn{4}{c}{\dotfill}\\
x_1&x_2&\ldots&x_n
\end{array}\right|=0.$$
The group $SG_n$ acts on $Q$ freely.
Assume that there is $0\not =g\in G_n$ such that $$gx=x\qquad for \ x\in Q\subset {{\mathbb A}}^n.$$ Then we can identify that $x$ is an eigen vector for the (linear) action $g$ with the eigen value $1$. Hence we can take $x=(1,0...,0)$ after some change of basis. Of course $e_n(1,0...,0)=0$ so $x\not \in Q$. This is a contradiction.
The group $SG_n$ acts freely on the (smooth) variety $Q$. Hence $Q/SG_n$ exists as a variety and we have $$Q/SG_n=Spec(A^{SG_n})\quad for\ A=k[x_1,...,x_n]/(e_n-1).$$
We have an isomorphism $$A^{SG_n}\cong k[c_{n,1},...,c_{n,n-1}]\quad i.e.,\ \ Q/SG_n\cong {{\mathbb A}}^{n-1}.$$
It is almost immediate $$k[c_{n,1},...,c_{n,n-1}]\subset A^{SG_n}.$$
The coordinate ring $\bar A$ of the Zariski closure $\bar Q$ of $Q$ in ${{\mathbb P}}^{n}$ is given as $$\bar A=k[x_0,...,x_n]/(e_n=x_0^{1+q+..+q^{n-1}}).$$ Of course, the coordinate ring $\bar B$ of the closure of $Spec(k[c_{n,1},...,c_{n,n-1}])$ in $\bar Q$ is given as $$\bar B=k[x_0,c_{n,1},...,c_{n,n-1}].$$
Next we compute the Poincare polynomials of $\bar A$ and $\bar A^{SG_n}$ ; $$PS(\bar A)=(1-t^{1+q+...+q^{n-1}})/(1-t)^{n+1}=
(1+t+...+t^{q+...+q^{n-1}})/(1-t)^n,$$ $$PS(\bar B)=1/(1-t)(1-t^{|c_{n,1}|})...(1-t^{|c_{n,n-1}|})$$ $$=(1+t+...+t^{|c_{n,1}|-1})^{-1}...(1+t+...+t^{|c_{n,n-1}|-1})^{-1}/(1-t)^n.$$ Hence we get $$PS(\bar A)/PS(\bar B)=(1+t+...t^{|c_{n,1}|-1})...(1+t+...t^{|c_{n,n-1}|-1})$$ $$\qquad \qquad \qquad \times
(1+t+...+t^{q+..+q^{n-1}}).$$ Thus we know $$rank(PS(\bar A)/PS(\bar B))=|c_{n,1}|\times ...\times |c_{n,n-1}|
\times (1+q+...+q^{n-1})$$ $$=(q^n-q^1)...(q^n-q^{n-1})((q^n-1)/(q-1))=|SG_n|.$$
On the other hand $c_{n,1},...,c_{n,n-1}$ is regular sequence in $\bar A$. Hence $\bar A$ is $\bar B$-free, that is $$\bar A=\bar B\{x_1,...,x_m\}$$ where $m=|SG_n|$ from the results using the Poincare polynomials above.
Let $\pi:Q\to Q/SG_n$ be the projection. Since $\pi$ is etale, for all $x\in Q$, the local ring $O_x$ is $O_{\pi(x)}$-free, and $rank_{O_{x'}}(O_x)=|SG_n|$. Thus we get the desired result $A^{SG_n}=k[c_{n,1},...,c_{n,n-1}]$.
Similarly, we can prove
Let $A'=k[x_1,...,x_n]/(e_n^{q-1}-1)$ and $Q'=Spec(A')$. Then we have an isomorphism $$(A')^{G_n}\cong k[c_{n,1},...,c_{n,n-1}]\quad i.e.,\ \ Q'/G_n\cong {{\mathbb A}}^{n-1}.$$
In $\S 7$ below, we give a complete different proof of the above theorem.
equivariant cohomology
======================
For a smooth algebraic variety $X$ over $k=\bar {{\mathbb F}}_p$, we consider the $mod$ $\ell$ etale cohomology for $\ell\not =p$. Let us write simply $$H^*(X)=H_{et}^*(X;{{\mathbb Z}}/\ell).$$ Let $\rho: G\to W={{\mathbb A}}^n$ a faithful representation. Let $V_n=W-S$ be an open set of $W$ such that $G$ act freely $V_n$ where $codim_WS> n\ge 2$. Then the classifying space $BG$ of $G$ is defined as $colim_{n\to \infty} (V_n/G)$. Let $X$ be a smooth $G$-variety. Then we can define the equivariant cohomology (\[Vi\], \[Mo-Vi\]) $$H_G^*(X)=lim_nH_{et}^*(V_n\times _GX;{{\mathbb Z}}/\ell).$$ Of course $H_G^*(pt.)=H^*(BG)=H^*_{et}(BG;{{\mathbb Z}}/\ell)$.
One of the most useful facts in equivariant cohomology theories is the following localized exact sequence. Let $i:Y\subset X$ be a regular closed inclusion of $G$-varieties, of $codim_X(Y)=c$ and $j:U=X-Y\subset X$. Then there is a long exact sequence $$\to H_G^{*-2c}(Y)\stackrel{i_*}{\to}
H_G^*(X)\stackrel{j^*}{\to} H_G^*(U)\stackrel{\delta}{\to}H_G^{*-2c+1}(Y)\to...$$
Now we apply the above exact sequence for concrete cases. We consider the case $G=G_n=GL_n({{\mathbb F}}_q)$. Recall $$F=Spec(k[x_1,..., x_n]/(e_n^{q-1}))=\cup _{\lambda=(\lambda_1,...,\lambda_n)\not =0}(F_{\lambda})$$ where $F_{\lambda}=\{(x_1,...,x_n)|\lambda_1x_1+...+\lambda_nx_n=0\}\subset {{\mathbb A}}^n.$
Let $F(1)=F$ and $F(2)$ be the ($codim=1$) set of singular points in $F(1)$, namely, $F(2)=\cup F_{\lambda,\mu}$ with $$F_{\lambda,\mu}=\begin{cases} F_{\lambda}\cap F_{\mu}\quad if \ F_{\lambda}\not =F_{\mu}
\\ \emptyset\quad if \ F_{\lambda}= F_{\mu}.
\end{cases}$$ Similarly, we define $F(i)$ as the variety defined by the set of $codim_{{{\mathbb A}}^n}F(i)=i$. Let us write $X(i)=X-F(i).$ Thus we have a sequence of the algebraic sets $$F(1)\supset F(2)\supset ...\supset F(n)=\{0\}\supset F(n+1)=\emptyset,$$ $$X-F(1)=X(1)\subset X(2)\subset...\subset X(n)={{\mathbb A}}^n-\{0\}\subset X(n+1)={{\mathbb A}}^n.$$ Therefore we have the long exact sequences $$\to H_{G_n}^{*-2}(F(1)-F(2))\stackrel{i_*}{\to} H_{G_n}^*(X(2))\stackrel{j^*}{\to}
H^*_{G_n}(X(1))
\stackrel{\delta}{\to}...,$$ $$...........................................$$ $$\to H_{G_n}^{*-2i}(F(i)-F(i+1))\stackrel{i_*}{\to}
H_{G_n}^*(X(i+1))\stackrel{j^*}{\to} H^*_{G_n}(X(i))
\stackrel{\delta}{\to}...,$$ $$...........................................$$ $$\to H_{G_n}^{*-2n}(F(n)-F(n+1))\stackrel{i_*}{\to}
H_{G_n}^*(X(n+1))\stackrel{j^*}{\to}
H^*_{G_n}(X(n))
\stackrel{\delta}{\to}...$$
We have $H^*_{G_n}(X(1))\cong\Lambda(f)$ with $|f|=1$.
>From the $G_n$ version (but not $SG_n$) of Lemma 2.1, we have $$X(1)\cong Q'\times_{\mu_{q^n-1}}{{\mathbb G}}_m.$$ Hence we can compute the equivariant cohomology from Theorem 2.5,Lemma 2.3 and Remark 2.1 $$H^*_{G_n}(X(1))\cong H^*(X(1)/G_{n})$$ $$\cong H^*(Q'/G_n\times_{\mu_{q^n-1}}{{\mathbb G}}_m)\cong
H^*({{\mathbb A}}^{n-1}\times_{\mu_{q^n-1}}{{\mathbb G}}_m)$$ $$\cong H_{\mu_{q^n-1}}({{\mathbb G}}_m)\cong \Lambda(f)\quad |f|=1.$$
We have an isomorphism $$H_{G_n}^*(F(i)-F(i+1))\cong H^*(BG_{i})\otimes \Lambda(f)$$
Each irreducible component of $F(i)$ is a $codim=i$ subspace, which is also identified an element of the Grassmannian. Hence we can write $$F(i)-F(i+1)\cong \amalg_{\bar g\in G_n/(P_{i,n-i})} g({{\mathbb A}}^{n-i}-F(1)')$$ where $g\in G_{n}$ is a representative element of $\bar g,$ $F(1)'=Spec(k[x_1,...,x_{n-i}]/(e_{n-i}^{q-1})$ and $P_{i,n-i}$ is the parabolic subgroup $$P_{i,n-i}=(G_i\times G_{n-i})\ltimes U_{i,n-i}({{\mathbb F}}_q)
\cong \{\left(\begin{array}{cc} G_i& *\\
0 & G_{n-i}\end{array}\right)
|*\in U_{i,n-i}({{\mathbb F}}_q)\}.$$ Since the stabilizer group of $X(1)'= {{\mathbb A}}^{n-i}-F(1)'$is $P_{i,n-i}$, we note from \[Vi\] that $H_{G_{n}}^{*}(F(i)-F(i+1))\cong H_{P_{i,n-1}}^*(X(1)')\cong H_{G_{i}\times G_{n-i}}^*(X(1)').$
Hence we can compute ( for $* < N$) $$H^*_{G_n}(F(i)-F(i+1))\cong H^*(V_{N}'\times V_{N}''\times _{G_{i}\times G_{n-i}}X(1)').$$ $$\cong H^*((V_N'/G_i)\times V_N''\times_{G_{n-i}}X(1)').$$ $$\cong H_{G_i}^*\otimes H_{G_{n-i}}^*(X(1)').$$ Here $X(1)'$ is the ($n-i$)-dimensional version of $X(1)$, and we identify $V_N\cong V_N'\times V_N''$ where $G_i$ acts freely on $V_N'$ and so on. >From the preceding lemma, we know $H_{G_{n-i}}^*(X(1)')\cong \Lambda(f)$.
Let $r$ be the smallest number such that $q^r-1=0\ mod(\ell)$. Recall that $$|G_n|=(q^n-1)(q^n-q)...(q^n-q^{n-1}).$$ Hence if $n<r$, then $H^*(BG_n)\cong{{\mathbb Z}}/\ell$, and hence $H_{G_n}^*(F(i)-F(i+1))\cong \Lambda(f)$ for $i\le n$.
The cohomology of $BGL_n$ is the same as that of $BGL_n({{\mathbb C}})$, i.e., $$H^*(BGL_n)\cong {{\mathbb Z}}/\ell[c_1,...,c_n].$$ The Frobenius map $F$ acts on this cohomology by $c_i\mapsto q^{i}c_i$. Recall that the Lang map induces a principal $G_{n}$-bundle $$G_n\to GL_n\stackrel{L}{\to}GL_n$$ where $L(g)=g^{-1}F(g)$. Thus we have a map $$H^*(BGL_n)/((q^i-1)c_i)\cong {{\mathbb Z}}/\ell[c_r,...,c_{[n/r]r}]
\to H^*(BG_n).$$
If $r=1$, then we have an isomorphism $$H^*(BG_n)
\cong
{{\mathbb Z}}/\ell[c_1,...,c_{n}]
\otimes
\Delta(e_1,...,e_{n}).$$
We prove by induction on $n$. Assume that $$H^*(BG_i)\cong
{{\mathbb Z}}/\ell[c_1,...,c_{i}]\otimes\Delta(e_1,...,e_{i})\quad for \ i<n.$$
We consider the long exact sequence $$\to H_{G_n}^{*-2i}(F(i)-F(i+1))\stackrel{i_*}{\to} H_{G_n}^*(X(i+1))\stackrel{j^*}{\to} H^*_{G_n}(X(i))
\stackrel{\delta}{\to}...$$ Here we use induction on $i$, and assume that $$H_{G_n}^*(X(i))\cong
H_{G_{i-1}}^*\otimes \Lambda(e_i)$$ $$\cong {{\mathbb Z}}/{\ell}[c_1,...,c_{i-1}]\otimes
\Delta(e_1,...,e_{i-1})\otimes \Lambda(e_i).$$ (Letting $e_1=f$, we have the case $i=1$ from Lemma 3.1.) >From the preceding lemma, we still see $$H_{G_n}^*(F(i)-F(i+1))\cong
H_{G_{i}}^*\otimes \Lambda(f)$$ $$\cong {{\mathbb Z}}/{\ell}[c_1,...,c_i]\otimes \Delta(e_1,...,e_i)\otimes \Lambda(f).$$
In the above long exact sequence,the map $j^*$ is an epimorphism for $*<2i-1$, because $H^{minus}(F(i)-F(i+1))=0$. But $H_{G_n}^*(X(i))$ is multiplicatively generated by the elements of $dim\le 2i-2$ and $e_{i}$. By dimensional reason, we see $$\delta(e_{i})=1\quad or\quad\delta(e_{i})=0.$$ Of course if $\delta(e_{i})=0$, then $\delta=0$ for all $*\ge 0$.
Consider the restriction map $H_{G_n}^*(X(i+1))\to H_{G_i}^*({{\mathbb A}}^i)$ which is induced from $X(i+1)={{\mathbb A}}^n-F(i+1)\supset {{\mathbb A}}^i$. Since $|G_i|=(q^{ir}-1)q|G_{i-1}|$, the $\ell$-Sylow subgroup of $G_i$ is different from that of $G_{i-1}$, (More precisely, $rank_{\ell}G_i>rank_{\ell}G_{i-1}$.) So from the Quillen theorem, the Krull dimension of $H_{G_n}(X(i+1))$ is larger than that of $H_{G_n}^*(X(i))$. This fact implies $i_*(1)=c_i$. ( Let $p:V\to X$ be a $j$-dimensional bundle and $i:X\to V$ a section. Then the Chern class $c_j$ is defined as $i^*i_*(1)$.) Thus we see $\delta(e_{i})=0$.
Therefore we have the short exact sequence $$0\to
H^*_{G_i}
\otimes\Lambda(f)
\stackrel{i_*}{\to}
H^*_{G_n}(X(i+1))
\stackrel{j^*}{\to}
H^*_{G_{i-1}}\otimes\Lambda(e_{i})\to 0,$$ namely, we have an isomorphism $$grH_{G_n}^*(X(i+1))\cong
{{\mathbb Z}}/{\ell}[c_1,...,c_{i-1}]\otimes \Delta(e_1,...,e_{i})$$ $$\otimes({{\mathbb Z}}/\ell[c_i]\{i_*(1)=c_i,i_*(f)\}\oplus {{\mathbb Z}}/\ell\{1\}).$$ Let us write $i_*(f)=e_{i+1}$. Then $H^*_{G_n}(X(i+1))$ is the desired form $$H_{G_n}^*(X(i+1))\cong
{{\mathbb Z}}/{\ell}[c_1,...,c_{i-1}]\otimes \Delta(e_1,...,e_{i})$$ $$\otimes({{\mathbb Z}}/\ell[c_i]\{c_i,e_{i+1}\}\oplus {{\mathbb Z}}/\ell\{1\})$$ $$\cong
{{\mathbb Z}}/{\ell}[c_1,...,c_{i}]
\otimes
\Delta(e_1,...,e_{i})
\otimes \Lambda(e_{i+1}).$$ Thus we can see the desired result $H_{G_n}^*(X(n+1))\cong H^*(BG_n)$.
[**Remark.**]{} In the above proof, to see $i_*(1)=c_i$ we used the Krull dimesion (by Quillen). However there is more natural argument (see Proposion 4.2 in the next section) where the properties of the maximal torus $T(\dot w)$ are used.
We have the isomorphism $$H^*(BG_n)
\cong
{{\mathbb Z}}/\ell[c_{r},...,c_{[n/r]r}]
\otimes
\Delta(e_r,...,e_{[n/r]r}).$$
We prove the theorem by induction on $n$. Assume that $$H^*(BG_i)\cong
{{\mathbb Z}}/\ell[c_r,...,c_{[n/i]r}]\otimes\Delta(e_r,...,e_{[n/r]r})\quad for \ i<n.$$ We also consider the long exact sequence $$\to H_{G_n}^{*-2i}(F(i)-F(i+1))\stackrel{i_*}{\to} H_{G_n}^*(X(i+1))\stackrel{j^*}{\to} H^*_{G_n}(X(i))
\stackrel{\delta}{\to}...$$ Here we use induction on $i$, and assume $ H_{G_n}^*(X(i))\cong
H_{G_{i-1}}^*\otimes \Lambda(e_i).$
>From Lemma 3.2, we still see $$H_{G_n}^*(F(i)-F(i+1))\cong
H_{G_{i}}^*\otimes \Lambda(f).$$ By dimensional reason, we see $ \delta(e_{i})=1\quad or\quad\delta(e_{i})=0.$
Now we consider the case $r\ge 2$ and $mr<i\le mr+r-1.$ This case we still assume $$H_{G_i}^*\cong H_{G_{i-1}}^*\cong H_{G_{mr}}^*
\cong {{\mathbb Z}}/\ell[c_r,...,c_{mr}]\otimes \Delta(e_r,...,e_{mr}).$$ Hence the above exact sequence is written as $$\to
H^*_{G_{mr}}
\otimes\Lambda(f)
\stackrel{i_*}{\to}
H^*_{G_n}(X(i+1))
\stackrel{j^*}{\to}
H^*_{G_{mr}}\otimes\Lambda(e_{i})\to ....$$
The $\ell$- Sylow subgroup of $G_i$ and $G_{i-1}$ are the same, and hence $c_i=0$ in $H^*_{G_i}$. (See also Proposition 4.2 below.) This means $ \delta(e_i)=1$ (Of course $\delta(1)=0$).
Hence we have the isomorphism $$H_{G_n}^*(X(i+1))\cong H^*_{G_{mr}}\{1,i_*(f)\}\cong H_{G_{mr}}^*\{1,e_{i+1}\}
\cong H_{G_i}^*\otimes \Lambda(e_{i+1}).$$ Other parts of the proof are almost the same as in the case $r=1$.
maximal torus and $SL_n$
========================
Let $r$ be the smallest positive integer such that $q^r-1=0\ (mod(\ell))$. Let $w=(1,2,..,r)\in S_r$ and $G_r=GL_r({{\mathbb F}}_q)=GL_r^F$ for the Frobenius map $F:x\mapsto x^q$. For a matrix $A=(a_{i,j})\in GL_n$, the adjoint action is given as $$ad(w)F(A)=wFw^{-1}(a_{i,j})=(b_{i,j})\quad with \ b_{i,j}=a_{i-1,j-1}^q.$$ Let $T(w)$ be the maximal torus $T^*\subset GL_r$, for which the Frobenius is given as $ad(w)F$ (see the next section for details) so that $$T(w)^F=\{t\in T^*|ad(w)F(t)=t\}$$ $$\cong \{x\in {{\mathbb F}}_{q^r}^*|(x,x^q,...,x^{q^{r-1}})\in T^*\}\cong {{\mathbb F}}_{q^r}^*.$$ Take $H^*(BT^*)\cong {{\mathbb Z}}/\ell[t_1,..,t_ r]$. Let $i:T(w)^F\subset T^*$. Then we can take the ring generator $t\in H^2(BT(w)^F)$ such that $i^*t_i=q^{i-1}t$.
The following map is injective $$H^*(BGL_r)/((q^i-1)c_i)\cong {{\mathbb Z}}/{\ell}[c_r]\to H^*(BG_r).$$
It is enough to prove that for the map $$i^*:H^*(BGL_r)\to H^*(BG_r)\to H^*(BT(w)^F)\cong H^*({{\mathbb F}}_{q^r}^*),$$ we can see $i^*c_1=...=i^*c_{r-1}=0$, and $ i^*c_r=(-1)^rt^r$.
Let $s_i$ be the $i$-th elementary symmetric function of variables $t_1,...,t_r$, namely, $$(X-t_1)(X-t_2)...(X-t_r)=X^r+s_1X^{r-1}+...+s_r.$$ Since $i^*(t_i)=q^{i-1}t$, we see that $$(X-t)(X-qt)...(X-q^{r-1}t)=X^n+i^*(s_1)X^{r-1}+...+i_*(s_r) .$$ On the other hand, the polynomial $X^r-t^r$ has its roots $X=t,qt,...,q^{r-1}t$. Hence we see that the above formula is $X^r-t^r.$ It implies the assertion above.
The following map is injective $$H^*(BGL_n)^F\cong {{\mathbb Z}}/{\ell}[c_r,...,c_{[n/r]r}]\to H^*(BG_r).$$
Let $k=[n/r]$. let us take $$w=(1,...,r)(r+1,...,2r)...((k-1)r+1,...,kr).$$ We consider the map $$i^*:H^*(BGL_n)\to H^*(BG_n)\to H^*(BT(w)^F)
\cong H^*(B({{\mathbb F}}_{q^r}^*\times ... \times{{\mathbb F}}_{q^r}^*)).$$ We chose $t_i\in H^2(BT)$ ($1\le i\le n$) and $t_j'\in H^2(BT(w)^F)$ $(1\le j\le k$) such as $i^*t_1=t_1', i^*t_2=qt_1',...$. Then the arguments similar to the proof of the preceding lemma, we have $$X^n+i^*(c_1)X^{r-1}+...+i_*(c_r)=(X^r\pm (t_1')^r)...(X^r\pm (t_k')^r) .$$ Thus we get the result.
Now we consider the case $G=SL_n$. Write $SL_n({{\mathbb F}}_{q})$ by $SG_n$.
If $r\ge 2$, then, the following map is injective $$H^*(BSL_r)^F\cong {{\mathbb Z}}/{\ell}[c_r]\to H^*(BSG_r).$$
Let $w=(1,...,r)$ and recall $q(r)=1+q+...+q^{r-1}$. Then the maximal torus of $SG_r$ is written $$ST^*(w)^F\cong \{t\in F_{q^r}^*|(x,...,x^{q^{r-1}})\in T^*,\ x^{q(r)}=1\}
\cong {{\mathbb Z}}/q(r).$$ We consider the map as the case $G_r$ $$i^*:H^*(BSL_r)\to H^*(BSG_r)\to H^*(BST(w)^F)$$ $$\cong H^*(B{{\mathbb Z}}/q(r))\cong {{\mathbb Z}}/\ell[t]\otimes \Lambda(v).$$ Let us write $H^*(BST^*)\cong {{\mathbb Z}}/\ell[t_1,...,t_r]/(t_1+..+t_r)$. Then we also see that $i^*(t_i)=q^{i-1}t$ ( note $\sum q^{i-1}=q(r)=0\in {{\mathbb Z}}/\ell$). The arguments in the proof of Lemma implies this lemma.
For the case $r\ge 2$, the following map is injective $$H^*(BGL_n)^F\cong {{\mathbb Z}}/{\ell}[c_r,...,c_{[n/r]r}]\to H^*(BSG_n).$$ When $r=1$, the map ${{\mathbb Z}}/\ell[c_2,...,c_n]\to H^*(BSG_n)$ is injective.
The maximal torus of $SG_n$ is written $$ST^*(w)^F\cong \{t\in F_{q^r}^*|(x_1,...,x_1^{q^{r-1}},...,x_k,...,x_k^{q^{r-1}}))\in T^*,
\ (x_1...x_k)^{q(r)}=1\}.$$ We can get the result as the case $G_n$. When $r=1$, note that $c_1=t_1+...+t_n=0$ still in $H^*(BST^*)$.
For the case $r\ge 2$, we have the isomorphism $H^*(BSG_n)\cong H^*(BG_n)$. When $r=1$, we have $$H^*(BG_n;{{\mathbb Z}}/{\ell})\cong {{\mathbb Z}}/\ell[c_2,...,c_{n}]
\otimes \Delta(e_2,...,e_{n}).$$
Almost arguments work as the case $G_n$. For example, in the proof of Lemma 3.2, for $G=G_n$, we showed $$F(i)-F(i+1)\cong G_n/(P_{i,n-i})\times ({{\mathbb A}}^{n-i}-F(1)')$$ where $P_{i,n-i}$ is the parabolic subgroup $(G_i\times G_{n-i})\ltimes U_{i,n-i}$. We must consider the $SG_n$-version $$SG_n/S(G_i\times G_{n-i})\ltimes U_{i,n-i}({{\mathbb A}}^{n-i}-F(1)').$$ Here we can reduce $S(G_i\times G_{n-i})$ to the case $G_i\ltimes SG_{n-i}$. Then the inductive arguments work also this case.
motivic cohomology
==================
In this section, we consider the motivic version of preceding section. Let us write $$H^{*,*'}_G(X)=H^{*,*'}_G(X;{{\mathbb Z}}/p)$$ the (equivariant) motivic cohomology over the field $k=\bar {{\mathbb F}}_p$. Then we have the long exact sequence $$\to H_{G_n}^{*-2i,*'-i}(F(i)-F(i+1))\stackrel{i_*}{\to}
H_{G_n}^{*,*'}(X(i+1))\stackrel{j^*}{\to} H^{*,*'}_{G_n}(X(i))
\stackrel{\delta}{\to}.$$ However we note the following fact: the projection $$\begin{aligned}
V_{N}''\times_{G_{n-i}}({{\mathbb A}}^{n-i}-F(1)')\to &{{\mathbb A}}^{n-i}-F(1)'/G_{n-i} \\
&\cong{{\mathbb A}}^{n-i-1}\times_{\mu_{{q}^{n-i}-1}} {{\mathbb G}}_{m}\to {{\mathbb G}}_{m}/\mu_{{q}^{n-i}-1}\cong {{\mathbb G}}_{m}\end{aligned}$$ is an ${{\mathbb A}}^1$-homotopy equivalence when we replace $V_{N}''$ as a suitable large $G_{n-i}$-vector space. Then Lemma 3.2 holds for the motivic cohomology. Then the most arguments in the preceding sections also work for the motivic cohomology with the degree $$deg(c_i)=(2i,i),\qquad deg(e_i)=(2i-1,i).$$ Thus we get Theorem 1.2 in the introduction.
the Deligne-Lusztig theory
==========================
Let $G$ be a connected reductive algebraic group defined over a finite field ${{{\mathbb F}}}_q$, $q=p^r$, let $F\colon G\rightarrow G$ be the Frobenius map and let $G^F$ be the (finite) group of fixed points of $F$ in $G$.
In the paper \[De-Lu\], Deligne and Lusztig studied the representation theory of $G^F$ over fields of characteristic 0. The main idea is to construct such representations in the $\ell$-adic cohomology spaces of certain algebraic varieties $\tilde X(\dot w)$ over ${{{\mathbb F}}}_q$, on which $G^F$ acts. Fix a Borel subgroup $B^*\subset G$ and a maximal ${{{\mathbb F}}}_q$-split torus $T^*\subset B^*$, both defined over ${{{\mathbb F}}}_q$. Let $W$ be the Weyl group of $T^*$ and $$G=\bigcup_{w\in W}B^*\dot wB^*\qquad (disjoint\ union)$$ be the Bruhat decomposition, $\dot w$ being a representative of $w\in W$ in the normalizer of $T^*$. Let $X$ be the variety of all Borel subgroups of $G$. This is a smooth scheme over ${{{\mathbb F}}}_q$, on which the Frobenius element $F$ acts. Any $B\in X$ is of the form $B=gB^*g^{-1}=adgB^*$, where $g\in G$ is determined by $B$ up to right multiplication by an element of $B^*$. Let $X(w)\subset X$ be the locally closed subscheme consisting of all Borel subgroups $B=gB^*g^{-1}$ such that $g^{-1}F(g)\in B^*\dot wB^*$, namely, $$(6.1)\quad X(w)=\{g\in G|g^{-1}F(g)\in B^*\dot wB^*\}/B^*$$ $$\cong \{g\in G|g^{-1}F(g)\in \dot wB^*\}/(B^*\cap ad\dot wB^*).$$ ( Borel groups $ad(g)B^*$ and $ad(g)FB^*$ are called in relative position $w$ if $g\in X(\dot w)$.)
For any $w\in W$, let $T(w)$ be the torus $T^*$, for which the Frobenius map is given by $ad(w)F$ so that $$(6.2)\quad T(w)^F=\{t\in T^*|ad(w)F(t)=t\}.$$ Hence $T(w)^F$ is isomorphic to the set of ${{{\mathbb F}}}_q$-points of a torus $T(w)\subset G$, defined over ${{{\mathbb F}}}_q$.
Let $U^*$ be the unipotent radical of $B^*$. For any $B\in X$ let $E(B)=\{g\in G|gB^*g^{-1}=B\}/U^*$. The Frobenius map induces a map $F\colon E(B)\rightarrow E(F(B))$. Let $E(B,\dot w)=\{u\in E(B)|F(u)=u\dot w\}$. For $B\in X(w)$ the sets $E(B,\dot w)$ are the fibers of a map $\pi\colon\tilde X(\dot w)\rightarrow X(w)$, where $\tilde X(\dot w)$ is a right principal homogeneous space of $T(w)^F$ over $X(w)$. The groups $G^F$ and $T(w)^F$ act on $\tilde X(\dot w)$ and these actions commute. Thus we have the isomorphism $$(6.3)\quad \tilde X(\dot w)\cong
\{g\in G|g^{-1}F(g)\in \dot wU^*\}/(U^*\cap ad\dot wU^*).$$
Now let $\ell$ be a prime distinct from $p$, and ${{\mathbb Q}}_{\ell}$ be the algebraic closure of the field of $\ell$-adic numbers. Deligne-Lusztig consider the actions of $G^F$ and $T(w)^F$ on the $\ell$-adic cohomology $H_c{}^*(\tilde X(\dot w),{{\mathbb Q}}_{\ell})$ with compact support. For any $\theta\in {Hom}(T(w)^F,{{\mathbb Q}}_{\ell})$, let $H_c^*(\tilde X(\dot w),{{\mathbb Q}}_{\ell})_\theta$ be the subspace of $H_c{}^\ast(\tilde X(\dot w),{{\mathbb Q}}_{\ell})$ on which $T(w)^F$ acts by $\theta$. This is a $G^F$-module.
The main subject of the paper \[De-Lu\] is the study of virtual representations $R^\theta(w)=\sum_i(-1)^iH_c{}^i(\tilde X(\dot w),{{\mathbb Q}}_{\ell})_\theta$ (it can be shown that the right hand side is independent of the lifting $\dot w$ of $w$). [**Example.**]{} (See 2.1 in \[De-Lu\].) Let $V$ be an $n$-dimensional vector space over $k$ and put $G=GL(V)$. We may take a basis such that a maximal torus $T\cong {{\mathbb G}}_m^n$ and the Weyl group $W\cong S_n$; the symmetric group of $n$-letters. Then $X=G/B$ is the space of complete flags $$D \ : \ D_0=0\subset D_1\subset ...\subset D_{n-1}\subset D_n=V$$ with $dimD_i=i$. The space $E=G/T$ is the space of complete flags marked by nonzero vector $e_i\in D_i/D_{i-1}$, where $T$ acts on $E$ by $(D,(e_i))(t_i)=(D,(t_ie_i))$.
Let $w=(1,...,n)$. Then two flags $D'$ and $D''$ are relative position $w$ (for details see 1.2 in \[De-Lu\]) if and only if $$D_i''+D_i'=D_{i+1}'\ (1\le i<n-1),\quad D_{n-1}''+D_1^i=V.$$ Hence $D$ and $FD$ are in relative position $w$, if and only if $$D_1\subset D_1+FD_1\subset D_1+FD_1+F^2D_1\subset ...$$ and $V=\oplus^{n+1}F^iD_1$. A marking $e$ of $F$ is given such that $F(e)=e\cdot \dot{w}$ if and only if $$e_2=F(e_1)(mod(e_1)),\ \ ...,\ \ e_n=F^{n-1}(e_1)(mod(e_1,...,F^{n-2}(e_1))$$ $$and \qquad e_1=F^n(e_1)(mod(e_1,...,F^{n-1}(e_1));$$ Hence the mark $e$ is defined by $e_1\in D_1$ with the condition that $$F( e_1\wedge F(e_1)\wedge...\wedge F^{n-1}(e_1))=
(-1)^{n-1}(e_1\wedge F(e_1)\wedge...\wedge F^{n-1}(e_1)).$$ If $(x_i)$ are the coordinate of $e_1$, the above condition can be rewritten $$(6.4)\quad (-1)^{n-1}(det(x_i^{q^{j-1}})_{1\le i,j\le n})^{q-1}=1.$$ Hence the map $(D_1,e_1)$ induces an isomorphism of $\tilde X(\dot{w})$ with the affine hypersurface (6.4). Note that this hypersurface is stable under $x\mapsto tx$ for $t\in F_{q^n}^*$, and this is the action of $T(w)^F$.
Recall that $(det(x_i^{q^{j-1}})_{1\le i,j\le n})$ is written by $e_n$ in $\S 2$. Thus we have
The variety $Q'$ in Theorem 2.5 in $\S 2$ is isomorphic to $\tilde X(\dot w)$.
In the next section, we will give a complete different proof of the above theorem.
The Deligne-Lusztig variety $\tilde X(\dot{w_n})$
=================================================
In 1.11.4 in \[De-Lu\], Deligne and Lustig prove the following theorem
$$G_n\setminus\tilde X(\dot w_n)\cong U^*/(U^*\cap ad(\dot w_n)U^*).$$
We will give a complete different proof of the above theorem and Theorem 2.4 by using Dickson invaraints for $G=GL_n({{\mathbb F}}_q)$ and $w_n=(1,...,n)$.
Take an adequate basis of the $n$ dimensional vector space such that $$w_n=\left( \begin{array}{cccc}
0 & 0 & \ldots & 1\\
1 & 0 & \ldots & 0\\
\multicolumn{4}{c}{\dotfill} \\
0 & \ldots & 1 & 0\\
\end{array} \right),
\qquad U^*=\{ \left( \begin{array}{cccc}
1 & * & \ldots & *\\
0 & 1 & \ldots & *\\
\multicolumn{4}{c}{\dotfill} \\
0 & \ldots & 0 &1
\end{array}\right) | *\in \bar F_p\}.$$ Let $x_{i,j}(a)=1+ae_{i,j}$ where $e_{i,j}$ is the elementary matrix with $1$ in $(i,j)$-entry and $0$ otherwise. Then $U^*$ is generated by $x_{i,j}(a)$, $$U^*={{\langle}}x_{i,j}(a)| 1\le i<j \le n\ |\ a\in \bar F_p{{\rangle}}$$ with the relation $$x_{i,j}(a)x_{i,j}(b)=x_{i,j}(a+b),\quad [x_{i,j}(a),x_{k,l}(b)]=\delta_{j,k}x_{i,l}(ab).$$ Note $ad(w)x_{i,j}(a)=wx_{i,j}(a)w^{-1}=x_{i+1,j+1}(a)$ identifying $i,j\in {{\mathbb Z}}/n$. Hence $$InU^*=U^*\cap ad(w)U^*={{\langle}}x_{i,j}|x_{1,j}=0{{\rangle}}$$ and $ad(w^{-1})InU^*={{\langle}}x_{i,j}|x_{i,n}=0{{\rangle}}$, that is $$InU^*=\left( \begin{array}{ccccc}
1 & 0 & 0&\ldots & 0\\
0 & 1 & *&\ldots & *\\
\multicolumn{4}{c}{\dotfill}&* \\
0 & 0 & \ldots & 0&1\\
\end{array} \right),
\qquad ad(w^{-1})InU^*= \left( \begin{array}{ccccc}
1 & * & \ldots&* & 0\\
0 & 1 & \ldots&* & 0\\
\multicolumn{4}{c}{\dotfill} &\cdot \\
0 & 0 & \ldots &0 &1
\end{array}\right) .$$
In Theorem 7.1, the $InU^*$ action on $U^*$ is given by the following $\rho$ (see 1.11.4 in \[De-Lu\]) $$\rho(u)v=ad(\dot w^{-1}_n)(u)vF(u^{-1})\quad for \ u\in InU^*,\ v\in U^*.$$
There is an isomorphism $$U^*/\rho(InU^*)\cong {{\langle}}x_{ij}(a)|x_{i,j}=0\ if \ j\not =n{{\rangle}}$$ $$=\{ \left( \begin{array}{cccccc}
1 & 0 &\ldots & 0&d_1\\
\multicolumn{4}{c}{\dotfill}&* \\
0 & 0&\ldots & 1&d_{n-1}\\
0 & 0 & \ldots & 0&1\\
\end{array} \right)\in U^*\ |\ d_1,...,d_{n-1}\in \bar F_p\}.$$
We consider the $\rho$-action when the case $u=x_{i,j}(a)$ and $v=x_{k,l}(b)$, namely, $$\rho(u)v=ad(\dot w^{-1})(x_{ij}(a))x_{k,l}(b)F(x_{i,j}(a)^{-1})$$ $$=x_{i-1,j-1}(a)x_{k,l}(b)x_{i,j}(-a^q).$$ For roots $x_{i,j}$ and $x_{i',j'}$, we define an order $x_{i,j}<x_{i',j'}$ if $i<i'$ or $i=i'$, $j<j'$. Then any $v\in U^*$ is uniquely written by the product $\Pi x_{i,j}(b_{i,j})$ when we fix the above order in the product. For any $a\in U^*$, let $x_{i_0,j_0}$ be the minimal root of $v$ such that $x_{i_0,j_0}(b_{i_0,j_0})\not =0, j_0<n$.
Take $i=i_0+1$, $j=j_0+1$ and $a=-b_{i_0,j_0}$. Then the equation $$\rho(u)v=ad(\dot w^{-1})(x_{ij}(a))(\Pi x_{k,l}(b))F(x_{i,j}(a)^{-1})$$ $$=x_{i_0,j_0}(-b_{i_0,j_0})(\Pi x_{k,l}(b_{i,j}))x_{i_0+1,j_0+1}(-a^q)$$ $$=(\Pi_{(i_0,j_0)<(k,l)} x_{k,l}(b_{i,j}))x_{i_0+1,j_0+1}(-a^q)$$ implies that a nonzero minimal root of $\rho(u)v$ is larger than $(i_0,j_0)$. Repeating this process, there exists $u\in InU^*$ such that $\rho(u)v=\Pi_{i=1}^{n-1}x_{i,n}(d_i)$.
But all nonzero elements in the right hand side group in this lemma are not in $Im(\rho(u))v$ for $u\not =1$. Hence we have the lemma.
Recall that we can identify $$Q'=\{x=\left( \begin{array}{cccc}
x_1 & x_1^q& \ldots & x_1^{q^{n-1}}\\
x_2 & x_2^q & \ldots & x_2^{q^{n-1}}\\
\multicolumn{4}{c}{\dotfill} \\
x_n & x_n^q & \ldots & x_n^{q^{n-1}}
\end{array} \right) \in GL_n| |x|^{q-1}=det(x)^{q-1}=1\}.$$
We can define the map $f:Q'\to U^*/(\rho(InU^*)$ by $x\mapsto \dot w^{-1}_nx^{-1}Fx$, in fact, $$f(x) =\left( \begin{array}{cccccc}
1 & 0 &\ldots & 0& c_{n,1}\\
\multicolumn{4}{c}{\dotfill}&* \\
0 & 0&\ldots & 1&{c_{n,n-1}}\\
0 & 0 & \ldots & 0&1\\
\end{array} \right)$$ where $c_{n,i}=c_{n,i}(x_1,...,x_n)$ is the Dickson element defined in $\S 2$. This map also induces the isomorphism $$G_n\setminus Q'\cong U^*/(\rho(InU^*)) \cong Spec(k[c_{n,1},...,c_{n,n-1}]) \qquad (so\ Q'\cong \tilde X(\dot w_n)).$$
Let us write $$e_n\left( \begin{array}{cccc}
i_1 & i_2& \ldots & i_n\\
j_1 & j_2 & \ldots & j_n\\
\end{array} \right)
=\left| \begin{array}{cccc}
x_{j_1}^{q^{i_1}} & x_{j_1}^{q^{i_2}}& \ldots & x_{j_1}^{q^{i_n}}\\
x_{j_2}^{q^{i_1} }& x_{j_2}^{q^{i_2}}& \ldots & x_{j_2}^{q^{i_n}}\\
\multicolumn{4}{c}{\dotfill} \\
x_{j_n}^{q^{i_1}} & x_{j_n}^{q^{i_2}}& \ldots & x_{j_n}^{q^{i_n}}\\
\end{array} \right|$$ so that $e_n\left( \begin{array}{cccc}
0 & 1& \ldots & n-1\\
1 & 2 & \ldots & n\\
\end{array} \right)=e(x)=|x|.$ Then the $(j,i)$ cofactor of the matrix $x$ is expressed as $$D_{j,i}=(-1)^{i+j}e_{n-1}\left( \begin{array}{cccccc}
0 & 1& \ldots& \hat{i-1}& \ldots & n-1\\
1 & 2& \ldots& \hat j& \ldots & n
\end{array} \right).$$ By Clamer’s theorem, we know $$x^{-1}=|x|^{-1}(D_{j,i})^t=|x|^{-1}(D_{i,j}).$$
Let us write $(B_{i,j})=|x|x^{-1}F(x)$. Then we can compute $$B_{s,t}=\sum D_{s,k}x(k,t)^q=\sum D_{s,k}x_k^{q^t}
\qquad (where\ x(k,l)=(k,l)-entry \ of\ x)$$ $$=\left| \begin{array}{ccccc}
x_1 &\ldots& \stackrel{s}{x_1^{q^t}}& \ldots & x_1^{q^{n-1}}\\
x_2 & \ldots&x_2^{q ^t}& \ldots & x_2^{q^{n-1}}\\
\multicolumn{5}{c}{\dotfill} \\
x_n & \ldots &x_n^{q^t} & \ldots & x_n^{q^{n-1}}
\end{array} \right|.$$ This element is nonzero only if $t=s-1$ or $t=n$. If $t=s-1$, then the above element is $|x|$. If $t=n$, then the above element is indeed, $(-1)^{n-s}|x|c_{n,s-1}$ by the definition of the Dickson elements as stated in $\S 2$. Thus we have $$x^{-1}F(x)=|x|^{-1}(B_{st})=
\left( \begin{array}{cccccc}
0 & 0 &\ldots & 0& c_{n,0}\\
1 & 0&\ldots & 0&{c_{n,1}}\\
\multicolumn{4}{c}{\dotfill}&* \\
0 & 0 & \ldots &1&c_{n,n-1}\\
\end{array} \right).$$ Here $c_{n,0}=1$ and acting $\dot w^{-1}_n$, we have the desired result for $f(x)$.\
We will show that $f$ is an isomorphism.\
We note that $f$ is decomposed into $$\xymatrix{
G_n\backslash GL_n \ar[r]^-{\bar{L}}
& GL_n \ar[r]^-{\dot w_n^{-1}}
& G_n \backslash GL_n \\
G_n\backslash Q' \ar[rr]_{f} \ar[u]^-{incl.} &
& U^*/\rho(InU^*) \ar[u]_-{incl.}
}$$ where $L(x)=x^{-1}F(x)$.
Since the Lang map is separable, so is $f$. We see that $f$ is injective from the diagram. To show that $f$ is an isomorphism, it is enough to see that $f:Q'\to U^*/\rho(InU^*)$ is surjective.\
When we consider $Q'$ as a subvariety of ${{\mathbb A}}^{n}$, the above $f$ is identified with a map $g|Q'$, where $g:{{\mathbb A}}^{n}\to{{\mathbb A}}^{n-1}$ is defined by $g(x)=(c_{n,1}(x),...,c_{n,n-1}(x)).$\
Then the surjectivity follows from the following lemma:
Let $(f_{1},...,f_{n})$ be a homogeneous regular sequence of $k[x_{1},...,x_{n}]$. Then the associated map $f:{{\mathbb A}}^{n}\to{{\mathbb A}}^{n}$ is surjective. It means that\
$f':V(f_{1}-a)\to{{\mathbb A}}^{n-1}$ is surjective for $a$$\in$ k where $f'=pr(f|V(f_{1}-a))$where $pr:{{\mathbb A}}^{n}\to{{\mathbb A}}^{n-1}$ is the projection $pr(x_{1},...,x_{n})=(x_{1},...,x_{n-1})$
We consider the inclusion $i:{{\mathbb A}}^{n}\subset{{\mathbb P}}^{n}$ defined by $i(x_{1},...,x_{n})=[x_{1},...,x_{n},1]$, and denote the coordinate of ${{\mathbb P}}^{n}$by$[u_{1},...u_{n},z]=[u,z]$.\
We denote by$\tilde f:{{\mathbb P}}^n\to {{\mathbb P}}^n$the rational map extended from $f$ and denote by $d_{i}$ the degree of $f_{i}$.\
For $\alpha \in {{\mathbb A}}^{n} $,we see that $\tilde f^{-1}(\alpha)$is given by $$V_{+}(f_{1}(u)-\alpha_{1}z^{d_{1}},...,f_{n}(u)-\alpha_{n}z^{d_{n}}),when \alpha = (\alpha_{1},...,\alpha_{n}).$$\
Then $\tilde f^{-1}(\alpha)\neq \phi$ by the Bezout theorem. Since $(f_{1},...,f_{n})$ is a homogeneous regular sequence, we see that $V(f_1,...,f_n)=\{0\}$. It implies that $$\tilde f^{-1}(\alpha)\cap V_{+}(z)=\{[u_{1},...,u_{n},1]|f_{1}(u)=\dots=f_{n}(u)=0\}=\phi.$$ We have $f^{-1}(\alpha)=\tilde f^{-1}(\alpha)\neq\phi$.Hence $f$ is surjective.\
Hence we know $$\tilde X(\dot w)\cong \{(x_1,...,x_n)\in {{\mathbb A}}^n|e(x_1,...,x_n)^{q-1}=|x|^{q-1}=1\}$$
There is an isomorphism of varieties $$X(1)\cong \tilde X(\dot w_n)\times_{T(\dot w_n)^F}{{\mathbb G}}_m.$$
We have isomorphisms $$G_n\backslash
X(1)\cong G_n\backslash
(\tilde X(\dot w_n)
\times_{T(\dot w)^F}{{\mathbb G}}_m\cong {{\mathbb A}}^{n-1}\times {{\mathbb G}}_m.$$
[R-W-Y]{}
P. Deligne and G. Lusztig. Represntations of reductive groups over finite fields. *Ann. Math.* **103** (1976), 103-161.
M. Kameko and M. Mimura. Mùi invariants and Milnor operations, Geometry and Topology Monographs [**11**]{}, (2007), 107-140.
L.Molina and A.Vistoli. On the Chow rings of classifying spaces for classical groups. *Rend. Sem. Mat. Univ. Padova* **116** (2006), 271-298.
H.Mui. *J.Fac.Sci.U. of Tokyo* **22** (1975), 319-369.
D.Quillen. On the cohomology and $K$-theory of general linear groups over a finite field. *Ann. Math.* **96** (1972), 552-586.
A.Vistoli. *J. Reine Angew. Math.* **610** (2007) 181-227.
V. Voevodsky. The Milnor conjecture. *www.math.uiuc.edu/K-theory/0170* (1996).
V.Voevodsky. Motivic cohomology with ${{\mathbb Z}}/2$-coefficients. *Publ.Math. IHES.* **98** (2003), 59-104.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: '[The task of [*data fusion*]{} is to identify the true values of data items ([[*e.g.*]{}]{}, the true date of birth for [*Tom Cruise*]{}) among multiple observed values drawn from different sources ([[*e.g.*]{}]{}, Web sites) of varying (and unknown) reliability. A recent survey [@LDL+12] has provided a detailed comparison of various fusion methods on Deep Web data. In this paper, we study the applicability and limitations of different fusion techniques on a more challenging problem: [*knowledge fusion*]{}. Knowledge fusion identifies true subject-predicate-object triples extracted by multiple information extractors from multiple information sources. These extractors perform the tasks of entity linkage and schema alignment, thus introducing an additional source of noise that is quite different from that traditionally considered in the data fusion literature, which only focuses on factual errors in the original sources. We adapt state-of-the-art data fusion techniques and apply them to a knowledge base with 1.6B unique knowledge triples extracted by 12 extractors from over 1B Web pages, which is three orders of magnitude larger than the data sets used in previous data fusion papers. We show great promise of the data fusion approaches in solving the knowledge fusion problem, and suggest interesting research directions through a detailed error analysis of the methods. ]{}'
author:
- |
Xin Luna Dong, Evgeniy Gabrilovich, Geremy Heitz, Wilko Horn,\
Kevin Murphy, Shaohua Sun, Wei Zhang\
bibliography:
- 'base.bib'
title: From Data Fusion to Knowledge Fusion
---
\[section\] \[definition\][Proposition]{} \[definition\][Lemma]{} \[definition\][Remark]{} \[definition\][Corollary]{} \[definition\][Claim]{} \[definition\][Theorem]{} \[definition\][Example]{}
\[subsection\]
Introduction {#sec:intro}
============
State of the Art in Data Fusion {#sec:fusion}
===============================
Fusing Extracted Knowledge {#sec:kv}
==========================
Applying Data Fusion Methods to Knowledge Fusion {#sec:performance}
================================================
Future directions {#sec:improvement}
=================
Conclusions {#sec:conclude}
===========
Acknowledgements
================
We thank Fernando Pereira, Divesh Srivastava, and Amar Subramanya for constructive suggestions that helped us improve this paper. We also thank Anish Das Sarma, Alon Halevy, Kevin Lerman, Abhijit Mahabal, and Oksana Yakhnenko for help with the extraction pipeline.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We propose a way to recover Lorentz invariance of the perturbative S matrix in the Discrete Light-Cone Quantization (DLCQ) in the continuum limit without spoiling the trivial vacuum.'
author:
- 'Masa-aki Taniguchi, Shozo Uehara, Satoshi Yamada and Koichi Yamawaki'
title: 'Recovering Lorentz Invariance of DLCQ [^1]'
---
DPNU-03-30\
hep-th/0309240
INTRODUCTION
============
The Discrete Light-Cone Quantization (DLCQ) was first introduced by Maskawa and Yamawaki (MY) in 1976 based on the canonical quantization for the constrained system due to Dirac and was also considered by Casher [@Casher] slightly later in 1976 and by Pauli and Brodsky in 1985 in different contexts. In the original paper of MY , the light-like coordinate $x^-$ was compactified $-L \leq x^- \leq L$ with periodic boundary condition thus discretizing the conjugate momentum, $p^+ = (n\pi)/L
\quad (n=0,\pm 1,\pm 2,...)$, in order to isolate the zero mode ($n=0$), $\phi_0\equiv \frac{1}{2L}\int_{-L}^{L}dx^-\phi $, from the non-zero mode, $\varphi \equiv \phi-\phi_0$.
The most important finding of MY is the discovery of the [*Zero-Mode Constraint (ZMC)*]{}: $$\Phi_3 \equiv \frac{1}{2L}\int^{L}_{-L}d x^-
\left[(m^2- \partial_{\bot}^2)\phi
+\frac{\partial V(\phi)}{\partial \phi}\right]
=\left[\frac{\partial \mathcal{H}}{\partial \phi}\right]_0
=0 \, \label{ZMC0}$$ in the 4-dimensional scalar theory with mass $m$ and self-interaction $V(\phi)$, where $\mathcal{H}= \frac{1}{2}\phi(m^2 -
\partial_{\bot}^2)\phi +V(\phi)$ is the light-cone Hamiltonian density and $[\mathcal{O}]_0 \equiv \frac{1}{2L}\int_{-L}^{L}dx^- \mathcal{O}$ is the zero mode of the operator $\mathcal{O}$. Thanks to ZMC, MY succeeded to compute the [*well-defined Dirac bracket*]{} at $x^+=y^+$: $$\{\phi(\vec{x}), \phi(\vec{y})\}_D \equiv
\{\phi(\vec{x}),\phi(\vec{y})\}- \int d\vec{z} d\vec{z}^{\prime}
\{\phi(\vec{x}), \Phi_i(\vec{z})\}
C^{-1}_{ij}(\vec{z},\vec{z}^\prime)
\{\Phi_j(\vec{z}^\prime),\phi(\vec{y})\}
\label{Dbracket}$$ where for convenience the primary second-class constraint $\Phi \equiv \pi -\partial_{-}\phi$ may be divided into the non-zero mode $\Phi_1 \equiv \pi_{\varphi} - \partial_{-}\varphi$ and the zero mode $\Phi_2 \equiv [\Phi]_0= \pi_0$, and $C^{-1}$ is the inverse of $C_{ij}(\vec{x},\vec{y}) \equiv \{\Phi_i(\vec{x}), \Phi_j(\vec{y})\}$, with $\vec{x}\equiv (x^-, x^{\bot})$ and $(i,j)=1,2,3$. Without ZMC, $\Phi_3$, which yields $C_{i,3}=-C_{3,i} \ne0 \quad
(i=1,2)$, we would have trouble to get the inverse $C^{-1}$ and hence the Dirac bracket, since $C_{1,2}=C_{2,1}=C_{2,2}=0$. Thus the ZMC is vital to a well-defined canonical light-cone commutator $[\phi(x),\phi(y)]_{x^+=y^+}$ which is given as $i\hbar$ times the Dirac bracket (up to operator ordering).
From the ZMC (\[ZMC0\]) and the Dirac bracket (\[Dbracket\]) (hence the canonical commutator), MY obtained two important physical consequences :
- Proof of the Trivial Vacuum\
Since the [*translation invariance*]{}, $[\phi,P^+]= i \partial_-\phi$ is manifest in DLCQ even for finite $L$, the vacuum is defined independently of the dynamics (thus “trivial”) as the lowest $p^+$ state ($p^+=0$). This trivial vacuum proved a unique $p^+=0$ state and hence the true vacuum, since the zero mode is [*not an independent degree of freedom*]{}, written in terms of the non-zero modes, $\phi_0=\phi_0(\varphi)$, by solving the ZMC (at least in perturbation), and hence can be removed from the physical Fock space.
- Violation of the Lorentz Invariance\
Based on the canonical commutator mentioned above, the Lorentz algebra $[\phi(x), M^{\mu\nu}]$ was explicitly computed; for the Lorentz generators $M^{+-}, M^{{\bot}-}$ which change the quantization plane (light-front), the Lorentz invariance is [*violated at operator level*]{} due to the unwanted surface term at $x^- =\pm L$ which does not vanish even in the continuum limit ($L\to\infty$).
Thus the incompatibility between the trivial vacuum and the Lorentz invariance was the very nature of DLCQ from the beginning back in 1976 and has been a long-standing problem. (See Ref. [@Yam].)
One might consider that the trivial vacuum could be fake, namely it could not be reconciled with the nonperturbative phenomena such as the confinement and the spontaneous symmetry breaking (SSB) which are attributed to the complicated vacuum structure in the usual quantization. However, it was demonstrated [@KTY; @Yam] that [*the operator solution of ZMC together with the trivial vacuum can in fact describe the SSB*]{} in four dimensions for continuous symmetry in a way to regularize the zero mode of the Nambu-Goldstone boson through explicit symmetry breaking which is taken to zero at the end of all calculations. (See also footnote \[ft1\].) What about the Lorentz invariance?
In this talk we shall demonstrate that
- The Lorentz invariance is violated [*at S matrix level*]{} as well as at operator level due to [*lack of the zero-mode loop*]{} in the perturbative DLCQ in the continuum limit [@TUYY].
- Lorentz invariance at perturbative S matrix level can be recovered by [*modifying the naive DLCQ action*]{} into the one with additional operator arising from the zero-mode loop in such a way that [*the trivial vacuum remains intact*]{} by this modification [@TUYY2].
DLCQ PERTURBATION
=================
To be definite we confine ourselves to two dimensional scalar theory with $V(\phi)=\frac{\lambda}{4 !} \phi^4$. Extension will be discussed in the end. The ZMC (\[ZMC0\]) in two dimensions reads $ m^2\phi_0 +\frac{\lambda}{3!}\left([\varphi^3]_0
+3[\varphi^2]_0\phi_0 +\phi_0^3\right) = 0\, .\label{ZMC2}$ This is a cubic equation for $\phi_0$ once $[\varphi^3]_0$ and $[\varphi^2]_0$ are given, and hence has three solutions (up to operator ordering):[@TUYY2] $$\phi_0^{(0)}= G_- - G_+, \quad \phi_0^{(\pm v)}
= \pm \left[ e^{\frac{4\pi i}{3}} G_{\mp} - e^{\frac{2\pi i}{3}}
G_{\pm}\right],\label{Sol}$$ where $$G_{\mp} \equiv \left(\mp \frac{1}{2}[\varphi^3]_0+\frac{1}{2}
\sqrt{\left([\varphi^3]_0\right)^2+4\left([\varphi^2]_0
+\frac{2m^2}{\lambda}\right)^3}\right)^{\frac{1}{3}}\, .$$ In the case of $m^2>0$, only $\phi_0^{(0)}$ is a real solution [^2] whose Taylor expansion for $\lambda\ll 1$ coincides with the perturbative solution (see Fig. 1) : $$\begin{aligned}
\phi_0^{(0)}(x^+)&=& -i\lambda \int d^2y\,
\Delta_{0}(x-y)\frac{\varphi^3(y)}{3!} \nonumber\\
&&+ (-i\lambda)^2 \int d^2y_1 \int d^2y_2\,
\Delta_{0}(x-y_1)\,\frac{\varphi^2(y_1)}{2!}\Delta_{0}(y_1-y_2)\,
\frac{\varphi^3(y_2)}{3!}\,+ \cdots.\, ,\end{aligned}$$ with $\int d^2 x \equiv \int d x^+ \int_{-L}^{L} d x^-$, where $\Delta_0(x-y)=\frac{1}{2iL}{\frac{1}{m^2}\delta(x^+-y^+)}$ is the zero-mode propagator (actually not “propagating” due to instantaneous delta function $\delta(x^+-y^+)$, reflecting the fact that the zero mode is not an independent degree of freedom). Note that the ZMC solution produces [*only the tree*]{} zero-mode graph.
![Perturbative expansion of the zero-mode induced parts of the interaction Hamiltonian $H_{int}^{(0)}(\varphi)$. The “external lines” are $p^+=0$ combinations of operators, $[\varphi^2]_0$ or $[\varphi^3]_0$.](Durham-fig1.eps){width="70mm"}
![Perturbative expansion of the zero-mode induced parts of the interaction Hamiltonian $H_{int}^{(0)}(\varphi)$. The “external lines” are $p^+=0$ combinations of operators, $[\varphi^2]_0$ or $[\varphi^3]_0$.](Durham-fig2.eps){width="70mm"}
Plugging the solution $\phi_0^{(0)}=\phi_0^{(0)}(\varphi)$ into the original Hamiltonian, $\mathcal{H}= \frac{1}{2}(\varphi
+\phi_0^{(0)})(m^2) (\varphi +\phi_0^{(0)}) + \frac{\lambda}{4!}
(\varphi +\phi_0^{(0)})^4=\frac{1}{2} m^2 \varphi^2 +
\mathcal{H}_{int}^{(0)}(\varphi) $, we obtain the effective interaction Hamiltonian $\mathcal{H}_{int}^{(0)}(\varphi)$ whose perturbative expansion reads : $$\begin{aligned}
{}-i\int d x^+ H_{int}^{(0)}(\varphi)
&=& -i\lambda\int d^2 x \frac{1}{4!} \varphi^4(x)
+\frac{1}{2!}(-i\lambda)^2 \int d^2 x \int d^2 y
\left(\frac{\varphi^3(x)}{3!} \Delta_0(x-y)
\frac{\varphi^3(y)}{3!} \right)\nonumber \\
&+&\frac{3}{3!}(-i\lambda)^3 \int d^2 x \int d^2 y \int d^2 z
\left(\frac{\varphi^3(y)}{3!} \Delta_0(y-x)
\frac{\varphi^2(x)}{2!} \Delta_0(x-z) \frac{\varphi^3(z)}{3!}
\right) \nonumber \\
&& +~O(\lambda^4)\, ,
\label{effint}\end{aligned}$$ which has the zero-mode-induced terms (see Fig. 2) in addition to the original term (first term). Note again that zero-mode appears only in the tree internal line coupled to $[\varphi^2]_0$ and $[\varphi^3]_0$.
Now, the perturbative S matrix is given by the expansion of $S=T\exp{\left[-i\int_{-\infty}^{\infty}dx^{+}
H_{int}^{(0)}(\varphi)\right]}$. It yields Feynman graphs which contain no closed single line of the zero-mode propagator, since there is no zero mode operator to be contracted in $H_{int}(\varphi)$, namely [*the zero-mode loop is absent*]{}.
NON-COVARIANT RESULT
====================
Let us now show that as a simplest example the two-body scattering amplitude (forward scattering amplitude) with zero $p^+$ transfer in DLCQ disagrees with the covariant one [@TUYY]. At one loop the covariant amplitude in this case is given by $$\begin{aligned}
A_{\rm Cov}^{\rm Forward} &=& \frac{1}{2}(-\lambda)^2 \int
\frac{dk^+dk^-}{(2\pi)^2i}\left(
\frac{1}{m^2 - 2k^+k^-}\right)^2 \nonumber\\
&=&\frac{1}{2}(-\lambda)^2\int_{0}^{\infty}\frac{dk^+}{4\pi(k^+)^2}
\int_0^{\infty}d\xi \,\xi\,\delta(\xi)\,
\exp\left(-\frac{\xi m^2}{2k^+}\right)\
=\frac{(-\lambda)^2}{8\pi m^2} \ne 0 \,,
\label{Covariant}\end{aligned}$$ which is [*non-vanishing*]{} result, where the integral $d k^+$ should be done before that of $d \xi$, since otherwise $d \xi$ first would give $0$ but $d k^+$ afterward would divergence and hence the integral would be ill-defined $\infty \times 0$. On the other hand, straightforward calculation via the DLCQ perturbation mentioned above yields for the same process the [*vanishing*]{} amplitude: $$\begin{aligned}
A_{\rm DLCQ}^{\rm Forward} = \frac{1}{2} (-\lambda)^2
\frac{1}{2L}\sum_{l>0}\frac{1}{2(\frac{l\pi}{L})^2}
\int_0^{\infty}d\xi\,\xi\,\delta(\xi)\,
\exp\left(-\frac{\xi m^2}{2 (l\pi/L)}\right)=0\,\end{aligned}$$ [*independently of*]{} $L$ and so does in the continuum limit $L \rightarrow \infty$. Note that the DLCQ amplitude differs from (\[Covariant\]) simply by the replacement $1/(2\pi)\int_0^{\infty}dk^+\rightarrow 1/(2L)\sum_{l>0}\, ,$ where [*the zero mode loop is absent*]{} $l\ne 0$ in the DLCQ amplitude, and this time the integral $d \xi$ can be done before $
\sum_{l>0}$, since $(2L)^{-1}\sum_{l>0}\,[2 (l\pi/L)^2]^{-1}
= L/24 < \infty$ and hence the calculation is well-defined, $({\rm finite}) \times 0=0$.
The discrepancy is just the forward scattering amplitude with measure zero but has serious physical consequences. Actually, the same arguments can apply to the loop diagrams attached with arbitrary number of sets of external lines with $p^+=0$ corresponding to $[\varphi^2]_0$ and $[\varphi^3]_0$. Vanishing of all these types of diagrams implies that the effective potential is zero, for example, as was emphasized in [@Heinzl].
The absence of the zero-mode loop can be seen more transparently in the path integral formalism [@TUYY2] where the second-class constraints $\Phi_1\equiv
\pi_{\varphi}-\partial_- \varphi, \Phi_2\equiv \pi_0, \Phi_3\equiv
[\frac{\partial \mathcal{H}}{\partial \phi}]_0$ are incorporated in the standard way: $$\begin{aligned}
Z&=&\int \left[D\pi_{\varphi} D\pi_0 D\varphi D\phi_0\right]
\delta(\Phi_1)\delta(\Phi_2) \delta(\Phi_3) (Det\, C)^{1/2}\,
\exp \left[i \int \left(\pi \partial_+\phi-\mathcal{H}\right)\right]
\nonumber\\
&=& \int \left[D\varphi D\phi_0 D B Dc D{\bar c}\right]\,
\exp\left[ i \int \left\{\mathcal{L} +\left(B \left[
\frac{\partial \mathcal{H}}{\partial \phi}\right]_0
+i {\bar c} \left[\frac{\partial^2 \mathcal{H}}{\partial\phi^2}
\right]_0 c\right)\right\}\right]\, ,\end{aligned}$$ where the integral $D\pi_{\varphi} D\pi_0 \delta(\Phi_1)\delta(\Phi_2)$ was done trivially, and $\delta(\Phi_3) = \int [D B]
\exp[i\int B [\frac{\partial \mathcal{H}}{\partial \phi}]_0] $ and $(Det \, C)^{1/2} = Det \left(\left[\frac{\partial^2
\mathcal{H}}{\partial \phi^2} \right]_0\right) =\int [D c D \bar c]
\exp \left[i \int i {\bar c} \left[\frac{\partial^2
\mathcal{H}}{\partial \phi^2}\right]_0 c\right]$, with $B,c, {\bar
c}$ being Nakanishi-Lautrup field, ghost and antighost, respectively.[^3]
Now, $\left(B\left[\frac{\partial\mathcal{H}}{\partial\phi}\right]_0
+i{\bar c}\left[\frac{\partial^2 \mathcal{H}}{\partial\phi^2}\right]_0
c \right)=\delta_{\rm BRS} \left(-i {\bar c} \left[
\frac{\partial \mathcal{H}}{\partial \phi}\right]_0 \right)$ is a BRS singlet, where the BRS transformation $\delta_{\rm BRS}$ is defined as $\delta_{\rm BRS} {\bar c} =iB, \delta_{\rm BRS} \phi_0 =c,
\delta_{\rm BRS} B=\delta_{\rm BRS} c =0$. Then $(\phi_0,B,{\bar c},c)$ [*are BRS-quartet and hence the loop of $(\phi_0, B, {\bar c}, c)$ cancel each other*]{} in much the same way as in the gauge theories . We have explicitly checked this cancellation of the zero-mode loop by the $B, {\bar c},c$ loops to two-loop order. In contrast, the covariant expression has no constraint and simply $Z=\int[D\phi]
\exp\left[ i \int \mathcal{L}\right]$ where the loop effects of the zero-mode (though not clearly separated) are not canceled out.
RECOVERING LORENTZ INVARIANCE
=============================
We have seen that the violation of Lorentz invariance is a real effect in DLCQ as far as we use the naive discretization $\int_{-\infty}^{+\infty} d x^-\mathcal{L} \rightarrow \int_{-L}^{+L}
d x^-\mathcal{L}$, with $\mathcal{L}$ in DLCQ being [*the same*]{} as that of the continuum theory. However we know in the lattice theory that the discretized action should be different in principle from that of the continuum theory: Lorentz covariant theory should be constructed by [*using the solution*]{} which has the second order phase transition and hence yields the sensible continuum limit.
In much the same spirit, here we propose the new DLCQ Lagrangian modified by adding extra operators $-\Delta H$ in such a way that [*the perturbative solution*]{} yields the covariant limit [@TUYY2]. The extra term is generated by the zero-mode loop which can be explicitly estimated by the covariant perturbation theory. Important point of our method is that the [*trivial vacuum is not destroyed*]{} by the extra term.
At one loop order of the zero mode loop we have found the compact form of the extra operators in the Hamiltonian $\Delta H$: $$\begin{aligned}
\frac{\Delta H}{2L}&=&\frac{1}{2}\int\frac{d k^+ d k^-}{(2\pi)^2i}
\log\left[\frac{m^2+\frac{\lambda}{2!}\left(
[\varphi^2]_0+(\phi_0^{(0, \pm v)})^2\right)-k^+ k^-}
{m^2-k^+k^-}\right] \nonumber \\
&=&-\sum_{n=1}^{\infty}
\left(-\frac{\lambda}{2!}\right)^n\frac{1}{2n}I_n
\left([\varphi^2]_0+(\phi_0^{(0,\pm v)})^2\right)^n\,,\label{deltah}\end{aligned}$$
{width="35mm"}
where $\phi_0^{(0, \pm v)}$ is an operator solution of ZMC in (\[Sol\]) and $I_n (n\geq1)
=\int\frac{dk^+dk^-}{(2\pi)^2i}(m^2-k^+k^-)^{-n}$. ($n=1$, the $\log$ divergence is regularized by Pauli-Villars regularization). This yields the correct Lorentz invariant scattering amplitude for any number of external legs with zero $p^+$ transfer, i.e., $[\varphi^2]_0$ and/or $[\varphi^3]_0$ (see Fig. 3). The term $n=2$, for example, indeed yields the Lorentz-invariant two-body forward scattering amplitude with zero $p^+$ exchange in (\[Covariant\]). It is obvious that the extra term $\Delta H$ yields the correct effective potential by simply replacing $[\varphi^2]_0+(\phi_0^{(0,
\pm v)})^2$ by a c-number constant $\phi_c^2$.
CONCLUSION
==========
We have shown that the DLCQ with naive discretization violates Lorentz invariance in the perturbative S matrix due to the lack of zero mode loop, while DLCQ guarantees that the trivial vacuum is a true vacuum even when the SSB takes place. We have proposed to modify the DLCQ action by adding extra operators in such a way that the continuum limit S matrix recovers Lorentz invariance without spoiling the trivial vacuum. We have given such a modification explicitly at one loop level. Study at higher loops is in progress. Similar arguments will be applied to four dimensional theories but meet with substantial complexity due to the transverse degrees of freedom. This is also under investigation. Although our arguments are only within perturbation, we expect that our strategy to modify the DLCQ action based on the solution of the dynamics should be the right way to recover the Lorentz invariance without spoiling the trivial vacuum even in the non-perturbative approach.
ACKNOWLEDGEMENTS {#acknowledgements .unnumbered}
================
This work is supported in part by JSPS Grant-in-Aid for the Scientific Research (B)(2) \#14340072 (S.Y. and K.Y.) and MEXT Grant-in-Aid for the Scientific Research \#13135212 (S.U.).
[99]{} T. Maskawa and K. Yamawaki, Prog. Theor. Phys. [**56**]{} (1976) 270. A. Casher, Phys. Rev. [**D14**]{} (1976) 452. H.C. Pauli and S.J. Brodsky, Phys. Rev. [**D32**]{} (1985) 1993; [*ibid*]{} (1985) 2001. K. Yamawaki, hep-th/9802037, in [*Proc. NuSS 97 “QCD, Lightcone Physics and Hadron Phenomenology”*]{}, Seoul, eds. C-R. Ji and D-P. Min (World Scientific, Singapore, 1998) 116-169. Y. Kim, S. Tsujimaru, and K. Yamawaki, Phys. Rev. Lett. [**74**]{} (1995) 4771 \[arXiv:hep-ph/9502214\]; S. Tsujimaru and K. Yamawaki, Phys. Rev. [**D57**]{} (1998) 4942 \[arXiv:hep-th/9704171\]. G. McCartor and D.G. Robertson, Z. Phys. [**C53**]{} (1992) 679; M. Maeno, Phys. Lett. [**B320**]{} (1994) 83 \[arXiv:hep-th/9307146\]; S. S. Pinsky, B. van de Sande and J. R. Hiller, Phys. Rev. [**D51**]{} (1995) 726 \[arXiv:hep-th/9409019\]. M. Taniguchi, S. Uehara, S. Yamada and K. Yamawaki, Mod. Phys. Lett. [**A16**]{} (2001) 2177 \[arXiv:hep-th/0106167\]. M. Taniguchi, S. Uehara, S. Yamada and K. Yamawaki, in preparation.
T. Heinzl, arXiv:hep-th/0212202. T. Kugo and I. Ojima, Prog. Theor. Phys. Suppl. [**66**]{} (1979) 1.
[^1]: Talk given at Light-Cone Workshop “Hadrons and Beyond” (LC03), Durham, 2003.
[^2]: In the case of double-well potential $m^2<0$, we can easily see [@TUYY2] that other two solution $\phi_0^{(\pm v)}$ also become real for weak coupling $\lambda \ll 1$ and their Taylor expansion coincides with the perturbation around the non-zero value $\pm v$ ($v = \sqrt{-6m^2/\lambda}$) [@KTY], where $\pm v$ is nothing but the vacuum expectation value $\langle 0| :\phi_0^{(\pm v)}: |0 \rangle $ [*on the trivial vacuum*]{} $|0 \rangle$ ($:\quad :$ is the normal ordering). This implies that [*the trivial vacuum is indeed the true vacuum even when SSB takes place*]{}. As was emphasized by Ref. [@Yam], the information of [*SSB is carried by the operator but not by the vacuum*]{} which is always trivial in contrast to the usual quantization. The three operator solutions $\phi_0^{(0)}, \phi_0^{(\pm v)}$ are plugged into the original Hamiltonian, yielding three [*different effective Hamiltonians*]{} (without zero mode) $\mathcal{H}^{(0, \pm v)}(\varphi)$ whose vacuum energy is $\langle 0| :\mathcal{H}^{(0,\pm v)}: |0 \rangle=0,
{}-\frac{3m^4}{2\lambda}$ [*for the trivial vacuum*]{} $|0\rangle$. Thus $\mathcal{H}^{(v)}$ or $\mathcal{H}^{(-v)}$, which is the [*operator non-invariant under $Z_2$*]{}, yields the physical (ground-state) solution, in perfect agreement with the result of the conventional SSB in a different language.\[ft1\]
[^3]: If we perform the integral of the zero mode $[D
\phi_0] \delta(\Phi_3) (Det \, C)^{1/2}$, we obtain $Z=\int [D\varphi] \exp \left[i\int\mathcal{L}^{(0)}(\varphi)\right]$, where $ \mathcal{L}^{(0)}(\varphi)$ is the effective Lagrangian written only in terms of non-zero modes, corresponding to (\[effint\]), which yields the same Feynman rule as in section 2. For the SSB solution $\phi_0^{(\pm v)}$, we make a simple replacement: $\mathcal{L}^{(0)}(\varphi) \rightarrow \mathcal{L}^{(\pm
v)}(\varphi)$.
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'Kouichi <span style="font-variant:small-caps;">Hagino</span>$^{1,2}$, Hua <span style="font-variant:small-caps;">Mei</span>$^{1,3}$, Jiangming <span style="font-variant:small-caps;">Yao</span>$^{1,3,4}$, and Toshio <span style="font-variant:small-caps;">Motoba</span>$^{5,6}$'
title: ' Beyond-mean-field approach to low-lying spectra of $\Lambda$ hypernuclei'
---
Introduction
============
The development in $\Lambda$-hypernuclear spectroscopy has enabled one to explore several aspects of hypernuclear structure [@Hashimoto06]. The measured energy spectra and electric multipole transition strengths in low-lying states have in fact provided rich information on the $\Lambda$-nucleon interaction in nuclear medium as well as on the impurity effect of $\Lambda$ particle. Many theoretical methods have been developed to investigate the spectroscopy of hypernuclei, such as the cluster model [@Motoba83; @Hiyama99], the shell model [@Millener], the ab-initio method [@abinitio], the antisymmetrized molecular dynamics (AMD) [@Isaka11], and self-consistent mean-field models [@Zhou07; @Win08; @Win11; @Lu11]. Among them, the self-consistent mean-field approach is the only method which can be globally applied from light to heavy hypernuclei.
Even though the self-consistent mean-field approach provides an intuitive view of nuclear deformation, it is a drawback of this method that it does not yield a spectrum in the laboratory frame, since the approach itself is formulated in the body-fixed frame. This can actually be overcome by going beyond the mean-field approximation, in particular, by carrying out the angular momentum projection. One can also take into account the quantum fluctuation of the mean-field wave function by superposing many Slater determinants with the generator coordinate method (GCM). When the pairing correlation is important, the particle number projection can also be implemented. Such scheme has been referred to as a beyond-mean-field approach, and has rapidly been developed in the nuclear structure physics for the past decade [@Bender03; @Yao14].
In this contribution, we present a new method for low-lying states of hypernuclei based on the beyond-mean-field approach [@Meihua1; @Meihua2]. In this method, we first apply the beyond-mean-field approach to a core nucleus. Low-lying states of hypernuclei are then constructed by coupling a $\Lambda$ particle to the core nucleus states. We thus call this approach the microscopic particle-rotor model, in which the rotor part is described with the microscopic beyond-mean-field method. See Refs. [@Meihua3; @Weixia15; @Zhou15; @Yao11] for other different types of application of the beyond-mean-field approach to hypernuclei, which are complementary to the present microscopic particle-rotor model. We shall apply the microscopic particle-rotor model to the $^{13}_{~\Lambda}$C hypernucleus and discuss the impurity effect in this hypernucleus.
Microscopic particle-rotor model
================================
We consider a hypernucleus, which consists of a $\Lambda$ particle and an even-even core nucleus. In the microscopic particle-rotor model, the wave function for the whole $\Lambda$ hypernucleus with the angular momentum $J$ and its $z$-component $M$ is given as $$\label{wavefunction}
\Psi_{JM}({\mbox{\boldmath $r$}},\{{\mbox{\boldmath $r$}}_N\})
=\sum_{n,j,\ell, I} R_{j\ell nI}(r)
[{\cal Y}_{j\ell}(\hat{{\mbox{\boldmath $r$}}})\otimes
\Phi_{nI}(\{{\mbox{\boldmath $r$}}_N\})]^{(JM)},$$ where ${\mbox{\boldmath $r$}}$ and ${\mbox{\boldmath $r$}}_N$ are the coordinates of the $\Lambda$ hyperon and the nucleons, respectively. In this equation, ${\cal Y}_{j\ell}(\hat{{\mbox{\boldmath $r$}}})$ is the spin-angular wave function for the $\Lambda$ hyperon, while $\vert\Phi_{nI}\rangle$ is the wave functions for the low-lying states of the nuclear core. The latter is constructed from the mean-field wave functions as, $$\label{GCM}
\vert \Phi_{nI M_I}\rangle
=\int d\beta\, f_{nI} (\beta)
\hat P^{I}_{M_IK} \hat P^N\hat P^Z\vert \varphi(\beta)\rangle,$$ where $\beta$ is the quadrupole deformation parameter and $|\varphi(\beta)\rangle$ is the mean-field wave function at $\beta$ obtained with the constrained mean-field approximation. Here we have assumed that the core nucleus has axial symmetric shape. $\hat P^{I}_{M_IK}$, $\hat P^N$, and $\hat P^Z$ are the projections operators for the angular momentum, the neutron number, and the proton number, respectively. The weight function $ f_{nI}(\beta)$ in Eq. (\[GCM\]) is determined by the variational principle, that is, by solving the Hill-Wheeler equation.
We assume that the total Hamiltonian for this system is given by, $$\hat H = \hat T_\Lambda + \sum^{A_c}_{i=1} v_{N\Lambda}({\mbox{\boldmath $r$}},{\mbox{\boldmath $r$}}_{N_i})
+\hat{H}_{\rm c},
\label{eq:H}$$ where $A_c$ is the mass number of the core nucleus. Here, the first term is the kinetic energy of $\Lambda$ hyperon and the second term denotes a nucleon-hyperon interaction. The last term, $\hat{H}_{\rm c}$, is the Hamiltonian for the core nucleus, which is solved with the beyond-mean-field approach. With the Hamiltonian, Eq. (\[eq:H\]), one can derive the coupled-channels equations for the radial wave functions, $R_{j\ell nI}(r)$, in which the coupling potentials are given in terms of the transition densities [@Meihua1; @Meihua2]. The solutions of the coupled-channels equations provide the spectrum of the hypernucleus as well as the transition probabilities among the low-lying states.
Application to the $^{13}_{~\Lambda}$C hypernucleus
===================================================
We now apply the microscopic particle-rotor model to the $^{13}_{~\Lambda}$C hypernucleus, even though the model can be applied also to even heavier hypernuclei, such as $^{155}_{~~\Lambda}$Sm. To this end, we use the relativistic point coupling model with the PC-F1 parameter set [@PC-F1] for the effective nucleon-nucleon interaction. For the nucleon-hyperon interaction, we use a simple relativistic zero-range interaction with a repulsive vector-type and an attractive scalar-type terms [@Meihua1; @Meihua2], $$v_{N\Lambda}({\mbox{\boldmath $r$}},{\mbox{\boldmath $r$}}_{N})=\alpha_V^{N\Lambda}
\delta({\mbox{\boldmath $r$}}-{\mbox{\boldmath $r$}}_N)
+\alpha_S^{N\Lambda}\gamma_\Lambda^0\delta({\mbox{\boldmath $r$}}-{\mbox{\boldmath $r$}}_N) \gamma_N^0,$$ where $\gamma^0$ is a Dirac matrix. The parameters $\alpha_V^{N\Lambda}$ and $\alpha_S^{N\Lambda}$ are determined so as to reproduce the empirical $\Lambda$ binding energy of $^{13}_{~\Lambda}$C. The coupled-channels equations for the radial wave functions, $R_{j\ell nI}(r)$, are solved by expanding $R_{j\ell nI}(r)$ on a spherical harmonic oscillator basis with 18 major shells. To this end, we include the 0$^+_1$, 2$^+_1$, 4$^+_1$, 0$^+_2$, 2$^+_2$, and 4$^+_2$ states in the core nucleus, $^{12}$C.
![The left panel: the spectrum of the $^{13}_{~\Lambda}$C hypernucleus obtained with the microscopic particle-rotor model. The experimental data are taken from Ref. [@Hashimoto06]. The right panel: a comparison of calculated $E2$ transition strengths, $B(E2)$, for $^{12}$C and $^{13}_{~\Lambda}$C, given in units of $e^2$ fm$^4$. ](fig1.eps "fig:") ![The left panel: the spectrum of the $^{13}_{~\Lambda}$C hypernucleus obtained with the microscopic particle-rotor model. The experimental data are taken from Ref. [@Hashimoto06]. The right panel: a comparison of calculated $E2$ transition strengths, $B(E2)$, for $^{12}$C and $^{13}_{~\Lambda}$C, given in units of $e^2$ fm$^4$. ](fig2.eps "fig:")
The left panel of Fig. 1 shows the low-lying spectrum of $^{13}_{~\Lambda}$C so obtained. One can see that the low-lying spectrum for $^{13}_{~\Lambda}$C is well reproduced, although the excitation energies are slightly overestimated. This calculation indicates the ground state rotational band of $^{13}_{~\Lambda}$C shown in the column (b), that is, the ground state 1/2$^+$ and the two doublets of $(5/2^+, 3/2^+)$ and $(9/2^+, 7/2^+)$, mainly consist of the configuration of $\Lambda s_{1/2}$ coupled to the ground rotational band (that is, 0$^+_1$, 2$^+_1$, and 4$^+_1$) of the core nucleus, $^{12}$C. The doublet states are almost degenerate in energy, the energy difference being only 10 keV for each of the two doublets. The levels in the column (c) correspond to the configuration of $\Lambda s_{1/2}$ coupled to the second rotational band ($n=2$) in $^{12}$C. These states share similar features as those in the ground state band shown in the column (b).
In the negative-parity states shown in the column (d), the dominant configuration in the wave functions is that with the $\Lambda$ particle in the $p$ orbitals coupled to the ground state rotational band of the core nucleus. That is, the first $3/2^-$ and $1/2^-$ states consist mainly of $0_1^+\otimes\Lambda_{p_{3/2}}$ and $0_1^+\otimes\Lambda_{p_{1/2}}$, respectively, as is indicated in the figure. The energy splitting between these states is as small as 199 keV, which reflects mainly the spin-orbit splitting of $\Lambda$ hyperon in the $p_{3/2}$ and $p_{1/2}$ states. The obtained splitting is in a good agreement with the empirical value, 152$\pm$54$\pm$36 keV [@Ajimura01].
In contrast to the first $1/2^-$ and $3/2^-$ states, the second $1/2^-$ and $3/2^-$ states in the column (e) show a large configuration mixing. That is, the fraction of the $0_1^+\otimes\Lambda_{p_{1/2}}$ and $2_1^+\otimes\Lambda_{p_{3/2}}$ configurations in the wave function for the 1/2$^-_2$ state is 0.60 and 0.38, respectively, while the fraction of the $2_1^+\otimes\Lambda_{p_{3/2}}$ and $2_1^+\otimes\Lambda_{p_{1/2}}$ configurations is 0.54 and 0.45, respectively, in the wave function for the 3/2$^-_2$ state [@Meihua2]. This large admixture of the configurations is due to the fact that there are two states whose unperturbed energy in the single-channel calculations is close to one another. A similar admixture occurs in other hypernuclei as well, such as $^9_\Lambda$Be and $^{21}_{~\Lambda}$Ne [@Meihua1; @Meihua2], which however show this feature already in the first $1/2^-$ and $3/2^-$ states. The difference between $^{13}_{\Lambda}$C and a pair of ($^9_\Lambda$Be, $^{21}_{~\Lambda}$Ne) originates mainly from the sign of the quadrupole deformation of the core nucleus, that is, an oblate deformation for $^{12}$C and a prolate deformation for $^8$Be and $^{20}$Ne [@Meihua2].
The right panel of Fig. 1 shows a comparison of the calculated $E2$ transition strengths for low-lying positive parity states of $^{13}_{~\Lambda}$C with those of the core nucleus, $^{12}$C. In general, these transition strengths cannot be compared directly due to different angular momentum factors. However, for the transition from the 3/2$^+$ and the 5/2$^+$ states to the 1/2$^+$ state, such factor becomes trivial, and the transition strength can be directly interpreted as the that for the core nucleus from the 2$^+$ to the 0$^+$ states [@Meihua2]. Our calculation indicates that the $E2$ transition strength for $2^+_1\rightarrow 0^+_1$ in $^{12}$C is significantly reduced, by a factor of $\sim14\%$, due to the addition of a $\Lambda$ particle. The main cause of the reduction in the $B(E2)$ value is the reduction in nuclear deformation. According to our calculation, the proton radius $r_p$ is reduced from 2.44 fm to 2.39 fm by adding a $\Lambda$ particle to the $^{12}$C nucleus, which leads to about 7.9% reduction in $r_p^4$. In contrast, the deformation parameter $\beta$ is altered from $-0.29$ to $-0.23$, leading to 37.1% reduction in $\beta^2$. This clearly indicates that the change in deformation is the most important impurity effect in $sd$-shell hypernuclei. A similar conclusion has been reached also in Ref. [@Yao11].
Summary
=======
We have presented the microscopic particle-rotor model for the low-lying states of single-$\Lambda$ hypernuclei. In this formalism, the wave functions for hypernuclei are constructed by coupling the $\Lambda$ hyperon to the low-lying states of the core nucleus. Applying this method to $^{13}_{~\Lambda}$C, we have well reproduced the experimental energy spectrum of this hypernucleus. We have also found that the deformation is reduced by adding a $\Lambda$ particle in the positive-parity states, leading to a reduction in the $B(E2)$ value from the first $2^+$ to the ground states in the core nucleus.
In this paper, for simplicity, we have assumed the axial deformation for the core nucleus. An obvious extension of our method is to take into account the triaxial deformation of the core nucleus. One interesting application for this is $^{25}_{~\Lambda}$Mg, for which the triaxial degree of freedom has been shown to be important in the core nucleus $^{24}$Mg.
Acknowledgment {#acknowledgment .unnumbered}
==============
This work was supported by JSPS KAKENHI Grant Number 2640263 and the NSFC under Grant Nos. 11575148 and 11305134.
[99]{}
O. Hashimoto and H. Tamura, Prog. Part. Nucl. Phys. **57**, 564 (2006).
T. Motoba, H. Bandō, and K. Ikeda, Prog. Theor. Phys. **70**, 189 (1983).
E. Hiyama, M. Kamimura, K. Miyazaki, and T. Motoba, Phys. Rev. C **59**, 2351 (1999).
D. J. Millener, Nucl. Phys. [**A804**]{}, 84 (2008); [**A914**]{}, 109 (2013).
R. Wirth [*et al.*]{}, Phys. Rev. Lett. **113**, 192502 (2014).
M. Isaka, M. Kimura, A. Doté and A. Ohnishi, Phys. Rev. C **83**, 044323 (2011).
X. R. Zhou , H.-J. Schulze, H. Sagawa, C. X. Wu, and E.-G. Zhao, Phys. Rev. C **76**, 034312 (2007).
M. T. Win and K. Hagino, Phys. Rev. C **78**, 054311 (2008).
Myaing Thi Win, K. Hagino, and T. Koike, Phys. Rev. C **83**, 014301 (2011).
B.-N. Lu, E.-G. Zhao, and S.-G. Zhou, Phys. Rev. C **84**, 014328 (2011).
M. Bender, P.-H. Heenen, and P.-G. Reinhard, Rev. Mod. Phys. [**57**]{}, 121 (2003).
J.M. Yao, K. Hagino, Z.P. Li, J. Meng, and P. Ring, Phys. Rev. C[**89**]{}, 054306 (2014).
H. Mei, K. Hagino, J.M. Yao, and T. Motoba, Phys. Rev. C[**90**]{}, 064302 (2014).
H. Mei, K. Hagino, J.M. Yao, and T. Motoba, Phys. Rev. C[**91**]{}, 064305 (2015).
H. Mei, K. Hagino, and J.M. Yao, arXiv:1511.02957 \[nucl-th\].
W.X. Xue, J.M. Yao, K. Hagino, Z.P. Li, H. Mei, and Y. Tanimura, Phys. Rev. C[**91**]{}, 024327 (2015).
J.-W. Cui, X.-R. Zhou, and H.-J. Schulze, Phys. Rev. C[**91**]{}, 054306 (2015).
J.M. Yao, Z.P. Li, K. Hagino, M. Thi Win, Y. Zhang, and J. Meng, Nucl. Phys. [**A868**]{}, 12 (2011).
T. Burvenich, D. G. Madland, J. A. Maruhn, and P.-G. Reinhard, Phys. Rev. C **65**, 044308 (2002).
S. Ajimura [*et al.*]{}, Phys. Rev. Lett. [**86**]{}, 4255 (2001).
| {
"pile_set_name": "ArXiv"
} |
---
author:
- |
Chia Min Leong$^1$, YaeEun Moon$^1$, Vicki Wu$^1$, Timothy Wei$^1$ and\
Steve Peters$^2$\
\
$^1$ Rensselaer Polytechnic Institute, Troy, NY 12180, USA\
$^2$ USA Bobsled & Skeleton Federation
title: 'DPIV Measurements of Olympic Skeleton Athletes (Fluid Dynamics Video)'
---
The Olympic sport of skeleton involves an athlete riding a small sled face first down a bobsled track at speeds up to 130 km/hr. In these races, the difference between gold and missing the medal stand altogether can be hundredths of a second per run. As such, reducing aerodynamic drag through proper body positioning is of first order importance. To better study the flow behavior and to improve the performance of the athletes, we constructed a mock section of a bobsled track which was positioned at the exit of an open loop wind tunnel. DPIV measurements were made along with video recordings of body position to aid the athletes in determining their optimal aerodynamic body position. In the fluid dynamics video shown, the athlete slowly raised his head while DPIV measurements were made behind the helmet in the separated flow region.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Geometric entanglement (GE), as a measure of multipartite entanglement, has been investigated as a universal tool to detect phase transitions in quantum many-body lattice models. In this paper we outline a systematic method to compute GE for two-dimensional (2D) quantum many-body lattice models based on the translational invariant structure of infinite projected entangled pair state (iPEPS) representations. By employing this method, the $q$-state quantum Potts model on the square lattice with $q \in \{2,3,4,5\}$ is investigated as a prototypical example. Further, we have explored three 2D Heisenberg models: the antiferromagnetic spin-1/2 XXX and anisotropic XYX models in an external magnetic field, and the antiferromagnetic spin-1 XXZ model. We find that continuous GE does not guarantee a continuous phase transition across a phase transition point. We observe and thus classify three different types of continuous GE across a phase transition point: (i) GE is continuous with maximum value at the transition point and the phase transition is continuous, (ii) GE is continuous with maximum value at the transition point but the phase transition is discontinuous, and (iii) GE is continuous with non-maximum value at the transition point and the phase transition is continuous. For the models under consideration we find that the second and the third types are related to a point of dual symmetry and a fully polarized phase, respectively.'
author:
- 'Qian-Qian Shi'
- 'Hong-Lei Wang'
- 'Sheng-Hao Li'
- Sam Young Cho
- 'Murray T. Batchelor'
- 'Huan-Qiang Zhou'
title: |
Geometric entanglement and quantum phase transitions\
in two-dimensional quantum lattice models
---
Introduction
============
Of the various measures of entanglement in quantum many-body systems [@review], geometric entanglement is a measure of the multipartite entanglement contained in a pure state. More precisely stated, the geometric entanglement (GE) quantifies the distance between a given quantum state wavefunction and the closest separable (unentangled) state [@review; @hba; @tcw]. GE has been shown to serve as an alternative marker to locate critical points for quantum many-body lattice systems undergoing quantum phase transitions. This was demonstrated explicitly for a number of one-dimensional (1D) quantum systems [@wei; @ow] for which the GE diverges near the critical point with an amplitude proportional to the central charge of the underlying conformal field theory at criticality [@central; @scopy]. Moreover, for 1D quantum systems at criticality, the leading finite-size correction to the GE per lattice site is universal [@qqs; @jean; @hu], and related to the Affleck-Ludwig boundary entropy [@AL]. For the Lipkin-Meshkov-Glick model, a system of mutually interacting spins embedded in a magnetic field for which analytic results can be derived, the global GE of the ground state and the single-copy entanglement behave as the entanglement entropy close to and at criticality [@LMG].
GE thus serves as a useful tool to investigate quantum criticality in quantum many-body lattice systems. Apart from some exceptions [@hl; @gehlw; @thesis; @roman; @z3pott], almost all work to date exploiting GE to study phase transitions has been restricted to quantum systems in 1D. This is mainly due to the difficulty to compute GE, because it involves a formidable optimization over all possible separable states. Indeed, the calculation of various entanglement measures has been shown recently to be NP-complete [@Ionnou; @Huang]. This is further compounded by the inherent computational difficulties posed by two-dimensional (2D) quantum systems. Nevertheless, significant progress has been made to develop efficient numerical algorithms to simulate 2D quantum many-body lattice systems in the context of tensor network representations [@verstraete; @vidal; @gv2d; @jwx; @bvt; @rosgv; @pwe; @sz; @hco; @wks; @orusreview; @tn_latest]. The algorithms have been successfully exploited to compute, for example, the ground-state fidelity per lattice site [@zhou; @zov; @jhz; @whl; @lsh], which has been established as a universal marker to detect quantum phase transitions in many-body lattice systems. Indeed, the ground state fidelity per lattice site is closely related to the GE. Therefore, it is natural to expect that there should be an efficient way to compute the GE in the context of tensor network algorithms. This has been achieved for quantum many-body lattice systems with periodic boundary conditions in one spatial dimension in the context of the matrix product state representation [@hu].
Quantum phase transitions in 2D quantum lattice models can be investigated using a number of different physical quantities, including local order parameters, fidelity per lattice site, single-copy entropy and multi-partite entropy measurement, and GE per lattice site. As is well known, local order parameters are defined relating to the spontaneous symmetry breaking of some symmetry group, which results in degeneracy of ground states. Different degenerate ground states can be distinguished from different values of the local order parameters, and continuous and discontinuous phase transitions can be identified from the continuous and discontinuous behavior of the local order parameters. Although demonstrated to be capable of detecting phase transitions, it is not entirely clear if GE can distinguish different degenerate ground states and identify both continuous and discontinuous quantum phase transitions.
To address this issue, we improve a systematic method [@gehlw] to efficiently compute the GE per lattice site for 2D quantum many-body lattice systems in the context of tensor network algorithms based on an infinite projected entangled pair state (iPEPS) representation. This method is used to evaluate the GE per lattice site for a number of distinct models defined on the infinite square lattice. These models are (i) the quantum Ising model in a transverse field, (ii) the $q$-state quantum Potts model with $q=3, 4$ and $5$, (iii) the spin-$\frac12$ antiferromagnetic XXX model in an external magnetic field, (iv) the spin-$\frac12$ antiferromagnetic XYX model in an external magnetic field, and (v) the spin-$1$ XXZ model. By comparing the behavior of the local order parameters and the GE per site for the different models, it is seen that GE can detect both continuous and discontinuous quantum phase transitions. However, we observe that the continuity of the GE per site does not necessarily match with the continuity of local order parameters.
This paper is arranged as follows. The definition of GE is given in Section II. Results using the GE per lattice site as a marker of quantum phase transitions in the various 2D quantum models on the infinite square lattice are given in Section III using the procedure outlined in Appendices A and B based on the iPEPS representation. Discussion of the results and concluding remarks are given in Section IV.
The geometric entanglement per lattice site
===========================================
For a pure quantum state $|\psi\rangle$ with $N$ parties, the GE, as a global measure of the multipartite entanglement, quantifies the deviation from the closest separable state $|\phi\rangle$. For a spin system each party could be a single spin but could also be a block of contiguous spins. The GE $E(|\psi\rangle)$ for an $N$-partite quantum state $|\psi\rangle$ is expressed as [@review; @hba; @tcw] $$E(|\psi\rangle)=-\log_2{\Lambda_{{\rm max}}^{2}}, \label{def}$$ where $\Lambda_{\rm max}$ is the maximum fidelity between $|\psi\rangle$ and all possible separable (unentangled) and normalized states $|\phi\rangle$, with $$\Lambda_{{\rm max} }={\rm \max_{|\phi\rangle }} \; |\langle\psi|\phi\rangle|.$$ The GE per party ${\mathcal E}_{N}(|\psi\rangle)$ is then defined as $${\mathcal E}_{N}(|\psi\rangle)=N^{-1}E(|\psi\rangle).$$ It corresponds to the maximum fidelity per party $\lambda^{{\rm
max}}_{N}$, where $$\lambda^{{\rm max}}_{N}=\sqrt[N]{{\Lambda_{{\rm max}}}},$$ or equivalently, $${\mathcal E}_{N}(|\psi\rangle)=-\log_2{{(\lambda^{{\rm
max}}_{N})}^{2}}.$$ The relation (\[def\]) is analogous to the relation between the free energy and the partition function. Note that for unentangled states the GE is zero.
For our purpose, we shall consider a quantum many-body system on an infinite-size square lattice, which undergoes a quantum phase transition at a critical point in the thermodynamic limit. In this situation, each lattice site constitutes a party, thus the GE per party is the GE per lattice site, which is well defined in the thermodynamic limit ($N \rightarrow \infty$), since the contribution to the fidelity from each party (site) is multiplicative. In the infinite size limit we thus denote $${\mathcal E}=\lim_{N\to\infty} {\mathcal E}_{N}$$ as the GE per site.
geometric entanglement in infinite square lattice systems
=========================================================
In this section we provide an analysis from the perspective of GE of different 2D quantum spin systems undergoing different types of quantum phase transitions. The GE per site has previously been applied to ground states of a number of 1D spin chains [@wei; @ow; @central; @scopy; @qqs; @jean; @hu; @hl] and some 2D models [@hl; @gehlw; @thesis; @roman; @z3pott] across primarily continuous phase transitions. Our analysis here extends these previous studies to a wider range of 2D quantum models and with more exotic situations. Previously, the GE of ground states for the quantum Ising model in a transverse field and the XYX model on the square lattice have been investigated with TRG [@hl] and iPEPS [@gehlw] approaches, both of which show a continuous behavior in the GE corresponding to a continuous phase transition. Our results are based on the application and improvement of an iPEPS algorithm for the calculation of the GE of 2D quantum systems [@gehlw], as detailed in Appendices A and B. Preliminary results along this direction were also reported in Ref. [@thesis]. The size of the truncation dimension $D$ controls the underlying accuracy of the iPEPS algorithm. For the purposes of this study, which focusses on the general behavior of the GE in the vicinity of quantum critical points, we consider sufficiently large values of $D$, compared to Refs. [@gehlw; @thesis], to be confident of the observed behavior. Refinements in the implementation of the basic iPEPS algorithm have been discussed recently [@tn_latest].
2D quantum Ising model in a transverse field
--------------------------------------------
We consider the hamiltonian $$H = -\sum_{( \vec r, \vec r')}
S^{[\vec r]}_x S^{[\vec r']}_x - \lambda \sum_{\vec r} S^{[\vec r]}_z,$$ where here and in later subsections $S^{[\vec r]}_{\alpha}(\alpha=x,y,z)$ are spin-$\frac12$ Pauli operators acting at site $\vec r$, with $( \vec r, \vec r')$ running over all possible nearest-neighbor pairs on the square lattice. The parameter $\lambda$ measures the strength of the transverse magnetic field, with a phase transition from a ferromagnetic phase with two degenerate ground states to a paramagnetic phase with a single ground state occurring at the critical value $\lambda_c$. This model has been widely studied via a number of different techniques. Of particular relevance here are the previous calculation of the GE using the tensor product state and tensor renormalization group approach [@hl] and iPEPS [@gehlw; @thesis].
![(color online) Numerical values for (a) the local order parameter $\langle S_x\rangle$ and (b) the GE per site ${\mathcal E}_{\infty}(\lambda)$ for the 2D quantum Ising model as a function of the transverse field strength $\lambda$ for the indicated values of the iPEPS truncation dimension $D$. []{data-label="fig1"}](q2a.eps "fig:"){width="0.6\linewidth"} ![(color online) Numerical values for (a) the local order parameter $\langle S_x\rangle$ and (b) the GE per site ${\mathcal E}_{\infty}(\lambda)$ for the 2D quantum Ising model as a function of the transverse field strength $\lambda$ for the indicated values of the iPEPS truncation dimension $D$. []{data-label="fig1"}](q2b.eps "fig:"){width="0.7\linewidth"}
For different values of the truncation dimension $D$, when $\lambda<\lambda_c(D)$, spontaneous symmetry breaking occurs with $S_x$ the unitary element of the broken $Z_2$ symmetry group. Thus two degenerate ground states are detected and distinguished by the sign of the local order parameter $\langle S_x\rangle$, with the amplitude associated with each of the two degenerate ground states having the same value. The order parameter $\langle S_x\rangle$ is shown in Fig. \[fig1\](a). Phase transition points are identified at the values shown in Table I. Near these points the amplitude of $\langle S_x\rangle$ is observed to continuously approach zero, indicative of a continuous phase transition.
On the other hand, the estimates for the GE per lattice site for the ground states are shown in Fig. \[fig1\](b). For each value of $D$ the GE curve has a maximal value at $\lambda_c(D)$, which indicate the phase transition points. These values, shown in Table I, agree with the values detected with the local order parameter. Our results are consistent with the previous studies [@hl; @gehlw; @thesis]. The GE is seen to be continuous near the maximal points. The cusp-like behavior at the critical point is a characteristic feature of the GE at a continuous phase transition. The two degenerate ground states for $\lambda < \lambda_c$ can also be distinguished via the initial random states.
$D$ $\lambda_c(D)$
--- ------- ----- ---------------- --
$q=2$ 2 3.28
6 3.235
8 3.23
$q=3$ 3 2.620
6 2.616
9 2.616
$q=4$ 4 2.430
8 2.428
10 2.426
$q=5$ 5 2.330
8 2.326
10 2.326
: \[table\] Estimates of critical points $\lambda_c$ for the $q$-state quantum Potts model on the square lattice. The values for $\lambda_c(D)$ follow from the behavior of the local order parameter and the GE per site, where $D$ is the iPEPS truncation dimension.
2D $q$-state quantum Potts model
--------------------------------
The quantum Ising model discussed above is the special $q=2$ case of the more general $q$-state quantum Potts model defined on the square lattice. For a regular lattice, classical mean-field solutions [@Wu] and extensive computations (see, e.g., Refs. [@Wu; @dyw] and references therein) have suggested that the 3D classical $q$-state Potts model, and thus the 2D $q$-state quantum version, undergo a continuous phase transition for $q \le 2$ and a discontinuous phase transition for $q > 2$. We now turn to this 2D $q$-state quantum model and examine the GE per site and local order parameters in the vicinity of the phase transition points for the values $q=3,4$ and $5$.
On the square lattice the hamiltonian can be written in the form $$H =
-\sum_{( \vec r, \vec r')}
\left( \sum_{p=1}^{q-1} M^{[\vec r]}_{x,p} M^{[\vec r']}_{x,q-p} \right)
-\lambda \sum_{\vec r} M^{[\vec r]}_z,$$ where $M^{[\vec r]}_{x, p}$ and $M^{[\vec r]}_{z, p}$, with $p = 1,\ldots,q-1$, are spin matrices of size $q \times q$ acting at site $\vec r$. The parameter $\lambda$ is the analogue of the transverse magnetic field in the Ising case. In this formulation the spin matrices acting at each site are given by [@qPotts] $$M_{x,1} = \left(\begin{array}{cc} 0 & I_{q-1} \\ 1 & 0 \end{array} \right), \quad
M_z = \left( \begin{array}{cc} q-1 & 0 \\ 0 & -I_{q-1} \end{array}
\right),$$ where $I_{q-1}$ is the $(q-1)\times (q-1)$ identity matrix with $M_{x,p} = \left(M_{x,1}\right)^p$ and $\left(M_{x,1}\right)^q = I_q$.
![(color online) Numerical values for the local order parameter $\langle M_{x_1}\rangle$ and the GE per site ${\mathcal E}(\lambda)$ for the 2D quantum $q$-state Potts model for (a) $q=3$, (b) $q=4$ and (c) $q=5$ as a function of the transverse field strength $\lambda$ for the indicated values of the iPEPS truncation dimension $D$. For each case the curves for both the local order parameter and the GE show a discontinuous behavior, indicative of a discontinuous phase transition. []{data-label="fig2"}](q3.eps "fig:"){width="1.0\linewidth"} ![(color online) Numerical values for the local order parameter $\langle M_{x_1}\rangle$ and the GE per site ${\mathcal E}(\lambda)$ for the 2D quantum $q$-state Potts model for (a) $q=3$, (b) $q=4$ and (c) $q=5$ as a function of the transverse field strength $\lambda$ for the indicated values of the iPEPS truncation dimension $D$. For each case the curves for both the local order parameter and the GE show a discontinuous behavior, indicative of a discontinuous phase transition. []{data-label="fig2"}](q4.eps "fig:"){width="1.0\linewidth"} ![(color online) Numerical values for the local order parameter $\langle M_{x_1}\rangle$ and the GE per site ${\mathcal E}(\lambda)$ for the 2D quantum $q$-state Potts model for (a) $q=3$, (b) $q=4$ and (c) $q=5$ as a function of the transverse field strength $\lambda$ for the indicated values of the iPEPS truncation dimension $D$. For each case the curves for both the local order parameter and the GE show a discontinuous behavior, indicative of a discontinuous phase transition. []{data-label="fig2"}](q5.eps "fig:"){width="1.0\linewidth"}
Denoting the phase transition point by $\lambda_c$ we expect to detect $q$-degenerate ground states for $\lambda<\lambda_c$, corresponding to a $Z_q$ broken symmetry phase. The local order parameter $\langle M_{x,1}\rangle$ can distinguish the different degenerate ground states, but with the same amplitude for each of the $q$ ground states. As for the $q=2$ case, the phase transition point $\lambda_c(D)$ is estimated with increasing truncation dimension $D$. Fig. \[fig2\](a) shows the amplitude of the local order parameter $\langle M_{x,1}\rangle$ for the 3-state Potts model. This plot shows a jump in the curve which indicates a possible discontinuous phase transition point. The successive estimates for $\lambda_c$ are given in Table I [@ZN]. This same type of discontinuous behavior is seen for the local order parameter $\langle M_{x,1}\rangle$ in Fig. \[fig2\](b) for $q=4$ and Fig. \[fig2\](c) for $q=5$. The successive estimates for $\lambda_c$ are given in Table I.
The GE of the ground states is also shown in Fig. \[fig2\]. For each of the values of $q$, a maximal value is detected for the GE curve, where a jump also occurs. The estimates for the transition points $\lambda_c$ are well matched with those obtained via the local order parameters and also via the observed multi-bifurcation points in the magnetization [@dyw]. It is clear that the measure of GE can distinguish between discontinuous and continuous phase transitions in the 2D quantum $q$-state Potts model. To further test the utility of this approach we turn now to the investigation of GE in other 2D quantum models.
2D spin-$\frac12$ XXX model in a magnetic field
-----------------------------------------------
The spin-$\frac12$ antiferromagnetic XXX model on the square lattice has hamiltonian $$H =
\sum_{( \vec r, \vec r')} \left( S^{[\vec r]}_x S^{[\vec r']}_x
+ S^{[\vec r]}_y S^{[\vec r']}_y
+ S^{[\vec r]}_z S^{[\vec r']}_z \right)-h \sum_{\vec r} S^{[\vec r]}_z,$$ where $h$ is an external magnetic field along the $z$ direction. This model has been studied via a barrage of different techniques [@sandvik]. Whereas magnetic order is normally ruled out in 1D Heisenberg models, this is not the case for 2D Heisenberg models [@mermin; @sandvik]. Thus in 2D the ground state can be non-magnetic, i.e., with a non-zero magnetization. In the absence of the magnetic field the ground state of the Heisenberg model has antiferromagnetic (Néel) order with a non-zero local staggered magnetization and infinitely degenerate ground states resulting from the breaking of global SU$(2)$ spin rotation symmetry. For $h\neq0$, infinitely degenerate ground states are detected resulting from the spontaneous symmetry breaking of U$(1)$ symmetry in the $x$-$y$ plane. As $h\rightarrow \infty$ it is anticipated the system will be fully polarized in the $z$ direction. In fact this transition to the fully polarized state occurs at $h=4$.
In Fig. \[fig3\](a) we show the local magnetizations for sublattices A and B defined by $m_{\alpha}^A=\langle S_{\alpha}^A\rangle$ and $m_{\alpha}^B=\langle S_{\alpha}^B\rangle$ for $\alpha=x,y,z$. For $h<4$, $m_\alpha^A = -m_\alpha^B$ for $\alpha=x,y$ with $m_z^A=m_z^B$. For $h>4$, $m_\alpha^A = m_\alpha^B = 0$ for $\alpha=x,y$ with $m_z^A=m_z^B =\frac12$. The latter is indicative of the fully polarized state.
In Fig. \[fig3\](b) we show the staggered magnetization parameters $M_{xy}=\sqrt{\langle \frac12(S_x^A-S_x^B)\rangle^2+\langle \frac12(S_y^A-S_y^B)\rangle^2}$ and $M_z=\langle \frac12(S_z^A+S_z^B)\rangle$ as functions of the magnetic field. The local order parameter $M_{xy}$ decreases with increasing $h$, with $M_{xy} \to 0$ when $h>4$. On the other hand, $M_z$ increases with increasing $h$, with $M_{z} \to \frac12$ for $h>4$, which corresponds to a continuous phase transition from a Néel phase to the fully polarized phase. As mentioned previously, for $h<4$, infinitely degenerate ground states exist corresponding to the $x$-$y$ plane U$(1)$ symmetry breaking, which is indicated from the random-like magnetization on the sublattices. As for the previous models, the numerical results indicate that the different degenerate ground states for fixed $h$ give the same value for the GE per site, as Fig. \[fig3\](c) shows. In Fig. \[fig3\](c), the GE per site clearly decreases with increasing $h$, indicating decreasing entanglement as $h\to 4$. In fact, zero GE per site indicates a factorizing field, in this case the simple polarized state. This factorizing field is discussed further below in the context of the more general $XYX$ model.
In contrast to the previous models studied here, the GE is not maximal at the phase transition point, because the phase transition point is not the critical point which would correspond to maximum entanglement. In this case zero or nonzero GE per site distinguishes between the two different phases in the model.
![(color online) Plots for the 2D spin-$\frac12$ XXX model with varying magnetic field $h$. (a) Local magnetization $m$ for sublattices A and B, obtained with iPEPS truncation dimension $D=2$. (b) Amplitude of the local order parameters $M$. (c) GE per site ${\mathcal E}(h)$. []{data-label="fig3"}](delta10a.eps "fig:"){width="0.8\linewidth"} ![(color online) Plots for the 2D spin-$\frac12$ XXX model with varying magnetic field $h$. (a) Local magnetization $m$ for sublattices A and B, obtained with iPEPS truncation dimension $D=2$. (b) Amplitude of the local order parameters $M$. (c) GE per site ${\mathcal E}(h)$. []{data-label="fig3"}](delta10b.eps "fig:"){width="0.8\linewidth"} ![(color online) Plots for the 2D spin-$\frac12$ XXX model with varying magnetic field $h$. (a) Local magnetization $m$ for sublattices A and B, obtained with iPEPS truncation dimension $D=2$. (b) Amplitude of the local order parameters $M$. (c) GE per site ${\mathcal E}(h)$. []{data-label="fig3"}](delta10c.eps "fig:"){width="0.8\linewidth"}
2D spin-$\frac12$ XYX model in a magnetic field
-----------------------------------------------
Tuning the anisotropy of the Heisenberg interactions leads to the more general spin-$\frac12$ XYX antiferromagnetic model in a uniform magnetic field, defined on the square lattice by the hamiltonian $$H =
\sum_{( \vec r, \vec r')} \left( S^{[\vec r]}_x S^{[\vec r']}_x
+\Delta_{y} S^{[\vec r]}_y S^{[\vec r']}_y
+S^{[\vec r]}_z S^{[\vec r']}_z \right)
-h \sum_{\vec r} S^{[\vec r]}_z.$$ Varying the anisotropic exchange interaction parameter $\Delta_y$ leads to different behavior, with the two cases $\Delta_{y}<1$ and $\Delta_{y}>1$ corresponding to an easy-plane and easy-axis behavior, respectively. The quantum criticality of this model is well understood [@tros], with an ordered phase below a critical field value $h_c$, above which is a partially polarized state with field-induced magnetization reaching saturation as $h \to \infty$. The ordered phase in the easy-plane (easy-axis) case arises by spontaneous symmetry breaking along the $x$ ($y$) direction, which corresponds to a finite value of the order parameter $M_x$ ($M_y$) below $h_c$. At the transition point $h_c$, long-range correlations are destroyed.
Previous studies of this model using GE [@gehlw; @hl] focussed on the value $\Delta_{y}=0.25$ with the magnetic field as control parameter. In this easy-plane region the model is known to undergo a continuous quantum phase transition in the same universality class as the transverse Ising model [@tros]. The GE was seen to have a cusp at the critical value $h_c$ [@gehlw; @hl] as observed for the transverse Ising model (recall Fig. \[fig1\](b)). Significantly, it is known that a factorizing field exists at the value $h_f = 2 \sqrt{2(1+\Delta_y)}$, with $h_f < h_c$, where the ground state becomes a separable product state [@tros]. At this point it follows that the GE vanishes, i.e., ${\mathcal E}_{\infty}(h_f) = 0$. This was observed in the simulations using GE [@gehlw; @hl]. Other entanglement measures were also confirmed to vanish at this point [@tros; @lsh].
In this study we consider the fixed parameter value $h=0.25$ and vary the coupling $\Delta_{y}$. The model is anticipated to undergo a transition at $\Delta_{y}=1$ from an antiferromagnetic phase in the $x$ direction to an antiferromagnetic phase in the $y$ direction, with local order parameter the staggered magnetization $M_x$ ($M_y$) corresponding to the phase $\Delta_{y}<1$ ($\Delta_{y}>1$).
![(color online) Plots for the spin-$\frac12$ XYX model with fixed magnetic field $h=0.25$ and varying anisotropy parameter $\Delta_y$. (a) Components of the local magnetization $m$ for sublattices A and B for truncation dimension $D=6$. (b) Components of the staggered magnetizations $M$ for truncation dimension $D=6$. (c) GE per site ${\mathcal E}(\lambda)$ with truncation dimension $D$.[]{data-label="figXYX"}](h0.250a.eps "fig:"){width="0.75\linewidth"} ![(color online) Plots for the spin-$\frac12$ XYX model with fixed magnetic field $h=0.25$ and varying anisotropy parameter $\Delta_y$. (a) Components of the local magnetization $m$ for sublattices A and B for truncation dimension $D=6$. (b) Components of the staggered magnetizations $M$ for truncation dimension $D=6$. (c) GE per site ${\mathcal E}(\lambda)$ with truncation dimension $D$.[]{data-label="figXYX"}](h0.250b.eps "fig:"){width="0.7\linewidth"} ![(color online) Plots for the spin-$\frac12$ XYX model with fixed magnetic field $h=0.25$ and varying anisotropy parameter $\Delta_y$. (a) Components of the local magnetization $m$ for sublattices A and B for truncation dimension $D=6$. (b) Components of the staggered magnetizations $M$ for truncation dimension $D=6$. (c) GE per site ${\mathcal E}(\lambda)$ with truncation dimension $D$.[]{data-label="figXYX"}](h0.250c.eps "fig:"){width="0.7\linewidth"}
Fig. \[figXYX\](a) shows the components of the local magnetization for sublattices A and B defined by $m_{\alpha}^A=\langle S_{\alpha}^A\rangle$ and $m_{\alpha}^B=\langle S_{\alpha}^B\rangle$ for $\alpha=x,y,z$. For $\Delta_y<1$, $m_x^A$ and $m_x^B$ have opposite values, with $m_y^A=m_y^B=0$ and $m_z^A=m_z^B$. This implies that a staggered magnetization $M_x=\langle \frac12 (S_x^A-S_x^B)\rangle$ exists. For $\Delta_y>1$, $m_y^A$ and $m_y^B$ have opposite values, with $m_x^A=m_x^B=0$ and $m_z^A=m_z^B$, implying the staggered magnetization $M_y=\langle \frac12 (S_y^A-S_y^B)\rangle$. These magnetizations are shown in Fig. \[figXYX\](b) for the two different phases for truncation dimension $D=6$. There is a clear jump discontinuity at $\Delta_y=1$ for each of the staggered magnetizations $M_x$ and $M_y$. This behavior persists with increasing truncation dimension $D$, indicating a discontinuous phase transition at $\Delta_y=1$.
The GE per site is shown for the same parameter range in Fig. \[figXYX\](c) with increasing truncation dimension. The characteristic cusp occurs as $\Delta_{y}$ varies across the critical point $\Delta_{y}=1$. In contrast to the local order parameter, which is discontinuous, the GE is continuous. In this case the GE thus does not detect the discontinuous behavior at the phase transition point.
2D spin-$1$ XXZ model
---------------------
The spin-$1$ XXZ model is defined on the square lattice by the hamiltonian $$H =
\sum_{( \vec r, \vec r')} \left( S^{[\vec r]}_x S^{[\vec r']}_x
+ S^{[\vec r]}_y S^{[\vec r']}_y
+ \Delta S^{[\vec r]}_z S^{[\vec r']}_z \right),$$ where now $S^{[\vec r]}_{\alpha} (\alpha=x,y,z)$ are spin-$1$ operators at site $\vec r$, with again the summation running over all nearest neighbor pairs on the square lattice. For anisotropic exchange interaction parameter $\Delta=1$ the hamiltonian has $SU(2)$ symmetry and the ground state is infinitely degenerate. For this model there is a quantum phase transition at $\Delta=1$ [@spin1]. This phase transition is clearly marked in Fig. \[figXXZ\_1\], which shows plots of the single-copy entanglement for the spin-$1$ XXZ model.
We also calculate the components of the local magnetization $m$ for sublattices $A$ and $B$: $m_{\alpha}^A=\langle S_{\alpha}^A\rangle$ and $m_{\alpha}^B=\langle S_{\alpha}^B\rangle$ for $\alpha=x,y,z$. These are shown in Fig. \[figXXZ\_2\](a) as a function of $\Delta$. For $\Delta<1$, the values of $m_{x}$ and $m_{y}$ have opposite signs on each sublattice, with $m_z^A=m_z^B$. For $\Delta>1$, $m_{x}=m_{y}=0$ on each sublattice, with $m_z^A=-m_z^B$. This indicates a local order with staggered magnetization in the $z$ phase characterized by $
M_{z}=\langle \frac12 (S_z^A-S_z^B)\rangle.
$ For different values of magnetic field strength $h$, the magnetization for the $x$ and $y$ directions does not change continuously for $\Delta<1$, while the staggered magnetization $
M_{xy}=\sqrt{\langle \frac12(S_x^A-S_x^B)\rangle^2+\langle \frac12(S_y^A-S_y^B)\rangle^2},
$ continuously changes with varying $\Delta$ (see Fig. \[figXXZ\_2\](b)). This indicates the existence of local order defined with the staggered magnetization in the $x$-$y$ plane.
Hence, for $\Delta<1$, the phase is characterized by the antiferromagnetic order parameter $M_{xy}$ in the $x$-$y$ easy plane. Infinitely degenerate ground states exist corresponding to the $x$-$y$ plane $U(1)$ symmetry breaking, indicated from the random-like magnetization of sublattices as shown in Fig. \[figXXZ\_2\](a). For $\Delta>1$, the phase is characterized by the antiferromagnetic order parameter $M_z$ along the easy axis, with doubly degenerate ground states corresponding to the spin-flip (or one-site translational invariant) $Z_2$ symmetry breaking. These order parameters are shown in Fig. \[figXXZ\_2\](b). For each truncation dimension, a distinct jump is detected at $\Delta=1$, indicating the phase transition is possibly discontinuous. The total antiferromagnetic order parameter $
M_{t}=\sqrt{\langle \frac12 (S_x^A-S_x^B)\rangle^2+\langle \frac12 (S_y^A-S_y^B)\rangle^2
+\langle \frac12 (S_z^A-S_z^B)\rangle^2}
$ is shown in Fig. \[figXXZ\_2\](c). It varies continuously.
For comparison we also consider the GE per site for this model. The GE per site is seen to take the same value for all of the degenerate ground states. As shown in Fig. \[figXXZ\_2\](d), the GE is continuous with a well defined cusp at the point $\Delta=1$, while the local order parameters in Fig. \[figXXZ\_2\](b) are discontinuous at $\Delta=1$. This behavior is similar to that observed in the previous subsection for the 2D spin-$\frac12$ XYX model with fixed magnetic field $h$.
![(color online) Single-copy entanglement (a) $S_1(\Delta)$ and (b) $S_2(\Delta)$ as a function of the anisotropy parameter $\Delta$ for the 2D spin-$1$ XXZ model with increasing iPEPS truncation dimension $D$. Here the single-copy entanglement is calculated based on the one-site and two-site reduced density matrix, obtained by using the iTEBD method.[]{data-label="figXXZ_1"}](xxzspin1sca.eps "fig:"){width="0.7\linewidth"} ![(color online) Single-copy entanglement (a) $S_1(\Delta)$ and (b) $S_2(\Delta)$ as a function of the anisotropy parameter $\Delta$ for the 2D spin-$1$ XXZ model with increasing iPEPS truncation dimension $D$. Here the single-copy entanglement is calculated based on the one-site and two-site reduced density matrix, obtained by using the iTEBD method.[]{data-label="figXXZ_1"}](xxzspin1scb.eps "fig:"){width="0.7\linewidth"}
![(color online) (a) Sublattice components of the local magnetization $m$, (b) amplitude of the local order parameters $M_{xy}$ and $M_z$, (c) total antiferromagnetic order parameter $M_t$ and (d) GE per lattice site ${\mathcal E}(\Delta)$ as a function of anisotropy parameter $\Delta$ for the 2D spin-$1$ XXZ model. Values of the iPEPS truncation dimension $D$ are as indicated, with $D=9$ for (a).[]{data-label="figXXZ_2"}](xxzspin10a.eps "fig:"){width="0.7\linewidth"} ![(color online) (a) Sublattice components of the local magnetization $m$, (b) amplitude of the local order parameters $M_{xy}$ and $M_z$, (c) total antiferromagnetic order parameter $M_t$ and (d) GE per lattice site ${\mathcal E}(\Delta)$ as a function of anisotropy parameter $\Delta$ for the 2D spin-$1$ XXZ model. Values of the iPEPS truncation dimension $D$ are as indicated, with $D=9$ for (a).[]{data-label="figXXZ_2"}](xxzspin10b.eps "fig:"){width="0.7\linewidth"} ![(color online) (a) Sublattice components of the local magnetization $m$, (b) amplitude of the local order parameters $M_{xy}$ and $M_z$, (c) total antiferromagnetic order parameter $M_t$ and (d) GE per lattice site ${\mathcal E}(\Delta)$ as a function of anisotropy parameter $\Delta$ for the 2D spin-$1$ XXZ model. Values of the iPEPS truncation dimension $D$ are as indicated, with $D=9$ for (a).[]{data-label="figXXZ_2"}](xxzspin10c.eps "fig:"){width="0.7\linewidth"} ![(color online) (a) Sublattice components of the local magnetization $m$, (b) amplitude of the local order parameters $M_{xy}$ and $M_z$, (c) total antiferromagnetic order parameter $M_t$ and (d) GE per lattice site ${\mathcal E}(\Delta)$ as a function of anisotropy parameter $\Delta$ for the 2D spin-$1$ XXZ model. Values of the iPEPS truncation dimension $D$ are as indicated, with $D=9$ for (a).[]{data-label="figXXZ_2"}](xxzspin10d.eps "fig:"){width="0.7\linewidth"}
summary and discussion
======================
We have demonstrated how to efficiently compute the GE of 2D quantum models defined on the infinite square lattice, by optimizing over all possible separable states, in the context of the tensor network algorithm based on the iPEPS representation. Our results, in line with previous studies [@hl; @gehlw; @thesis; @roman; @z3pott], demonstrate that GE is able to detect continuous quantum phase transitions and factorized fields for different 2D quantum systems. Continuous phase transitions are marked by a characteristic continuous cusp-like behavior of the GE per site at the critical point. This behavior is evident in Fig. \[fig1\](b) for the 2D quantum Ising model in a transverse field. Additionally we have demonstrated that the GE can detect discontinuous phase transitions in the $q$-state quantum Potts model for $q \ge 3$. For the values of $q$ considered, a maximal value is detected for the GE curve, where the discontinuous transition occurs (see Fig. \[fig2\]). It is thus clear that the measure of GE can distinguish between continuous and discontinuous phase transitions in the 2D quantum $q$-state Potts model.
This overall picture is not so simple, however. It has been demonstrated recently that quantum phase transitions should be treated with caution, at least with regard to the ground state entanglement spectrum [@CKS2014]. In the present study, we have seen an example where the GE is continuous with non-maximum value at the transition point and the phase transition is continuous. This is the case for the spin-$\case12$ XXX model, for which the GE is not maximal at the phase transition point $h=4$ at which the GE vanishes (see Fig. \[fig3\](c)). We have also seen examples where the GE per site shows continuous behavior at discontinuous phase transition points. For example, in the spin-$\case12$ XYX model and the spin-$1$ XXZ model, the GE per site shows continuous behavior with well defined cusps at the transition point (Fig. \[figXYX\](c) and Fig. \[figXXZ\_2\](d)) but the corresponding local staggered magnetizations (Fig. \[figXYX\](b) and Fig. \[figXXZ\_2\](b)) are discontinuous. We have also considered the single-copy entanglement $S_L$, in which the system under consideration is divided into two parts, one with $L$ lattice sites and the other with the other lattice sites. It is known that the single-copy entanglement $S_L$ sets a bound for the GE [@scopy], i.e., ${\cal E}<S_L=-\log_2{\mu_L^1}$, with $\mu_L^1$ the largest eigenvalue of the reduced density matrix $\rho_L$ for an $L$-site subsystem. It is found that the single-copy entanglements $S_1$ and $S_2$ show the same continuous behavior near the phase transition points for the models considered here, as shown, e.g., in Fig. \[figXXZ\_1\] for the spin-$1$ XXZ model. It appears then for such systems, for which discontinuous phase transitions occur corresponding to different types of symmetry breaking and the total magnetization is continuous, both multi-partite and bi-partite entanglement measures are continuous.
This situation can be further understood as follows. First, as already mentioned, due to the spontaneous symmetry breaking, including the spontaneous breaking of continuous symmetry, different degenerate ground states are detected in the broken symmetry phase by different values of the local order parameter. However, the GE takes the same value for all degenerate ground states. In terms of entanglement, this can be understood from the GE being a global measure of entanglement, which is the same for all degenerate ground states. From a different perspective, we can suppose the closest separable state to the ground state and the ground states share some property, i.e., we suppose the closest separable state of different degenerate ground states breaks the same symmetry of the ground states [^1].
Also from the perspective of symmetry, it could be argued for the spin-$\case12$ XYX model and the spin-$1$ XXZ model that the discontinuity due to the isometry cannot be detected with the GE measure because the hamiltonian/ground state obeys a dual symmetry either side of the phase transition point. For the spin-$\case12$ XYX model this symmetry is in $\Delta_y$, with $\Delta_y=1$ the “self-dual" point [^2]. Given such a duality of the hamiltonian, both the ground state and its closest separable state are also each dual under this transformation, so the fidelity between the ground state and its closest separable state also obeys the same relation, leading to continuous behavior across the phase transition point. Similarly for the spin-$1$ XXZ model, the hamiltonian is symmetric at $\Delta=1$, with $\Delta \neq 1$ breaking the symmetry into easy-plane and easy-axis. In this case such a duality in the total staggered magnetization of the easy-plane and easy-axis would imply a duality of a local property of the ground state wavefunction, leading to continuity in the GE.
We have seen then three different types of continuous GE across a phase transition point:
\(i) GE is continuous with maximum value at the transition point and the phase transition is continuous,
\(ii) GE is continuous with maximum value at the transition point but the phase transition is discontinuous, and
\(iii) GE is continuous with non-maximum value at the transition point and the phase transition is continuous.
For the models under consideration the second and third types are related to a point of dual symmetry and a fully polarized phase, respectively. Given this refinement in our understanding of GE as a marker of quantum phase transitions, and the development of powerful tensor network algorithms, we can be confident that GE can be used as an alternative route to explore quantum criticality in quantum lattice models.
[*Acknowledgements.*]{} MTB gratefully acknowledges support from Chongqing University and the 1000 Talents Program of China. This work is supported in part by the National Natural Science Foundation of China (Grant Numbers 11575037, 11374379 and 11174375).
The infinite projected entangled pair state algorithm
=====================================================
Our aim is to compute the GE per lattice site for a quantum many-body lattice system on an infinite-size square lattice in the context of the iPEPS algorithm [@gv2d]. Here we follow the presentation given in Ref. [@gehlw].
Suppose we consider a system characterized by a translation-invariant Hamiltonian $H$ with nearest-neighbor interactions: $H=\sum_{\langle ij \rangle} h_{\langle ij \rangle}$, with $h_{\langle ij \rangle}$ being the Hamiltonian density. Assume that a quantum wavefunction $|\psi\rangle$ is translation-invariant under two-site shifts, then one only needs two five-index tensors $A^s_{lrud}$ and $B^s_{lrud}$ to express the iPEPS representation. Here, each tensor is labeled by one physical index $s$ and four bond indices $l$, $r$, $u$ and $d$, as shown in Fig. \[FIG1\](i). Note that the physical index $s$ runs over $1,\ldots,\mathrm{d}$, and each bond index takes $1,\ldots,D$, with $\mathrm{d}$ being the physical dimension, and $D$ the bond dimension. Therefore, it is convenient to choose a $2\times 2$ plaquette as the unit cell (cf. Fig. \[FIG1\](ii)). The ground state wavefunction is well approximated by $|\psi_{\tau}\rangle$, which is obtained by performing an imaginary time evolution [@gv2d] from an initial state $|\psi_{0}\rangle$, with $|\psi_{\tau}\rangle=e^{-H\tau}|\psi_{0}\rangle/||e^{-H\tau}|\psi_{0}\rangle||$ [@gv2d], as long as $\tau$ is large enough.
A key ingredient of the iPEPS algorithm is to take advantage of the Trotter-Suzuki decomposition that allows to reduce the (imaginary) time evolution operator $e^{-H \delta\tau}$ over a time slice $\delta \tau$ into the product of a series of two-site operators, where the imaginary time interval $\tau$ is divided into $M$ slices: $\tau = M \delta \tau$. Therefore, the original global optimization problem becomes a local two-site optimization problem. With an efficient contraction scheme available to compute the effective environment for a pair of the tensors $A^s_{lrud}$ and $B^s_{lrud}$ [@gv2d], one is able to update the tensors $A^s_{lrud}$ and $B^s_{lrud}$. Performing this procedure until the energy per lattice site converges, the ground state wavefunction is produced in the iPEPS representation.
![(color online) (i) A five-index tensor $A^s_{lrud}$ labeled by one physical index $s$ and four bond indices $l$, $r$, $u$ and $d$. (ii) The iPEPS representation of a wavefunction on the square lattice. Copies of the tensors $A^s_{lrud}$ and $B^s_{lrud}$ are connected through four types of bonds. (iii) A one-index tensor $\tilde{A}^{s}$ labeled by one physical index $s$. (iv) The iPEPS representation of a separable state in the square lattice. (v) A reduced four-index tensor $a_{lrud}$ from a five-index tensor $A^s_{lrud}$ and a one-index $\tilde{A}^{s*}$. (vi) The tensor network representation for the fidelity between a quantum wavefunction (described by $A^s_{lrud}$ and $B^s_{lrud}$) and a separable state (described by $\tilde{A}^s$ and $\tilde{B}^s$), consisting of the reduced tensors $a_{lrud}$ and $b_{lrud}$. []{data-label="FIG1"}](peps){width="48.00000%"}
![(color online) Key ingredients for obtaining the gradient of the fidelity between a given ground state wavefunction $|\psi \rangle$ and a separable state $|\phi \rangle$ in the iPEPS representation. (i) Zero-dimensional transfer matrix $E_0$ and its dominant eigenvectors $V_{L}$ and $V_{R}$. Here the infinite matrix product state representation of the dominant eigenvectors for the one-dimensional transfer matrix $E_1$ follows from Ref. [@rosgv], and $V_{L}$ and $V_{R}$ may be evaluated using the Lanczos method. The contraction of the entire tensor network is the dominant eigenvalue $\eta_{\langle\phi|\psi\rangle}$ of the zero-dimensional transfer matrix $E_0$ for $\langle\phi|\psi\rangle$. (ii) A half-filled square denotes $a_{-}$, the derivative of the four-index tensor $a_{lrud}$ with respect to $\tilde{A}^{s*}$, which is nothing but the five-index tensor $A^s_{lrud}$. Similarly, we may define $b_{-}$, the derivative of the four-index tensor $b_{lrud}$ with respect to $\tilde{B}^{s*}$. (iii) and (iv) Pictorial representation of the contributions to the derivative of $\eta_{\langle\phi|\psi\rangle}$ with respect to $\tilde{A}^{s*}$, with different relative positions between filled circles and half-filled squares. []{data-label="FIG2"}](peps2){width="48.00000%"}
![(color online) Key ingredients in the diagonal contraction scheme for obtaining the gradient of the fidelity between a given ground state wavefunction $|\psi \rangle$ and a separable state $|\phi \rangle$ in the iPEPS representation. (i) Tensor network representation for the fidelity between a quantum wavefunction $|\psi \rangle$ and a separable state $|\phi \rangle$. (ii) Tensor network representation for a zero-dimensional transfer matrix $E_0$ and its dominant eigenvectors $V_{L}$ and $V_{R}$. Here the infinite matrix product state representation of the dominant eigenvectors for the one-dimensional transfer matrix $E_1$ follows from Ref. [@rosgv], and $V_{L}$ and $V_{R}$ are evaluated using the Lanczos method. The contraction of the entire tensor network is the dominant eigenvalue $\eta_{\langle\phi|\psi\rangle}$ of the zero-dimensional transfer matrix $E_0$ for $\langle\phi|\psi\rangle$. (iii) The pictorial representation of the contributions to the derivative of $\eta_{\langle\phi|\psi\rangle}$ with respect to $\tilde{A}^{s*}$, with different relative positions between filled circles and half-filled squares. []{data-label="diagnal"}](addtwo2){width="48.00000%"}
Efficient computation of the GE in the iPEPS representation
===========================================================
Once the iPEPS representation for the ground state wavefunction is generated, we are ready to evaluate the GE per lattice site. Here we begin by outlining the scheme developed in Ref. [@gehlw].
First, we need to compute the fidelity between the ground state wavefunction and a separable state. The latter is represented in terms of one-index tensors $\tilde A^s$ and $\tilde B^s$. To this end, we form a reduced four-index tensor $a_{lrud}$ from the five-index tensor $A^s_{lrud}$ and a one-index tensor $\tilde A^s$, as depicted in Fig. \[FIG1\](iii). As such, the fidelity is represented as a tensor network in terms of the reduced tensors $a_{lrud}$ and $b_{lrud}$ (cf. Fig. \[FIG1\](iv)).
The tensor network may be contracted as follows. First, form the 1D transfer matrix $E_1$, consisting of two consecutive rows of the tensors in the checkerboard tensor network. This is highlighted in Fig. \[FIG1\](vi) by the two dashed lines. Second, compute the dominant eigenvectors of the transfer matrix $E_1$, corresponding to the dominant eigenvalue. This can be done, following a procedure described in Ref. [@rosgv]. Here, the dominant eigenvectors are represented in the infinite matrix product states. Third, choose the zero-dimensional transfer matrix $E_0$ (Fig. \[FIG2\](ii)), and compute its dominant left and right eigenvectors, $V_{L}$ and $V_{R}$. This may be achieved by means of the Lanczos method. In addition, one also needs to compute the norms of the ground state wavefunction $|\psi\rangle$ and a separable state $|\phi\rangle$ from their iPEPS representations. Putting everything together, we are able to obtain the fidelity per unit cell between the ground state $|\psi\rangle$ and a separable state $|\phi\rangle$: $$\lambda=\frac{|\eta_{\langle\phi|\psi\rangle}|}
{\sqrt{\eta_{\langle\psi|\psi\rangle}\eta_{\langle\phi|\phi\rangle}}},$$ where $\eta_{\langle\phi|\psi\rangle}$, $\eta_{\langle\phi|\phi\rangle}$ and $\eta_{\langle\phi|\phi\rangle}$ are, respectively, the dominant eigenvalue of the zero-dimensional transfer matrix $E_0$ for the iPEPS representation of $\langle\phi|\psi\rangle$, $\langle\psi|\psi\rangle$ and $\langle\phi|\phi\rangle$.
We then proceed to compute the GE per site, which involves the optimization over all the separable states. For our purpose, we define $F=\lambda^{2}$. The optimization amounts to computing the logarithmic derivative of $F$ with respect to $\tilde{A}^{*}$, which is expressed as $$G \equiv \frac{\partial \ln F}{\partial
\tilde{A}^{*}}=\frac{1}{\eta_{\langle\phi|\psi\rangle}}\frac{\partial\eta_{\langle\phi|\psi\rangle}}{\partial
\tilde{A}^{*}}-\frac{1}{\eta_{\langle\phi|\phi\rangle}}\frac{\partial\eta_{\langle\phi|\phi\rangle}}{\partial
\tilde{A}^{*}}.$$
The problem therefore reduces to the computation of $G$ in the context of the tensor network representation. First, note that a pictorial representation of the derivative $\partial a_{lrud}/
\partial \tilde{A}^{s*}$ of the four-index tensor $a_{lrud}$ with respect to $\tilde{A}^{s*}$ is shown in Fig. \[FIG2\](ii), which is nothing but the five-index tensor $A^s_{lrud}$. Similarly, we may define the derivative of the four-index tensor $b_{lrud}$ with respect to $\tilde{B}^{s*}$. Then, we are able to represent the contributions to the derivative of $\eta_{\langle\phi|\psi\rangle}$ with respect to $\tilde{A}^{s*}$ in Fig. \[FIG2\](iii) and Fig. \[FIG2\](iv). In our scheme, we update the real and imaginary parts of $\tilde{A}^s$ separately: $$\begin{aligned}
\Re(\tilde{A}^s)&=&\Re(\tilde{A}^s)+\delta \Re(G)^{s},\\
\Im(\tilde{A}^s)&=&\Im(\tilde{A}^s)+\delta \Im(G)^{s}.\end{aligned}$$ Here $\delta\in[0,1)$ is the step size in the parameter space, which is tuned to be decreasing during the optimization process. In addition, we have normalized the real and imaginary parts of the gradient $G$ so that their respective largest entries are unity. The procedure to update the tensor $\tilde{B}^s$ is the same. If the fidelity per unit cell converges, then the closest separable state $|\phi\rangle$ is achieved, thus the geometric entanglement per lattice site for the ground state wavefunction $|\psi\rangle$ follows.
The above scheme has some basic limitations for the practical calculation of the GE per site. In particular, it can only be readily achieved for relatively small truncation dimensions. Larger truncation dimensions can be achieved by changing the transfer matrix direction in Fig. \[FIG1\](vi) to the diagonal direction, as shown in Fig. \[diagnal\](i). This leads to a simpler calculation of the derivatives necessary for the calculation of the GE (see Fig. \[diagnal\](ii) and Fig. \[diagnal\](iii)). For both schemes, the leading computational time scale during contraction is the same, namely $O(D^4)$. However, in the diagonal scheme, the computation time scale relative to truncation dimension $D$ can be reduced. For example, to calculate the largest eigenvalue $\eta_{\langle\phi|\psi\rangle}$ of $E_0$ the computation time for each Lanczos step can be reduced from $O(D^4)$ to $O(D^2)$ by storing tensors whose computation time scales as $O(D^4)$. Since the time cost to obtain $\eta_{\langle\phi|\psi\rangle}$ is mainly due to iterations in the Lanczos algorithm, this improvement allows us to deal with relatively larger truncation dimensions and thus obtain higher accuracy for the GE per site.
[10]{}
For a review, see L. Amico, R. Fazio, A. Osterloh and V. Vedral, Rev. Mod. Phys. **80**, 517 (2008).
H. Barnum and N. Linden, J. Phys. A **34**, 6787 (2001).
T.-C. Wei and P. M. Goldbart, Phys. Rev. A **68**, 042307 (2003).
T.-C. Wei, D. Das, S. Mukhopadyay, S. Vishveshwara and P.M. Goldbart, Phys. Rev. A **71**, 060305 (2005).
R. Orús and T.-C. Wei, Phys. Rev. B **82**, 155120 (2010); R. Orús and T.-C. Wei, Quantum Inf. Comput. **11**, 0563 (2011); R. Orús, T.-C. Wei and H.-H. Tu, Phys. Rev. B **84**, 064409 (2011).
A. Botero and B. Reznik, arXiv:0708.3391.
R. Orús, Phys. Rev. Lett. **100**, 130502 (2008).
Q.-Q. Shi, R. Orús, J. O. Fj[æ]{}restad, and H.-Q Zhou, New J. Phys. **12**, 025008 (2010).
J.-M. Stéphan, G. Misguich and F. Alet, Phys. Rev. B **82**, 180406R (2010).
B.-Q. Hu, X.-J. Liu, J.-H. Liu and H.-Q. Zhou, New J. Phys. **13**, 093041 (2011).
I. Affleck and A.W.W. Ludwig, Phys. Rev. Lett. **67**, 161 (1991).
R. Orús, S. Dusuel and J. Vidal, Phys. Rev. Lett. **101**, 025701 (2008).
C.-Y. Huang and F.-L. Lin, Phys. Rev. A **81**, 032304 (2010).
H.-L. Wang, Q.-Q. Shi, S.-H. Li and H.-Q. Zhou, arXiv:1106.2129.
J. Jordan, Studies of Infinite Two-Dimensional Quantum Lattice Systems with Projected Entangled Pair States, PhD Thesis, University of Queensland (2011) http://www.romanorus.com/JordanThesis.pdf
R. Orús, T.-C. Wei, O. Buerschaper and A. Garcia-Saez, Phys. Rev. Lett. **113**, 257202 (2014).
R. Mohseninia, S. S. Jahromi, L. Memarzaeh and V. Karimipour, Phys. Rev. B **91**, 245110 (2015).
L. M. Ionnou, Quantum Inf. Comput. **7**, 335 (2007).
L. Huang, New J. Phys. **16**, 033027 (2014).
F. Verstraete, D. Porras, and J. I. Cirac, Phys. Rev. Lett. **93**, 227205 (2004); F. Verstraete and J. I. Cirac, arXiv:cond-mat/0407066; V. Murg, F. Verstaete and J. I. Cirac, Phys. Rev. A **75**, 033605 (2007).
G. Vidal, Phys. Rev. Lett. **91**, 147902 (2003); **93**, 040502 (2004); **98**, 070201 (2007).
J. Jordan, R. Orús, G. Vidal, F. Verstraete and J. I. Cirac, Phys. Rev. Lett. **101**, 250602 (2008).
H. C. Jiang, Z. Y. Weng and T. Xiang, Phys. Rev. Lett. **101**, 090603 (2008).
B. Bauer, G. Vidal and M. Troyer, J. Stat. Mech. P09006(2009).
R. Orús and G. Vidal, Phys. Rev, B **78**, 155117 (2008).
P. Pippan, S. R. White and H. G. Evertz, Phys. Rev. B **81**, 081103(R) (2010).
Q.-Q. Shi and H.-Q. Zhou, J. Phys. A **42** 272002 (2009).
J. Haegeman, J. I. Cirac and T. J. Osborne, I. Pižorn, H. Verschelde and F. Verstraete, Phys. Rev. Lett. **107**, 070601 (2011).
L. Wang, Y.-J. Kao and A.W. Sandvik, Phys. Rev. E **83**, 056703 (2011).
For recent reviews, see R. Orús, Ann. Phys. (N.Y.) **349**, 117 (2014); Eur. Phys. J. B **87**, 280 (2014).
H. N. Phien, J. A. Bengua, H. D. Tuan, P. Corboz and R. Orús, Phys. Rev. B **92**, 035142 (2015).
H.-Q. Zhou and J. P. Barjaktarevi$\check{\rm c}$, J. Phys. A **41**, 412001 (2008); H.-Q. Zhou, J.-H. Zhao and B. Li, J. Phys. A **41**, 492002 (2008).
H.-Q. Zhou, R. Orús and G. Vidal, Phys. Rev. Lett. **100**, 080602 (2008).
B. Li, S.-H. Li and H.-Q. Zhou, Phys. Rev. E **79**, 060101(R) (2009).
J.-H. Zhao, H.-L. Wang, B. Li and H.-Q. Zhou, arXiv:0902.1669; Phys. Rev. E. **82**, 061127 (2010).
H.-L. Wang, J.-H. Zhao, B. Li and H.-Q. Zhou, J. Stat. Mech. L10001(2011).
F-Y. Wu, Rev. Mod. Phys. **54**, 235 (1982).
Y.-W. Dai, S. Y. Cho, M. T. Batchelor and H.-Q. Zhou, Phys. Rev. E **89**, 062142 (2014).
The result for $q=3$ can be compared with previous estimates. For a special direction of the field, the $Z_N$ toric code in a magnetic field corresponds to the $N$-state quantum Potts model in a transverse field. This model was studied using iPEPS [@toric] to compute the critical field value $1/(0.126 \times 3) \simeq 2.64$ for $N=3$. Note also from series expansion [@series], one gets $1/0.3806 \simeq2.627$. These results are to be compared with our estimate of $\lambda_c=2.616$. We thank Julien Vidal for these remarks.
M. D. Schulz, S. Dusuel, R. Orús, J. Vidal and K. P. Schmidt, New J. Phys. **14**, 025005 (2012).
C. J. Hamer, J. Oitmaa and Z. Weihong, J. Phys. A **25**, 1821 (1992).
J. Solyom and P. Pfeuty, Phys. Rev. B **24**, 218 (1981).
See, e.g., A. W. Sandvik, AIP Conf. Proc. **1297**, 135 (2010) and references therein.
N. D. Mermin and H. Wagner, Phys. Rev. Lett. **17**, 1133 (1966).
J. D. Reger and A. P. Young, Phys. Rev. B **37**, 5978 (1988); A. W. Sandvik and H. G. Evertz, Phys. Rev. B **82**, 024407 (2010);
T. Roscilde, P. Verrucchi, A. Fubini, S. Haas and V. Tognetti, Phys. Rev. Lett. **94**, 147208 (2005).
H.-Q. Lin and V. J. Emery, Phys. Rev. B **40**, 2730 (1989); R. F. Bishop, J. B. Parkinson and Y. Xian, Phys. Rev. B **46**, 880 (1992); V. Y. Irkhin, A. A. Katanin and M.I. Katsnelson, J. Phys.:Condens. Matter **4**, 5227 (1992); D. J. J. Farnell, K. A. Gernoth and R. F. Bishop, Phys. Rev. B **64**, 172409 (2001).
A. Chandran, V. Khemani and S. L. Sondhi, Phys. Rev. Lett. **113**, 060501 (2014).
[^1]: The GE by definition is given from the fidelity between two states. However, in this case the two states are related, i.e., the closet separable state to the ground state and the ground states share some property. For two degenerate ground states $|\Psi_g^1\rangle$ and $|\Psi_g^2\rangle$ satisfying $|\Psi_g^2\rangle=U|\Psi_g^1\rangle$, the closest separable state of different degenerate groundstate $|\phi^1\rangle$ and $|\phi^2\rangle$ has the same relationship, i.e., the two closest separable states to the two states would correspondingly satisfy $|\phi^2\rangle=U|\phi^1\rangle$. Then the fidelity between the groundstate $|\Psi_g^1\rangle$ and its closest separable state $|\phi^1\rangle$ can be written as $\langle \Psi_g^1 |\phi^1\rangle=\langle \Psi_g^1 |UU^{\dag}|\phi^1\rangle=\langle \Psi_g^2 |\phi^2\rangle$, which thus gives the same value for the GE.
[^2]: Supposing the coupling in front of the two-body interaction term $S_z^r S_z^{r'}$ is $\Delta_z$, with $\Delta_z=1$ in the XYX model, the Hamiltonian obeys a duality relation under the transformation $\sigma_x \leftrightarrow\sigma_y$, $\Delta_y \rightarrow 1/\Delta_y$, $\Delta_z \rightarrow\Delta_z/\Delta_y$, $h \rightarrow h/\Delta_y$.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In most practical situations, the compression or transmission of images and videos creates distortions that will eventually be perceived by a human observer. Vice versa, image and video restoration techniques, such as inpainting or denoising, aim to enhance the quality of experience of human viewers. Correctly assessing the similarity between an image and an undistorted reference image as subjectively experienced by a human viewer can thus lead to significant improvements in any transmission, compression, or restoration system. This paper introduces the Haar wavelet-based perceptual similarity index (HaarPSI), a novel and computationally inexpensive similarity measure for full reference image quality assessment. The HaarPSI utilizes the coefficients obtained from a Haar wavelet decomposition to assess local similarities between two images, as well as the relative importance of image areas. The consistency of the HaarPSI with the human quality of experience was validated on four large benchmark databases containing thousands of differently distorted images. On these databases, the HaarPSI achieves higher correlations with human opinion scores than state-of-the-art full reference similarity measures like the structural similarity index (SSIM), the feature similarity index (FSIM), and the visual saliency-based index (VSI). Along with the simple computational structure and the short execution time, these experimental results suggest a high applicability of the HaarPSI in real world tasks.'
author:
- 'Rafael Reisenhofer[^1]'
- 'Sebastian Bosse[^2]'
- 'Gitta Kutyniok[^3]'
- ' Thomas Wiegand[^4]'
bibliography:
- 'refsHaarPSI.bib'
title: 'A Haar Wavelet-Based Perceptual Similarity Index for Image Quality Assessment'
---
Introduction {#sec:intro}
============
Digital images and videos are omnipresent in daily life and the importance of visual data is still growing: According to [@Cisco], by 2020, nearly a million minutes of video content is estimated to cross the internet every second.
Typically, video and image signals are intended to be ultimately viewed by humans. For transmission or storage, most signals are compressed in order to meet today’s channel and/or storage demands. Compression as well as transmission errors can introduce distortions to video or image signals that are visible to human viewers. For evaluating or optimizing a transmission system or parts of it, e.g. by controlling the rate-distortion trade-off of a video encoder, it is crucial to measure the severity of distortions in a perceptually meaningful way. Quality ’in a perceptually meaningful way’ can only be measured reliably in psychometric tests. In such tests, participants are asked to rate the subjectively perceived quality of images or videos that have previously been subject to some kind of distortion introducing processing. The quality ratings of individual participants can eventually be averaged to obtain a single mean opinion score (MOS) for each stimulus. However, although being the gold standard for assessing perceived quality such studies are expensive and time-consuming and not feasible at all for real-time tasks like optimizing or monitoring transmission systems. This has been motivating research in computational image quality assessment for decades.
Image quality assessment methods typically belong to one of three categories with different challenges and scopes of applications: Full reference (FR) image quality assessment approaches require and utilize the availability of a reference image. Reduced reference (RR) methods exploit a small set of features extracted from the reference image. No reference (NR) approaches estimate the perceived quality of a possibly distorted image solely from the image itself [@Lin2011]. Unconstrained NR IQA has the notion of being the holy grail of IQA and, when successful, essentially replicates human abilities. It is, however, not a feasible approach for some applications such as, for example, encoder control for video compression. An NR quality metric used for rate-distortion optimization in a video encoder would steer the optimization towards coding decisions that remove any type of noise or artifacts. However, there are videos in which noise and artifacts were intentionally added to create a certain visual effect. As an example, the reader is invited to imagine a video encoder that removes film grain from the Quentin Tarantino movie *The Hateful Eight* due to the application of an NR quality metric that penalizes ”noisy” coding decisions. Such an encoder would change a deliberate artistic decision made by the filmmakers and thus deteriorate the viewing experience. The simplest FR image quality metric is the mean squared error (MSE), which is defined as the average of the squared differences of the reference and the distorted image. Although being widely used, it does not correlate well with perceived visual quality [@Gir93]. More sophisticated approaches towards perceptually accurate image quality assessments (IQA) typically follow one of three strategies. *Bottom-up* approaches explicitly model various processing mechanisms of the human visual system (HVS), such as masking effects [@Watsona1997], contrast sensitivity [@daly1990], or just-noticeable-distortion [@Lubin1997; @jia2006estimating] in order to assess the perceived quality of images. For instance, the adaptivity of the HVS to the magnitude of distortions is modeled explicitly by the concept of most apparent distortion (MAD) [@LaCh2010] in order to apply two different assessment strategies for supra- and near-threshold distortions.
However, the method proposed in this paper as well as most image quality metrics developed recently follow a *top-down* approach. There, general functional properties of the HVS (considered as a black box) are assumed in order to identify and to exploit image features corresponding to the perceived quality. Prominent examples are the structural similarity index (SSIM) [@WBSS2004], visual information fidelity (VIF) [@ShBo2006], the gradient similarity measure (GSM) [@LLN2012], spectral residual based similarity (SR-SIM) [@ZhLi2012], and the visual saliency-induced index (VSI) [@ZSL2014]. The SSIM [@WBSS2004] aims at taking into account the sensitivity of the human visual system towards structural information. This is done by pooling three complementary components, namely luminance similarity (comparing local mean luminance values), contrast similarity (comparing local variances) and structural similarity, which is defined as the local covariance between the reference image and its perturbed counterpart. Although being criticized [@DoYa2011], it is highly cited and among the most popular image quality assessment metrics. The SSIM was generalized for a multi-scale setting by the multi-scale structural similarity index (MS-SSIM) [@WSB2003]. One of the first information theoretic approaches to FR IQA was presented as visual information fidelity (VIF) [@ShBo2006]. VIF models the wavelet coefficients as Gaussian Scale Mixtures and quantifies the mutual information shared between reference and test images. The information theoretic measure of mutual information is shown to be correlated to perceived image quality. Changes in contrast and structure are captured by considering local gradients in [@LLN2012], while the squared difference in pixel values between the reference image and the distorted image is used to measure luminance variations. This approach thus follows the basic framework of combining complementary feature maps originally introduced in [@WBSS2004]. Additionally, masking effects are estimated, based on the local gradient magnitude of the reference image and incorporated when the two feature maps are combined. Spectral residual-based similarity (SR-SIM) [@ZhLi2012] takes into account changes in the local horizontal and vertical gradient magnitudes. Additionally, it incorporates changes in a spectral residual-based visual saliency estimate. The visual saliency-induced index (VSI) [@ZSL2014] follows the same line as SR-SIM by combining similarities in the gradient magnitude and the visual saliency. However, it further exploits the visual saliency map for weighting the spatial similarity pooling. Furthermore, [@ZSL2014] also explores the influence of different saliency models on the performance of the proposed image quality measure. A combination of two feature maps is also applied successfully by the feature similarity index (FSIM) [@ZZMZ2011]. Due to its conceptual similarity to the proposed method, it will be discussed in more detail in a later section.
Adopting the advances in machine learning and data science, IQA methods following a third, purely *data driven* strategy have been proposed recently. So far, data driven approaches were mainly developed for the domain of NR IQA [@Kang2014; @Ye2012; @Zhang2015; @bosse2016dnnNrIqa], but they have also been adapted in the context of FR IQA [@bosse2016dnnFrIqa].
Contributions
-------------
This work introduces the Haar wavelet-based perceptual similarity index (HaarPSI), a novel and computationally inexpensive measure yielding FR image quality assessments. The HaarPSI utilizes the magnitudes of high-frequency Haar wavelet coefficients to define local similarities and low-frequency Haar wavelet coefficients to weight the importance of (dis)similarities at specific locations in the image domain.
[The six discrete two-dimensional Haar wavelet filters used in the definition of the HaarPSI respond to horizontal and vertical edges on different frequency scales. The HaarPSI is thus based on elementary implementations of functional properties known to be exhibited by neurons in the primary visual cortex, namely orientation selectivity and spatial frequency selectivity. We aim to demonstrate that such a simple model already suffices to define a similarity measure that yields state-of-the-art correlations with human opinion scores.]{}
The HaarPSI can also be seen as a drastic simplification of the FSIM [@ZZMZ2011], which is based on a similar combination of similarity and weight maps. In the definition of the FSIM, both local similarities and weights rely on the phase congruency measure [@Kov2000], whose computation requires images to be convolved with 16 complex-valued filters and contains several non-trivial steps such as adaptive thresholding. For the HaarPSI on the other hand, the two maps are computed from the responses of only six discrete Haar wavelet filters and are cleanly separated in the sense that local similarities and weights are based on different frequency scales. Surprisingly, these simplifications not only decrease the required computational effort but also lead to consistently higher correlations with human mean opinions scores.
In Section \[sec:results\], we evaluate the consistency of the HaarPSI with the human quality of experience and compare its performance to state-of-the-art similarity measures like SSIM [@WBSS2004], FSIM [@ZZMZ2011], and VSI [@ZSL2014]. As depicted in Tables \[tab:sroccdatabases\] and \[tab:overallperf\], the HaarPSI achieves higher correlations with human opinion scores than all other considered FR quality metrics in all test cases except one, where it only comes second to the VSI. In addition, the HaarPSI can be computed significantly faster than the metrics yielding the second and third highest correlations with human opinion scores, namely VSI and FSIM. In order to facilitate reproducible research, our Matlab implement of the HaarPSI is publicly available at <http://www.haarpsi.org/>.
It is both convenient and surprising that the promising experimental results of the HaarPSI are based on the responses of Haar filters, which are arguably the simplest and computationally most efficient wavelet filters existing. The results of a numerical analysis of the applicability of other wavelet filters in the newly proposed similarity measure can be found in Table \[tab:otherwavelets\].
The Feature Similarity Index (FSIM)
-----------------------------------
The feature similarity index (FSIM) [@ZZMZ2011], proposed in 2011, is currently one of the most successful and influential FR image quality metrics. The FSIM combines two feature maps derived from the phase congruency measure [@Kov2000] and the local gradients of the reference and the distorted image to assess local similarities between two images. For a grayscale image $f\in\ell^2({\mathbb{Z}}^2)$, the gradient map is defined by $$\label{eq:grad}
\operatorname{G_f}[x] = \sqrt{\left((g^{\text{hor}}*f)[x]\right)^2 + \left((g^{\text{ver}}*f)[x]\right)^2},$$ where $g^{\text{hor}}$ and $g^{\text{ver}}$ denote horizontal and vertical gradient filters (e.g. Sobel or Scharr filters), and $*$ denotes the two-dimensional convolution operator. The method used in the implementation of the FSIM to compute the phase congruency map was developed by Peter Kovesi [@KovONLINE] and contains several non-trivial operations, such as adaptive soft thresholding. However, in its essence, the phase congruency map of a grayscale image $f$ is given by $$\label{eq:pc}
\operatorname{PC_{f}}[x] \approx \frac{{\left\vert\sum_n (g^\text{c}_n*f)[x]\right\vert}}{\sum_n{\left\vert(g^\text{c}_n*f)[x]\right\vert}},$$
where $g^\text{c}_n$ denotes differently scaled and oriented complex-valued wavelet filters. The idea behind is that if the obtained complex-valued wavelet coefficients have the same phase at a location $x$, taking the absolute value of the sum is the same as taking the sum of the absolute values. If this is the case, $\operatorname{PC_{f}}[x]$ will be close to or precisely $1$.
To assess local similarities between two images with respect to the maps defined in and , the FSIM - like many other image quality metrics - uses a simple similarity measure for scalar values that already appeared in [@WBSS2004], namely $$\label{eq:scalarsim}
{\operatorname{S}}(a,b,C) = \frac{2ab + C}{a^2 + b^2 + C},$$ with a constant $C > 0$. The graph of ${\operatorname{S}}(a,b,C)$ for values ranging from $0$ to $100$ and $C=30$ is shown in Figure \[fig:scalarsim\]. The local feature similarity map for two grayscale images $f_1,f_2\in\ell^2({\mathbb{Z}}^2)$ is defined by $$\label{eq:localfs}
\operatorname{FS_{f_1,f_2}}[x]= {\operatorname{S}}\left(\operatorname{G_{f_1}}[x],\operatorname{G_{f_2}}[x],C_1\right)^\beta {\cdot} {\operatorname{S}}\left(\operatorname{PC_{f_1}}[x],\operatorname{PC_{f_2}}[x],C_2\right)^\gamma,$$ with constants $C_1,C_2>0$ and exponents $\beta,\gamma > 0$. Based on the assumption that the human visual system is especially sensitive towards structures at which the phases of the Fourier components are in congruency (see e.g. [@MRBO1986]), the phase congruency map is not only used in but also applied to determine the relative importance of different image areas with respect to human perception. Eventually, the feature similarity index is computed by taking the weighted mean of all local feature similarities, where the phase congruency map is used as a weight function, that is $$\label{eq:fsim}
\operatorname{FSIM_{f_1,f_2}} = \frac{\sum_x \operatorname{FS_{f_1,f_2}}[x] \cdot \operatorname{PC_{f_1,f_2}}[x]}{\sum_x \operatorname{PC_{f_1,f_2}}[x]},$$ where $$\operatorname{PC_{f_1,f_2}}[x] = \max\left(\operatorname{PC_{f_1}}[x],\operatorname{PC_{f_2}}[x]\right).$$
The original publication of the FSIM proposes a generalization to color images defined in the YIQ color space, named FSIMC. In the YIQ space, the Y channel encodes luminance information, while the I and Q channels encode chromatic information. Color images defined in the RGB color space can easily be transformed to the YIQ space with a linear mapping, namely
$$\begin{bmatrix} f^{\text{Y}} \\ f^{\text{I}} \\ f^{\text{Q}} \end{bmatrix}
\approx
\begin{bmatrix} 0.299 & 0.587 & 0.114 \\ 0.596 & -0.274 & -0.322 \\ 0.211 & -0.523 & 0.312 \end{bmatrix}\cdot
\begin{bmatrix} f^{\text{R}} \\ f^{\text{G}} \\ f^{\text{B}} \end{bmatrix}.$$
FSIMC simply incorporates the chroma channels I and Q into the local feature similarity measure . The gradient maps as well as the phase congruency maps are purely derived from the luminance channel Y in FSIMC and FSIM alike.
The Haar Wavelet-Based Perceptual Similarity Index {#sec:HaarPSI}
==================================================
The basic idea of the HaarPSI is to construct feature maps in the spirit of as well as a weight function similar to by considering a single wavelet filterbank. The response of any high-frequency wavelet filter will look similar to the response yielded by a gradient filter like the Sobel operator. Furthermore, the phase congruency measure used as a weight function in the FSIM is computed directly from the output of a multi-scale complex-valued wavelet filterbank, as illustrated by Equation . This gives a strong intuition that it should be possible to define a similarity measure derived from the response of a single set of discrete wavelet filters that at least matches the performance of the FSIM on benchmark databases but requires significantly less computational effort.
The wavelet chosen for this endeavor is the so-called Haar wavelet, which was already proposed in 1910 by Alfred Haar [@Haar1910] and is arguably the simplest and computationally most efficient wavelet there is. The one-dimensional Haar filters are given by $$\label{eq:haarfilters}
h^{\text{1D}}_1 = \frac{1}{\sqrt{2}}\cdot[1,1] \text{ and } g^{\text{1D}}_1 = \frac{1}{\sqrt{2}}\cdot[-1,1],$$ where $h^{\text{1D}}_1$ denotes the low-pass scaling filter and $g^{\text{1D}}_1$ the corresponding high-pass wavelet filter. For any scale $j\in{\mathbb{N}}$, we can construct two-dimensional Haar filters by setting $$\begin{aligned}
g^{\text{(1)}}_j &= g^{\text{1D}}_j \otimes h^{\text{1D}}_j,\\
g^{\text{(2)}}_j &= h^{\text{1D}}_j \otimes g^{\text{1D}}_j,\end{aligned}$$ where $\otimes$ denotes the outer product and the one-dimensional filters $h^{\text{1D}}_j$ and $g^{\text{1D}}_j$ are given for $j>1$ by $$\begin{aligned}
g^{\text{1D}}_j &= h^{\text{1D}}_{1}*(g^{\text{1D}}_{j-1})_{\uparrow 2}, \\
h^{\text{1D}}_j &= h^{\text{1D}}_{1}*(h^{\text{1D}}_{j-1})_{\uparrow 2},\end{aligned}$$ where $\uparrow2$ is the dyadic upsampling operator, and $*$ denotes the one-dimensional convolution operator. Note that $g^{\text{(1)}}_j$ responds to horizontal structures, while $g^{\text{(2)}}_j$ picks up vertical structures. The six Haar filters used to define the HaarPSI are shown in Figure \[fig:haarfilters\].
The local similarity map $\operatorname{FS_{f_1,f_2}}$ multiplicatively combines gradient-based and phase congruency-based similarities whose contributions are weighted by the exponents $\alpha,\beta>0$. The HaarPSI does not consider different types of similarities. However, to correctly predict the perceptual similarity experienced by human viewers, it can be useful to apply an additional non-linear mapping to the local similarities obtained from high-frequency Haar wavelet filter responses. This non-linearity is chosen to be the logistic function, which is widely used as an activation function in neural networks [for modeling thresholding in biological neurons]{} and [is]{} given for a parameter $\alpha >0$ as $$\label{eq:logistic}
l_\alpha(x) = \frac{1}{1 + e^{-\alpha x}}.$$
For two grayscale images $f_1,f_2\in\ell^2({\mathbb{Z}}^2)$, the local similarity measure used to compute the HaarPSI is based on the first two stages of a two-dimensional discrete Haar wavelet transform and given by $$\label{eq:localhs}
\operatorname{HS^\text{(k)}_{f_1,f_2}}[x]=l_\alpha\left(\frac{1}{2}\sum_{j= 1}^2{\operatorname{S}}\left({\left\vert(g^\text{(k)}_j*f_1)[x]\right\vert},{\left\vert(g^\text{(k)}_j*f_2)[x]\right\vert},C\right)\right),$$ where $C > 0$, $k\in\{1,2\}$ selects either horizontal or vertical Haar wavelet filters, ${\operatorname{S}}$ denotes the similarity measure , and $*$ is the two-dimensional convolution operator. The local similarity measure $\operatorname{HS^\text{(k)}_{f_1,f_2}}$ can be seen as an analog to $\operatorname{FS_{f_1,f_2}}$. However, $\operatorname{HS^\text{(k)}_{f_1,f_2}}$ does not mix different different concepts like gradients and phase congruency and is computed straightforwardly on the responses of two high-frequency discrete Haar wavelet filters. A visualization of the local similarity map $\operatorname{HS^\text{(k)}_{f_1,f_2}}$ is shown in Figure \[fig:haarmaps\].
Analogous to the phase congruency map $\operatorname{PC_{f}}$ in the definition of the FSIM, the HaarPSI considers a weight map which is derived from the response of a single low-frequency Haar wavelet filter: $$\label{eq:haarweight}
\operatorname{W^\text{(k)}_{f}}[x] = {\left\vert(g^{\text{(k)}}_3*f)[x]\right\vert},$$ where $k\in\{1,2\}$ again differentiates between horizontal and vertical filters. Figure \[fig:haarmaps\] shows an example of the weight map $\operatorname{W^\text{(k)}_{f}}$ computed from a natural image.
The Haar-wavelet based perceptually similarity index for two grayscale images $f_1,f_2$ is eventually given as the weighted average of the local similarity map $\operatorname{HS^\text{(k)}_{f_1,f_2}}$ , that is, $$\label{eq:haarpsi}
\operatorname{HaarPSI_{f_1,f_2}} = l_\alpha^{-1}\left(\frac{\sum\limits_x \sum\limits_{k=1}^2\operatorname{HS^\text{(k)}_{f_1,f_2}}[x] \cdot \operatorname{W^\text{(k)}_{f_1,f_2}}[x]}{\sum\limits_x \sum\limits_{k=1}^2\operatorname{W^\text{(k)}_{f_1,f_2}}[x]}\right)^2,$$ with $$\operatorname{W^\text{(k)}_{f_1,f_2}}[x] = \max(\operatorname{W^\text{(k)}_{f_1}}[x],\operatorname{W^\text{(k)}_{f_2}}[x])$$ for $k\in \{1,2\}$. The function $l_\alpha^{-1}(\cdot)$ maps the weighted average from the interval $[\frac{1}{2},l_\alpha(1)]$ back to $[0,1]$. Applying $(\cdot)^2$ further spreads the HaarPSI in the unit interval and helps to linearize the relationship between the HaarPSI and human opinion scores. In particular, this procedure aims to increase the readability of the HaarPSI in the sense that a single value should be ’meaningful on its own’ and not only relative to other HaarPSI values. Please note that, due to the monotonicity of the logistic function, applying $l_\alpha^{-1}(\cdot)^2$ cannot improve or worsen the rank order-based correlations with human opinion scores reported in Section \[sec:results\].
Analogous to the FSIM, the HaarPSI can be extended to color images in the YIQ color space by considering a third local similarity map based on the chroma channels I and Q. The map $\operatorname{HS^\text{(3)}_{f_1,f_2}}$ is computed analogous to by averaging local similarities obtained from comparing $f^{\text{I}}_1$ with $f^{\text{I}}_2$ and $f^{\text{Q}}_1$ with $f^{\text{Q}}_2$. In contrast to $\operatorname{HS^\text{(1)}_{f_1,f_2}}$ and $\operatorname{HS^\text{(2)}_{f_1,f_2}}$, the chromatic information used for $\operatorname{HS^\text{(3)}_{f_1,f_2}}$ is not based on orientation sensitive filters. The corresponding weight map $\operatorname{W^\text{(3)}_{f^\text{Y}_1,f^\text{Y}_2}}$ is thus also computed by averaging $\operatorname{W^\text{(1)}_{f^\text{Y}_1,f^\text{Y}_2}}$ and $\operatorname{W^\text{(2)}_{f^\text{Y}_1,f^\text{Y}_2}}$. Formally, the generalization of the HaarPSI to color images is given by $$\label{eq:haarpsic}
\operatorname{HaarPSIC_{f_1,f_2}} = l_\alpha^{-1}\left(\frac{\sum\limits_x \sum\limits_{k=1}^3\operatorname{HS^\text{(k)}_{f_1,f_2}}[x] \cdot \operatorname{W^\text{(k)}_{f^\text{Y}_1,f^\text{Y}_2}}[x]}{\sum\limits_x \sum\limits_{k=1}^3\operatorname{W^\text{(k)}_{f^\text{Y}_1,f^\text{Y}_2}}[x]}\right)^2,$$ with $\operatorname{HS^\text{(1)}_{f_1,f_2}}$ and $\operatorname{HS^\text{(2)}_{f_1,f_2}}$ defined as in , $$\label{eq:localhs3}
\resizebox{.9\hsize}{!}{$\operatorname{HS^\text{(3)}_{f_1,f_2}}[x]=l_\alpha\left(\frac{1}{2}\left({\operatorname{S}}\left({\left\vert(m*f^{\text{I}}_1)[x]\right\vert},{\left\vert(m*f^{\text{I}}_2)[x]\right\vert},C\right) + {\operatorname{S}}({\left\vert(m*f^{\text{Q}}_1)[x]\right\vert},{\left\vert(m*f^{\text{Q}}_2)[x]\right\vert},C)\right)\right)$,}$$ with a $2\times2$ mean filter $m$ and $$\operatorname{W^\text{(3)}_{f^\text{Y}_1,f^\text{Y}_2}}[x] = \frac{1}{2}\left(\operatorname{W^\text{(1)}_{f^\text{Y}_1,f^\text{Y}_2}}[x]+\operatorname{W^\text{(2)}_{f^\text{Y}_1,f^\text{Y}_2}}[x]\right).$$
\
\
\
Parameter Selection
-------------------
The HaarPSI as well as the HaarPSIC require only two parameters to be selected, namely $C$ and $\alpha$. Both parameters were optimized on randomly chosen subsets of four large publicly available databases, where each subset was a quarter the size of the original database. Each of the databases, which will be described in more detail in Section \[sec:results\], contains large numbers of differently distorted images and their corresponding MOS values. The parameters $C$ and $\alpha$ were selected to maximize the mean of the four Spearman rank order correlation coefficients (SROCC) obtained from comparing HaarPSIC and MOS values from subsets of the TID 2008 [@PLZECB2009], TID 2013 [@Ponomarenko2015], LIVE [@SWCBOnline] and CSIQ [@LaCh2010] image databases. The optimization was carried out in two steps. First, a grid search was performed in which the parameter $C$ took values in the interval $[5,100]$ and $\alpha$ in the range between $2$ and $8$. The best $(C,\alpha)$ pair was then used as the initial value of the Nelder-Mead algorithm. The thus refined parameters were eventually rounded to the nearest integer in the case of $C$ and to the nearest tenth in the case of $\alpha$. This procedure resulted in the choices of $C=30$ and $\alpha = 4.2$. To verify the generality of the HaarPSI, the same optimization procedure was repeated once only considering the TID 2008 and TID 2013 databases and once restricted to the LIVE and the CSIQ image databases. The results of all three optimizations are compiled in Figure \[fig:optimization\].
Experimental Results {#sec:results}
====================
The consistency of the HaarPSI with the human perception of image quality was evaluated and compared with most of the image quality metrics discussed in Section \[sec:intro\] on four large publicly available benchmark databases of quality-annotated images. Those databases differ in the number of reference images, the number of distortion magnitudes and types, the number of observers, the level of control of the viewing conditions, and the stimulus presentation procedure.
The LIVE database [@SWCBOnline] contains 29 reference color images and 779 distorted images that were perturbed by JPEG compression, JPEG 2000 compression, additive Gaussian white noise, Gaussian blurring as well as JPEG 2000 compressed images that have been transmitted over a simulated Rayleigh fading channel. Each distortion is introduced at five to six different levels of magnitude. On average, about 23 subjects evaluated the quality of each image with respect to the reference image. The viewing conditions were fairly controlled for in terms of viewing distance. Ratings were collected in a double stimulus manner.
The TID 2008 database [@PLZECB2009] comprises 25 colored reference images and 1700 degraded images, that had been subject to a wide range of distortions, including various types of noise, blur, JPEG and JPEG 2000 compression, transmission errors, local image distortions, as well as luminance and contrast changes. Subjective ratings were gathered by comparisons. The results from several viewing conditions of experiments in three different labs and on the internet were averaged. TID 2008 was later extended to TID 2013 [@Ponomarenko2015], which added new types of distortions, which are mostly of a chromatic nature. In total, TID 2013 contains 3000 differently distorted images.
The CSIQ database [@LaCh2010] is based on 30 reference color images and contains 866 distorted images. Six different types of distortions (JPEG compression, JPEG 2000 compression, global contrast decrements, additive pink Gaussian noise, and Gaussian blurring) at four to five different degradation magnitudes were applied to the reference images. The viewing distance was controlled. Images were presented on a monitor array and subjects were asked to place all distorted versions of one reference image according to its perceived quality.
The main goal of most computational image similarity measures is to yield a monotonic relationship with human mean opinion scores across different databases and distortion types. To ensure a fair evaluation, different computational measures are typically compared with respect to rank order-based correlations or after performing nonlinear regression. Throughout the numerical evaluation of the HaarPSI, we apply the rank order-based SROCC to measure correlations between human mean opinion scores and different computational similarity and distortion indexes. We also considered applying Kendall’s $\tau$ and the Pearson product-moment correlation after performing a four parameter logistic regression as alternatives for the SROCC. We found that these correlation coefficients essentially duplicate the results reported in this section. The corresponding versions of Tables \[tab:sroccdatabases\] and \[tab:details\] were thus not included here but can be found at [www.haarpsi.org](www.haarpsi.org).
Following the ITU guidelines for evaluating quality prediction models [@ITUTP1401], we also tested the statistical significance of the results reported in this section. Correlation coefficients for which the $H_0$ hypothesis that they are not significantly different than the respective HaarPSI correlation can be refuted with $p < 0.05$ are highlighted in color in Tables \[tab:sroccdatabases\], \[tab:details\] and \[tab:otherwavelets\]. In accordance with [@fieller1957tests], the variance of the z-transforms were approximated by $1.06/(N - 3)$, where $N$ denotes the degrees of freedom (i.e. the number of samples in the considered database or distortion specific subset).
[\*[12]{}[c]{}]{}\
& & PSNR & VIF & SSIM & MS-SSIM & GSM & MAD & SR-SIM & FSIM & VSI & HaarPSI\
& LIVE & 0.8756 & 0.9636 & 0.9479 & 0.9513 & 0.9561 & 0.9672 & 0.9619 & 0.9634 & 0.9534 & **0.9690**\
& TID2008 & 0.5531 & 0.7491 & 0.7749 & 0.8542 & 0.8504 & 0.8340 & 0.8913 & 0.8804 & 0.8830 & **0.9043**\
& TID2013 & 0.6394 & 0.6769 & 0.7417 & 0.7859 & 0.7946 & 0.7807 & 0.8075 & 0.8022 & 0.8048 & **0.8094**\
& CSIQ & 0.8058 & 0.9195 & 0.8756 & 0.9133 & 0.9108 & 0.9466 & 0.9319 & 0.9242 & 0.9372 & **0.9546**\
\
\
& & PSNR & VIF & SSIM & MS-SSIM & GSM & MAD & SR-SIM & FSIM & VSI & HaarPSI\
& LIVE & 0.8756 & 0.9636 & 0.9479 & 0.9513 & 0.9561 & 0.9672 & 0.9619 & 0.9645 & 0.9524 & **0.9683**\
& TID2008 & 0.5531 & 0.7491 & 0.7749 & 0.8542 & 0.8504 & 0.8340 & 0.8913 & 0.8840 & 0.8979 & **0.9097**\
& TID2013 & 0.6394 & 0.6769 & 0.7417 & 0.7859 & 0.7946 & 0.7807 & 0.8075 & 0.8510 & **0.8965** & 0.8732\
& CSIQ & 0.8058 & 0.9195 & 0.8756 & 0.9133 & 0.9108 & 0.9466 & 0.9319 & 0.9310 & 0.9423 & **0.9604**\
The highest correlation in each row is written in **boldface**.
The four databases used in the numerical evaluation only contain color images. However, out of the metrics considered in our experiments, only the FSIM and the HaarPSI are defined for both grayscale and color images, while the visual saliency-based index (VSI) was specifically designed for color images. All other similarity measures considered in our experiments only accept grayscale images as input or perform an RGB to grayscale conversion as a first processing step. To reflect these differing designs, all methods were tested on all databases once with the original color images and once with grayscale conversions obtained from the Matlab *rgb2gray* function. To obtain the VSI for pairs of grayscale images, corresponding RGB images were created by setting the values for all three color channels to the values of the given grayscale channel. The correlation coefficients of all ten considered similarity measures with the human mean opinion scores for the LIVE image database, TID 2008, TID 2013 and the CSIQ database are compiled in Table \[tab:sroccdatabases\].
Table \[tab:overallperf\] provides a quick impression of the overall performance of each metric. It depicts the average SROCC of each metric with respect to all four databases as well as the mean execution time in milliseconds. The average execution time was measured on a Intel Core i7-4790 CPU clocked at $3.60$ GHz. To measure the execution time, each quality measure was computed ten times for ten different pairs of randomly generated $512\times512$ pixel images. All computations and measurements were carried out in Matlab using implementations made freely available by the respective authors. Note that due to an additional conversion step, metrics that are only defined for grayscale images can have slightly higher execution times when evaluated on color images.
[\*[6]{}[c]{}]{} & & &\
& & SROCC & Time (ms) & SROCC & Time (ms)\
& HaarPSI & 0.9279 & 24 & 0.9093 & 10\
& VSI & 0.9223 & 79 & 0.8946 & 80\
& FSIM & 0.9076 & 142 & 0.8925 & 121\
& SRSIM & 0.8982 & 10 & 0.8982 & 10\
& MAD & 0.8821 & 892 & 0.8821 & 891\
& GSM & 0.8780 & 8 & 0.8780 & 7\
& MSSSIM & 0.8762 & 30 & 0.8762 & 24\
& SSIM & 0.8350 & 6 & 0.8350 & 5\
& VIF & 0.8273 & 459 & 0.8273 & 453\
& PSNR & 0.7185 & 2 & 0.7185 & 1\
A high correlation with the mean opinion scores annotated to the distorted images of a large database containing many different types and degrees of distortions is arguably the best indicator of an image quality measure’s consistency with human perception. However, for certain applications like compression or denoising, it could be more important to know if an image quality metric has a high correlation with the human experience *within* a single distortion class. Table \[tab:details\] depicts the SROC coefficients for all image quality metrics when only subsets of databases containing specific distortions like Gaussian blur or JPEG transmission errors are considered.
Single correlation coefficients provide a useful means of objectively evaluating and comparing different computational models of image quality. However, they only measure a specific aspect of the relationship between an image similarity metric and human opinion scores, like linearity in the case of the Pearson correlation coefficient or monotonicity in the case of the SROCC. In an attempt to better visualize the relationship between the HaarPSI and human opinion scores, Figure \[fig:scatterplots\] shows scatter plots of the HaarPSI against difference mean opinion scores (DMOS) for all four databases. To provide as much insight as possible, the plots are categorized by specific distortion types.
[\*[12]{}[c]{}]{}\
& & PSNR & VIF & SSIM & MS-SSIM & GSM & MAD & SR-SIM & FSIM & VSI & HaarPSI\
& jpg2k & 0.8954 & 0.9696 & 0.9614 & 0.9627 & 0.9700 & 0.9692 & 0.9700 & **0.9724** & 0.9604 & 0.9684\
& jpg & 0.8809 & **0.9846** & 0.9764 & 0.9815 & 0.9778 & 0.9786 & 0.9823 & 0.9840 & 0.9761 & 0.9832\
& gwn & 0.9854 & 0.9858 & 0.9694 & 0.9733 & 0.9774 & **0.9873** & 0.9812 & 0.9716 & 0.9835 & 0.9845\
& gblur & 0.7823 & **0.9728** & 0.9517 & 0.9542 & 0.9518 & 0.9510 & 0.9660 & 0.9708 & 0.9527 & 0.9676\
& ff & 0.8907 & **0.9650** & 0.9556 & 0.9471 & 0.9402 & 0.9589 & 0.9466 & 0.9519 & 0.9430 & 0.9527\
\
& gwn & 0.9070 & 0.8797 & 0.8107 & 0.8086 & 0.8606 & 0.8386 & 0.8989 & 0.8758 & **0.9229** & 0.9177\
& gwnc & 0.8995 & 0.8757 & 0.8029 & 0.8054 & 0.8091 & 0.8255 & 0.8957 & 0.8931 & **0.9118** & 0.8982\
& scn & 0.9170 & 0.8698 & 0.8145 & 0.8209 & 0.8941 & 0.8678 & 0.9084 & 0.8711 & **0.9296** & 0.9271\
& mn & 0.8515 & **0.8683** & 0.7795 & 0.8107 & 0.7452 & 0.7336 & 0.7881 & 0.8264 & 0.7734 & 0.7909\
& hfn & **0.9270** & 0.9075 & 0.8729 & 0.8694 & 0.8945 & 0.8864 & 0.9195 & 0.9156 & 0.9253 & 0.9155\
& in & **0.8724** & 0.8327 & 0.6732 & 0.6907 & 0.7235 & 0.0650 & 0.7678 & 0.7719 & 0.8298 & 0.8269\
& qn & 0.8696 & 0.7970 & 0.8531 & 0.8589 & 0.8800 & 0.8160 & 0.8348 & 0.8726 & 0.8731 & **0.8842**\
& gblr & 0.8697 & 0.9540 & 0.9544 & 0.9563 & **0.9600** & 0.9196 & 0.9551 & 0.9472 & 0.9529 & 0.9001\
& den & 0.9416 & 0.9161 & 0.9530 & 0.9582 & **0.9725** & 0.9433 & 0.9666 & 0.9618 & 0.9693 & 0.9711\
& jpg & 0.8717 & 0.9168 & 0.9252 & 0.9322 & 0.9393 & 0.9275 & 0.9393 & 0.9294 & **0.9616** & 0.9417\
& jpg2k & 0.8132 & 0.9709 & 0.9625 & 0.9700 & 0.9758 & 0.9707 & 0.9809 & 0.9780 & 0.9848 & **0.9860**\
& jpgt & 0.7516 & 0.8585 & 0.8678 & 0.8681 & 0.8790 & 0.8661 & 0.8881 & 0.8756 & **0.9160** & 0.8921\
& jpg2kt & 0.8309 & 0.8501 & 0.8577 & 0.8606 & 0.8936 & 0.8394 & 0.8902 & 0.8555 & 0.8942 & **0.8963**\
& pn & 0.5815 & 0.7619 & 0.7107 & 0.7377 & 0.7386 & **0.8287** & 0.7659 & 0.7514 & 0.7699 & 0.8010\
& bdist & 0.6193 & 0.8324 & 0.8462 & 0.7546 & **0.8862** & 0.7970 & 0.7798 & 0.8464 & 0.6295 & 0.8026\
& ms & 0.6957 & 0.5096 & 0.7231 & **0.7338** & 0.7190 & 0.5163 & 0.5704 & 0.6554 & 0.6714 & 0.6051\
& ctrst & 0.5859 & **0.8188** & 0.5246 & 0.6381 & 0.6691 & 0.2723 & 0.6475 & 0.6510 & 0.6557 & 0.6209\
\
& gwn & 0.9291 & 0.8994 & 0.8671 & 0.8646 & 0.9064 & 0.8843 & 0.9251 & 0.9101 & **0.9460** & 0.9368\
& gwnc & **0.8981** & 0.8299 & 0.7726 & 0.7730 & 0.8175 & 0.8019 & 0.8562 & 0.8537 & 0.8705 & 0.8593\
& scn & 0.9200 & 0.8835 & 0.8515 & 0.8544 & 0.9158 & 0.8911 & 0.9223 & 0.8900 & **0.9367** & 0.9311\
& mn & 0.8323 & **0.8450** & 0.7767 & 0.8073 & 0.7293 & 0.7380 & 0.7855 & 0.8094 & 0.7697 & 0.7858\
& hfn & 0.9140 & 0.8972 & 0.8634 & 0.8604 & 0.8869 & 0.8876 & 0.9131 & 0.9040 & **0.9200** & 0.9069\
& in & **0.8968** & 0.8537 & 0.7503 & 0.7629 & 0.7965 & 0.2769 & 0.8280 & 0.8251 & 0.8741 & 0.8656\
& qn & 0.8808 & 0.7854 & 0.8657 & 0.8706 & 0.8841 & 0.8514 & 0.8497 & 0.8807 & 0.8748 & **0.8893**\
& gblr & 0.9149 & 0.9650 & 0.9668 & 0.9673 & **0.9689** & 0.9319 & 0.9622 & 0.9551 & 0.9612 & 0.9149\
& den & 0.9480 & 0.8911 & 0.9254 & 0.9268 & 0.9432 & 0.9252 & 0.9398 & 0.9330 & **0.9484** & 0.9456\
& jpg & 0.9189 & 0.9192 & 0.9200 & 0.9265 & 0.9284 & 0.9217 & 0.9396 & 0.9339 & **0.9541** & 0.9512\
& jpg2k & 0.8840 & 0.9516 & 0.9468 & 0.9504 & 0.9602 & 0.9511 & 0.9672 & 0.9589 & **0.9706** & 0.9704\
& jpgt & 0.7685 & 0.8409 & 0.8493 & 0.8475 & 0.8512 & 0.8283 & 0.8543 & 0.8610 & **0.9216** & 0.8938\
& jpg2kt & 0.8883 & 0.8761 & 0.8828 & 0.8889 & 0.9182 & 0.8788 & 0.9165 & 0.8919 & **0.9228** & 0.9204\
& pn & 0.6863 & 0.7720 & 0.7821 & 0.7968 & 0.8130 & **0.8315** & 0.7967 & 0.7937 & 0.8060 & 0.8154\
& bdist & 0.1552 & 0.5306 & 0.5720 & 0.4801 & **0.6418** & 0.2812 & 0.4722 & 0.5532 & 0.1713 & 0.4471\
& ms & 0.7671 & 0.6276 & 0.7752 & **0.7906** & 0.7875 & 0.6450 & 0.6562 & 0.7487 & 0.7700 & 0.7152\
& ctrst & 0.4400 & **0.8386** & 0.3775 & 0.4634 & 0.4857 & 0.1972 & 0.4696 & 0.4679 & 0.4754 & 0.4382\
& ccs & 0.0766 & 0.3099 & 0.4141 & 0.4099 & 0.3578 & 0.0575 & 0.3117 & **0.8359** & 0.8100 & 0.6735\
& mgn & 0.8905 & 0.8468 & 0.7803 & 0.7786 & 0.8348 & 0.8409 & 0.8781 & 0.8569 & **0.9117** & 0.8902\
& cn & 0.8411 & 0.8946 & 0.8566 & 0.8528 & 0.9124 & 0.9064 & 0.9259 & 0.9135 & 0.9243 & **0.9275**\
& lcni & 0.9145 & 0.9204 & 0.9057 & 0.9068 & 0.9563 & 0.9443 & 0.9608 & 0.9485 & 0.9564 & **0.9622**\
& icqd & **0.9269** & 0.8414 & 0.8542 & 0.8555 & 0.8973 & 0.8745 & 0.8810 & 0.8815 & 0.8839 & 0.8953\
& cha & 0.8872 & 0.8848 & 0.8775 & 0.8784 & 0.8823 & 0.8310 & 0.8758 & **0.8925** & 0.8906 & 0.8599\
& ssr & 0.9042 & 0.9353 & 0.9461 & 0.9483 & **0.9668** & 0.9567 & 0.9613 & 0.9576 & 0.9628 & 0.9651\
\
& gwn & 0.9363 & 0.9575 & 0.8974 & 0.9471 & 0.9440 & 0.9541 & 0.9628 & 0.9359 & 0.9636 & **0.9666**\
& jpeg & 0.8881 & **0.9705** & 0.9546 & 0.9634 & 0.9632 & 0.9615 & 0.9671 & 0.9664 & 0.9618 & 0.9695\
& jpg2k & 0.9362 & 0.9672 & 0.9606 & 0.9683 & 0.9648 & 0.9752 & 0.9773 & 0.9704 & 0.9694 & **0.9815**\
& gpn & 0.9339 & 0.9511 & 0.8922 & 0.9331 & 0.9387 & 0.9570 & 0.9520 & 0.9370 & **0.9638** & 0.9594\
& gblr & 0.9291 & 0.9745 & 0.9609 & 0.9711 & 0.9589 & 0.9682 & 0.9767 & 0.9729 & 0.9679 & **0.9783**\
& ctrst & 0.8621 & 0.9345 & 0.7922 & 0.9526 & 0.9354 & 0.9207 & **0.9528** & 0.9438 & 0.9504 & 0.9450\
The highest correlation in each row is written in **boldface**.
\
It should be noted that for all results reported in this section, the HaarPSI, as well as other image quality metrics such as the SSIM, the FSIM or the VSI, were preprocessing each image by convolving it with a $2\times2$ mean filter as well as a subsequent dyadic subsampling step. This preprocessing approximates the low-pass characteristics of the optical part of the human visual system [@palmer1999vision] by a simple model.
Conclusion
==========
The HaarPSI is a novel and computationally inexpensive image quality measure based solely on the coefficients of three stages of a discrete Haar wavelet transform. Its validity with respect to the human perception of image quality was tested on four large databases containing more than 5000 differently distorted images, with very promising results. In a comparison with 9 popular state-of-the-art image similarity metrics, the HaarPSI yields significantly higher or statistically indistinguishable Spearman correlations when restricted to grayscale conversions. For color images, it only comes second to the VSI when tested on the TID 2013 (see Table \[tab:sroccdatabases\]). Along with its simple computational structure and its comparatively short execution time, this suggests a high applicability of the HaarPSI in real world optimization tasks. In particular, image quality metrics like PSNR, SSIM, or SR-SIM, that outperform the HaarPSI with respect to speed achieve considerably inferior correlations with human opinion scores (see Table \[tab:overallperf\]). Regarding the applicability of the HaarPSI in specific optimization tasks, we would like to mention that the HaarPSI has consistently high correlations with human opinion scores throughout all databases with respect to distortions caused by the JPEG and JPEG 2000 compression algorithms (see Table \[tab:details\]).
The results reported in Tables \[tab:sroccdatabases\] and \[tab:details\] might seem contradictory at first glance. In many cases, the HaarPSI yields the highest SROCC for a complete database but is outperformed by other metrics like the VSI when restricting the same database to a single distortion type. However, taking into account statistical significance, it is apparent that only when tested on the TID databases restricted to Gaussian blur, the performance of the HaarPSI is consistently lower than the performance of other similarity metrics. This particular shortcoming can be explained by the fact that the HaarPSI is almost exclusively relying on high-frequency information and thus maybe too sensitive in the case of distortions purely based on low-pass filtering.
When only considering a specific type of distortion, the correlations yielded by the HaarPSI might be improved by tuning the constants $C$ and $\alpha$, which have originally been selected to optimize the overall performance. Increasing $C$ decreases the sensitivity of the HaarPSI to changes in the high-frequency components measured by the similarity maps $\operatorname{HS^\text{(1,2)}_{f_1,f_2}}$ relative to the weights $\operatorname{W^\text{(1,2)}_{f}}$, which are based on a lower frequency band and serve as a rough model of attention-like processes. The effect of the parameter $\alpha$ on the HaarPSI is qualitatively similar when it is approaching zero. This could explain the roughly negative linear relationship between $C$ and $\alpha$ in Figure \[fig:optimization\]. However, for larger choices of $\alpha$, the function $l_\alpha(\cdot)$ is increasingly mimicking the behavior of a thresholding operator in the sense that only severe changes in the high-frequency components will have a significant effect on the HaarPSI. To also provide a quantitative analysis of these relationships, Figure \[fig:corSpearVsAlphaC1\] depicts the influence of $C$ and $\alpha$ on the correlation with human opinion scores in the case of TID 2013 with respect to six different distortions. Figure \[fig:corSpearVsAlphaC1\_blur\] indeed suggests that in the case of Gaussian blur, the performance of the HaarPSI can be improved by attenuating its sensitivity to changes in the high-frequency components via increasing $C$ and choosing $\alpha$ close to 0. In contrast, Figure \[fig:corSpearVsAlphaC1\_jpeg\] indicates that the HaarPSI achieves the highest correlations in the case of JPEG compression artifacts when it is tuned to be sensitive to severe changes in the high frequency components at highly salient locations.
\
[It is surprising that the extremely simple computational model of orientation and spatial frequency selectivity used in the HaarPSI suffices to obtain comparatively high correlations with human opinion scores. Additionally, these correlations are stable with respect to a wide range of parameters $C$ and $\alpha$ (cf. Figure \[fig:optimization\]). This could indicate that the computational structure of the HaarPSI succeeds at reproducing the *functional essence* of at least some parts of the human visual system. It is, however, quite likely that the HaarPSI owes some of its experimental success to the limitations of the used benchmark databases, which only consider a limited number of reference images and specific types of distortions. Certainly, orientation selectivity in the primary visual cortex is not restricted to horizontal and vertical edges.]{}
[Another computational principle that plays an important role in natural neural systems and that was recently successfully applied in the context of perceptual image similarity measurement is *divisive normalization* [@laparra2017perceptually]. While the similarity measure ${\operatorname{S}}(a,b,C)$ introduces some kind of normalization, divisive normalization is not included in any of the computational stages of the HaarPSI. It remains an open question if and how the HaarPSI could be further improved by incorporating divisive normalization in a similar fashion as the concepts of orientation selectivity and spatial frequency selectivity.]{}
Many practical applications demand image similarity metrics to yield values that are easy to interpret. Ideally, an image similarity of $0.9$ would in fact indicate that the average human would also assess a similarity of $\SI{90}{\%}$ between two images or that a decrease in similarity to $0.8$ corresponds to a $\SI{10}{\%}$ decrease in perceived quality for a human viewer. Due to the generality and difficulty of this task, computational models of image similarity typically only aim at establishing a monotonic relationship with human mean opinion scores, which is also reflected in the choice of the SROCC as a measure of consistency. In the case of the HaarPSI, applying $l_\alpha^{-1}(\cdot)^2$ to the final similarity score significantly linearizes its relationship with human opinion scores, thereby leading to the strong linear correlations depicted in the scatter plots in Figure \[fig:scatterplots\]. While $l_\alpha^{-1}(\cdot)^2$ is monotonically increasing on $[\frac{1}{2},1)$ and therefore not affecting the SROCC, we hope that this improves the readability and applicability of the HaarPSI. To also provide an objective measure of linear correlation, we repeated the numerical evaluation from Section \[sec:results\] with the Pearson product-moment correlation instead of the SROCC (see Table \[tab:detailpearson\] in Appendix \[app:pearsoncorrelations\]). The results of this analysis indicate that even without additional nonlinear regression, the HaarPSI has a highly linear relationship with human mean opinion scores from different databases and across varying types of distortion.
The HaarPSI can conceptually be understood as a simplified version of the FSIM. Both metrics rely on the construction of two maps, where one map measures local similarities between a reference image and a distorted image and the other map assesses the relative importance of image areas. However, in the HaarPSI, these maps are defined only in terms of a single Haar wavelet filterbank, while the FSIM utilizes an implementation of the phase congruency measure that requires the input images to be convolved with 16 complex-valued filters and contains several non-trivial computational steps, like adaptive thresholding. Another difference is that the FSIM uses the phase congruency measure both as a weight function in and as a part of the local similarity measure . In the HaarPSI, the weight function and the local similarity measure are strictly separated in the sense that they are based on distinct bands of the frequency spectrum.
These conceptual simplifications lead to a significant decrease in execution time (see Table [\[tab:overallperf\]]{}) and enable a better understanding of how single elements of the measure and properties of the input images contribute to the final similarity score. In the case of the HaarPSI, it is clear that the local similarity measure is based on high-frequency information, while the weight map, which provides a crude measure of visual saliency, is using filters that are tuned to lower frequencies. We suspect that a similar principle plays an important role in the FSIM, where additional high-frequency filters are applied to obtain the gradient map used in the local similarity measure . However, for the FSIM, it is difficult to verify this, as filters that are tuned to lower frequencies are only implicitly used in the computation of the phase congruency measure, which is in turn part of both the local similarity measure and the weight map.
We do not have a straightforward explanation as to why the HaarPSI outperforms the FSIM with respect to correlations with human opinion scores (see Table [\[tab:sroccdatabases\]]{}). After all, both measures have a similar overall structure and implement similar principles such as frequency and orientation selectivity. We assume that the reduced complexity of the HaarPSI also limits uncontrollable side effects when accentuating different aspects of the input images by varying the parameters $C$ and $\alpha$. This could improve the chance of successfully fitting subsets of benchmark databases when only considering two free parameters, but also decrease the generalization error. Furthermore, the principle of orientation selectivity is implemented differently in the HaarPSI in the sense that measurements regarding horizontal and vertical structures are only combined at the very end, that is, when taking the weighted average. It is well known that orientation selectivity is a strong organization principle in the primary visual cortex, where neurons that are tuned to similar orientations are grouped together in so-called orientation columns [[@hubel1974sequence]]{}. It thus seems reasonable that a consistent separation of the information yielded by vertical and horizontal filters has a positive effect on the correlations with human opinion scores.
From a computational point of view, it is very beneficial to apply discrete Haar wavelet filters instead of other wavelet filters. However, by changing $h^{\text{1D}}_1$ and $g^{\text{1D}}_1$ in to the respective filters, the measure given in can easily be defined for other wavelets. Table \[tab:otherwavelets\] depicts the performance of such measures based on selected Daubechies wavelets [@Daub1988], symlets [@Daub1992], coiflets [@Daub1993] and the Cohen-Daubechies-Feauveau wavelet [@CDF1992] with respect to the four databases considered in Section \[sec:results\]. It is interesting to see that Haar filters not only seem to be the computationally most efficient but also the qualitatively best choice for the measure .
[\*[7]{}[c]{}]{}\
& Daub2PSI & Daub4PSI & Sym4PSI & CDFPSI & Coif1PSI & HaarPSI\
LIVE & 0.9620 & 0.9530 & 0.9552 & 0.9604 & 0.9603 & **0.9690**\
TID2008 & 0.8971 & 0.8796 & 0.8915 & 0.8836 & 0.8965 & **0.9043**\
TID2013 & 0.8064 & 0.7982 & 0.8022 & 0.7965 & 0.8055 & **0.8094**\
CSIQ & 0.9492 & 0.9442 & 0.9454 & 0.9404 & 0.9485 & **0.9546**\
\
\
& Daub2PSI & Daub4PSI & Sym4PSI & CDFPSI & Coif1PSI & HaarPSI\
LIVE & 0.9659 & 0.9610 & 0.9630 & 0.9675 & 0.9644 & **0.9683**\
TID2008 & 0.8992 & 0.8804 & 0.8950 & 0.8932 & 0.8986 & **0.9097**\
TID2013 & 0.8724 & 0.8643 & 0.8696 & 0.8633 & 0.8716 & **0.8732**\
CSIQ & 0.9603 & 0.9577 & 0.9592 & 0.9596 & 0.9593 & **0.9604**\
The highest correlation in each row is written in **boldface**.
A Matlab function implementing the HaarPSI can be downloaded from [www.haarpsi.org](www.haarpsi.org).
Acknowledgements {#acknowledgements .unnumbered}
================
[R. Reisenhofer would like to thank Eero Simoncelli, Johannes Ballé and the members of the Laboratory for Computational Vision at NYU for their kind hospitality and very insightful discussions. S. Bosse and G. Kutyniok would like to thank Anthony Norcia for fruitful discussions. G. Kutyniok would also like to thank Eero Simoncelli for interesting discussions on the topic and to acknowledge support by the Einstein Foundation Berlin, the Einstein Center for Mathematics Berlin (ECMath), the European Commission-Project DEDALE (contract no. 665044) within the H2020 Framework Program, DFG Grant KU 1446/18, DFG-SPP 1798 Grants KU 1446/21 and KU 1446/23, the DFG Collaborative Research Center TRR 109 Discretization in Geometry and Dynamics, and by the DFG Research Center Matheon Mathematics for Key Technologies in Berlin.]{}
Pearson Product-Moment Correlations {#app:pearsoncorrelations}
===================================
[^1]: R. Reisenhofer is with the Working Group Computational Data Analysis, Universität Bremen, Fachbereich 3, Postfach 330440, 28334 Bremen, Germany (e-mail: [email protected]).
[^2]: S. Bosse is with the Fraunhofer Heinrich Hertz Institute (Fraunhofer HHI), 10587 Berlin, Germany (e-mail: [email protected]).
[^3]: G. Kutyniok is with the Department of Mathematics, Technische Universität Berlin, 10623 Berlin, Germany (e-mail: [email protected])
[^4]: T. Wiegand is with the Fraunhofer Heinrich Hertz Institute (Fraunhofer HHI), 10587 Berlin, Germany, and with the Image Communication Laboratory, Berlin Institute of Technology, 10587 Berlin, Germany (e-mail: [email protected]).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'High $T_c$ superconductors show a rich variety of phases associated with their charge degrees of freedom. Valence charges can give rise to charge ordering or acoustic plasmons in these layered cuprate superconductors. While charge ordering has been observed for both hole- and electron-doped cuprates, acoustic plasmons have only been found in electron-doped materials. Here, we use resonant inelastic X-ray scattering (RIXS) to observe the presence of acoustic plasmons in two families of hole-doped cuprate superconductors \[La$_{2-x}$Sr$_{x}$CuO$_4$ (LSCO) and Bi$_2$Sr$_{1.6}$La$_{0.4}$CuO$_{6+\delta}$ (Bi2201)\], crucially completing the picture. Interestingly, in contrast to the quasi-static charge ordering which manifests at both Cu and O sites, the observed acoustic plasmons are predominantly associated with the O sites, revealing a unique dichotomy in the behaviour of valence charges in hole-doped cuprates.'
author:
- Abhishek Nag
- 'M. Zhu'
- Matías Bejas
- 'J. Li'
- 'H. C. Robarts'
- Hiroyuki Yamase
- 'A. N. Petsch'
- 'D. Song'
- 'H. Eisaki'
- 'A. C. Walters'
- 'M. García-Fernández'
- Andrés Greco
- 'S. M. Hayden'
- 'Ke-Jin Zhou'
title: 'The nature of plasmon excitations in hole-doped cuprate superconductors'
---
High temperature superconductivity emerges in cuprates when electrons or holes are introduced to the CuO$_2$ planes of the parent antiferromagnetic insulators [@keimer2015nat]. Their electronic structure may be understood in terms of a hybridisation between the Cu $3d_{x^2-y^2}$ and O $2p_{\sigma}$ orbitals, and a strong on-site Coulomb repulsion between electrons on the Cu sites [@zaanen1985prl; @emery1987prl; @varma1987ssc; @armitage2010rmp]. When holes are introduced (see Fig. \[fig1\](a, b)), they reside preferentially in the so-called “charge-transfer band” (CTB) which is composed primarily of O orbitals [@chen1992prl]. In contrast, doped electrons enter the upper Hubbard band (UHB) and primarily reside on the Cu sites [@armitage2010rmp]. Despite this asymmetry in the electronic structure, charge order, a complex phase of periodically modulated charge-carrier density, has been observed ubiquitously on both the electron- and hole-doped sides of the phase diagram [@comin2016arcmp].
Surprisingly, a more widely observed mode of collective charge density oscillation, the acoustic plasmon [@diaconescu2007nat], has been rather elusive for the cuprates. In contrast to three-dimensional (3D) metals, where long-range Coulomb interactions give rise to isotropic long-wavelength optical-like gapped plasmons, out-of-phase oscillations of charges in neighbouring planes of two-dimensional (2D) layered electron gases, form acoustic plasmons, whose energy tends to zero for small in-plane wavevectors (see Fig. \[fig1\](c)) [@fetter1973ap; @fetter1974ap]. Due to the confinement of the doped-charges to the CuO$_2$ planes and the poor screening of out-of-plane Coulomb interactions by the intervening dielectric blocks (see Fig. \[fig1\](d)), acoustic plasmons are also expected in the layered cuprates [@kresin1988prb; @greco2019cp; @markiewicz2008prb].
A remarkable discovery has been the recent observation of acoustic plasmons in electron-doped La$_{2-x}$Ce$_x$CuO$_4$ (LCCO) and Nd$_{2-x}$Ce$_x$CuO$_4$ (NCCO) [@hepting2018nat; @lin2020npj] using Cu $L_{3}$-RIXS, owing to improvements in energy-resolution and the ability to probe in- and out-of-plane momenta independently. The excitations detected in the electron-doped superconductors showed the strong out-of-plane dispersion expected for plasmons in layered systems. The situation in hole-doped cuprates, however, has remained rather controversial. While Cu $L_3$-RIXS did not detect plasmons in several hole-doped families [@lee2014np; @dellea2017prb; @miao2017pnas], O $K$-RIXS did detect excitations in La$_{2-x}$(Br,Sr)$_x$CuO$_4$ that were interpreted as incoherent intra-band transitions [@ishii2017prb]. Also recently, electron energy-loss spectroscopy, a traditional probe for studying plasmons, found only featureless non-dispersive charge excitations in hole-doped Bi$_{2.1}$Sr$_{1.9}$CaCu$_{2}$O$_{8+\delta}$ [@mitrano2018pnas]. There is much evidence for the existence of the low-energy quasiparticle states in cuprates, for example, from angle-resolved photoemission studies [@hashimoto2008prb]. Since plasmons originate from the existence of conduction electrons, long-range nature of the Coulomb interaction, their absence in hole-doped cuprates conflicts with our general understanding of the collective behaviour of the doped-charges. Here we report that acoustic plasmons are indeed present in hole-doped cuprates. We make an extensive O $K$-RIXS study of the LSCO and Bi2201 systems over a wide range of in- and out-of-plane momenta.
Spectroscopically, Cu $L_3$-and O $K$-RIXS directly probe the charge and magnetic excitations associated with the Cu 3$d$ and O 2$p$ orbitals, respectively, at the corresponding absorption peaks (Fig. \[fig1\](a)). In order to compare the excitations associated with the two orbitals, a survey was made at both edges near the in-plane zone-centre with a fixed scattering angle on Bi2201 (Fig. \[fig1\](e, f)). The low-energy inelastic spectra at Cu $L_3$ resonance are dominated by paramagnon excitations without any noticeable signs of plasmon-like excitations, similar to previous reports on other hole-doped systems [@lee2014np; @dellea2017prb; @miao2017pnas]. At the O $K$-edge *hole*-peak however [@chen1992prl], a mode is found below 1 eV, dispersing towards the zero-energy in both materials (see also Extended Data Fig. 1).
We next collected RIXS spectra by varying the incident energy ($E_i$) across the hole-peak in O $K$-edge X-ray absorption (XAS) of LSCO (Fig. \[fig1\](b)) [@chen1992prl], as shown in Fig. \[fig1\](g). We find a broad feature at $\sim0.5$ eV shifting towards higher energies with increasing $E_i$. With doping, the probability of scattering from doped-charges in the intermediate state of RIXS increases [@bisogni2012prb]. Moreover, energy-shift of the magnetic excitations associated with incoherent charge excitations is enhanced in $\sigma$-polarised RIXS (see Methods A) [@minola2017prl]. Thus this feature, can be ascribed to bimagnon excitations, with an itinerant character [@bisogni2012prb]. Owing to the improved energy-resolution and low $h$-value (0.03), we find an additional sharp mode, at $\sim0.13$ eV, whose energy remains constant with $E_i$. This is a signature of its coherent nature [@lin2020npj] and is in contrast with previous O $K$-RIXS results [@ishii2017prb]. This feature cannot be due to two-particle electron-hole like excitations, which are incoherent in nature. Neither can it be due to single-magnons or paramagnons since $\Delta S = 1$ spin-flip processes are forbidden at the O $K$-edge [@bisogni2012prb]. To ascertain its origin, we explored further its dispersion in energy-momentum space.
The broad feature seen in Fig. \[fig1\](g) is almost non-dispersive in the $h$-direction further confirming its assignment as bimagnons [@bisogni2012prb]. This can be seen from $(h,E)$-maps collected at constant $l$-values for LSCO and Bi2201 (Fig. \[fig2\](a-d)). In contrast, the sharp mode disperses towards zero energy near the in-plane zone-centre in both systems. The dispersion and reduction of amplitude and width towards the in-plane zone-centre of this mode, is reminiscent of the acoustic plasmon behaviour observed in electron-doped LCCO [@hepting2018nat]. However, the mode is strongly damped in comparison (Extended Fig. 3(d)), reflecting the stronger correlations (e.g. pseudogap) near optimal doping in hole-doped cuprates [@armitage2010rmp; @lin2020npj].
The most stringent test for identifying the modes as plasmons in these systems is their $l$-dispersion. In the out-of-plane direction, plasmons in layered electron systems have a periodicity of $\pi/d$ (where $d$ is the interlayer spacing between the layers), which corresponds to $l=2$ in these systems, with a minimum in energy at $l=1,3,5...$ (Fig. \[fig1\](c)). LSCO and Bi2201 have interlayer spacings which differ by a factor of $\sim$2, allowing us to probe a significant portion of this period. The sharp mode observed in Fig. \[fig2\], is found to disperse to a minimum energy value at $l = 1$ for both the systems. This can be seen in the ($l, E)$-maps collected at fixed $h$-values shown in Fig. \[fig3\](a-d). This behaviour fundamentally proves the presence of acoustic plasmons in hole-doped cuprates. We can exclude the previous interpretation of these excitations as incoherent intra-band charge or electron-hole excitations which are 2D [@ishii2017prb], without significant $l$-dependence [@greco2019cp]. Our LSCO results are consistent with a limited out-of-plane dispersion study done recently [@singh2020arx].
The cuprates are strongly correlated electron systems [@keimer2015nat]. As such, it is interesting to compare our experimental results with the recently developed calculations of plasmons in the framework of a $t$-$J$-$V$ model (see Methods C and Supplementary Information IV) [@greco2019cp], although generic plasmon behaviour can also be described within random-phase-approximation calculations [@kresin1988prb; @markiewicz2008prb]. The model is given 3D character by incorporating long-range bare Coulomb interaction $V(\mathbf{Q})$ and interlayer hopping $t_z$. The imaginary part of charge susceptibility $\mathbf{\chi}_c''$($\mathbf{Q}$, $\omega$), obtained from the model resembles well the spectral shape of the plasmons for both systems, as shown in Fig. \[fig3\](f) at different ($h$, $l$)-values. This demonstrates that the charge excitations in RIXS, although influenced by resonance and polarisation effects, can fundamentally be related to the charge-density response function. At $l$-values close to 2, the much larger suppression of plasmons compared to theory could be due to their decay through the incoherent charge channels associated with bimagnons or through Umklapp scattering [@hepting2018nat]. Severe suppression of the charge excitations in this region forbids us from even detecting the optical plasmon branch in Bi2201 (see Extended Data Fig. 2). In Fig. \[fig4\](a-c) we show that both the $h$- and $l$-direction plasmon dispersions extracted by fitting the RIXS spectra for LSCO and Bi2201 are also represented well by the $t$-$J$-$V$ model. The higher acoustic plasmon velocities in Bi2201 than LSCO seen in Fig. \[fig4\](a, b), arise mainly due to the larger interlayer spacing, considering the two systems have similar carrier densities and Fermi velocities (see Supplementary Information Fig. S11) [@fetter1973ap; @hepting2018nat]. Nevertheless, the remarkable similarity of the plasmon dispersions in the two different families suggests their ubiquitous existence in hole-doped cuprates.
In order to shed light on the possible existence of plasmons at Cu sites, we use the model described above to calculate the expected plasmon energies (see solid lines) for both systems along the ($h$, $l$)-paths corresponding to the data shown in Fig. \[fig1\](e, f) and Extended Fig. 1. For Bi2201, the experiment probes near $l=3$, where the plasmons are expected to be strong. There is no evidence for plasmon excitations (Fig. \[fig1\](e)). Likewise, for LSCO, no plasmon spectral weight is observed near $l=1.8$ (Extended Fig. 1(a)). However, plasmons were observed in electron-doped LCCO using Cu $L_3$-RIXS at similar $l$-values as in LSCO [@hepting2018nat]. Our high-resolution data close to the in-plane zone-centre and wide range of $l$-values therefore indicate the absence of plasmons at the Cu sites in hole-doped cuprates [@lee2014np; @dellea2017prb]. This contrasts with the strong signals observed at the O sites.
It is interesting to discuss our findings in the context of the charge order type of density modulation observed in hole-doped cuprates using both Cu $L_3$- and O $K$-edge resonant X-ray scattering [@abbamonte2005np; @achkar2013prl; @li2020pnas]. In common with other charge-density waves, the order has a valence charge modulation and associated atomic displacements [@johannes2008prb], making it possible to be observed by non-resonant X-ray scattering techniques[@forgan2015natcomm]. As such, it is likely that, in these systems, the charge ordering signal observed at the Cu $L_3$ absorption peak is primarily owing to atomic displacements caused by electron-phonon coupling, while the dominant signal at the O $K$ hole-peak reflects directly the valence charge modulation [@achkar2013prl; @fink2009prbr]. Due to the much higher frequencies of the dynamic plasmons, it may be that they couple weakly to the phonons, thereby reducing the possibility to observe any signature directly from the Cu 3$d$ orbitals.
The discovery of acoustic plasmons in the hole-doped systems remarkably illustrates the universal existence of low-energy collective excitations besides phonons and spin-fluctuations across the cuprate phase diagram. The general existence of acoustic plasmons may stimulate more studies of charge dynamics in connection with the pseudogap phase, non-Fermi liquid behaviour and perhaps the superconductivity in cuprates [@keimer2015nat]. Our results suggest that the charge dynamics in hole-doped cuprates are mostly associated with the O sites, highlighting the importance of the three band model in the cuprates [@varma1987ssc; @emery1987prl]. Going beyond, it would be interesting to utilise the site-sensitivity of the RIXS technique to characterise plasmon behaviour in other layered superconductors, like iron-pnictides [@paglione2010nat], having strong out-of-plane band dispersions, or the newly-found nickelates in which 2D Ni-3$d$ states strongly hybridise with 3D rare-earth 5$d$ states [@hepting2020nm].
Acknowledgements {#acknowledgements .unnumbered}
================
We thank W.-S. Lee, V. Kresin, J. Lorenzana and S. Johnston for insightful discussions. All data were taken at the I21 RIXS beamline of Diamond Light Source (United Kingdom) using the RIXS spectrometer designed, built and owned by Diamond Light Source. We acknowledge Diamond Light Source for providing the beamtime on Beamline I21 under proposals SP20709 and MM24587. Work at Bristol was supported by EPSRC Grant EP/R011141/1. H. Y. was supported by JSPS KAKENHI Grant No.JP18K18744 and JP20H01856. We acknowledge T. Rice for the technical support throughout the beamtimes. We also thank G. B. G. Stenning and D. W. Nye for help on the Laue instrument in the Materials Characterisation Laboratory at the ISIS Neutron and Muon Source.
Author Contributions {#author-contributions .unnumbered}
====================
K.-J.Z., S.M.H., and A.N. conceived the project; A.N., K.-J.Z., M.Z., S.M.H., A.N.P., J.L., A.C.W., and M.G.F. performed RIXS measurements on LSCO; A.N., K.-J.Z., and M.G.F. performed RIXS measurements on Bi2201; A.N. analysed the data with support from K.-J.Z., and S.M.H.; H.C.R., M.Z. and S.M.H. prepared and characterised LSCO samples; H.E. and D.S. synthesised and characterised Bi2201 samples; M.B., H.Y., and A.G. performed theoretical calculations; A.N., K.-J.Z., and S.M.H. wrote the manuscript with input from all the authors.
Competing interests {#competing-interests .unnumbered}
===================
The authors declare no competing financial interests.
Additional information {#additional-information .unnumbered}
======================
**Correspondence and requests for materials** should be addressed to A.N., S.M.H., or K.-J.Z.
Methods
=======
RIXS measurements
-----------------
High-quality single crystals of La$_{1.84}$Sr$_{0.16}$CuO$_4$ (LSCO) and Bi$_2$Sr$_{1.6}$La$_{0.4}$CuO$_{6+\delta}$ (Bi2201) were used to collect the RIXS spectra. Superconducting transition temperatures $T_c$s from magnetisation measurements were found to be 38 K and 34.5 K for LSCO and Bi2201 respectively, consistent with optimal hole-doping of $~\sim0.16$. Samples were cooled to their respective $T_c$s prior to measurement. Lattice constants used for LSCO (Bi2201) are $a=b=3.77$ (3.86) Å and $c=13.1$ (24.69) Å. See Supplementary Information I for details of sample growth, preparation and characterisation.
RIXS spectra at O $K$-edge were collected with an energy resolution of $\bigtriangleup E\simeq$ 0.043 eV, at I21-RIXS beam line, Diamond Light Source, United Kingdom [@i21web]. We denote momentum transfers along $h$-, $k$- and $l$-directions in reciprocal lattice units where $\mathbf{Q}=ha^*+kb^*+lc^*$ ($a^*=2\pi/a$, $b^*=2\pi/b$, $c^*=2\pi/c$), $k=0$ if not stated explicitly. Additional spectra were also collected at the Cu $L_3$-edge for both the samples with an energy resolution of $\bigtriangleup E\simeq$ 0.045 eV. Samples were mounted such that the $a$-axis and $c$-axis lay in the horizontal scattering plane while the $b$-axis was perpendicular to the scattering plane (Fig. \[fig1\](d)). To map the dispersion of plasmon excitations in the ($h,l$)-plane in RIXS, one needs to continuously change the incident and scattering angles, a newly achieved capability at I21 RIXS beamline, Diamond light source, UK [@i21web]. The O $K$-edge XAS shown in Fig. \[fig1\](b) were collected in total electron yield mode. All RIXS data presented here were obtained with incident $\sigma$ polarisation (perpendicular to the scattering plane) to enhance the charge excitations [@hepting2018nat]. The zero-energy transfer position and resolution of the RIXS spectra were determined from subsequent measurements of elastic peaks from an adjacent carbon tape.
RIXS data fitting
-----------------
RIXS data were normalised to the incident photon flux, and subsequently corrected for self-absorption effects using the procedure described in [@minola2015prl] prior to fitting. A Gaussian lineshape with the experimental energy resolution was used to fit the elastic line. Gaussian lineshapes were also used to fit the low energy phonon excitations at $\sim 0.045$ eV and their overtones. Anti-symmetrised Lorentzian functions were used to fit the plasmon and the bimagnon peaks (See Supplementary Information III). First we extracted the zone-centre energy, amplitude and width of the broad incoherent mode at $h=0.01$ and concluding this to be a bimagnon, fixed its amplitude and width for the whole ($h$, $l$)-range [@bisogni2012prb; @vernay2007prb]. The energy values of the bimagnons were allowed to vary within $\pm20$ meV for the RIXS spectra along $h$-direction. For the RIXS spectra along $l$-direction, bimagnon energies were kept fixed to the values obtained for corresponding $h$-values from the $h$-direction scans. This allowed us to decompose the inelastic spectra into two components with less ambiguity, especially for the $h$-values where energies of the two modes were nearby. Significant correlations were however found below $h<0.02$, between the elastic, phonon and plasmon amplitudes and energies, and hence the plasmon energy values determined in these regions are less conclusive. A high energy quadratic background was also included in the fitting model to account for the tailing contribution from $dd$-excitations above 1.5 eV. Representative RIXS spectra fits with the different peaks are shown at the top of Fig. \[fig2\](e, f), while rest are presented in Supplementary Information III.
$t$-$J$-$V$ model
-----------------
For discussing the nature and origin of three-dimensional charge excitations in LSCO and Bi2201 we employed the minimal layered $t$-$J$-$V$ model [@greco2019cp; @zhang1988prb]:
$$H = -\sum_{i, j,\sigma} t_{i j}\tilde{c}^\dag_{i\sigma}\tilde{c}_{j\sigma}
+ \sum_{\langle i,j \rangle} J_{ij} \left( \vec{S}_i \cdot \vec{S}_j - \frac{1}{4} n_i n_j \right)
+\frac{1}{2} \sum_{i,j} V_{ij} n_i n_j.
\label{tJV}$$
See Supplementary Information IV for the details of the hopping parameters $t_{i j}$ and the exchange parameter $J_{i j}$ used in the model. The interlayer hopping $t_z$ was also included in $t_{i j}$. The form of long-range Coulomb interaction $V_{i j}$ used in Eq. \[tJV\] in momentum space is [@becca1996prb]: $$V(\mathbf{Q})=\frac{V_c}{A(q_x,q_y) - \cos q_z},
\label{LRC}$$
where $V_c= e^2 d(2 \epsilon_{\perp} a^2)^{-1}$ and $A(q_x,q_y)= \alpha (2 - \cos q_x - \cos q_y)+1$ with $\alpha=\frac{\epsilon_\parallel/\epsilon_\perp}{(a/d)^2}$, $e$ the elementary charge and high frequency in-($\epsilon_{\parallel}$) and out-of-plane ($\epsilon_{\perp}$) dielectric constants.
The model Hamiltonian Eq. \[tJV\] contains several material dependant parameters. To reduce the number of tuning parameters in the model, we have used, if available, the most common values from literature. The optical plasmon ($\omega_{\textrm{opt}}$) energy was fixed at 0.85 eV for both LSCO and Bi2201 as reported from optical measurements which are sensitive to only the $l=0$ momentum transfers [@heumen2009njp; @uchida1991prb; @suzuki1989prb]. The exchange coupling parameter $J$ was taken as $0.3t$ for both the materials given their similar values in the parent compounds [@hybertsen1990prb; @peng2017np; @lee2006rmp]. It is important to mention here that plasmons are nearly unaffected by the value of $J$. The nearest neighbour hopping parameter $t/2$ was fixed at 0.35 eV for both the materials [@ishii2017prb; @horio2018prl]. We took the second-nearest neighbour hopping parameter $t'$, whose exact value is not known apart from reports that $|t'_{\textrm{LSCO}}|<|t'_{\textrm{Bi2201}}|$, to be $-0.2t$ for LSCO and $-0.35t$ for Bi2201 [@hashimoto2008prb; @pavarini2001prl; @greco2019cp; @horio2018prl]. We however found that the overall nature of plasmon dispersions does not depend on a precise choice of the band parameters.
We found the optical plasmon energy to be proportional to $\sqrt{V_c/\alpha}$ of Eq. \[LRC\]. Depending on the band parameters, we obtained the proportionality constant to be 0.41 for LSCO and 0.40 for Bi2201. With the band parameters fixed, the value of $V_c/\alpha$ was tuned to get $\omega_{\textrm{opt}}=0.85$ eV. Next, $V_c$ and $\alpha$ values were optimised to best match the plasmon dispersions observed in the experiment. For LSCO (Bi2201), we obtained $V_c$ and $\alpha$ to be 18.9 eV (52.5 eV) and 3.47 (8.14) respectively. This gave for LSCO (Bi2201), $\epsilon_\perp$ as $2.21\epsilon_0$ ($1.43\epsilon_0$) and $\epsilon_\parallel$ as $1.62\epsilon_0$ ($1.14\epsilon_0$). Microscopically therefore the relatively large values of $V_c$ and $\alpha$ in Bi2201 come mainly from a large interlayer spacing $d$. The interlayer hopping $t_z$ was tuned to match the plasmon energies for $h\rightarrow0.0$ and an upper limit of $0.01t$ was found for both LSCO and Bi2201 (See Supplementary Information Fig. S10).
[100]{} Keimer, B., Kivelson, S. A., Norman, M. R., Uchida, S. & Zaanen, J. From quantum matter to high-temperature superconductivity in copper oxides. *Nature* [**518**]{}, 179 (2015). Zaanen, J., Sawatzky, G. A. & Allen, J. W. Band gaps and electronic structure of transition-metal compounds, *Phys. Rev. Lett.* **55**, 418 (1985). Emery, V. J. Theory of high-$T_c$ superconductivity in oxides, *Phys. Rev. Lett.* **58**, 2794 (1987). Varma. C. M., Schmitt-Rink S. & Abrahams, E. Charge transfer excitations and superconductivity in“ionic” metals, *Sol. State. Comm.* **62**, 681 (1987). Armitage, N. P., Fournier, P. & Greene, R. L. Progress and perspectives on electron-doped cuprates, *Rev. Mod. Phys.* **82**, 2421 (2010). Chen, C. T. et al. Out-of-plane orbital characters of intrinsic and doped holes in La$_{2-x}$Sr$_x$CuO$_4$. *Phys. Rev. Lett.* **68**, 2543 (1992). Comin, R. & Damascelli, A. Resonant X-ray scattering studies of charge order in cuprates. *Ann. Rev. Cond. Matter Phys.* **7**, 369 (2016). Diaconescu, B. et al. Low-energy acoustic plasmons at metal surfaces. *Nature* **448**, 57 (2007). Fetter, A. Electrodynamics of a layered electron gas. I. Single layer. *Ann. Phys.* **81**, 367 (1973). Fetter, A. L. Electrodynamics of a layered electron gas II. Periodic array. *Ann. Phys.* **88**, 1 (1974). Kresin, V. Z. & Morawitz, H. Layer plasmons and high-$T_c$ superconductivity. *Phys. Rev. B* **37**, 7854 (1988). Greco, A., Yamase, H. & Bejas, M. Origin of high-energy charge excitations observed by resonant inelastic X-ray scattering in cuprate superconductors. *Commun. Phys.* **2**, 3 (2019). Markiewicz, R. S., Hasan, M. Z., & Bansil, A. Acoustic plasmons and doping evolution of Mott physics in resonant inelastic X-ray scattering from cuprate superconductors. *Phys. Rev. B* **77**, 094518 (2008). Hepting, M. et al. Three-dimensional collective charge excitations in electron-doped copper oxide superconductors. *Nature* **563**, 374 (2018). Lin, J. et al. Doping evolution of the charge excitations and electron correlations in electron-doped superconducting La$_{2-x}$Ce$_x$CuO$_4$. *npj Quantum Mater.* **5**, 4 (2020). Lee, W. et al. Asymmetry of collective excitations in electron- and hole-doped cuprate superconductors. *Nat. Phys.* **10**, 883 (2014). Dellea, G. et al. Spin and charge excitations in artificial hole- and electron-doped infinite layer cuprate superconductors. *Phys. Rev. B* **96**, 115117 (2017). Miao, H. et al. High-temperature charge density wave correlations in La$_{1.875}$Ba$_{0.125}$CuO$_4$ without spin–charge locking. *Proc. Natl Acad. Sci. USA* **114**, 12430 (2017). Ishii, K. et al. Observation of momentum-dependent charge excitations in hole-doped cuprates using resonant inelastic X-ray scattering at the oxygen *K* edge. *Phys. Rev. B* **96**, 115148 (2017). Mitrano, M. et al. Anomalous density fluctuations in a strange metal. *Proc. Natl. Acad. Sci. USA* **115**, 5392 (2018). Hashimoto, M. et al. Doping evolution of the electronic structure in the single-layer cuprate Bi$_2$Sr$_{2-x}$La$x$CuO$_{6+\delta}$: comparison with other single-layer cuprates. *Phys. Rev. B* **77**, 094516 (2008). Bisogni, V. et al. Bimagnon studies in cuprates with resonant inelastic X-ray scattering at the O *K* edge. II. Doping effect in La$_{2-x}$Sr$_x$CuO$_4$. *Phys. Rev. B* **85**, 214528 (2012). Minola, M. et al. Crossover from collective to incoherent spin excitations in superconducting cuprates probed by detuned resonant inelastic X-ray scattering. *Phys. Rev. Lett.* **119**, 097001 (2001). Singh, A. et al. Acoustic plasmons and conducting carriers in hole-doped cuprate superconductors. *arXiv:2006.13424v1* (2020).
Abbamonte, P. et al. Spatially modulated ‘Mottness’ in La$_{2-x}$Ba$_x$CuO$_4$. *Nature Phys.* **1**, 155 (2005). Achkar, A. J. et al. Resonant X-ray scattering measurements of a spatial modulation of the Cu 3$d$ and O 2$p$ energies in stripe-ordered cuprate superconductors. *Phys. Rev. Lett.* **110**, 017001 (2013). Li, J. et al. Multi-orbital charge density wave excitations and concomitant phonon anomalies in Bi$_2$Sr$_2$LaCuO$_{6+\delta}$ *Proc. Natl. Acad. Sci. USA* ****, (2020). Johannes, M. D. & Mazin, I. I. Fermi surface nesting and the origin of charge density waves in metals, *Phys. Rev. B* **77**, 165135 (2008). Forgan, E. M. et al., The microscopic structure of charge density waves in underdoped YBa$_2$Cu$_3$O$_{6.54}$ revealed by X-ray diffraction, *Nat. Comms.* **6**, 10064 (2015). Fink, J. et al. Charge ordering in La$_{1.8-x}$Eu$_{0.2}$Sr$_x$CuO$_4$ studied by resonant soft X-ray diffraction. *Phys. Rev. B* **79**, 100502(R) (2009). Paglione, J. & Greene, R. L. High-temperature superconductivity in iron-based materials. *Nat. Phys.* **6**, 645 (2010). Hepting, M. et al. Electronic structure of the parent compound of superconducting infinite-layer nickelates. *Nat. Mater.* **19**, 381 (2020).
https://www.diamond.ac.uk/Instruments/Magnetic-Materials/I21.html Minola, M. et al. Collective Nature of Spin Excitations in Superconducting Cuprates Probed by Resonant Inelastic X-Ray Scattering. *Phys. Rev. Lett.* **114**, 217003 (2015). Vernay, F. H., Gingras, M. J. P. & Devereaux, T. P. Momentum-dependent light scattering in insulating cuprates. *Phys. Rev. B* **75**, 020403(R) (2007). Zhang, F. C. & Rice, T. M. Effective Hamiltonian for the superconducting Cu oxides. *Phys. Rev. B* **37**, 3759 (1988). Becca, F., Tarquini, M., Grilli, M. & Di Castro, C. Charge-density waves and superconductivity as an alternative to phase separation in the infinite-U Hubbard-Holstein model. *Phys. Rev. B* **54**, 12443 (1996). van Heumen, E., Meevasanam W., Kuzmenko, A. B., Eisaki, H. & van der Marel, D. Doping-dependent optical properties of Bi2201. *New J. Phys.* **11**, 055067 (2009). Uchida, S. et al. Optical spectra of La$_{2-x}$Sr$_x$CuO$_4$: Effect of carrier doping on the electronic structure of the CuO$_2$ plane. *Phys. Rev. B* **43**, 7942 (1991). Suzuki, M. Hall coefficients and optical properties of La$_{2-x}$Sr$_x$CuO$_4$ single-crystal thin films. *Phys. Rev. B* **39**, 2312 (1989). Hybertsen, M. S., Stechel, E. B., Schluter, M. & Jennison, D. R. Renormalization from density-functional theory to strong-coupling models for electronic states in Cu-O materials. *Phys. Rev. B* **41**, 11068 (1990). Peng, Y. Y. et al. Influence of apical oxygen on the extent of in-plane exchange interaction in cuprate superconductors. *Nat. Phys.* **13**, 1201 (2017). Lee, P. A., Nagaosa, N. & Wen, X.-G. Doping a Mott insulator: physics of high-temperature superconductivity. *Rev. Mod. Phys.* **78**, 17 (2006). Horio, M. et al. Three-dimensional Fermi surface of overdoped La-based cuprates. *Phys. Rev. Lett.* **121**, 077004 (2018). Pavarini, E., Dasgupta, I., Saha-Dasgupta, T., Jepsen, O. & Andersen O. K. Band-structure trend in hole-doped cuprates and correlation with $T_{c}$$_{\textrm{max}}$. *Phys. Rev. Lett.* **87**, 047003 (2001).
![**Probing plasmons in hole-doped cuprates using RIXS.** **a,** Schematic electronic stucture of cuprates in the Zaanen-Sawatzky-Allen scheme [@zaanen1985prl; @emery1987prl; @varma1987ssc; @armitage2010rmp]. Holes are doped into the charge transfer band and electons are doped into the upper Hubbard band. Vertical arrows represent the principal excitation processes occurring at the O $K$-and Cu $L_3$-XAS peaks. **b,** Peak feature attributed to the doped-holes observed through O $K$-XAS for the two systems [@chen1992prl]. The arrow shows the incident energy chosen to probe the plasmon dispersions in this work. **c,** Acoustic plasmon branches dispersing towards zero-energy in the $h$-direction while maintaining a periodicity of $l=2$ in the $l$-direction, due to the presence of a CuO$_2$ layer in the middle of a unit cell in single-layered systems. **d,** Representative unit cell of single-layered high-$T_c$ cuprates, and the RIXS scattering geometry showing incoming ($k_{in}$) and scattered ($k_{out}$) X-ray beams. **e** and **f,** RIXS intensity maps for Bi2201 at a fixed scattering angle of 114$^{\circ}$ at Cu $L_3$- and O $K$-edges respectively. **f,** Incident energy detuned, vertically stacked RIXS spectra, showing their decomposition into a sharp Raman-like peak and a broad energy-shifting peak for LSCO at ( $h=0.03$, $l=1.00$). Colours of the spectra correspond to incident energies shown in **b**. Solid black lines are fits to the RIXS spectra (see Methods B and Supplementary Information III). Vertical black bars are least-square-fit peak positions. []{data-label="fig1"}](Fig1_red.pdf)
![**In-plane plasmon dispersions in LSCO and Bi2201.** RIXS intensity maps of **a, b,** LSCO for momentum transfer along the $h$-direction at $l=0.60$ and $l=1.00$ and **c, d,** of Bi2201 at $l=1.75$ and $l=1.50$. Green circle symbols indicate the least-square-fit peak positions of the plasmon excitations. Brown square symbols show the fitted bimagnon peak positions. **e, f,** Top, representative RIXS spectra (blue and red circle symbols) at mentioned ($h$, $l$)-values for LSCO and Bi2201. Shaded areas represent the different peak profiles in the energy transfer region below 1.5 eV. Black solid lines are fits to the RIXS spectra. See Methods B for details of the decomposition procedure. **e, f,** Bottom, vertically stacked RIXS spectra at indicated ($h$, $l$)-values. Vertical black bars are least-square-fit plasmon peak positions. The insets indicate the directions in the $h$, $l$-plane, along which the shown RIXS spectra have been collected. []{data-label="fig2"}](Fig2.pdf)
![**Out-of-plane plasmon dispersions in LSCO and Bi2201.** RIXS intensity maps of **a, b,** LSCO and of **c, d,** Bi2201 for momentum transfer along the $l$-direction at $h=0.03$ and $h=0.05$. Green circle symbols indicate the least-square-fit peak positions of the plasmons. Brown dashed line shows the extracted bimagnon energies from the $h$-direction scans. **e,** Representative RIXS spectra at mentioned ($h$, $l$)-values for LSCO (blue circle symbols) and Bi2201 (red circle symbols). From the RIXS spectra, fitted peak profiles from bimagnons have been subtracted to show the evolution of the plasmon peak (green shaded area). The inset indicates the direction in the ($h$, $l$)-plane, along which the shown RIXS spectra have been collected. **f,** $\mathbf{\chi}_c''$($\mathbf{Q}$, $\omega$) calculated from $t$-$J$-$V$ model for corresponding ($h$, $l$)-values in **e** for LSCO (Bi2201) are shown by the dotted blue (red) lines. Shaded green areas are the plasmon peak profiles obtained from RIXS data fitting. Different broadening factors ($\Gamma=0.29t$ for Bi2201 and $0.2t$ for LSCO) were chosen to replicate the lineshapes of the two materials. []{data-label="fig3"}](Fig3.pdf)
![**Comparison of plasmon dispersions obtained from RIXS with $t$-$J$-$V$ model LSCO and Bi2201.** **a, b,** Plasmon energies in LSCO (blue symbols) and Bi2201 (red symbols) for momentum transfer along the $h$-direction for different $l$-values and **c,** for momentum transfer along the $l$-direction for different $h$-values, summarised from least-square-fits of RIXS spectra. Continuous lines are plasmon dispersions obtained from optimised $t$-$J$-$V$ model. Also compared are the plasmon dispersions from electron-doped La$_{2-x}$Ce$_x$CuO$_4$ ($x$=0.175) (green lines with symbols) [@hepting2018nat]. []{data-label="fig4"}](Fig4.pdf)
![**Comparison of Cu $L_3$-and O $K$-RIXS for probing plasmons in LSCO and Bi2201.** **a, b,** RIXS intensity maps of LSCO for momentum transfer along the $h$-direction and **c, d,** of Bi2201 at Cu $L_3$ absorption and O $K$ hole-peaks. The spectra are collected with a fixed scattering angle (2$\theta$) for each map and hence the ($h, l$)-values vary simultaneously. Green solid lines show the expected plasmon dispersions from $t$-$J$-$V$ model. **e, f,** Representative RIXS spectra for LSCO and Bi2201 at given ($h$, $l$)-values for Cu $L_3$- (square symbols) and O $K$- (circle symbols) edges. **g,** RIXS spectra for Bi2201 at given ($h$, $l$)-values for Cu $L_3$-edge. []{data-label="ex_fig1"}](Edge_comparison_v2.pdf){width="\textwidth"}
![**Optical plasmon dispersions in Bi2201.** **a, b,** RIXS intensity maps of Bi2201 for momentum transfer along the $h$-direction at $l=1.50$ and $l=2.00$. **c, d,** RIXS intensity maps of Bi2201 after subtraction of bimagnon excitations from **a** and **b**. **e,** Zoomed in map of **d**. Continuous line in **d, e,** are the expected optical plasmon dispersion. For both $h$ close to 0 and $l$ close to 2 the plasmon amplitudes reduce substantially (see Extended Data Fig. 3). Although some residual spectral weight is visible for slightly higher $h$-values at $l=2.00$, the optical plasmon branch still remains intangible. []{data-label="ex_fig3"}](Optical_plasmon.pdf)
![**Plasmon amplitude and width variation along $h$- and $l$-directions in LSCO and Bi2201.** **a, b,** and **d, e,** Plasmon amplitudes and widths in LSCO (blue symbols) and Bi2201 (red symbols) for momentum transfer along the $h$-direction for different $l$-values and **c,** and **f,** for momentum transfer along the $l$-direction for different $h$-values, summarised from least-square-fits of RIXS spectra. Continuous lines are integrated spectral weights of $\mathbf{\chi}_c''$($\mathbf{Q}$, $\omega$) calculated within the $t$-$J$-$V$ model. To make an appropriate comparison, all the calculated amplitude values were scaled to the experimental values at $h=0.05, l=1.00$ for LSCO and $h=0.05, l=1.50$ for Bi2201. Also compared are the plasmon widths from electron-doped La$_{2-x}$Ce$_x$CuO$_4$ ($x$=0.175) (green line with symbols) [@hepting2018nat]. []{data-label="ex_fig3"}](Amplitude_and_fwhm.pdf)
[ **** ]{}
Sample growth, preparation and characterisation.
================================================
High-quality single crystals of La$_{1.84}$Sr$_{0.16}$CuO$_4$ (LSCO) and Bi$_2$Sr$_{1.6}$La$_{0.4}$CuO$_{6+\delta}$ (Bi2201) were grown by floating-zone method. Bi2201 sample was annealed at 650 $^{\circ}$C in O$_2$ atmosphere for two days to improve sample homogeneity. Fig. \[Sample\_info\](a) and (b) show the Laue diffraction patterns obtained from LSCO and Bi2201 respectively, for preliminary orientation. The orientations were further refined using in-situ diffraction, charge-order and superstructure (present in Bi2201 due to structural distortions in Bi-O layers) peaks prior to the collection of RIXS spectra. Superconducting transition temperatures $T_c$s extracted from magnetisation measurements of LSCO (38 K) and Bi2201 (34.5 K) are shown in Fig. \[Sample\_info\](c) and (d). While LSCO was cleaved in vacuum, Bi2201 was cleaved in air and immediately transferred to vacuum. LSCO and Bi2201 samples were cooled down to their respective $T_c$s prior to measurement and the pressure inside the sample vessel was maintained at $\sim5\times10^{-10}$ mbar. Negative and positive values of $h$ in the RIXS maps presented in this work represent the grazing-incident and grazing-exit geometries respectively (Fig. \[Sample\_info\](e,f)). Fig. \[Sample\_info\](g) shows the accessible ($h$, $l$)-values at O $K$-for LSCO and Bi2201 at I21. For each material, reduction in $l$-value forces transition from a backward to forward scattering experimental geometry which enhances the elastic line, thereby limiting the lowest usable $l$-to study plasmons.
![**Sample properties.** **a, b,** Laue diffraction patterns of LSCO and Bi2201 samples. **c, d,** Zero-Field-Cooled (ZFC) and Field-Cooled (FC) magnetisation curves for LSCO and Bi2201. The onset of the superconducting transition $T_c$ is shown by the vertical dotted lines. **e,** Grazing-incident and **f,** grazing-exit scattering geometries. **g,** Shows the accessible ($h$, $l$)-values at O $K$-for LSCO and Bi2201 at I21. Negative and positive values of $h$ represent the grazing-in and grazing-exit configurations respectively. Vertical and horizontal lines show the ($h$-, $l$)-trajectories along which RIXS spectra were collected in this work. []{data-label="Sample_info"}](Sample_info.pdf)
Additional RIXS data
====================
RIXS intensity map of LSCO for momentum transfer along the $h$-direction at $l=0.80$ is shown in Fig. \[Extended\_hscans\](a). The plasmon energies from these scans have been used in Fig. 4(a) of the main paper. Extended $h$-scans are shown in Fig. \[Extended\_hscans\](b, c) till $h=0.15$, showing the continuous rise of the plasmon energies towards the $dd$ excitations in Bi2201. RIXS intensity map of LSCO for momentum transfer along the $l$-direction at $h=0.08$ is shown in Fig. \[Extended\_lscans\](a). The plasmon energies from these scans have been used in Fig. 4(c) of the main paper. In Fig. \[Extended\_lscans\](b) RIXS intensity map of Bi2201 for momentum transfer along the $l$-direction at $h=0.02$ is shown. Although plasmon dispersion is visible for $h=0.02$ in Bi2201, the energies could not be extracted by fitting the data, due to either proximity to elastic line below $l=1.5$ or weak spectral weight above $l=1.5$. The over-plotted continuous dispersion line obtained from $t$-$J$-$V$ model seemingly follows the plasmon spectral weight. Plasmon excitations are also observed along the $k$ directions in LSCO (see Fig. \[Kscan\_LSCO\](a)) as expected from the 4-fold symmetry of orbitals in the CuO$_2$ planes. As such, using the same parameters of the $t$-$J$-$V$ model optimised for plasmon dispersions along $h$- and $l$-directions, the plasmon excitations along the $k$ direction can be reproduced (see Fig. \[Kscan\_LSCO\](b)).
Close to the zone-centre, high energy-resolution scans are need to differentiate the plasmon peak from the elastic and the phonon peaks. This is important if one wants to study the validity of the $t$-$J$-$V$ model with interlayer hopping which predicts a zone-centre gap for the acoustic plasmons. From the current results we can only estimate the upper limit of the acoustic plasmon energies at the zone-centre to be $\approx0.075$ eV for LSCO, setting interlayer hopping $t_z\lesssim0.007$ eV in this material (see Fig. \[gamma\]).
![**Additional in-plane plasmon dispersions in LSCO and Bi2201.** **a,** RIXS intensity map of LSCO for momentum transfer along the $h$ direction at $l=0.80$ and **b, c,** of Bi2201 at $l=1.50$ and $l=1.75$. Green circle symbols indicate the least-square-fit peak positions of the plasmon excitations. Fitted plasmon energies for LSCO have been used in Fig. 4(a) of the main paper. []{data-label="Extended_hscans"}](Extended_hscans.pdf)
![**Additional out-of-plane plasmon dispersions in LSCO and Bi2201.** **a,** RIXS intensity map of LSCO for momentum transfer along the $l$-direction at $h=0.80$ and **b,** of Bi2201 at $h=0.02$. Green circle symbols indicate the least-square-fit peak positions of the plasmon excitations. Error bars are smaller than the symbols. Continuous line shows the expected plasmon dispersion from $t$-$J$-$V$ model. []{data-label="Extended_lscans"}](Extended_lscans.pdf)
![**Plasmon dispersion in LSCO along $k$-direction.** **a,** RIXS intensity map of LSCO for momentum transfer along the $k$ direction at ($h=0.04$, $l=1.0$). **b,** Intensity map of the charge susceptibility corresponding to the spectra shown in **a**, calculated using $t$-$J$-$V$ model. []{data-label="Kscan_LSCO"}](Kscan_LSCO.pdf)
RIXS data fitting
=================
All RIXS data shown in this paper were fitted following the procedure described in Methods B. The RIXS spectra fits of LSCO for incident energy detuned scans are shown in Fig. \[LSCO\_edfits\]. Also in Fig. \[LSCO\_edfits\](b) are shown the change in plasmon and bimagnon amplitudes and widths as the incident energy is varied. The non-resonant behaviour of the bimagnon amplitude implies an incoherent character in sharp contrast to the plasmon. RIXS spectra fits of LSCO and Bi2201 for momentum transfer along the $h$-direction are shown in Fig. \[LSCO\_hfits\] and \[Bi2201\_hfits\], and along the $l$-direction in Fig. \[LSCO\_lfits\] and \[Bi2201\_lfits\] respectively. The scattering intensities $S$($\mathbf{Q}$, $\omega$) of the plasmons and bimagnons at given values of $\mathbf{Q}=ha^*+kb^*+lc^*$ ($a^*=2\pi/a$, $b^*=2\pi/b$, $c^*=2\pi/c$), dependent on the imaginary part of their respective dynamic susceptibilities $\mathbf{\chi}''$($\mathbf{Q}$, $\omega$) were modelled as: $$S(\mathbf{Q}, \omega) \propto \frac{\mathbf{\chi}''(\mathbf{Q}, \omega)}{1- e^{-\hbar\omega/k_BT} },$$ where $k_B$, $T$ and $\hbar$ are the Boltzmann constant, temperature and the reduced Planck constant. A generic damped harmonic oscillator model can be used for the response function
$$\mathbf{\chi}''(\mathbf{Q}, \omega) \propto \frac{\gamma\omega}{\left[\omega^2-\omega^2_0\right]^2+4\omega^2\gamma^2},
\label{dho}$$
where $\omega_0$ and $\gamma$ are the undamped frequency and the damping factor respectively. Eq. \[dho\] can be equivalently written using an anti-symmetrised Lorentzian function, $$\frac{1}{\omega_p}\left[\frac{\gamma}{(\omega-\omega_p)^2+\gamma^2}-\frac{\gamma}{(\omega+\omega_p)^2+\gamma^2}\right],$$
with peaks at $\pm\omega_p$ for $\omega_p^2=\omega_0^2-\gamma^2$, given that $\gamma\le\omega_0$, which was found to hold for the plasmon excitations observed in this study (see Extended Data Fig. 3). In the results, we plotted the plasmon propagation energy as the peak $\omega_p$ of this function.
![**RIXS spectra fits of LSCO for varying incident energies.** **a,** Fits to the RIXS spectra of LSCO at ($h=0.03$, $l=1.00$) for mentioned incident energies about the *hole*-peak in O $K$-edge XAS presented in Fig. 1(b) of main paper are shown by using the model described in Section III. The extracted plasmon and bimagnon peak energies have been used in Fig. 1(g) of main paper. **b,** The plasmon and bimagnon amplitudes (circle and square red symbols) and widths (circle and square blue symbols) as a variation of incident energy is plotted. []{data-label="LSCO_edfits"}](LSCO_edfits.pdf)
![**RIXS spectra fits of LSCO for momentum transfer along the $h$ direction.** **a, b, c,** and **d,** Fits to the RIXS spectra of LSCO presented in Fig. 2 (a, b, e) of main paper at mentioned ($h$, $l$)-values are shown by using the model described in Section III. The extracted plasmon peak energies have been used in Fig. 2(a, b, e) and Fig. 4(a) in the main paper and Fig. \[Extended\_hscans\]. The plasmon amplitudes and widths have been used in Extended Data Fig. 3 (a, d) of the main paper. []{data-label="LSCO_hfits"}](LSCO_hfits.pdf)
![**RIXS spectra fits of Bi2201 for momentum transfer along the $h$ direction.** **a, b, c,** and **d,** Fits to the RIXS spectra of Bi2201 presented in Fig. 2 (c, d, f) of main paper at mentioned ($h$, $l$)-values are shown by using the model described in Section III. The extracted plasmon peak energies have been used in Fig. 2(c, d, f) and Fig. 4(b) in the main paper and Fig. \[Extended\_hscans\]. The plasmon amplitudes and widths have been used in Extended Data Fig. 3 (b, e) of the main paper. []{data-label="Bi2201_hfits"}](Bi2201_hfits.pdf)
![**RIXS spectra fits of LSCO for momentum transfer along the $l$-direction.** **a, b,** and **c,** Fits to the RIXS spectra of LSCO presented in Fig. 3 (a, b) of main paper at mentioned ($h$, $l$)-values are shown by using the model described in Section III. The extracted plasmon peak energies have been used in Fig. 3(a, b) and Fig. 4(c) in the main paper and Fig. \[Extended\_lscans\]. In Fig. 3(e) of the main paper, RIXS spectra subtracted by the fitted bimagnon contribution have been shown. The plasmon amplitudes and widths have been used in Extended Data Fig. 3 (c, f) of the main paper. []{data-label="LSCO_lfits"}](LSCO_lfits.pdf)
![**RIXS spectra fits of Bi2201 for momentum transfer along the $l$-direction.** **a, b, c,** and **d,** Fits to the RIXS spectra of Bi2201 presented in Fig. 3 (c, d) of main paper at mentioned ($h$, $l$)-values are shown by using the model described in Section III. The extracted plasmon peak energies have been used in Fig. 3(c, d) and Fig. 4(c) in the main paper. In Fig. 3(e) of the main paper, RIXS spectra subtracted by the fitted bimagnon contribution in this manner have been shown. The plasmon amplitudes and widths have been used in Extended Data Fig. 3 (c, f) of the main paper. []{data-label="Bi2201_lfits"}](Bi2201_lfits.pdf)
$t$-$J$-$V$ model
=================
We used the following form of the layered $t$-$J$-$V$ model [@greco2016prb]:
$$H = -\sum_{i, j,\sigma} t_{i j}\tilde{c}^\dag_{i\sigma}\tilde{c}_{j\sigma}
+ \sum_{\langle i,j \rangle} J_{ij} \left( \vec{S}_i \cdot \vec{S}_j - \frac{1}{4} n_i n_j \right)
+\frac{1}{2} \sum_{i,j} V_{ij} n_i n_j.
\label{tJV}$$
On each plane the hopping $t_{i j}$ takes a value $t$ $(t')$ between the first (second) nearest-neighbours sites on the square lattice and $J$ is the exchange interaction between the nearest-neighbours. Since hole-doped cuprates are correlated electron systems the $t$-$J$ model is believed to be a minimal model of the CuO$_2$ planes [@zhang1988prb]. The fact that we deal with a correlated system is contained in $\tilde{c}^\dag_{i\sigma}$ and $\tilde{c}_{i\sigma}$ which are the creation and annihilation operators, respectively, of electrons with spin $\sigma(=\uparrow, \downarrow)$ in the Fock space without any double occupancy. $n_i$ is the electron density operator and $\vec{S}_i$ the spin operator. The three-dimensional nature of the model is considered by the presence of a hopping $t_z$ between the adjacent planes, and the long-range Coulomb interaction $V_{ij}$ for a layered system. The form of $V_{i j}$ in Eq. \[tJV\] in momentum space is given in Eq. 2 of main paper [@becca1996prb]. The exchange interaction $J$ is considered only inside the plane. The out-of-plane exchange term is much smaller than $J$ [@theo1988prb]. Finally, the indices $i$ and $j$ run over the sites of a three-dimensional lattice, and $\langle i,j \rangle$ indicates a pair of nearest-neighbour sites.
A theoretical treatment of this model is non-trivial because the Hamiltonian is defined in the restricted Hilbert space that prohibits non-double occupancy on each site, which complicate the commutation rules of the operators. In addition, there is no small parameter for perturbation. We implement a large-$N$ expansion [@foussats2004prb; @greco2016prb; @greco2019cp] where the spin index $\sigma$ is extended to a new index $p$ running from $1$ to $N$. In order to get a finite theory in the limit $N\rightarrow \infty$, we rescale the hopping $t_{ij}$ to $t_{ij}/N$, $J$ to $J/N$ and $V_{ij}$ to $V_{ij}/N$, and $1/N$ is used as the small parameter to control the expansion. $N$ is put to $N=2$ in the end. Although the physical value is $N=2$, the large-$N$ expansion has several advantages over usual perturbations theories. Applying the large-$N$ treatment [@greco2016prb] the quasiparticles disperse in momentum space as
$$\epsilon_{\mathbf{k}} = \epsilon_{\mathbf{k}}^{\parallel} + \epsilon_{\mathbf{k}}^{\perp}
\label{Ek}$$
where the in-plane $\epsilon_{\mathbf{k}}^{\parallel}$ and the out-of-plane $\epsilon_{\mathbf{k}}^{\perp}$ dispersions are given by, respectively, $$\epsilon_{\mathbf{k}}^{\parallel} = -2 \left( t \frac{\delta}{2}+\Delta \right) (\cos k_{x}+\cos k_{y})
- 4t' \frac{\delta}{2} \cos k_{x} \cos k_{y} - \mu
\label{Epara}$$ and $$\epsilon_{\mathbf{k}}^{\perp} = 2 t_{z} \frac{\delta}{2} (\cos k_x-\cos k_y)^2 \cos k_{z}.
\label{Eperp}$$ The functional form $ (\cos k_x-\cos k_y)^2$ in $\epsilon_{\mathbf{k}}^{\perp}$ is frequently invoked for cuprates [@andersen1995jpcs]. Other forms for $\epsilon_{\mathbf{k}}^{\perp}$, however, do not change the qualitative features. Although the electronic dispersion looks similar to that in a free electron system, the hopping integrals $t$, $t'$, and $t_z$ are renormalised by doping $\delta$ because of electron correlation effects. For both the optimally-doped materials we use $\delta=0.16$.
The term $\Delta$ in Eq. \[Epara\], which is proportional to $J$, is the mean-field value of the bond variables introduced to decouple the exchange term through a Hubbard-Stratonovich transformation [@foussats2004prb; @greco2016prb]. The value of $\Delta$ is computed self-consistently together with the chemical potential $\mu$ for a given $\delta$ by using
$$\Delta = \frac{J}{4{N_sN_z}} \sum_{\mathbf{k}, \eta} \cos(k_\eta) n_F(\epsilon_{\mathbf{k}}),
\label {Delta}$$
and $$(1-\delta)=\frac{2}{N_s N_z} \sum_{\mathbf{k}} n_F(\epsilon_{\mathbf{k}}),$$
where $n_F$ is the Fermi function, and $N_s$ and $N_z$ are the total number of lattice sites on the square lattice and the number of layers along the $c$ direction respectively. We take the number of layers $N_z$ equal to 30, which should be large enough, and set the temperature to zero.
As shown previously [@foussats2004prb; @greco2016prb; @greco2019cp], the charge-charge correlation function $\mathbf{\chi}_c (\mathbf{r}_i -\mathbf{r}_j, \tau)=\left\langle T_\tau n_i(\tau) n_j(0)\right\rangle $ can be computed in the $\mathbf{q}$-$\omega$ space as $$\begin{aligned}
\mathbf{\chi}_c(\mathbf{q},\omega)= N \left ( \frac{\delta}{2} \right )^{2} D_{11}(\mathbf{q},\omega).\end{aligned}$$ Thus, $\mathbf{\chi}_c$ is the element $(1,1)$ of the $6 \times 6$ bosonic propagator $D_{ab}$ where
$$D^{-1}_{ab}(\mathbf{q},\mathrm{i}\omega_n)
= [D^{(0)}_{ab}(\mathbf{q},\mathrm{i}\omega_n)]^{-1} - \Pi_{ab}(\mathbf{q},\mathrm{i}\omega_n),
\label{dyson}$$
and the matrix indices $a$ and $b$ run from 1 to 6. $D^{(0)}_{ab}(\mathbf{q},\mathrm{i}\omega_n)$ is a bare bosonic propagator
$$\label{D0inverse}
[D^{(0)}_{ab}(\mathbf{q},\mathrm{i}\omega_n)]^{-1} = N
\left(
\begin{array}{llllll}
\frac{\delta^2}{2} \left[ V(\mathbf{q})-J(\mathbf{q})\right]
& \frac{\delta}{2} & 0 & 0 & 0 & 0\\
\frac{\delta}{2} & 0 & 0 & 0 & 0 & 0\\
0 & 0 & \frac{4\Delta^2}{J} & 0 & 0 & 0\\
0 & 0 & 0 & \frac{4\Delta^2}{J} & 0 & 0\\
0 & 0 & 0 & 0 & \frac{4\Delta^2}{J} & 0\\
0 & 0 & 0 & 0 & 0 & \frac{4\Delta^2}{J}
\end{array}
\right),$$
and $\Pi_{ab}$ are the bosonic self-energies, $$\begin{aligned}
&& \Pi_{ab}(\mathbf{q},\mathrm{i}\omega_n)
= -\frac{N}{N_s N_z}\sum_{\mathbf{k}} h_a(\mathbf{k},\mathbf{q},\varepsilon_\mathbf{k}-\varepsilon_{\mathbf{k}-\mathbf{q}})
\frac{n_F(\varepsilon_{\mathbf{k}-\mathbf{q}})-n_F(\varepsilon_\mathbf{k})}
{\mathrm{i}\omega_n-\varepsilon_\mathbf{k}+\varepsilon_{\mathbf{k}-\mathbf{q}}}
h_b(\mathbf{k},\mathbf{q},\varepsilon_\mathbf{k}-\varepsilon_{\mathbf{k}-\mathbf{q}}) \nonumber \\
&& \hspace{25mm} - \delta_{a\,1} \delta_{b\,1} \frac{N}{N_s N_z}
\sum_\mathbf{k} \frac{\varepsilon_\mathbf{k}-\varepsilon_{\mathbf{k}-\mathbf{q}}}{2}n_F(\varepsilon_\mathbf{k}) \; ,
\label{Pi}\end{aligned}$$ where the six-component vertex $h_a$ is given by
$$\begin{split}
h_a(\mathbf{k},\mathbf{q},\nu) = \left\{
\frac{2\varepsilon_{\mathbf{k}-\mathbf{q}}+\nu+2\mu}{2}+
2\Delta \left[ \cos\left(k_x-\frac{q_x}{2}\right)\cos\left(\frac{q_x}{2}\right) +
\cos\left(k_y-\frac{q_y}{2}\right)\cos\left(\frac{q_y}{2}\right) \right];1;
\right. \nonumber \\
\left. -2\Delta \cos\left(k_x-\frac{q_x}{2}\right); -2\Delta \cos\left(k_y-\frac{q_y}{2}\right);
2\Delta \sin\left(k_x-\frac{q_x}{2}\right); 2\Delta \sin\left(k_y-\frac{q_y}{2}\right)
\right\}.
\end{split}$$
Here $\mathbf{q}$ and $\mathbf{k}$ are three dimensional wavevectors and $\omega_n$ is a bosonic Matsubara frequency. The factor $N$ comes from the sum over the $N$ fermionic channels. The $6$ channels involved in $D_{ab}(\mathbf{q},\mathrm{i}\omega_n)$ come from on-site charge fluctuations, fluctuations of a Lagrangian multiplier introduced to impose the non-double occupancy at any site, and fluctuations of the four bond variables which depend on $J$ [@greco2016prb; @foussats2004prb].
To describe plasmon excitations we compute the spectral weight of the density-density correlation function Im$\mathbf{\chi}_c(\mathbf{q},\omega)$ after analytical continuation $${\rm i} \omega_n \rightarrow \omega + {\rm i} \Gamma.
\label{continuation}$$
Since $\Gamma$ influences the width of the plasmon, its effect on the plasmon peak position is strongest when it becomes comparable to undamped plasmon energy (overdamped condition). As observed in the experiment (see Extended Data Fig. 3(d, e) and (f)), this condition may only be true close to the zone-centre for the acoustic plasmons. Plasmon energies calculated from the optimised models, as function of $\Gamma$, plotted in Fig. \[gamma\](a) clearly demonstrates that plasmon dispersions have negligible effect at larger $h$ values. Here a small positive value $\Gamma=0.1t$ was chosen for all the plasmon dispersion simulations. However to replicate the plasmon peak profiles observed in Fig. 3(f) of main paper, $\Gamma$ values of $0.2t$ and $0.29t$ were chosen for LSCO and Bi2201 respectively. A finite value of $\Gamma$ contains information of an extrinsic broadening due to the instrumental resolution, and an intrinsic broadening due to incoherent effects due to electronic correlations [@prelovsek1999prb].
![**Effect of interlayer hopping on gap formation in acoustic plasmon branches** **a,** Continuous lines are plasmon energies calculated for optimised parameters for LSCO and Bi2201 at given ($h$, $l$)-values as a function of broadening parameter $\Gamma$. **b,** and **c,** The effect of interlayer hopping parameter $t_z$, on the plasmon energy gap at $h=0.00$, $l=1.0$ for $\Gamma=0.001t$ and $\Gamma=0.1t$ respectively. For comparison, fitted plasmon energy values from RIXS spectra collected on LSCO for $l=1.00$ is also shown. Two values of $\Gamma$ are shown since it affects the plasmon energies close to the zone-center for the acoustic branches as shown in panel (a). Black dotted lines are $h$-values below which plasmon energies are not reliable due to large correlation with the elastic and phonon intensities. []{data-label="gamma"}](Gamma_gap.pdf)
![**Comparison of acoustic plasmon velocities in LSCO and Bi2201** Acoustic plasmon velocities ($v_p$) for LSCO and Bi2201 at $l=1.00$ are estimated using linear fits to extracted $h$-direction plasmon energies. The Bi2201 to LSCO $v_p$ ratio is found to be around 1.6, roughly matching the ratio of the interlayer spacing between them ($\sim$ 1.88) [@fetter1973ap]. The difference observed could be due to the slightly different dielectric constants of the two materials (see Methods C). []{data-label="velocities"}](Velocities.pdf)
[100]{} Greco, A., Yamase, H. & Bejas, M. Plasmon excitations in layered high-$T_c$ cuprates. *Phys. Rev. B* **94**, 075139 (2016). Zhang, F. C. & Rice, T. M. Effective Hamiltonian for the superconducting Cu oxides. *Phys. Rev. B* **37**, 3759(R) (1988). Becca, F., Tarquini, M., Grilli, M. & Di Castro, C. Charge-density waves and superconductivity as an alternative to phase separation in the infinite-U Hubbard-Holstein model. *Phys. Rev. B* **54**, 12443 (1996). Thio, T. et al. Antisymmetric exchange and its influence on the magnetic structure and conductivity of La$_2$CuO$_4$. *Phys. Rev. B* **38**, 905(R) (1988). Greco, A., Yamase, H. & Bejas, M. Origin of high-energy charge excitations observed by resonant inelastic X-ray scattering in cuprate superconductors. *Commun. Phys.* **2**, 3 (2019). Foussats, A. & Greco Large-$N$ expansion based on the Hubbard operator path integral representation and its application to the $t-J$ model. II. The case for finite $J$. *Phys. Rev. B* **70**, 205123 (2004). Andersen, O. K., Liechtenstein, A. I., Jepsen, O. & Paulsen, F. LDA energy bands, low-energy hamiltonians, $t'$, $t''$, $t_\perp(k)$), and $J_\perp$. *J. Phys. Chem. Solids* **56**, 1573 (1995). Prelovsek, P. & Horsch, P. Electron-energy loss spectra and plasmon resonance in cuprates. *Phys. Rev. B* **60**, R3735(R) (1999). Fetter, A. Electrodynamics of a layered electron gas. I. Single layer. *Ann. Phys.* **81**, 367 (1973).
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'Ruichu Cai, Zhenjie Zhang, and Zhifeng Hao, [^1]'
bibliography:
- 'IEEEabrv.bib'
- 'ref.bib'
title: 'SADA: A General Framework to Support Robust Causation Discovery with Theoretical Guarantee'
---
[Cai : SADA: A General Framework to Support Robust Causation Discovery with Theoretical Guarantee]{}
Introduction
============
Causation discovery plays an important role on a variety of scientific domains. Different from the mainstream statistical learning approaches, causation discovery tries to understand the data generation procedure, rather than characterizing the joint distribution of the observed variables only. It turns out that understanding causality in such procedures is essential to predict the consequences of interventions, which is the key to a large number of applications, such as genetic therapy, advertising campaign design, etc.
From computational perspective, causation discovery is usually formulated over a probabilistic graphical model on the variables under the assumption of faithfulness [@pearl2009causality], in which the directed edges indicate causal relations. When it is unlikely to manipulate the samples in experiments, conditional independence tests are commonly employed to detect local causal structures among the variables [@pearl2009causality; @spirtes2011]. Despite of the successes of these approaches on small problem domains and large sample bases, they usually fail to find true causalities, when huge number of equivalent structures over the graphical probabilistic models render exactly the same conditional independence.
To tackle the difficulties of causation discovery under non-experimental setting, researchers are recently resorting to asymmetrical relations between the cause-effect pairs under assumptions on the data generation process. The discovery ability is dramatically improved, by exploiting linear non-Gaussian assumption [@ShimizuJMLR06; @ShimizuJMLR2011Direct], nonlinear assumption [@Janzing2008nipsnonlinear], discrete property [@Janzing2010jmlrdiscrete], deterministic mechanism [@Daniusis2010Deterministic] and so on. When the variables are correlated under linear relations and the noises follow non-Gaussian distributions, for example, LiNGAM [@ShimizuJMLR06] and its variants [@ShimizuJMLR2011Direct] are known as the best causation discovery algorithms. However, the scalability of LiNGAM and its variants is still questionable, since they heavily depend on the independent component analysis (ICA) during the computation. To return robust results from ICA, it is necessary to feed a large bulk of samples, which are expected to be no smaller than the number of variables. Similar problems arise to other well known methods, e.g., [@Janzing2008nipsnonlinear], [@Janzing2010jmlrdiscrete], which are usually used to infer the causal directions over individual variable pairs.
Motivated by the common observations on the sparsity of causal structures, i.e., each variable usually only depends on a small number of parent variables, we derive a Scalable cAusation Discovery Algorithm (named SADA) in this article. SADA helps the existing causation algorithms to get rid of difficulties on small sample size in practice. Well designed conditional independence tests are conducted to partition the problem domain into small subproblems. With the same number of samples, existing causation discovery algorithm (taken as basic causal solvers) could generate more robust and accurate results on these small subproblems. Partial results from all subproblems are finally merged together, to return a complete picture of causalities among all the variables.
This framework is generic, as only faithfulness condition and causal sufficiency assumption are employed, so that it works well with different basic causal solvers by exploiting additional compatible data generation assumptions, such as linear non-Gaussian data and additive noise data. This framework is also theoretically solid, as it always returns correct and complete result under the optimal setting on conditional independence tests and basic causal solvers. Even when the conditional independence tests are vulnerable, the framework is capable of improving the recall and precision of the overall results, when the basic causal solver on the subproblems achieves sufficient enhancement on accuracy. Our experiments on synthetic and real datasets verify the genericity, superior scalability and effectiveness of our proposal, when applied together with two mainstream causation discovery algorithms.
The outline of the paper is listed as follows. Section \[sec:related\] reviews existing studies on causation discovery problem. Section \[sec:framework\] introduces the framework and algorithms to tackle the problem of small sample size. Section \[sec:theory1\] provides theoretical analysis on the proposed framework with the assumption of ideal conditional independence tests. Section \[sec:theory2\] extends the analysis to more general assumptions on vulnerable conditional independence tests. Section \[sec:exp\] reports experimental results on both synthetic and real datasets, and Section \[sec:concl\] finally concludes the paper.
Related Work {#sec:related}
============
Causal Bayesian network (CBN) is part of the theoretical background of this work. Different from the traditional Bayesian network, each edge in a CBN is interpreted as a direct causal influence from the parent node to the child node [@pearl2009causality]. CBN has been used to model the causal structure in many real-world applications, e.g., the gene regulatory network [@yoo2002discovery; @ellis2008learning] and causal feature selection [@aliferis_local_2010].
A large number of works try to explore the conditional independence tests to learning the local structures of CBN, e.g. the well known PC algorithm [@kalisch2007pc; @cai2011bassum], Markov Blanket discovery methods [@zhu2007markov; @spirtes2011]. These methods provide the elements of causal structures, and are usually considered as start points of the causation discovery methods [@pearl1991ic; @aliferis_local_2010; @cai2013causal].
Pearl is one of the pioneers of the causal theory [@pearl2009causality]. Since Pearl’s Inductive Causality [@pearl1991ic], a large number of extensions are proposed by exploring the $V$-structure to determine the causal directions. Most of the extensions assume the acquisition of a sufficiently large sample set [@aliferis_local_2010]. Though there are studies aiming at the causation discovery under small sample size [@bromberg_improving_2009], the actual number of the samples used in their empirical evaluations remains significantly larger than the number of variables. Cai’s study [@cai2013causal] is another attempt under this category to extend the method to the high dimensional gene expression domain by exploiting the conflict relations among the local sub-structures. Recently, some partition based approaches are also proposed to improve the scalability of the structure learning based methods, such as Geng’s recursive decomposition strategy [@XieJMLR08] and Yehezkel’s autonomy identification based partition [@YehezkelJMLR09]. However, all these approaches, based on conditional independence tests, cannot distinguish two causality structures if they come from a so-called *Markov equivalence class* [@pearl2009causality], in which expensive intervention experiments are previously considered essential [@he_active_2008].
Recently, a lot of methods are proposed to break the limitations of the methods purely under conditional independence tests, by exploiting the asymmetric properties in the generative progress, which brings a gleam of dawn to resolve the causal equivalence problem. Existing studies on this line can be categorized based on the adopted assumption on the noise type or data generation mechanism. *Additive Noise Model* [@Janzing2008nipsnonlinear] and its variants highly depend on the independence relation between the causal variable and the noises, including, its generalization to post-linear [@zhang2009postnonlinear], its variants on the discrete data [@Janzing2010jmlrdiscrete]. Information Geometry based method is developed for the deterministic causal relations [@Janzing2012AIInfoGeo] by exploring the asymmetric relation between the data distribution and the generation mechanism. Its extension exploits the Kernel Hilbert space embedding based measure to infer the asymmetric properties [@ChenZCS2014]. LiNGAM and its variants [@ShimizuJMLR06; @ShimizuJMLR2011Direct], assume that the data generating process is linear and the noise distributions are non-Gaussian. There are other studies relat to this topic, such as explaining the underlying theoretical foundation behind asymmetric property based methods [@Bastian2010COLT; @Janzing2010Causal; @Janzing2012AIInfoGeo], addressing the latent variable problem[@tashiro2014parcelingam], regression-based inference method [@Mooij2009ICMLRegression] and kernel independence test based causation discovery methods [@Zhang2012CoRRKCIT]. Inference the direction between a causal-effect pair is focus of these methods. Though, there are some works try to generalize the model to the case with more than two variables, for example [@ShimizuJMLR06; @peters2012identifiability], there is no existing work to address the sample size problem to the best of our knowledge.
Granger’s causality [@Granger1969Econometrica] is another important subfield of causality, which uses Granger’s causality test [@hacker2006tests] to determine whether one time series is useful in forecasting another. Recently, Granger’ causality is extended to infer the gene regulatory networks from the time series gene expression data [@mukhopadhyay2007causality; @lozano2009grouped]. Granger’s work differs from traditional causation discovery techniques on two aspects. Firstly, compared with the conventional definition of causality, Granger’s causality is more likely a regression method and does not reflect the true causality mechanism. Secondly, the temporal information is essential for Granger’s causation discovery algorithms, which is expensive and some times impossible to collect.
SADA Framework {#sec:framework}
==============
Preliminaries
-------------
Assume that all samples from the problem domain contain information on $n$ different variables, i.e., $V=\{v_1,v_2\dots,v_n\}$. Let $D=\{x_1,x_2,\cdots, x_m\}$ denote an observation sample set. Each sample $x_i$ is a vector $x_i=\left(x_{i1},x_{i2},\ldots,x_{in},y_i\right)$, where $x_{ij}$ indicates the value of the sample $x_i$ on variable $v_j$ and $y_i$ is the target variable under investigation.
If $\mathcal{P}$ is a distribution over the domain of variables in $V$, we assume that there exists a causal Bayesian network $N$ faithful to the distribution $\mathcal{P}$. The network $N$ includes a directed acyclic graph $G$, each edge in which indicates a dependent relation between two variable nodes. Each edge is also associated with a conditional probability function which presents conditional probability distribution of the variables given the values of their parent variables. Following the common assumption of existing studies, we only consider problem domain meeting *Faithfulness Condition* [@koller2009pgm]. Specifically, $\mathcal{P}$ and $N$ are faithful to each other, *iff* every conditional independence entailed by $N$ corresponds to some Markov condition present in $\mathcal{P}$. Beside the faithfulness condition, *Causal Sufficiency* [@koller2009pgm] is another assumption taken in this work, which assumes that there are no latent confounders of any two observed variables.
Due to the probabilistic nature, it is likely to find a huge number of equivalent Bayesian networks. Two different Bayesian Networks, $N_1$ and $N_2$, are Markov equivalent, if $N_1$ and $N_2$ entail exactly the same conditional independence relations among the variables. In all these Bayesian networks, Causal Bayesian network (CBN) is a special one in which each edge is interpreted as a direct causal relation between a parent node and a child node.
Generally speaking, it is difficult to distinguish CBN from independence equivalent Bayesian networks, unless additional assumptions are made. When the variables are correlated in linear relations and the noises follow non-Gaussian distributions independently, LiNGAM and its variants [@ShimizuJMLR06; @ShimizuJMLR2011Direct] are known to return more accurate causations from uncontrollable samples. In particular, such assumption can be formulated by an equation, such that every variable $v_i = \sum_{v_j \in P\left(v_i\right)}{ A_{ij}\cdot v_j} + e_i$, where $P\left(v_i\right)$ contains all the parent variables of $v_i$, $A_{ij}$ is the linear dependence weight w.r.t. $v_i$ and its parent $v_j$, and $e_i$ is an non-Gaussian noise over $v_i$. Assume that the variables in $V$ are organized based on a topological order in the causal structure. The generation procedure of a sample could be written as $V = A\cdot V + E$. LiNGAM aims to find such a topological order and reconstructs the matrix $A$ by exploiting *independence component analysis* (ICA) over the sample.
When assuming non-linear generation procedure [@Janzing2008nipsnonlinear] and discrete data domain [@Janzing2010jmlrdiscrete], additive noise model provides another approach to utilize the asymmetric relations between causal variables and consequence variables. A regression model $v_i = f\left(v_j\right) + e_i$ is trained for each pair of variables $v_i$ and $v_j$. If the noise variable $e_i$ is independent of $v_j$, variable $v_j$ is returned as the cause of variable $v_i$. Note that algorithms under additive noise model are usually run over pairs of variables independently.
A common observation on the CBNs in real-world domains is the sparsity on the causal relations. Specifically, a variable usually only has a small number of causal variables in the CBN, regardless of the underlying true generative procedure. This property, however, is not fully exploited by the existing causation algorithms.
Framework
---------
In SADA, the variables are partitioned into subsets, by utilizing *causal cuts* on the variables based on conditional independence relations over the domain with sparse causal structure. To begin with, we present the definitions of *causal cut* and *causal cut set*.
*Causal Cut.* Let $G=\left(V,E\right)$ denote a causal structure on the variable set $V$. Three disjoint variable subsets $\left(C, V_1, V_2\right)$ of $V$ forms a *causal cut* over $G$, *if* (1) $C \cup V_1 \cup V_2 =V$; (2) there is no edge between $V_1$ and $V_2$ in $E$.
*Causal Cut Set.* In a causal cut $\left(C, V_1, V_2\right)$, the variable set $C$ is a *causal cut set* if it ensures there is no edge between $V_1$ and $V_2$.
Based on the above definitions, given a causal cut $\left(C, V_1, V_2\right)$ over the problem $G=\left(V,E\right)$, one of the two following cases must hold for each directed edge $u \rightarrow v$ in $E$: (1) intra-causality: $\{u,v\}\subset V_1$, $\{u, v\}\subset V_2$ or $\{u, v\}\subset C$; and (2) inter-causality: $u \in V_1\cup V_2$ and $v \in C$, or $u \in C$ and $v \in V_1\cup V_2$. This intuition guarantees the independence between two subproblems on the variable sets $V_1 \cup C$ and $V_2 \cup C$, which paves the foundation for the ’Split-and-Merge’ framework.
![An example probabilistic graphical model over 9 variables and two causal cuts with causal cut sets $C$ and $C'$.[]{data-label="fig:example"}](figs/example.eps){width=".6\columnwidth"}
In Figure \[fig:example\], for example, $C=\{v_4\}$ is a causal cut set, which separates the variables into the causal cut $\left(C=\{v_4\}, V_1=\{v_1, v_3, v_6,v_7\}, V_2=\{ v_2, v_5, v_8, v_9\}\right) $. Given a directed graph $G$, there could be different valid causal cut sets satisfying the above conditions. In the example graph, $C'=\{v_4, v_7, v_8\}$ is another causal cut set with $V_1=\{v_1, v_3, v_6\}$ and $V_2=\{ v_2, v_5, v_9\}$.
Please note that causal cut set is closely related to the concept of $d$-separation but it may not lead to $d$-separation. For example, in the causal cut $\left(C=\{v_4\}, V_1=\{v_1, v_3, v_6,v_7\}, V_2=\{ v_2, v_5, v_8, v_9\}\right)$ , the variable $v_7$ is not independent of $v_2$ given any subset of the causal cut set $C=\{v_4\}$; while in the causal cut $\left( C'=\{v_4, v_7, v_8\}, V_1=\{v_1, v_3, v_6\}, V_2=\{ v_2, v_5, v_9\}\right)$, the causal cut set $C'$ lead to the $d$-separation between $V_1$ and $V_2$. The connection between causal cut and $d$-separation will be formalized in the next section.
Given a causal cut $\left(C,V_1,V_2\right)$ on variable set $V$, we are able to transfer the causation discovery problem on $V$ into two smaller ones over the variable sets $V_1\cup C$ and $V_2\cup C$ respectively. This partitioning operation could be recursively called, until the number of variables involved in the subproblem is below a specified threshold $\theta$. The complete pseudocodes are available in Algorithm \[algo:framework\]. The inputs of SADA include the sample set $D$, the variables $V$, a threshold $\theta$ and an underlying causation discovery algorithm $A$. Here, $\theta$ is used to terminate the recursive partitioning when the variable set is sufficiently small, and $A$ is an arbitrary basic causal solver invoked to find the actual causal structure on the subset of variables. $A$ is usually taken as a basic causal solver in this work.
In the rest of the section, we will discuss how to effectively and efficiently find causal cut on a variable set $V$ by exploiting the corresponding observational samples. We will also present the details of the merging operator, which tackles the problem of inconsistency and redundancy on the partial results from the subproblems. Note that only the faithfulness condition and causal sufficiency assumption are employed in this proposed framework (except the basic causal solver), thus SADA can work with different basic causal solvers by exploiting additional data generation assumptions, such as linear non-Gaussian data, additive noise data and so on.
Input: sample set $D$, variable set $V$, variable threshold $\theta$ and a basic causal solver $A$\
Output: $G$: causal structure\
Return the result $G$ by running algorithm $A$ on $D$ and $V$. Find a causal cut $\left(C,V_1,V_2\right)$ on $D$ and $V$. $G_1=$**SADA**$\left(D,V_1\cup C,\theta,A\right)$. $G_2=$**SADA**$\left(D,V_2\cup C,\theta,A\right)$. Return $G$ by merging $G_1$ and $G_2$.
Finding Causal Cut
------------------
The searching of the causal cut is crucial to the partitioning operation in SADA. To identify potential causal cut, our algorithm resorts to conditional independence testing over variables in the Bayesian network. The following lemma formalizes the connection.
\[lemma:properyofcvs\] $\left(C,V_1,V_2\right)$ is a causal cut over causal structure $G$, *if* (1) $C \cup V_1\cup V_2 =V$; (2) $C$, $V_1$ and $V_2$ are disjoint with each other; and (3) $\forall u\in V_1$ and $\forall v \in V_2$, there exists a variable set $C_{uv} \subset C$ such that $u \bot v | C_{uv}$.
For all pairs of variables $\left(u,v\right)$ that $u\in V_1$ and $v \in V_2$ are *d*-separated by $C$, there is no directed edge between $V_1$ and $V_2$. Combining condition (1) and (2), we have that $\left(C,V_1,V_2\right)$ is a causal cut over causal structure $G$.
Note that Lemma \[lemma:properyofcvs\] is a sufficient condition of causal cut, but not a necessary condition. That is, a causal cut $\left(C,V_1,V_2\right)$ may not satisfy the above three conditions. For example, the triple $\left(C=\{v_4\}, V_1=\{v_1, v_3, v_6,v_7\}, V_2=\{ v_2, v_5, v_8, v_9\}\right) $ is a causal cut of the example given in Figure \[fig:example\], but it dose not satisfy the condition (3) of Lemma \[lemma:properyofcvs\], because the variable $v_7$ is not independent of $v_2$ given any subset of the causal cut set $C=\{v_4\}$. In detail, when $C_{uv}=\emptyset$, $v_7$ is dependent on $v_2$ because of the directed path $v_2 \rightarrow v_4\rightarrow v_7$; when $C_{uv}=\{v_4\}$, $v_7$ is dependent on $v_2$ because of the directed path $v_2 \rightarrow v_4\leftarrow v_1 \rightarrow v_3 \rightarrow v_7$ ( the path $v_2 \rightarrow v_4\leftarrow v_1$ is connected given the variable $v_4$).
By exploiting the sufficient condition given in Lemma \[lemma:properyofcvs\], we derive a new algorithm to find a causal cut set $C$, and the corresponding causal cut $\left( C, V_1, V_2 \right)$. In the search algorithm, each variable is heuristically assigned to one of the set $V_1$, $V_2$ and $C$. Besides of casual cut property, we also want to optimize the following two objectives during the assignment procedure: (1)minimizing the size of $C$. Because $C$ appears in both subproblems $V_1 \cup C$ and $V_1 \cup C$, i.e., smaller $C$ is preferred; (2) minimizing the size difference of $V_1$ and $V_2$. According to the principle of divide-and-conquer, the causal cut with similar sizes of $V_1$ and $V_2$ is preferred. The details of the algorithms are listed in Algorithm \[algo:split\]. The algorithm runs with $k$ different initial variable pairs. The algorithm greedily adds the variable $w$ into $V_1$ (or $V_2$), if $w$ is independent of all the variables of $V_2$ (or $V_1$) given some subset of $C$. Only the $w$ can not added to neither of $V_1$ and $V_2$, $w$ is added to $C$. After completing all assignments, the algorithm also tries to move the variables from $C$ to $V_1$ or $V_2$ to maximize the partitioning effect. Finally, the causal cut with largest $\min\{|V_1|,|V_2|\}$ are returned as final result. We leave the discussion on the parameters $k$ and $\theta$ to next section. Please note that the sample size needed in the cut algorithm highly depends on the local connectivity of the causal structure but not on the number of variables. This is an important advantage of the algorithm to applications in large scale sparse causation discovery problems.
Input: sample set $D$, variable set $V$, number of initial variable pairs $k$ Output: a causal cut $\left(C,V_1,V_2\right)$ Randomly pick up two variables $u$ and $v$ such that $\exists V'\subset V-\{u,v\}$ satisfies $u\bot v| V'$. Find the smallest $\hat{V}\subseteq V-\{u,v\}$ to make $u\bot v | \hat{V}$. Initialize $V_1=\{u\}$, $V_2=\{v\}$ and $C=\hat{V}$. Remove variables in $V_1$, $V_2$ and $C$ from $V$. Add $w$ into $V_2$. Add $w$ into $V_1$. Add $w$ into $C$. Move $s$ from $C$ to $V_2$. Move $s$ from $C$ to $V_1$. Let $\Phi_j=\left(C,V_1,V_2\right)$ Return $\Phi_j$ with the largest $\min\{|V_1|,|V_2|\}$.
Given the example structure of Figure \[fig:example\], the running step is given in Table \[tab:divide\] under the assumption that $V_1$, $V_2$ and $C$ are initialized as $\{v_1\}$ , $\{v_2\}$ and $\emptyset$, respectively. Among the steps, $v_3$ and $v_6$ are marginally independent of any variable of the current $V_2$. In another word, $C'=\emptyset$ is used in the conditional independence test. The similar cases happen in the assignment of $v_5$ and $v_9$. In the checking of $v_7$, $v_7$ is dependent of the variables of $v_2$ given any sub set of the current causal cut set $C=\{v_4\}$ and added to the causal cut set $C$. $v_8$ is processed similarly.
Step $V$ $V_1$ C $V_2$
------------- -------------------------------- --------------- --------------- ---------------
Initial $v_3,v_4,v_5,v_6, v_7,v_8,v_9$ $v_1$ $\emptyset$ $v_2$
Check $v_3$ $v_4,v_5,v_6, v_7,v_8,v_9$ $v_1,v_3$ $\emptyset$ $v_2$
Check $v_4$ $v_5,v_6, v_7,v_8,v_9$ $v_1,v_3$ $v_4$ $v_2$
Check $v_5$ $v_6, v_7,v_8,v_9$ $v_1,v_3$ $v_4$ $v_2,v_5$
Check $v_6$ $v_7,v_8,v_9$ $v_1,v_3,v_6$ $v_4$ $v_2,v_5$
Check $v_7$ $v_8,v_9$ $v_1,v_3,v_6$ $v_4,v_7$ $v_2,v_5$
Check $v_8$ $v_9$ $v_1,v_3,v_6$ $v_4,v_7,v_8$ $v_2,v_5$
Check $v_9$ $\emptyset$ $v_1,v_3,v_6$ $v_4,v_7,v_8$ $v_2,v_5,v_9$
: Running Example of Split[]{data-label="tab:divide"}
Merging Partial Results {#sec:merge}
-----------------------
As is shown in Algorithm \[algo:framework\], two partial results $G_1$ and $G_2$ are combined as a single casual graph as on variables in $V$. Since $G_1$ and $G_2$ are calculated independently but contain overlap over the causal cut set $C$. Thus conflict and redundancy need to be carefully handled in the merging operation. Recall the example given in Figure \[fig:example\], assume $C=\{v_4, v_7, v_8\}$, $V_1=\{v_1, v_3, v_4, v_6, v_7\}$ and $V_2=\{v_2, v_4, v_5, v_8, v_8, v_9\}$, the edges $v_7\rightarrow v_8$ and $v_8\rightarrow v_7$ may be appear in results from $V_1$ and $V_2$, generating conflicts. Similarly, the basic solver may return $v_4\rightarrow v_7$ and $v_7 \rightarrow v_8$ on $V_1$, and $v_4\rightarrow v_8$ on $V_2$. It is easy to see, the edge $v_7\rightarrow v_8$ is redundant. Generally speaking, such conflicts and redundancy depend on the assumption of the causal structures. Under the directed acyclic graph assumption, the examples shown in Fig 2 are all the patterns we can detect.
The general form of a conflict is a cycle of directed edges among a group of variables, as shown in Figure \[fig:merge\_conflict\]. Given two nodes $v_1$ and $v_2$, there are two paths co-existing, such as $v_1 \rightarrow \cdots \rightarrow v_2$ and $v_1 \leftarrow v_2$. These two paths form a cycle and violate the acyclic constraints. To resolve such conflict, we simply remove the least reliable edge in the cycle, whenever a cycle is found. Here the reliability of the edge $v_1\rightarrow v_2$ is measured by the significance level, $sig\left(v_1\rightarrow v_2\right)$, which is returned by the basic causal solvers. For example, the $p$-value of the Wald test is used as the significance level for edges returned by LiNGAM [@ShimizuJMLR06], and the $p$-value of the noise’s independence of the causal variable is used as the significance level for edges returned by additive noise model [@Janzing2011tpamidiscrete].
Figure \[fig:merge\_redundancy\] illuminates a potential redundancy case. Given two variables $v_1$ and $v_2$, if both $v_1 \rightarrow \cdots \rightarrow v_2$ and $v_1 \rightarrow v_2$ are discovered, $v_1 \rightarrow v_2$ may be redundant. Because the dependency relation $v_1 \rightarrow v_2$ could be blocked by certain variables in the variable set $Path\left(v_1\rightarrow v_2\right)$. Here, $Path\left(v_1\rightarrow v_2\right)$ refers to the variable set involved in the directed path $v_1 \rightarrow \cdots \rightarrow v_2$. Such redundancy raises when the following two conditions are satisfied: (1) the source and destination variables are both in the causal cut set, i.e., $v_1, v_2 \in C$, (2) there is another variable set $V_3\subset V_1$ (or $V_3\subset V_2$), such that $v_1 \rightarrow V_3 \rightarrow v_2$. If the above two conditions are met, one path $v_1 \rightarrow v_2$ will be returned from the subproblem over $V_1 \cup C$, while another path $v_1 \rightarrow v_2$ turns up from the other subproblem over $V_2 \cup C$. To tackle this problem, our merging algorithm runs the following conditional independence tests to verify if $\exists V' \subset Path\left(v_1\rightarrow v_2\right)$ that $v_1 \bot v_2 | V'$.
\[h\]
To summarize, the merging operation works as follows. Firstly, all directed edges from both solutions are simply added into a single edge set. Secondly, edges are ranked according to the associated significance measure, calculated by the basic causal solver employed by SADA. Thirdly, a sequential conflict testings are run over the ordered edges non-decreasingly on the significance. An edge is removed if it is conflicted with any of the previous edges. Finally, the redundancy edges are discovered and removed based on results of the conditional independence tests. A complete description is available in Algorithm \[algo:merge\].
Input: [$G_1$, $G_2$: solutions to $V_1\cup C$ and $V_2\cup C$]{}\
Output:[$G$: solution for $C \cup V_1\cup V_2$]{}\
//basic merging $G=G_1 \cup G_2$;\
//conflict removal Sort edges in $G$ in descending order of significance;\
Mark all variable pairs as unreachable; //redundancy removal $G$;\
Analysis under Reliable Conditional Independence Test {#sec:theory1}
=====================================================
In this section, we study the theoretical properties of SADA, especially on the effectiveness on problem scale reduction and consistency on causal results, under the assumption that no error is introduced by any conditional independence tests. The assumption is mathematically formulated as follows.
\[assum:simple\] For any variables $v_1$, $v_2$ and variable set $V$, the conditional independence tests always return true, *iff* $v_1\bot v_2|V$.
Intuitively, when the above assumption holds, the causal cut finding operator never generates wrong partitions which divides a causal-effect variable pair into two separate sets $V_1$ and $V_2$. Although such requirement is unlikely to meet in practice, it simplifies the model and allows us to derive accurate analysis on SADA in the rest of the section.
Effectiveness on Scale Reduction
--------------------------------
In this part of the section, we aim to verify the effectiveness of the causal cut finding algorithm. In particular, we try to prove that the scale of the subproblem is significantly reduced by applying the randomized causal cut finding algorithm.
\[theorem:convergence\] If every variable has no more than $d_m$ parental variables in CBN, by setting $k=\left(2d_m+2\right)^2$, Algorithm \[algo:split\] returns a causal cut $\left(C,V_1,V_2\right)$ with probability at least 0.5, such that $$\min\{|V_1|,|V_2|\}\geq \frac{|V|}{2d_m+2}$$
Since the causal structure must be a DAG, there is at least one topological order on the variables. Here topological order of a DAG is a linear order of its vertices such that for every directed edge $v_i\rightarrow v_j$, $v_i$ comes before $v_j$ in the order. Let $V=\{v_1,v_2,\ldots,v_{|V|}\}$ be the topological order, then $v_i$’s parental variables are ahead of $v_i$ in the order. When randomly picking up variable pairs in $V$, i.e., $u$ and $v$ from $V$, we will first show that $u$ and $v$ generate a causal cut with $\min\{|V_1|,|V_2|\}\geq\frac{|V|}{2d_m+2}$ with probability at least $1/\left(2d_m+2\right)^2$.
Without loss of generality, we assume $n=|V|$ and the variable $u$ is behind $v$ in the topological order over $V$. With probability $\eta$, $u$ is one of the variables between $v_{0.5n}$ and $v_{\left(0.5+\eta\right)n}$. Consider all the $\eta n$ variables between $v_{0.5n}$ and $v_{\left(0.5+\eta\right)n}$. We simply put all these variables in $V_1$, and put all parental variables of $V_1$, denoted by $P\left(V_1\right)$, and all variables behind $v_{\left(0.5+\eta\right)n}$ into $C$. The rest of the variables are inserted into $V_2$. In the configuration $\left(C,V_1,V_2\right)$, $C$ contains all the parental variable of $V_1$ by adding $P\left(V_1\right)$ into $C$, and all the possible children variables of $V_1$ by adding all variables behind $v_{\left(0.5+\eta\right)n}$ into $C$, because the children variable must be ordered behind $V_1$. Thus, there is no direct edge between $V_1$ and $V_2$, and the configuration $\left(C,V_1,V_2\right)$ is a causal cut.
In the above causal cut $\left(C,V_1,V_2\right)$, $|V_1|=\eta n$ and $|V_2|\geq \frac{n}{2}-\eta n d_m$. The inequality is because $V_2=\{v_i| i\leq 0.5n \}-P(V_1)$ and $|P(V_1)| \leq \eta n d_m$. By picking $\eta=\frac{1}{2d_m+2}$, $\min\{|V_1|,|V_2|\}\geq \frac{n}{2d_m+2}$. When $v$ is selected in $V_2$, Algorithm \[algo:split\] must converge to a solution better than the artificial configuration above. Because in the above analysis $v$ could be place into $V_2$ or $C$, when $v$ is put into $V_2$, the size of $V_2$ is larger than the above expectation. This happens with probability at least $\frac{1}{\left(2d_m+2\right)^2}$ when $\eta=\frac{1}{2d_m+2}$.
By running the randomized causal cut finding algorithm $k=\left(2d_m+2\right)^2$ times, the probability of finding a causal cut with $\min\{|V_1|,|V_2|\}\geq \frac{n}{2d_m+2}$ is larger than $ 1- \left(1 - \frac{1}{\left(2d_m+2\right)^2} \right) ^ {\left(2d_m+2\right)^2}$. Since $\left(1 - \frac{1}{\left(2d_m+2\right)^2} \right) ^ {\left(2d_m+2\right)^2} \approx e^{-1}$ when $\left(2d_m+2\right)^2$ is sufficiently large, the probability of finding a causal cut with $\min\{|V_1|,|V_2|\}\geq \frac{n}{2d_m+2}$ is at least $1-e^{-1}$, i.e., larger than 1/2.
The last theorem implies that the causal cut finding algorithm is effective on reducing the scale of the subproblems. Another implication is on the selection of the parameter $\theta$. To guarantee there is a reduction on problem size, the parameter $\theta$ should be no smaller than $2d_m+2$, since such $\theta$ ensuring that $\frac{\theta}{2d_m+2}\geq 1$.
Recall and Precision on Result Causal Edges
-------------------------------------------
The accuracy of the causation discovery is measured based on the recall and precision on the result causal edges, i.e., the percentage of accurate causal edges and the percentage of causal edges returned. In this section, we show that SADA always finds fully accurate results in terms of recall and precision, if the invoked basic causal solver and conditional independence tests are both reliable.
A basic causal solver A is reliable, if A always outputs accurate causal edges on any variable set V even with latent confounders.
\[theorem:correctcomplete\] Assume $D$ is a set of samples generated from the causal structure $G$ over the variable set $V$. If the basic causal solver $A$ and conditional independence tests used in SADA are both reliable, SADA always finds the true causal structure $G$.
Assume $G'$ is the causal structure discovered by SADA. We only need to prove the correctness and completeness of $G'$. The correctness and completeness are equivalent to $\forall v_1\rightarrow v_2 \in G'$, $v_1 \rightarrow v_2 \in G$, and $\forall v_1\rightarrow v_2 \in G$, $v_1 \rightarrow v_2 \in G'$, respectively. The details of the proof are given as follows:
**Completeness:** Assume $v_1\rightarrow v_2 \in G$, firstly, according to the causal cut finding step, both $v_1$ and $v_2$ must be in one subproblem, $V_1\cup C$ or $V_2 \cup C$, but not acrose the two subproblems. Otherwise, $v_1$ and $v_2$ is conditional independent of each other given some subset of $C$, conflicts with the condition $v_1 \rightarrow v_2 \in G$ and the assumption that the conditional independence tests are reliable. Secondly, according to the following two conditions: ’$v_1$ and $v_2$ are in the same subproblem’ and ’basic causal solver is reliable’, $v_1 \rightarrow v_2 \in G'$ will be discovered in one of the subproblems. Finally, the edge $v_1 \rightarrow v_2$ will not be removed in the merging step. If the edge is removed by either conflict or redundancy reason, it will conflict with the condition $v_1\rightarrow v_2 \in G$ and the assumption that the condition independence test is reliable. Thus, $v_1\rightarrow v_2$ must be contained in the result of SADA, in anther word, $v_1\rightarrow v_2 \in G'$.
**Correctness:** Assume $v_1\rightarrow v_2 \in G'$, firstly we will show $v_1\rightarrow v_2$ is the correct result of the subproblem. According to the framework of SADA, $v_1$ and $v_2$ must be discovered in one of the subproblem $V_1 \cup C$ and $V_2 \cup C$. Without loss of generality, assume $v_1\rightarrow v_2$ is discovered in the subproblem $V_1 \cup C$ by the basic causal solver. According to the condition that the basic causal solver is reliable, $v_1 \rightarrow v_2$ must be the correct result of the subproblem $V_1 \cup C$. Secondly, we will show $v_1\rightarrow v_2 \in G$. If $v_1\rightarrow v_2$ is the correct result of $V_1 \cup C$ but not contained in $G$, then there must exist a variable set $V' \subset V$ satisfies $v_1 \bot v_2 | V'$. Thus, there must be a path $v_1 \rightarrow \cdots \rightarrow v_2$ which contains $V'$ as intermediate nodes. If such path exists, according to the merging step, $v_1\rightarrow v_2$ will be removed from the result set $G'$, and conflicts with the condition that $v_1\rightarrow v_2 \in G'$. Thus, $v_1\rightarrow v_2 \in G$.
Basically, the theorem above claims that the recall and precision on the causal edges returned by SADA are always satisfiable. However, as we emphasized at the beginning of the section, the assumption on reliable conditional independence tests is impractical, since randomness and noises always exist in the samples. In next section, we relax the assumption and show that SADA remains effective in a class of much more general settings.
Analysis under Vulnerable Conditional Independence Test {#sec:theory2}
=======================================================
![Workflow of SADA with single partitioning.[]{data-label="fig:sada_framework"}](figs/sada_framework.eps){width="\columnwidth"}
In this section, we analyze the performance of SADA under more general assumptions, taking errors incurred by causal cut finding, merging, redundancy removal and conflict removal into consideration. To accomplish the goals of the analysis, we investigate the impact of each step in the algorithm on the recall and precision of the results one at a time. As is shown in Figure \[fig:sada\_framework\], there are five key steps taken in SADA, including (1) finding causal cut $\left(C, V_1, V_2\right)$ over variable set $V$; (2) solving two subproblem $V_1\cup C$ and $V_2\cup C$ using basic causal solvers independently; (3) merging two sub-solutions $G_1$ and $G_2$; (4) removing conflict edges after the merging; and (5) detecting and removing redundancy edges.
To improve the readability of the paper, we summarize all the notations used in the rest of the section in Table \[tab:notation\]. Basically, $n$s indicate sizes of different subgraphs, $e$s indicate numbers of actual causal edges in the subgraphs, $f$s denote the number of non-causal ordered pairs, $G$s represent resulting causal structures of the problems, $R$s and $P$s are recalls and precisions on the causal edges in the results, and $r$s are the probabilities of returning a particular false causal edge in the results. We will also utilize the following equations, $e+f=n^2-n$, $\frac{eR}{eR+fr}=P$, $e_c+f_c=\left(n_c\right)^2-n_c$, which are trivial extensions of the definitions.
Notations Description
------------------------ ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
$n, n_1, n_2, n_c$ \# of variables in $V$, $V_1\cup C$, $V_2\cup C$ and $C$
$e, e_1, e_2, e_c$ \# of causal edges in $V\!\times\! V$, $\left(V_1\!\cup\! C\right)\! \times\! \left(V_1\!\cup\!C\right)\! - \!C \!\times\! C $, $\left(V_2\!\cup\!C\right) \!\times \!\left(V_2\!\cup\!C\right) \!- \!C \!\times\! C $ and $C \!\times \!C$
$f, f_1, f_2, f_c$ \# of non-causal ordered pairs in $V\!\times\! V$, $\left(V_1\!\cup\! C\right)\! \times\! \left(V_1\!\cup\!C\right)\! - \!C \!\times\! C $, $\left(V_2\!\cup\!C\right) \!\times \!\left(V_2\!\cup\!C\right) \!- \!C \!\times\! C $ and $C \!\times \!C$
$d$ average in-degree of the causal structure
$G, G_1, G_2$ solution of $V$, $V_1\cup C$ and $V_2\cup C$
$R, R_1, R_2$ recall of $G$, $G_1$ and $G_2$
$P, P_1, P_2$ precision of $G$, $G_1$ and $G_2$
$P_m, P_{co}, P_{re}$ precision after merging, conflict removal and redundancy removal
$r, r_1, r_2$ falsely discovered probability of non-cause edges in $G$, $G_1$ and $G_2$
$e_{m},f_{m}$ \# of causal edges and non-causal ordered pairs discovered in basic merging step
$e_{ce},e_{co},e_{re}$ \# of causal edges falsely removed in causal cut finding, conflict and redundancy removal
$\Lambda(V_1,V_2)$ \# the event that there is no causal edge across $V_1$ and $V_2$ detected by the conditional independence tests
$\delta$ the largest positive constant such that $R_1 \geq R+\delta$ and $R_2 \geq R+\delta$ always hold
$\gamma$ the largest positive constant such that $r \leq r_1-\gamma$ and $r \leq r_1-\gamma$ always hold
$\alpha$ the error probability of conditional independence tests on returning $\left(v_1,v_2,V\right)$ that $v_1\not\bot v_2|V$
$\beta$ the error probability of conditional independence tests on *not* returning $\left(v_1,v_2,V\right)$ that $v_1\bot v_2|V$
$\varepsilon$ the probability of a falsely discovered causal edge has higher significance than a true causal edge
The analysis is derived based on the following general assumptions, which are commonly satisfied in real world settings. They are motivated by our observations on SADA in empirical evaluations. To begin with, the first assumption addresses the property of the causal structure.
\[assum:edgedistr\] In the causal structure, the edges are uniformly distributed on the nodes.
The above assumption ensures the local causal structures are independent of each other, which is reasonable in most of the application scenarios. Under this assumption, each edge $v_i \rightarrow v_j$ appears with probability $c/\left(n-1\right)$ when the in-degree of the variable $v_j$ is $c$.
The following assumption attempts to build the connection between recall/precision of the origin problem and recall/precision of the sub-problems.
\[assum:randp\] There exist global constants $\delta>0$ and $\gamma>0$, such that $R_1 \geq R+\delta$, $R_2 \geq R+\delta$, $r_1 \leq r-\gamma$ and $r_2 \leq r-\gamma$.
The above assumption is based on the observation that the scale of $V_1 \cup C$ and $V_2 \cup C$ is usually significantly smaller than that of $V$. Given the fixed sample set used for training, it is common to gain accuracy improvement when the basic causal solver is run on problems of smaller scale. This assumption is also empirically validated in the experiments with results available in Figure \[fig:sim\_split\].
Next assumption is used to model the significance of the discovered edges, which is crucial to the analysis of conflict and redundancy removal.
\[assum:sig\] Given a true discovered edge $v_1 \rightarrow v_2$ and a falsely discovered edge $v_3 \rightarrow v_4$, there exists a global constant $\varepsilon >0$, such that $\Pr\left(sig\left(v_1 \rightarrow v_2\right) > sig\left(v_3 \rightarrow v_4\right)\right)>1-\varepsilon$.
As defined in the partial result merging algorithm (in section \[sec:merge\]), the significance measure is the $p$-value of the edge’s reliability. It is thus reasonable that the correctly discovered causal edges are more likely to get higher significance than the falsely discovered edges.
Finally, the last assumption regards the reliability of the conditional independence tests, on two types of errors in the results of conditional independence tests.
\[assum:cit\] In the conditional independence tests, the probability that the independent relation is correctly identified as independent is at least $1-\alpha$, and the probability that the dependent relation is falsely identified as independent is at most $\beta$.
In practice, the error bounds $\alpha$ and $\beta$ could be tuned by the users by specifying appropriate confidence interval. In the rest of the paper, without other specification, we use $0.05$ as the default values for $\alpha$ and $\beta$.
Effects of Causal Cut
---------------------
Causal cut benefits SADA algorithm by improving the recall and precision of the basic causal solver when applied on sub-problems with much smaller scale. The side effect of the causal cut is the additional error overhead caused by undetected causal variable pairs which are separated in the causal cut. In this part of the section, we aim to give an analysis on the expectation of the causal cutting error in the partitioning step of SADA. To simplify the analysis, instead of using the original randomized causal cut finding algorithm, we uniformly divides the variables into $V_1$, $C$ and $V_2$, given the specific sizes $n_1$, $n_2$, $n_c$. Note that such partitioning is unaware of the actual causal structure. The causal cutting error incurred by SADA algorithm is thus definitely smaller than the estimation.
We assume that $V_1$, $C$ and $V_2$ are random variable sets output by the uniform assignment. Based on the assumptions, it is equivalent to assign the causal edges into the graph, from a null causal structure on the fixed partitioning result $V_1$, $C$ and $V_2$. We thus derive all the probabilities by simulating the random edge assignment process as following.
Let $\Psi$ denote the set of all causal structures over the current variable set $V$, and $\Psi_i$ denote a subset of $\Psi$ with exactly $i$ edges between $V_1$ and $V_2$. Given Assumption \[assum:edgedistr\], the probability of having an actual structure $\psi\in\Psi_i$ could be evaluated using Equation (\[eqn:cut1\]), as the edges are independently assigned to the variables in $V$ and there are $e$ actual causal edges and $f$ non-causal ordered pairs,
$$\label{eqn:cut1}
\Pr\left(\psi\in\Psi_i\right) = {i \choose n_1n_2} \left(\frac{e}{f+e}\right)^i \left(\frac{f}{f+e}\right)^{n_1n_2-i}$$
Intuitively, in the equation, $\frac{e}{f+e}$ denotes the probability that there is a direct edge between a particular pair of variables. Similarly, $\frac{f}{f+e}$ denotes the probability that there is no edge between a particular pair of variables.
Moreover, we can further evaluate the probability of generating a valid partitioning in terms of the algorithmic condition in SADA. Particularly, SADA does not accept a partitioning if it finds a potential causal edge between any $v_1\in V_1$ and $v_2\in V_2$. We thus derive Equation (\[eqn:cut2\]) below to evaluate the joint probability of $\psi \in \Psi_i$ and $\Lambda(V_1,V_2)$. $\Lambda(V_1,V_2)$ refers to the event that there is no causal edge across $V_1$ and $V_2$ detected by the conditional independence tests:
$$\label{eqn:cut2}
\Pr\left(\psi \in \Psi_i, \Lambda(V_1,V_2)\right) = {i \choose n_1n_2} \left(\frac{e}{f+e}\right)^i \left(\frac{f}{f+e}\right)^{n_1n_2-i}\beta^i\left(1-\alpha\right)^{n_1n_2-i}$$
Given Equation (\[eqn:cut1\]) and Equation (\[eqn:cut2\]), we apply Bayesian rule to estimate the probability of generating a partition $\left( C, V_1, V_2\right)$ by the causal cut finding algorithm, under the condition of $\Lambda(V_1,V_2)$, i.e.,
$$\begin{aligned}
\label{eqn:cut3}
\Pr\left(\psi \in \Psi_i|\Lambda(V_1,V_2)\right) & = & \frac{P\left(\psi \in \Psi_i, \Lambda(V_1,V_2)\right)}{P\left(\Lambda(V_1,V_2)\right)} \nonumber \\
& = & \frac{{i \choose n_1n_2} \left(\frac{e}{f+e}\right)^i \left(\frac{f}{f+e}\right)^{n_1n_2-i}\beta^i \left(1-\alpha\right)^{n_1n_2-i}} {\sum_{j=0}^{n_1n_2} {j \choose n_1n_2} \left(\frac{e}{f+e}\right)^j \left(\frac{f}{f+e}\right)^{n_1n_2-j}\beta^j \left(1-\alpha\right)^{n_1n_2-j}}\nonumber \\
& = & \frac{{i \choose n_1n_2} \left(\frac{e \beta}{f \left(1-\alpha\right)}\right)^i} {\sum_{j=0}^{n_1n_2} {j \choose n_1n_2} \left(\frac{e\beta}{f\left(1-\alpha\right)}\right)^j}\end{aligned}$$
The expectation of the error $e_{ce}$ caused by the causal cut finding, i.e., the number of undetected edges across $V_1$ and $V_2$, could be calculated as
$$\begin{aligned}
\label{eqn:cut_bound}
& & \mbox{Exp}\left(e_{ce}\right) \nonumber\\
& = & \sum_{i=0}^{n_1n_2} i\Pr\left(\psi \in \Psi_i|\Lambda(V_1,V_2)\right)\nonumber\\
& = & \frac{\sum_{i=0}^{n_1n_2} i {i \choose n_1n_2} \left(\frac{e\beta}{f\left(1-\alpha\right)}\right)^i} {\sum_{j=0}^{n_1n_2} {j \choose n_1n_2} \left(\frac{e\beta}{f\left(1-\alpha\right)}\right)^j} \nonumber\\
& = & \frac{\phi\sum_{i=0}^{n_1n_2}{i \choose n_1n_2} \left(\frac{e\beta}{f\left(1-\alpha\right)}\right)^i+\sum_{i=0}^{n_1n_2} \left(i-\phi\right) {i \choose n_1n_2} \left(\frac{e\beta}{f\left(1-\alpha\right)}\right)^i} {\sum_{j=0}^{n_1n_2} {j \choose n_1n_2} \left(\frac{e\beta}{f\left(1-\alpha\right)}\right)^j}\nonumber\\
& = & \frac{\phi\sum_{i=0}^{n_1n_2}{i \choose n_1n_2} \left(\frac{e\beta}{f\left(1-\alpha\right)}\right)^i+\sum_{i=0}^{n_1n_2} \left(i-\phi\right) {i-\phi \choose n_1n_2} \left(\frac{e\beta}{f\left(1-\alpha\right)}\right)^{i-\phi}\frac{{i \choose n_1n_2}}{C_{n_1n_2}^{\left(i-\phi\right)}}\left(\frac{e\beta}{f\left(1-\alpha\right)}\right)^\phi} {\sum_{j=0}^{n_1n_2} {j \choose n_1n_2} \left(\frac{e\beta}{f\left(1-\alpha\right)}\right)^j}\nonumber \\
&\leq & \phi+1\nonumber\\
&\leq & \left\lceil\frac{n^2e\beta}{4f\left(1-\alpha\right)}\right\rceil+1\end{aligned}$$
in which $\phi$ is the smallest positive integer satisfying the condition $ \frac{{i \choose n_1n_2}}{{i-\phi \choose n_1n_2}}\left(\frac{e\beta}{f\left(1-\alpha\right)}\right)^\phi \leq 1$ for any integer $i\in [\phi+1, n_1n_2]$. Because $\frac{{i \choose n_1n_2}}{{i-\phi \choose n_1n_2}}\leq \left(\frac{n_1n_2}{\phi}\right)^\phi$ holds for $ \forall i\in [\phi+1, n]$, $\phi\leq\lceil\frac{n^2e\beta}{4f\left(1-\alpha\right)}\rceil+1$ and $E\left(e_{ce}\right)$ is no larger than $\phi+1$, correspondingly.
The causal cutting error is usually small, since $\frac{e\beta}{f\left(1-\alpha\right)}$ is not large in most cases. Under a typical setting with variable number $n=100$, in-degree $d=1.25$, $\alpha=0.05$ and $\beta=0.05$, the expectation of the causal cutting error is no larger than $\left\lceil\frac{n^2e\beta}{4f\left(1-\alpha\right)}\right\rceil+1=3$, which is much smaller than the expected number of causal edges at 125.
Effects of Result Merging
-------------------------
In the merging step of SADA, the algorithm simply includes all the resulting edges from the solutions to the subproblems, i.e., $G_1$ and $G_2$. The key to our analysis in this part is to understand the recall and precision on the causal edges within the variable set $C$, because they are calculated in both subproblems on $V_1\cup C$ and $V_2\cup C$. To make the analysis possible, we try to evaluate the accuracy on these edges in $C\times C$ by estimating the number of true causal edges and false causal edges returned in the merging step.
Since the recalls of $G_1$ and $G_2$ are $R_1$ and $R_2$ respectively, and $G_1$ and $G_2$ are solved independently, the number of actual causal edges identified in $C$ is $e_c(1-\left(1-R_1\right)\left(1-R_2\right))=e_c(R_1+R_2-R_1R_2)$. Similarly, the number of falsely discovered edges $C$ is $f_c(1-\left(1-r_1\right)\left(1-r_2\right))=f_c(r_1+r_2-r_1r_2)$.
Therefore, we could derive the number of true causal edges and false causal edges by the following two equations:
$$\label{eqn:merge_true}
e_{m}=e_1R_1+e_2R_2+e_c\left(R_1+R_2-R_1R_2\right)$$
and
$$\label{eqn:merge_false}
f_{m}=f_1r_1+f_2r_2+f_c\left(r_1+r_2-r_1r_2\right)$$
Based on Equation (\[eqn:merge\_true\]), we can further derive the lower bound on the number of returned causal edges. Note that the third inequality is due to Assumption \[assum:randp\], the fourth inequality is based on the fact $e_1+e_2+e_c+e_{ce}=e$, and the last inequality applies the rule $R+\delta \leq 1$.
$$\begin{aligned}
e_{m} &=& e_1R_1+e_2R_2+e_c\left(R_1+R_2-R_1R_2\right)\nonumber\\
&\geq& e_1R_1+e_2R_2+e_cR_1 \nonumber\\
&\geq& \left(e_1+e_2+e_c\right)R+\left(e_1+e_2+e_c\right)\delta+e_c\left(R+\delta\right)\nonumber\\
&\geq& \left(e-e_{ce}\right)R+\left(e-e_{ce}\right)\delta\nonumber\\
&\geq& eR+e\delta-e_{ce}\end{aligned}$$
Thus, the lower bound on the expectation $\mbox{Exp}\left(e_{m}\right)$ could be derived as follows, in which the inequality is based on the upper bound of $\mbox{Exp}\left(e_{ce}\right)$ available in Equation (\[eqn:cut\_bound\]).
$$\label{eqn:merge_bound}
\begin{split}
\mbox{Exp}\left(e_{m}\right) &\geq eR+e\delta-\mbox{Exp}\left(e_{ce}\right)\\
&\geq eR+e\delta-\left\lceil\frac{n^2e\beta}{4f\left(1-\alpha\right)}\right\rceil-1
\end{split}$$
The following lemma provides a sufficient condition to generate higher precision on the causal edges in $C$ than that of the basic causal solver directly applied on the original problem.
\[lemma:merge\_precision\] If $\delta > \frac{Pf_c\left(r-r^2\right)}{\left(1-P\right)\left(e_1+e_2+e_c\right)}$ or $\gamma > \frac{f_cr}{f_1+f_2+2f_c}$, $P_{m}\geq P$ holds.
Basically, Equation (\[eqn:merge\_bound\]) implies that the precision is higher, i.e., $P_{m}\geq P$, if $\delta > \frac{Pf_c\left(r-r^2\right)}{\left(1-P\right)\left(e_1+e_2+e_c\right)}$. When the condition is satisfied, we have
$$\begin{aligned}
\label{eqn:merge_precision1}\small
P_{m} &=& \frac{e_m}{e_m+f_m}\nonumber\\
&=& \frac{e_1R_1+e_2R_2+e_c\left(R_1+R_2-R_1R_2\right)} {e_1R_1+e_2R_2+e_c\left(R_1+R_2-R_1R_2\right)+ f_1r_1+f_2r_2+f_c\left(r_1+r_2-r_1r_2\right)}\nonumber\\
&\geq& \frac{e_1R_1+e_2R_2+e_c\left(R_1+R_2-R_1R_2\right)} {e_1R_1+e_2R_2+e_c\left(R_1+R_2-R_1R_2\right)+ f_1r+f_2r+2f_cr}\nonumber\\
&\geq& \frac{\left(e_1+e_2+e_c\right)R+\left(e_1+e_2+e_c\right)\delta+e_c\left(R+\delta-\left(R+\delta\right)^2\right)} {\left(e_1\!+\!e_2\!+\!e_c\right)R+\left(f_1\!+\!f_2\!+\!f_c\right)r+\left(e_1\!+\!e_2\!+\!e_c\right)\delta+f_c\left(r\!-\!r^2\right)\!+\!e_c\left(R\!+\!\delta\!-\!\left(R+\delta\right)^2\right)}\nonumber\\
&\geq& \frac{\left(e_1+e_2+e_c\right)R+\left(e_1+e_2+e_c\right)\delta} {\left(e_1+e_2+e_c\right)R+\left(f_1+f_2+f_c\right)r+\left(e_1+e_2+e_c\right)\delta+f_c\left(r-r^2\right)}\nonumber\\
&\geq& P\nonumber
\end{aligned}$$
The first equality is based on the definition of precision. The first inequality is because of the facts $r_1\leq r$ and $r_2\leq r$ given in Assumption \[assum:randp\]. The second inequality is derived based on $R_1\geq R+\delta$ and $R_2\geq R+\delta$ given in Assumption \[assum:randp\]. And the last inequality is due to $\frac{\left(e_1+e_2+e_c\right)R} {\left(e_1+e_2+e_c\right)R+\left(f_1+f_2+f_c\right)r}=P$ and $\delta \geq \frac{pf_c\left(r-r^2\right)}{\left(1-P\right)\left(e_1+e_2+e_c\right)}$.
Similarly, when $\gamma \geq \frac{f_cr}{f_1+f_2+2f_c}$, we can derive the bounds on $P_{m}$ by another way as: $$\begin{aligned}
\label{eqn:merge_precision2}
& & P_{m}\nonumber\\
&=& \frac{e_1R_1+e_2R_2+e_c\left(R_1+R_2-R_1R_2\right)} {e_1R_1+e_2R_2+e_c\left(R_1+R_2-R_1R_2\right)+f_1r_1+f_2r_2+f_c\left(r_1+r_2-r_1r_2\right)}\nonumber\\
&\geq& \frac{e_1R+e_2R+e_c\left(R+R-R^2\right)} {e_1R+e_2R+e_c\left(R+R-R^2\right)+f_1r_1+f_2r_2+f_c\left(r_1+r_2-r_1r_2\right)}\nonumber\\
&\geq& \frac{\left(e_1+e_2+e_c\right)R} {\left(e_1+e_2+e_c\right)R+\left(f_1+f_2+f_c\right)r+f_cr-\left(f_1+f_2+2f_c\right)\gamma}\nonumber\\
&\geq& P\nonumber
\end{aligned}$$
The first equality is based on the definition of precision. The first inequality is because of $R_1\geq R$ and $R_2\geq R$, given in Assumption \[assum:randp\]. The second inequality is because of $r_1\leq r-\gamma$ and $r_2\geq r-\gamma$, given in Assumption \[assum:randp\]. The last inequality is because of $\frac{\left(e_1+e_2+e_c\right)R} {\left(e_1+e_2+e_c\right)R+\left(f_1+f_2+f_c\right)r}=P$ and $\gamma \geq \frac{f_cr}{f_1+f_2+2f_c}$. This completes the proof of the lemma.
Effects of Conflict Removal
---------------------------
The step of conflict removal is expected to eliminate the false causal edges returned by the merging step, under the potential risk of falsely removing actual causal edges. As is shown in Algorithm \[algo:merge\], the selection of the removal edges heavily depends on the significance measure employed on candidate edges. In this part of the section, we analyze how the randomness on the significance measure affects the accuracy of results after conflict removal.
Given an edge $v_i \rightarrow v_j$, there are two types of conflicts to address, including (1) conflict between two edges, e.g. $v_i \rightarrow v_j$ against $v_i \leftarrow v_j$; and (2) conflict between an edge and a path, e.g. $v_i \rightarrow v_j$ against $v_i \ldots \leftarrow \ldots v_j$.
In the first type of conflict, the variable pair $v_i$ and $v_j$ exist on both $G_1$ and $G_2$. Thus, the number of conflict edge pairs between $v_i \rightarrow v_j$ and $v_i \leftarrow v_j$ can be estimated as $e_c\left(r_2R_1+R_2r_1\right)+f_cr_1r_2$. In the estimation, $e_cr_2R_1$ denotes the number of actual causal edges correctly discovered in $G_1$ with a corresponding reversed edge included in $G_2$. Similarly, $e_cr_1R_2$ denotes the number of actual causal edges correctly discovered in $G_2$, while a reversed one is available in $G_1$ at the same time. Finally, $f_cr_1r_2$ is the number of edge pairs, which are both false and reversed to each other. Based on Assumption \[assum:sig\], the expected number of actual edges removed by the current step is $\varepsilon e_c\left(r_2R_1+R_2r_1\right)$, by only considering pairs with at least one actual causal edge.
The second type of conflict is in the form $v_i \rightarrow v_j$ and $v_i \ldots \leftarrow \dots v_j$. Because the solutions to the subproblems, i.e., $G_1$ and $G_2$, are acyclic and there is no direct edge across the variable set $V_1$ and $V_2$, the conflict of second type are definitely triggered by the edges within $C\times C$. When merging results from $G_1$ and $G_2$ in terms of the edges in $C\times C$, there are $e_cR_2\left(1-R_1\right)$ additional true causal edges and $\left(n_c^2-n_c-e_c\right)r_2\left(1-r_1\right)$ additional false causal edges incurred by $G_2$. Similarly, there are $e_cR_2\left(1-R_1\right)+f_cr_2\left(1-r_1\right)$ extra edges from the results $G_1$ when merging the $G_2$’s results on $C$ to $G_1$. Thus, there are $e_c\left(R_1+R_2-2R_1R_2\right)+f_cr_2\left(r_1+r_2-2r_1r_2\right)$ edges potentially triggering conflicts between the edges.
Consider a particular edge $v_i \rightarrow v_j$ and the counter-result with path $v_i \ldots \leftarrow \dots v_j$. If there are $k$ intermediate variables on the path, the path appears with probability at most ${n-2\choose k}\left(\frac{d}{n}\right)^{k+1}$, in which $d$ is the maximal in-degree in the variables. By iterating on all possible lengths from 1 to $n-2$, the expected number of conflicted paths triggered by $v_i \rightarrow v_j$ is at most $\sum_{k=1}^{n-2} {n-2\choose k}\left(\frac{d}{n}\right)^{k+1}=\frac{d}{n}\left(\left(1+\frac{d}{n}\right)^{n-2}-1\right)= \frac{d}{n}\left(1+\frac{d}{n}\right)^{n-2}-\frac{d}{n}$.
During the conflict removal step, it is necessary and sufficient to remove exactly one edge with the lowest significance on the path to break the conflict. Such a removed edge is either an actual causal edge with the lowest significance or a false causal edge with the lowest significance. According to Assumption \[assum:sig\], the probability of generating lower significance for an actual causal edge against a false causal edge is as small as $\varepsilon$. It facilitates us to calculate an upper bound on the removed actual causal edges by $\left(e_c\left(R_1+R_2-2R_1R_2\right)+f_c\left(r_1+r_2-2r_1r_2\right)\right)\varepsilon$.
Combing both two types of conflicts, the expected number of actual edges removed in the conflict removal step over all conflict cases are upper bounded in Equation (\[eqn:conflict\_bound\]). The first inequality is derived by the fact that $R_1$, $R_2$, $r_1$ and $r_2$ are no greater than $1$. The second inequality is because $e_c+f_c=\left(n_c\right)^2-n_c < \left(n_c\right)^2$.
$$\begin{aligned}
\label{eqn:conflict_bound}
& & \mbox{Exp}\left(e_{co}\right)\nonumber\\
&=& \varepsilon \left(e_c\left(r_2R_1\!+\!R_2r_1\right)+ \left(\frac{d}{n}\left(1\!+\!\frac{d}{n}\right)^{n-2}\!\!-\!\frac{d}{n}\right)\left(e_c\left(R_1\!+\!R_2\!-\!2R_1R_2\right)+f_c\left(r_1\!+\!r_2\!-\!2r_1r_2\right)\right)\right)\nonumber\\
&\leq& \varepsilon \left(2e_cr+ \frac{d}{n}\left(1+\frac{d}{n}\right)^{n-2}e_c+2f_cr\right)\nonumber\\
&\leq& \varepsilon \left(2\left(n_c\right)^2r+ \frac{d^2n_c}{n}\left(1+\frac{d}{n}\right)^{n-2}\right)\end{aligned}$$
The following lemma gives the sufficient condition to ensure that the precision never drops after the conflict removal step in SADA.
\[lemma:conflict\_precision\] When the error of significance measure $\varepsilon$ is no larger than $1-P$, the precision on the returned causal edges after conflict removal never drops.
To prove the lemma, we take each conflict into consideration and update the precision on the results once at a time.
Due to the acyclic property of the actual causal structure, each conflict cycle must contain at least one false causal edge. We consider two different type of cases in this proof. The first type includes cases of conflict containing false causal edges only. Since at least one false causal edge is removed in the step, the conflict removal definitely improves the precision.
The second type of conflicts contains at least one actual causal edge in each conflict. With probability no larger than $\epsilon$, an actual causal edge is removed from the result, otherwise a false causal edge is deleted. Assume that there are $e'$ actual causal edges and $f'$ false causal edges in the result at this particular moment. When $\epsilon\leq 1-P$, the expectation of the new precision after breaking this conflict is no smaller than $P$.
$$\label{eqn:conflict_precision2}
\begin{split}
\varepsilon \frac{e'-1}{e'+f'-1} + \left(1-\varepsilon\right) \frac{e'}{e'+f'-1}= \frac{e'-\varepsilon}{e'+f'-1}\geq P.
\end{split}$$
The last inequality is derived by $\frac{e'}{e'+f'}\geq P$ and $\varepsilon \leq 1-P$. This completes the proof of the lemma.
Given the conclusion of the lemma, when $\varepsilon$ is sufficiently smaller, the conflict removal always brings benefit to the precision of the results, i.e., $\mbox{Exp}\left(P_{co}\right)\geq P$ holds.
Effects of Redundancy Removal
-----------------------------
We apply similar analysis strategy on the redundancy removal step as is done on the conflict removal step. Since the redundancy between $v_i\rightarrow v_j$ from $G_1$ and $v_i\rightarrow v_j$ from $G_2$ is already broken in the basic merging step, we only need to consider the redundancy between $v_i\rightarrow v_j$ and $v_i \rightarrow \cdots \rightarrow v_j$ in this step of SADA algorithm.
Similar to the results on the conflict removal step, the path $v_i \dots \rightarrow \dots v_j$ with $k$ intermediate variables appears with probability ${{n-2} \choose k}\left(\frac{d}{n}\right)^{k+1}$. Considering all the paths with length within the range $k\in[1, n-2]$, the expected number of redundancy path for $v_i\rightarrow v_j$ is $\sum_{k=1}^{n-2} {n-2\choose k}\left(\frac{d}{n}\right)^{k+1}=\frac{d}{n}\left(\left(1+\frac{d}{n}\right)^{n-2}-1\right)= \frac{d}{n}\left(1+\frac{d}{n}\right)^{n-2}-\frac{d}{n}$. Moreover, there are $e_c\left(R_1+R_2-2R_1R_2\right)+f_c\left(r_1+r_2-2r_1r_2\right)$ extra edges potentially triggering redundancy cycles. To eliminate the redundancy for each of the case, it is to remove $\beta$ actual causal edges in average, because the conditional independence tests are used to detect such redundancy. The expected number of actual causal edges removed in the redundancy step over all redundancy cases is thus upper bounded by the following formula, with limited impact on the recall of the results.
$$\label{eqn:redundancy_bound}
\begin{split}
\mbox{Exp}\left(e_{re}\right)& \leq \beta \left(e_c\left(R_1+R_2-2R_1R_2\right)+f_c\left(r_1+r_2-2r_1r_2\right)\right)\\
&\leq \beta \left(\frac{d}{n}\left(1+\frac{d}{n}\right)^{n-2}e_c+2\left(n_c\right)^2r\right)\\
&\leq \beta \left(2\left(n_c\right)^2r+ \frac{d^2n_c}{n}\left(1+\frac{d}{n}\right)^{n-2}\right)\\
\end{split}$$
Regarding the precision, the following lemma gives the sufficient condition to the improvement on precision by the redundancy removal step.
\[lemma:redundancy\_precision\] If the error probability $\beta$ is no larger than $1-P$, the expected precision $\mbox{Exp}\left(P_{re}\right)\geq P$ holds.
Let $P'=\frac{e'}{e'+f'}$ denote the current precision $P'=\frac{e'}{e'+f'}$ and $P''$ is the precision after removing one potential redundant edge. Given the condition $P'\geq P$, We have the following inequality about $P''$.
$$\label{eqn:redundancy_precision}
\begin{split}
\mbox{Exp}\left(P''\right)&=\beta \frac{e'-1}{e'-1+f'} + \left(1- \beta\right) \frac{e'}{e'+f'-1}= \frac{e'-\beta}{e'+f'-1}\geq P
\end{split}$$
The last inequality is because of $\frac{e'}{e'+f'}=P' \geq P$ and $\beta \leq 1-P$.
In the potential redundancy removal, the initial precision is $P'=P_{co}$ and $P'\geq P$ holds. Thus, $\mbox{Exp}\left(P_{re}\right)\geq P$ holds.
Overall Evaluation of SADA
--------------------------
In this section, we combine all the results in previous subsection and provide an overall evaluation on the recall/precision of SADA.
Equation (\[eqn:recall\]) gives an estimation on the recall after each recursive partitioning step. Theorem \[theorem:recall\] gives the sufficient condition of SADA’s recall is higher than that of the basic causal solver.
$$\label{eqn:recall}
\mbox{Exp}\left(R_{SADA}\right)=\frac{\mbox{Exp}\left(e_{m})-\mbox{Exp}\left(e_{co}\right)-\mbox{Exp}\left(e_{re}\right)\right)}{e}\\$$
\[theorem:recall\] When $\delta \geq e^{-1}\left(\lceil\frac{n^2e\beta}{4f\left(1-\alpha\right)}\rceil+1+\left(\varepsilon+\beta\right)n_c\left(2n_cr+ \frac{d^2}{n}\left(1+\frac{d}{n}\right)^{n-2}\right)\right)$, we always have $\mbox{Exp}\left(R_{SADA}\left(G\right)\right) \geq R\left(G\right)$.
Combing the lower bound on the number of discovered edges in the merging step (in Equation (\[eqn:merge\_bound\])), the upper bound of falsely removed causal edges in the conflict removal (in Equation (\[eqn:conflict\_bound\])) and redundancy removal step (in Equation (\[eqn:redundancy\_bound\])), we come to the following conclusion on the expectation of $R_{SADA}$:
$$\label{eqn:recall2}
\begin{split}
\mbox{Exp}\left(R_{SADA}\right)&=\frac{\mbox{Exp}\left(e_{m}\right)-\mbox{Exp}\left(e_{co}\right)-\mbox{Exp}\left(e_{re}\right)}{e}\\
&\geq \frac{eR+e\delta-\lceil\frac{n^2e\beta}{4f\left(1-\alpha\right)}\rceil-1 -\left(\varepsilon+\beta\right)n_c\left(2n_cr+ \frac{d^2}{n}\left(1+\frac{d}{n}\right)^{n-2}\right) }{e} \\
&\geq R
\end{split}$$
The last inequality is because of $\delta \geq \frac{1}{e}\left(\lceil\frac{n^2e\beta}{4f\left(1-\alpha\right)}\rceil+1+\left(\varepsilon+\beta\right)n_c\left(2n_cr+ \frac{d^2}{n}\left(1+\frac{d}{n}\right)^{n-2}\right)\right)$.
The above theorem gives the sufficient condition to generate higher recall than the basic causal solver directly applied on the original problem. In the following, we demonstrate that this sufficient condition can be easily satisfied in real applications. Given the typical setting with variable number $n=100$, average in-degree $d=1.25$, $\alpha=0.05$, $\beta=0.05$, $\varepsilon=0.05$ and $n_c=10$, the minimal $\delta$ required is $0.0404$. We also illustrate the minimal $\delta$s under different average in-degree $c$ and causal cut set size $n_c$ in Figure \[fig:delta\]. The results in the figure shows that $5\%$ improvement on subproblems with smaller domain is enough to help improve the accuracy by employing SADA. Even when the size of causal cut set is as large as 20, SADA is capable of achieving better accuracy if the basic causal solver is able to improve $10\%$ on the subproblems. Figure \[fig:delta\_nc\] also reflects the fact that it is important to control the size of causal cut set, under which SADA could guarantee more performance enhancements.
\[h\]
\[theorem:precision\] When the following three conditions hold: (1) $\delta > \frac{Pf_c\left(r-r^2\right)}{\left(1-P\right)\left(e_1+e_2+e_c\right)}$ or $\gamma > \frac{f_cr}{f_1+f_2+2f_c}$, (2) $\varepsilon$ is no larger than $1-P$ and (3) $\beta$ is no larger than $1-P$, we always have $\mbox{Exp}\left(P_{SADA}\right)\geq P$.
Based on Lemma \[lemma:merge\_precision\], $P_{m}\geq P$ holds after the merging step. According to Lemma \[lemma:conflict\_precision\] and \[lemma:redundancy\_precision\], the precision is not also reduced in each process of conflict removal and redundancy removal. Thus, $\mbox{Exp}\left(P_{SADA}\right)\geq P$ holds.
Given the same setting as used in Figure \[fig:delta\], when the basic causal solver achieves precision $P=0.5$ on the original problem, SADA improves the precision when $\delta >0.08$ or $\gamma >0.002$. Although the minimal requirement on $\delta$ for better precision is higher than that for better recall, both the increase on the recall on true causal edges and the decrease on the number of falsely discovered edges could contribute to the improvement of the precision.
Note that both the conditions given in Theorem \[theorem:recall\] and Theorem \[theorem:precision\] are *only* sufficient conditions to the accuracy improvements in SADA, and with a number of loose inequalities are in the proof of the theorems. Thus, SADA improves the performance of the basic causal solvers under much more general conditions in practice, i.e., SADA still improves the performance of the basic causal solvers even when the above conditions are not fully satisfied. In the experiments, we empirically evaluate the effects and verify the advantages of SADA.
Experiments {#sec:exp}
===========
Experiment Settings
-------------------
We evaluate our proposal on datasets generated by simulated and different real-world Bayesian network structures[^2], under linear non-Gaussian model and discrete additive noise model. Because of the non-existence of large scale causal inference problem with ground truth, simulated data on the given structures is used in most of causal structure learning methods [@aliferis_local_2010; @kalisch2007pc]. Please note that only faithfulness condition and causal sufficiency assumption are employed in the generic SADA framework. Additional compatible data generation assumptions, linear non-Gaussian assumption and additive noise assumption, are employed for linear non-Gaussian model and discrete additive noise model, respectively.
**Linear Non-Gaussian Acyclic Model**\
Under the assumption of linear non-Gaussian acyclic model, the samples are generated based on linear functions as $v_i=\sum_{v_j\in P\left(v_i\right)}w_{ij}v_j+e_i$. When randomly generating these linear functions, we restrict that $\sum_{P\left(v_i\right)}w_{ij}=1$ and the variance $Var\left(e_i\right)=1$ for every variable $v_i$.
On the linear non-Gaussian acyclic model, our algorithm is compared with LiNGAM [@ShimizuJMLR06], DLiNGAM [@ShimizuJMLR2011Direct] and Sparse-ICA LiNGAM [@zhang2009sica]. The implementation of LiNGAM and DLiNGAM are from the authors of the paper. The implementation of Sparse-ICA LiNGAM is based on the sparse-ICA of [@zhang2009sica], and the pruning algorithm of [@ShimizuJMLR06]. For SADA, we employ the conditional independence tests following the method proposed in [@Baba2004PartialCorrelation], with threshold at 95%. LiNGAM [@ShimizuJMLR06] with Wald test is appointed as the basic casual solver $A$ after SADA reaches the minimal scale threshold $\theta$ at subproblems. On all datasets, SADA stops the partitioning when the subproblem reaches the size $\theta=10$. The recursive partitioning is also terminated when Algorithm \[algo:split\] fails to find any causal cut. LiNGAM without applying any division is also used as the baseline approach, when reporting recall, precision and F1 score. Note that, when the variable size is larger than 100, LiNGAM cannot perform Wald test due to memory consumption constraint (i.e., one Wald test cannot be finished on a sever with 64GB memory for a whole day).
**Discrete Additive Noise Model**\
The generation process of the discrete data follows the cyclic method used in [@Janzing2011tpamidiscrete] under additive noise model (ANM) for causation discovery on discrete data. Each variable is restricted to 3 different values and values are randomly generated based on conditional probability tables. The implementation of SADA for discrete domain is slightly different from that for continuous domain. $G^2$ test [@spirtes2011] is employed as the conditional independence test, with the threshold at $95\%$. The basic causal solver $A$ called by SADA is a brute force method to find all causalities on problems of small scale. Again, the brute-force ANM without variable division is also employed as a baseline approach. The ANM algorithm checks every possible pair of variables following the method proposed in [@Janzing2011tpamidiscrete], and the redundancy and conflict edges are removed using the similar method as the merging step of SADA.
In all the experiments, the evaluation metric includes, *causal cutting error*, *recall*, *precision* and *F1 score*. The causal cutting error ratio is $e_{ce}/e$, i.e., the number of falsely removed causal edges in the causal cut finding step to the number of all causal edges. F1 score is calculated as $\frac{2P\times R}{P+R}$, in which $R$ and $P$ are recall and precision on the causal edge results respectively. Causal cutting error is evaluated on SADA, and the other three metric are evaluated for both SADA and the baseline method. The experiments are compiled and run with Matlab 2009a on a windows PC equipped with a dual-core 2.93GHz CPU and 2GB RAM, and a Linux sever with a 16-core 2.0GHz CPU and 60GB RAM. All Matlab codes of the causation discovery and the generator for linear non-Gaussian data are available at <https://sites.google.com/site/cairuichu/SADA.zip> for testing.
Results on Simulated Structure
------------------------------
The main purpose of this group of experiments is to evaluate the scale effect on the basic causal solvers and the sensitivity of our proposal to the variable size, sample size, connectivity, and other characteristics of the causal structures.
The simulated structures are randomly generated under control of a few parameters, including the variable size and average in-degree. The average in-degree reflects the local connectivity of the causal structure. In the structure generation process, all variables are sorted in topological order of the simulated causal structure, so that parent variables are always ahead of children variables. The samples are generated exactly in the order, ensuring that the values of parent variables are generated before the children nodes. The details of the causal structure generation can be found in Algorithm \[algo:csg\] in the appendix. Given the causal structures, the data is generated using linear non-Gaussian model or discrete model as described earlier. For linear non-Gaussian model, an additional parameter, called noise weight, is used to control the ratio of noise in the data generation process, please refer to Algorithm \[algo:ldg\] for the details. The parameters settings, including both structure generation and data generation phrases, are given in Table \[tab:para\], with default values highlighted in bold font.
Paremeter Range
------------------------ ---------------------------------
Variable Size {25, 50, **100**, 200, 400 }
Sample Size {50, 100, **200**, 400, 800 }
In-degree {0.75, 1, **1.25**, 1.5, 1.75 }
Noise Weight in LiNGAM {0.1, 0.2, **0.3**, 0.4, 0.5 }
: Ranges and Defaults of the Parameters in Simulated Structures[]{data-label="tab:para"}
**Effects of Partitioning on the Basic Causal Solver’s Recall and Precision**
In the theoretical analysis of the framework, we have shown that the improvement of recall and precision highly depends on the amount of accuracy gain on subproblems with smaller scales. While the conditions theoretically guarantee effectiveness, the actual impact is hardly reflected in the fairly loose bounds. To understand the actual effects of the partitioning, we test the splitting procedure under careful control on the sample size, in-degree and noise ratio, and vary on the number of variables.
The effects of partitioning on the basic causal solver’s recall and precision are summarized in Figure \[fig:sim\_split\], on the subproblems with size{100,50,25,13,6}. On the linear non-Gaussian data, the precision increases after each split, and the benefits on recall emerge when the variable size is small enough. Since SADA is capable of generating smaller subproblems for the basic causal solvers, the split brings benefits to both recall and precision. On discrete data, the recall is relative stable for it only checks each pair of variables independently and has less connection to the domain size, i.e., the number of variables. The precision increases after each split, as the suspicions causal pairs can be removed in the split step. The phenomena reflect that the two sufficient condition for the improvement of recall and precision can be more easily satisfied on the linear non-Gaussian data, while only the sufficient condition for the improvement of precision can be easily satisfied on the discrete data.
\[h\]
**Sensitivity to Repeat Time of Finding Causal Cut**
The repeat time of finding causal cut is another parameter of SADA. In Theorem \[theorem:convergence\], we provide a bound on the causal cut size when $k=(2d_m+2)^2$. It is interesting to investigate the effect of the different setting of this parameter. In this experiment, we following parameters are tested, $k=\{1, 5, 10, 15, 20\}$, where 20 is setting based on the $k=(2d_m+2)^2$.
Figure \[fig:sim\_sens\_repeat\] shows the sensitivity of SADA to the repeat time of finding causal cut. It is interesting to find that the algorithm works well when $k=1$, and the improvement is trivial with increasing of the repeat time. Because of the high computational complexity of the causal cut finding, the repeat time is 1 for all the following experiments.
\[h\]
**Scalability to Domain Size**
Figure \[fig:sim\_sens\_var\_lingam\] and Figure \[fig:sim\_sens\_var\_disc\] report the effects on recall, precision and F1 score, under varying number of variables, on linear non-Gaussian data and discrete data. Generally, SADA works much better on all different numbers of variables. Note that the gap between the methods grows when more variables are in the data domain. This property ensures SADA’s scalability to large domains. On the linear non-Gaussian data, LiNGAM fails to work when the variable size is larger than 100, while SADA still achieves good accuracy performance. The figures further strengthens the conclusions of Theorem \[theorem:recall\] and Theorem \[theorem:precision\]. In particular, on linear non-Gaussian data, both the sufficient conditions for the improvement of precision and recall are satisfied (as illuminated in Figure \[fig:sim\_split\_lingam\]). Compared against the LiNGAM, both the recall and precision are lifted. On discrete data, the sufficient condition for the improvement of precision is satisfied as well (as illuminated in Figure \[fig:sim\_split\_disc\]), the improvement mainly stems from the enhanced precision on the results.
\[h\]
**Sensitivity to the Sample Size**
Figure \[fig:sim\_sens\_sample\] analyzes the sensitivity of SADA and the compared methods to the sample size. SADA works better than the compared methods, regardless of the sample size on both linear non-Gaussian data and discrete data. Moreover, SADA also works well even when LiNGAM fails to work on the linear non-Gaussian data, as show in Figure \[fig:sim\_sens\_sample\_lingam\]. When sample size is 50 or 100, the sample size is smaller than the number of variables. In such case, LiNGAM fails to work, while SADA framework performs well. This is because SADA can effectively split the original problem into small subproblems solvable to the basic causal solver with small sample size. This is a fundamental advantage of the SADA framework. The performance of SADA improves slightly with the increase on sample size. There are two reasons behind the improvement. Firstly, large sample size improves the reliability of the conditional independence tests (i.e., smaller $\alpha$ and $\beta$) used to find the causal cut and reduce the causal cutting error of SADA. Secondly, large sample size helps basic causal solvers on accuracy, as there are more observations to identify reliable causal relations.
\[h\]
**Sensitivity to the Connectivity**
Figure \[fig:sim\_con\_sample\] shows the sensitivity of the algorithms to the average in-degree, an important metric to reflect the connectivity of causal structures. The performance of SADA drops with the growth of average in-degree, caused mainly by the large causal cut set size in the dense causal structures. As analyzed in Theorem \[theorem:convergence\], the causal cut finding strategy is highly dependent on the in-degree. Thus, the increasing average in-degree will reduce the quality of the causal cut and increase the causal cut set size. Recall the conclusion in Theorem \[theorem:recall\] and Theorem \[theorem:precision\], the sufficient condition of the improvement will be difficult to be satisfied in the partition with large causal cut set. Though SADA’s advantage over LiNGAM is small when the average in-degree is 1.75 on the linear non-Gaussian data, SADA is still competitive, for most of the real world causal structures are sparse as discussed in Table \[tab:dataset\]. Similar conclusions could be drawn on the discrete data.
\[h\]
**Sensitivity to Noise**
This set of experiments are only conducted on the linear non-Gaussian data, as it is way too difficult to control the noise ratio in the generation of discrete data. The noise has several effects in SADA, firstly moderate ratio of noise contributes to determined the direction of the causality in the basic causal solvers; secondly the noise will decrease the reliability of the conditional independence tests and further reduce the quality of the partition; thirdly, too much noise also reduces the quality of the basic causal solvers. When the noise weight is less than 0.3, both SADA and LiNGAM’s performance is insensitive to the noise because of the trade-off between the first two effects. When the noise ratio is larger than 0.3, both recall and precision reduce with the increasing of noise because of the third effect.
\[h\] ![Sensitivity to the noise on linear non-Gaussian model.[]{data-label="fig:sim_noise"}](figs/sim_sens_noise_lingam.eps "fig:"){width="47.00000%"}
Results on Real-World Structures
--------------------------------
It generally, real-world Bayesian network structures cover a variety of applications, including, medicine (*Alarm* dataset), weather forecasting (*Hailfinder* dataset), printer troubleshooting (*Win95pts* dataset), pedigree of breeding pigs (*Pigs* dataset) and linkage among genes (*Link* dataset). The structural statistics of these Bayesian networks are summarized in [Table \[tab:dataset\]]{}. In all the Bayesian networks, the maximal degrees, i.e., the maximal number of parental variables in the networks, are no larger than 6, regardless of the total number of variables. This verifies the correctness of our sparsity assumption.
Dataset Variable \# Avg degree Max degree
-------------- ------------- ------------ ------------
*Alarm* 37 1.2432 4
*Hailfinder* 56 1.1786 4
*Win95pts* 76 0.9211 6
*Pigs* 441 1.3424 2
*Link* 724 1.5539 3
: Statistics on the Datasets[]{data-label="tab:dataset"}
**On Linear Non-Gaussian Model**
The causal cutting errors are reported in Figure \[fig:lingam:de\], on varying the number of samples generated by the Bayesian networks. Even when the samples size is $2|V|$, the highest causal cutting error is within 0.12. Moreover, the causal cutting errors generally decrease with the growth of sample size. These results reveal the fundamental advantage of SADA, such that the sufficient number of samples only depends on the sparsity of the causal structure but not the number of variables. Note that the baseline approach LiNGAM does not work when the number of samples are as small as $2|V|$.
\[!t\]
![causal cutting error ratio on linear non-Gaussian models.[]{data-label="fig:lingam:de"}](figs/lingam_de.eps){width="47.00000%"}
In the following experiments, we compare SADA against the baseline approach by fixing the sample size at $2|V|$. As shown in Table \[tab:lingam:rp\], SADA achieves significantly better F1 score on all of the five datasets. SADA is particularly doing well on precision, i.e., returning more accurate causal relations. SADA’s division strategy is the main reason behind the improvement of precision on SADA. Specifically, the division on variables allows SADA to remove a large number of candidate variable pairs if they are assigned to $V_1$ and $V_2$. The basic causal solver is run on subproblem of much smaller scale, thus generating more reliable results. The Recall of SADA is comparable to LiNGAM on four of the datasets, and slightly worse on the other one. This shows that the unavoidable causal cutting error does not affect the recall under linear non-Guassian models.
-------------- ---------- ---------- ------ ------ ---------- ------ ------ ------ ---------- ------ ------ ------
*Alarm* **0.41** 0.24 0.02 0.20 **0.36** 0.30 0.20 0.33 **0.38** 0.27 0.04 0.25
*Hailfinder* **0.52** 0.24 0.23 0.42 **0.46** 0.13 0.15 0.45 **0.49** 0.17 0.18 0.39
*Win95pts* **0.57** 0.41 0.07 0.43 **0.42** 0.23 0.10 0.45 **0.48** 0.30 0.08 0.44
*Pigs* 0.56 **0.57** N.A. N.A. **0.23** 0.12 N.A. N.A. **0.33** 0.19 N.A. N.A.
*Link* **0.62** 0.53 N.A. N.A. **0.25** 0.07 N.A. N.A. **0.36** 0.13 N.A. N.A.
-------------- ---------- ---------- ------ ------ ---------- ------ ------ ------ ---------- ------ ------ ------
**On Discrete Additive Noise Model**
The causal cutting error of SADA on the discrete data is presented in Figure \[fig:disc:de\], which shows similar property of the result on linear non-Gaussian models. This further verifies the generality of SADA on different data domains.
\[!t\]
![Causal cutting error ratio on discrete models.[]{data-label="fig:disc:de"}](figs/disc_de.eps){width="47.00000%"}
In this group of experiments, we fix the sample size at 2000, and report recall, precision and F1 score in Table \[tab:disc:rp\]. Note that ANM is only applicable to domain with small number of variables. Because it cannot finish the computation on *Pigs* and *Link* in one week. This proves the improv ement of SADA on scalability in terms of the variables. Generally speaking, the results in the table also verify the effectiveness of SADA, especially the enhancement on precision and F1 score.
-------------- ---------- ---------- ---------- ------ ---------- ------
*Alarm* **0.67** 0.65 **0.72** 0.60 **0.70** 0.63
*Hailfinder* 0.71 **0.76** **0.57** 0.45 **0.63** 0.56
*Win95pts* 0.68 **0.71** **0.41** 0.38 **0.51** 0.49
*Pigs* **0.68** N.A. **0.50** N.A. **0.58** N.A.
*Link* **0.69** N.A. **0.46** N.A. **0.56** N.A.
-------------- ---------- ---------- ---------- ------ ---------- ------
: Results on Discrete Model[]{data-label="tab:disc:rp"}
As a conclusion, SADA shows excellent performance on 5 different domains with real-world Bayesian networks. SADA returns accurate causal structure when combined with two well known causal inference algorithms. The causal cut used to partition the problem does incur certain error on incorrect partitioning. Despite of the errors, SADA still outperforms ANM without partitioning on almost all settings.
Conclusion {#sec:concl}
==========
In this paper, we present a general and scalable framework, called SADA, to support causal structure inference, using a *split-and-merge* strategy. In SADA, causal inference problem on a large variable set is partitioned into subproblems with overlapping subsets of variables, utilizing the concept of causal cut. Our proposal facilitates existing causation discovery algorithms to handle problem domains with more variables and less samples, which extend the application scenarios of causation discovery. Strong theoretical analysis proves the effectiveness, correctness and completeness guarantee of SADA under a general setting. Experimental results further verifies the usefulness of the new framework with two mainstream causation algorithms on linear non-Gaussian model and discrete additive noise model. Theoretical and experimental analysis of SADA reveal the fundamental advantage of our approach, that the required sample depends on the generating graph connectivity and not the size of the variable set; this yields up to exponential savings in sample relative to previously known algorithms.
While our methods haven shown improvement over existing methods, we believe there remains room for further enhancement. One possible direction is to attempt other existing randomized division strategies commonly used to tackle combinatorial problems on graph data. Another interesting problem is how to reduce the computational cost when subproblems have a large overlap on variables.
Pseudocodes of Causal Structure Generation
==========================================
($n$, $d$)\
Set $G$ to an $n\times n$ false matrix;\
Pseudocodes of Linear Non-Gaussian Data Generation
==================================================
($CS$, $w$, $n$)\
[^1]:
[^2]: [www.cs.huji.ac.il/site/labs/compbio/Repository/](www.cs.huji.ac.il/site/labs/compbio/Repository/)
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We suggest a scalar model for deformation and flow of an amorphous material such as a foam or an emulsion. To describe elastic, plastic and viscous behaviours, we use three scalar variables: elastic deformation, plastic deformation rate and total deformation rate; and three material specific parameters: shear modulus, yield deformation and viscosity. We obtain equations valid for different types of deformations and flows slower than the relaxation rate towards mechanical equilibrium. In particular, they are valid both in transient or steady flow regimes, even at large elastic deformation. We discuss why viscosity can be relevant even in this slow shear (often called “quasi-static") limit. Predictions of the storage and loss moduli agree with the experimental literature, and explain with simple arguments the non-linear large amplitude trends.'
author:
- Philippe Marmottant
- François Graner
title: 'An elastic, plastic, viscous model for slow shear of a liquid foam'
---
Introduction
============
Elastic materials deform reversibly [@lan86]; plastic materials can be sculpted, that is, they can be deformed into a new shape and keep it [@cha87]; and viscous materials flow [@bat00]. A wide variety of materials display a combination of these properties, such as elasto-plastic metals and rocks, visco-elastic polymer solutions or visco-plastic mineral suspensions [@mac94; @fra93; @fra95].
Liquid foams, that is gas bubbles separated by liquid walls, are visco-elasto-plastic [@kra88; @wea99; @hoh05]: they are elastic at low strain, plastic at high strain and flow under high shear rate. This is also the case for other concentrated suspensions of deformable objects in a liquid [@mac94; @lar99; @mas96], such as droplets in emulsions, vesicles suspensions, or red blood cells in blood.
Despite a large literature on experiments and simulations (see [@hoh05] for a review), we lack an unified theoretical description of foams. There is no consensus yet on a central question: what are the physically relevant variables? A series of statistical models focus on fluctuations and their correlations [@sol97; @fal98; @kab03; @pic05; @cou05]. Conversely, recent contributions [@mar06; @jan06; @miy06; @wys06; @lab04; @hoh06; @sar06] focus on average macroscopic quantities to obtain a more classical continuous description.
Here we choose to group three macroscopic quantities which are measurable as averages on microscopical details [@mar06]: (i) Elastic deformation is a state variable [@por97] reversibly stored by the foam’s microstructure, that is, the shape of bubbles [@aub03; @asi03]; it determines the elastic contribution to the stress. (ii) Plastic deformation results in energy dissipation analogous to solid friction. (iii) Large scale velocity gradients are associated with a viscous friction. Each of the three mechanical behaviors is associated with a material specific parameter: elastic modulus, yield deformation and viscosity.
For simplicity, we assume here that these parameters are constant and the equations are linear. We consider here homogeneous deformation of a material, not depending on space coordinates. We consider only the magnitude of deformation, but not spatial orientation: the material state variables are all scalars. This represents an incompressible liquid foam, where the deformation is a pure shear. We assume that this shear is slow enough so that the foam is always close to mechanical equilibrium, but quick enough to neglect coarsening such as due to gas diffusion between bubbles, or bubble coalescence due to soap film breakage. Although this model is minimal, it is written with enough generality to enable for extensions to higher dimensions using tensors (the correspondance with tensors introduces a factor $1/2$, see section \[sec:Perspectives\]), to higher shear rates, and to other ingredients such as external forces (to be published).
This article is organised as follows. Section \[sec:qstat\_model\] introduces a visco-elasto-plastic model (eqs. \[eq-evol-u-scalar\],\[eq:stresselasticviscous\]) based on two scalar variables: the elastic deformation and the (slow) shear rate (Fig. \[diagramme\_vep\]). The rate of plastic deformation is determined by both the applied shear rate, and the current state of the elastic deformation (or equivalently the elastic part of the stress) rather than by the total stress [@cri82; @mil87]. Section \[predict\] presents scalar predictions of creep and oscillatory responses. The storage and loss moduli predicted as a function of the strain amplitude agree with experimental data without any adjustable parameters, using only the three model-independent parameters determined by experiments (yield point, shear modulus, viscosity). The agreement becomes very good if we describe the plastic yielding as a gradual transition spreading between an onset value of deformation and a saturation value (eq. \[eq-evol-u-smooth\]). Section \[disc\] summarises and discusses our model, and opens some perspectives.
Model {#sec:qstat_model}
=====
Kinetics
--------
### Elastic and plastic strain
The elastic deformation $U$ is a [*state variable*]{}, that is an intrinsic property of the foam’s current deformation state. We note its time derivative $dU/dt$. Conversely, we use a dot for the total strain rate $\dot{\scalardeformation}$ and the plastic strain rate $\P$, emphasising that they are not the time derivative of a state variable. For instance, the time integral $\scalardeformation = \int \dot{\scalardeformation}\; {\rm d}t$ of the velocity gradient is the gradient of displacement (more generally, for large deformations, it is a function of the displacement): it is extrinsic and [*explicitly depends*]{} on the sample’s past history.
The total applied deformation rate is shared between elastic deformation $U$ and the plastic deformation rate: $$\dot{\scalardeformation} =\frac{{\rm d}U}{{\rm d}t} + \P.
\label{def_P}$$
In the particular case of an elastic regime, $\P=0$, the elastic deformation $U$ is equal to the total applied deformation on the material $\scalardeformation$. Thus, in an elastic regime, no intrinsic definition of $U$ is necessary.
However, as soon as $\P \neq 0$, the situation changes. $U$ and $\dot{\scalardeformation}$ become independent variables, and $\scalardeformation=\int \dot{\scalardeformation}dt$ does not define the elastic deformation. In the extreme example of a steady flow, ${\rm d}U/{\rm d}t=0$, then $\dot{\scalardeformation} =\P$: $U$ and $\dot{\scalardeformation}$ are no longer correlated.
These variables are macroscopic: $U$ is related to the elastic contribution to macroscopic stress and $\P$ to the irreversibility of the stress [*versus*]{} total strain curve. In the specific case of foams, they can be traced back to detailed properties of the bubbles pattern: independent, intrinsic definition [@por97] based on geometry (shape of bubbles [@asi03]) for $U$; and topological rearrangements called “T1 processes" [@lau02; @gop03; @kab03] (using their rate and orientation [@mar06]) for $\P$.
### Sharing the total strain
The problem now is to express how, in eq. (\[def\_P\]), $\dot{\scalardeformation}$ is shared between ${\rm d}U/{\rm d}t$ and $\P$. We must write a closure relation between these variables, for instance by expressing how $\P$ depends on the current state of elastic deformation and on the applied deformation rate: $\P(U,\dot{\scalardeformation})$. We use the following three hypotheses leading to eq. \[kinematicHeaviside\].
First we describe an abrupt transition from elastic to plastic regime, as could be the case for an ordered foam [@prin83]. To indicate that T1s appear when the absolute value of deformation $\vert U\vert $ exceeds the yield deformation $U_Y$, we introduce the discontinuous Heaviside function $\cal H$ (which is zero for negative numbers, and 1 for numbers greater than or equal to zero). This hypothesis can be relaxed in the section \[sec:Hdefinition\], introducing a more progressive transition.
Secondly, we account for the hysteresis. Plastic rearrangements occur when the deformation rate $\dot\scalardeformation$ and the current deformation $U$ have the same sign, and again we express it using $\cal H$. Else, the deformation rate results in elastic unloading, and the deformation gets smaller than the yield deformation.
Thirdly, we use the fact that, in a slowly sheared motion, the only relevant time scale to fix the rate of plastic rearrangements is $\dot\scalardeformation$.
Eventually, the plasticity equation writes: $$\P=
{\cal H}(\vert U\vert -U_Y)\;{\cal H}(U\dot\scalardeformation) \;
\dot{\scalardeformation} .
\label{kinematicHeaviside}$$ Eq. (\[kinematicHeaviside\]) can be used to close the system of equations. Injecting it in eq. (\[def\_P\]) yields an evolution equation of $U$ as a function of the applied shear rate $\dot{\scalardeformation}$: $$\frac{{\rm d}U}{{\rm d}t}
= \dot{\scalardeformation}\;
\left[1 -
{\cal H}(\vert U\vert -U_Y)\;{\cal H}(U\dot\scalardeformation) \right].
\label{eq-evol-u-scalar}$$
In eq. (\[eq-evol-u-scalar\]) $U_Y$ appears as the stable value for $U$, that is, a fixed point, at least if $\dot{\scalardeformation}>
0$; else, the stable fixed point is $-U_Y$.
![Analog scalar system: an elastic brush whose flexion is $U$ stick/slipping on wall. We represent several states, for an imposed oscillatory “painting-like” motion of the handle $\scalardeformation$, from rest position 0: (a) onset of sliding to the right, (b) far-right position, (c) far-left position. []{data-label="fig:brush.eps"}](brush.eps){width="6cm"}
To visualise the direction and the amplitude of the deformation $U$, we suggest an analogy with the motion of a brush on a wall (Fig. \[fig:brush.eps\]). The handle of the brush moves with an oscillatory position $\scalardeformation$ parallel to the wall (analog of the imposed scalar deformation of the material), while the displacement of the handle with respect to the brush tip is $U$ (the analog of the internal elasticity of the material). The sliding velocity of the contact point is therefore $\P$ according to equation (\[def\_P\]) and is the analog of plasticity in a material.
### Gradual transition to plasticity {#sec:Hdefinition}
In a disordered foam, for instance with a wide distribution of bubble sizes, topological rearrangements do not necessarily occur for the same value of deformation.
We therefore distinguish two different yield deformations. First, a [*plasticity yield*]{} $U_{y}$, where deformation ceases to be reversible, as defined in material sciences. It is the highest deformation for which there is no T1. It is characteristic of the microstructure, and can even be close to zero for a very disordered foam.
Second, a [*saturation yield*]{} $U_{Y}$, the saturation value of elastic deformation at which the material can flow with arbitrary large total deformations (for instance in Bingham fluids). It is the lowest deformation for which the T1s convert the whole total strain into plastic strain. That is, $U_{Y}$ is the collapse limit at which a material structure cannot sustain stress.
We interpolate between $U_{y}$ and $U_Y$ using a function $h(U)$ which we call a [*yield function*]{}. It should be a growing (or at least non-decreasing) function of $U$ for $U>0$, and $h(-U)=h(U)$. Moreover, $h(0)=0$, so that $h(U)\geq 0$ for all $U$. Beside that, there is no special requirement on $h$, which even does not need to be continuous. Now, $U_{y}$ is defined as the largest value of $U$ for which $h(U)=0$, and $U_Y$ as the smallest value of $U$ for which $h(U)= 1$. They do not necessarily correspond to any singularity in $h$. We show in Appendix that the precise shape of $h$ is unimportant: only $U_{y}$ and $U_Y$ determine material’s behaviour. However it is useful for theory to derive analytical equations.
The yield function $h$ depends on the material under consideration, and can in principle be measured experimentally. By definition, $h=0$ corresponds to a purely elastic state where the elastic deformation follows the applied deformation. Conversely, for $h=1$ the plasticity rate is equal to the deformation rate. Such a smooth transition from elasticity to plasticity generalises the postulate (\[kinematicHeaviside\]) as: $$\P= h(U)\,{\cal
H}(U\dot\scalardeformation)\,
\dot{\scalardeformation}.
\label{def_h}$$ Note that we could in principle smoothen out the remaining Heaviside function too: depending on microscopical details, it could be conceivable that some T1s appear during the unloading. We do not explore this possibility here, because we seldom observe this effect and it does not seem to improve significantly the predictions presented below.
Injecting eq. (\[def\_h\]) into eq. (\[def\_P\]) we obtain: $$\frac{{\rm d}U}{{\rm d}t}
= \dot{\scalardeformation}\;
\left[1 -
h(U)\,{\cal H}(U\dot\scalardeformation) \right].
\label{eq-evol-u-smooth}$$ Again, the fixed points are $U=\pm U_{Y}$ according to the sign of $\dot\scalardeformation$.
Dynamics
--------
### Slow shear: foam close to equilibrium
In a foam, bubbles can swap neighbours giving rise to T1 topological rearrangements. A T1 is an infinitely short event during which the energy is continuous. Thus it does not dissipate energy by itself, but it brings the foam in an out-of-equilibrium state. It is thus followed by a dissipation of energy during the relaxation towards another equilibrium configuration, with a microscopical relaxation time $\tau_{\rm relax}$.
The average life time of a contact between two bubbles is $f^{-1}$, where $f$ is the average frequency of T1s per bubble contact. If $f\tau_{\rm relax} \ll 1$, the foam evolves (it is not static) but spends most of the time at or very close to mechanical equilibrium states. Thus Plateau rules of local mechanical equilibrium [@wea99] are (almost) always satisfied, up to corrections of order $f\tau_{\rm relax} $.
The frequency $f$ can be determined by various causes of perturbations, for instance coarsening [@wea99]. In rheology, it is determined by the plastic deformation rate $\P$ [@mar06]. For dimensional reasons, $f$ is proportional to $\P$. Since the plasticity amplitude $\P$ is always smaller than the deformation rate $\dot{\scalardeformation}$ (see eqs \[kinematicHeaviside\] and \[def\_h\]), the regime close to equilibrium is obtained in the slow shear limit: $$\dot{\scalardeformation}\tau_{\rm relax} \ll 1.
\label{slowshear}$$
### Contributions to total stress
We now include an additional viscous dissipation from the global deformation of the network of bubbles. This contribution is not linked to the relaxation of rearrangements, and does not modify the slow evolution of deformation.
We consider two separate contributions to stress, under the following hypotheses. According to experimental tests [@asi03], we consider an elastic contribution to the stress ${\sigma}^\mathrm{el}=\mu {U}$ proportional to the elastic deformation $U$, where $\mu$ is the shear elastic modulus. It describes a classical elastic behaviour, with a reversible restoring force.
According to the model proposed by Kraynik and co-workers [@kra87; @kra88], we consider a viscous contribution to the stress due to large scale velocity gradients: $\sigma^\mathrm{vis}=\eta\dot{\scalardeformation}$, where $\eta$ is a macroscopic viscosity. It describes a classical fluid behaviour: the corresponding dissipated power is quadratic, proportional to $\dot{\scalardeformation}^2$.
(0,0)![A linear elasto-visco-plastic rheological model.[]{data-label="fig-rheology"}](fig-rheologyPhilippe6.eps "fig:")
(5804,4896)(3039,-4549) (5941,164)[$\varepsilon$]{}(4681,-4471)[$\varepsilon_P$]{}(5500,-3400)[$\sigma_Y$]{}(6841,-4471)[$U$]{}(5750,-556)[$\eta$]{}(7000,-2000)[$\mu$]{}
In the spirit of a polymeric model [@mac94], we assume that the stresses add up (Fig. \[fig-rheology\]): $$\begin{aligned}
{\sigma}&=&\sigma^\mathrm{el}+\sigma^\mathrm{vis}\nonumber\\
&=& \mu {U}+\eta \dot{\scalardeformation}.
\label{eq:stresselasticviscous}\end{aligned}$$
The material is characterised by the coefficients $\eta$, $\mu$ and $U_Y$ (and $U_{y}$ in the case of gradual plasticity). Measuring experimentally, and understanding theoretically the physical origin of these coefficients, requires specific studies for each material considered: this is beyond the scope of the present paper. In principle, they can be rank-four tensors (anisotropic material). They can even vary with the material’s state (non-linear material), for instance in a shear-thinning case.
![ Scalar phase diagram for a slowly sheared foam or an emulsion. Axes are experimentally measurable [@mar06] local variables: shear rate $\dot{\scalardeformation}$ and elastic deformation $U$. The crossover from elastic to plastic is defined as the onset of the first isolated topological rearrangements; it occurs around $U_Y$, with possible precursors around $U_{y}$. The yield deformation $U_Y$ corresponds to a macroscopic rate of topological rearrangements. The crossover from solid to fluid is defined by the equality of viscous and elastic stresses. The slowly sheared regime presented here ceases to be valid when $\dot{\scalardeformation}$ becomes comparable to $\tau_{\rm
relax}^{-1}$, inverse of the microscopical relaxation time. []{data-label="diagramme_vep"}](VEP_qstat_blank.eps){width="8cm"}
As opposed to the cross-over from elastic to plastic regimes, which is topological and is visible on images, here the crossover from the elastic to the fluid regime can be detected only by measuring forces. It occurs when the viscous contribution to the stress becomes larger than the elastic one. Fig. (\[diagramme\_vep\]) thus plots the line corresponding to the crossover: $
\mu U = \eta \dot{\scalardeformation}.
$
Defining the macroscopic local Weissenberg number as: $$\mathrm{Wi}_{M} \equiv \frac{\eta\dot{\scalardeformation}}{\mu},
\label{def_CaM}$$ the cross-over between elastic and fluid regime occurs at: $
\mathrm{Wi}_{M} = U.
$
Upon increasing $\dot{\scalardeformation}$ from the plastic regime where $U$ is close to $U_{Y}$, we predict a cross-over from a stress bounded by a constant value $\mu U_{Y}$ (with a dissipated power linear in $\dot{\scalardeformation}$, see next section) to a stress linear in $\eta\dot{\scalardeformation}$ (with a dissipated power quadratic in $\dot{\scalardeformation}$), characteristic of a viscous friction. The cross-over from a plastic regime to a fluid regime occurs when $\eta \dot{\scalardeformation}$ is equal to $\mu U_{Y}$, i.e. when $\mathrm{Wi}_{M} = U_{Y}$.
Since the plastic deformation is not bound, in the plastic regime the foam can flow indefinitely. As in hydrodynamics, the displacement field itself is no longer relevant. The plastic flow [@fra93; @fra95] and the viscous flow [@bat00] look the same; their difference is not kinematic but dynamic: stresses are independent on and proportional to $\dot{\scalardeformation}$, respectively.
### Dissipation
The close to equilibrium criterion (eq. \[slowshear\]) regards time scales, and is not a criterion based on the absence or presence of dissipation. Viscous dissipative effects can indeed be observed when considering measurements of the loss modulus at very low amplitude oscillations, and hence at very slow shear rate (as presented below in Figs. \[fig:GprimeGsecondfigure\_general.eps\]-\[fig:GprimeGsecondRouyer\]). In fact, dissipation is absolutely necessary to relax towards equilibrium: it damps oscillations and decreases the energy. Note that a “quasi-static" regime, that is a succession of equilibrium states, necessarily obeys the equilibrium criterion; but it is not sure that the reverse is true. In fact, Ref. [@jan06b] claims that in the slowly sheared Couette flow by [@deb01] the velocity profile is determined by the ratio of velocity-dependent forces (internal viscosity and external friction on the plates of glass): static simulations are inappropriate.
As already mentioned, a T1 by itself, that is a side swapping, is an infinitesimally short topological event, during which the foam energy is continuous: there is no instantaneous dissipation. However, the T1 puts the foam in an out-of-equilibrium state. During a time $\tau_{\rm relax}$ the foam relaxes to a local energy minimum by dissipating an energy $\delta E$. A smaller microscopic dissipation yields a shorter relaxation time $\tau_{\rm relax}$, and a larger [*instantaneous*]{} dissipated power, of order ${\cal P}_{\rm diss} = \delta E/ \tau_{\rm relax}$. But the amount of energy dissipated, $\delta E$, is independent on the dissipation. Thus the dissipated power [*averaged*]{} over a long time (longer than $\dot{\scalardeformation}^{-1}$) is of order: $$\left \langle {\cal P}_{\rm diss}\right \rangle
=
f \delta E \sim \P \delta E \sim \dot{\scalardeformation} \delta E.
\label{power}$$ This dissipated power is proportional to $\dot{\scalardeformation}$, and not quadratic as in viscous flows, although the microscopical origin is a local viscous dissipation [@pug06].
![Schematic impact of individual microscopic rearrangements on the stored elastic deformation $U$, for a constant loading rate $\dot{\scalardeformation}$. Rearrangements relax exponentially the deformation over a time $\tau_{\rm relax}$, with here $\dot{\scalardeformation}\tau_{\rm relax}=0.02 \ll 1$. In the present macroscopic model, rearrangements are coarse-grained. []{data-label="sawtooth"}](sawtooth3.eps)
The elastic deformation is almost independent on the shear rate $
\dot{\scalardeformation}$ (see figure \[sawtooth\]). To obtain a steady shear in a solid regime, when $U$ saturates at the value $U_Y$, an experimentalist has to apply an constant external force which balances the average elastic stress, and does not depend on $\dot{\scalardeformation}$.
Such a dissipated power linear in $\dot{\scalardeformation}$, and a steady force which does not depend on $\dot{\scalardeformation}$, are characteristic of a solid friction [@cou79; @bau99].
Prediction and tests {#predict}
====================
We model the foam response in one type of mechanical experiment, imposed deformation, and in two types of rheometrical experiments, creep flow and oscillating shear.
Imposed shear
-------------
Here, we calculate the transient response during a shearing experiment, that is, the relation $U(\scalardeformation)$ between applied strain $\scalardeformation$ and elastic deformation $U$. For simplicity we take here $\scalardeformation=U=0$ at the start of the experiment, but that assumption is easy to relax.
By direct integration, see appendix, we show the material’s response: the elastic deformation $U$ is close to the imposed strain $\scalardeformation$ at low applied strain, and tends to a saturation value at large applied strain. This robust behaviour does not depend much on the chosen yield function (see fig. \[fig:Fromrest.eps\]).
![Response to imposed shear for two examples of yield functions. Dashed solid line: abrupt transition, $h(U)={\cal H}( U - U_Y)$ (eq. \[kinematicHeaviside\]). Thick line: finite $U_{y}$, and linear interpolation $h(U)=({ U -U_{y}})/({U_Y
-U_{y}}){\cal H}( U - U_{y})$ (eq. \[h\_interpolate\]). Here $U_{y}=0.75\; U_Y$. See fig. \[fig:HandFromrest\] for more examples. []{data-label="fig:Fromrest.eps"}](Fromrest2.eps)
Thus the distribution of bubble sizes does not affect much the foam’s transient response (as opposed to the liquid fraction, which drastically affects $U_Y$ [@prin83]). This explains why in the literature the function $U(\varepsilon)$ is sometimes taken for simplicity as a piece-wise linear function or as a hyperbolic tangent [@jan06].
This provides both the physical origin for the function $\sigma(\varepsilon)$ of the model by Janiaud [*et al.*]{} [@jan06], and a justification for their (up to now arbitrary) expression $\sigma = \sigma_Y \;
f(\varepsilon/\varepsilon_Y)$: the function $f$ corresponds to the present elastic deformation $U$, while $\varepsilon_Y$ is the yield deformation they chose equal to 1 for simplification.
Creep under constant applied stress
-----------------------------------
[![Response to constant applied stress: (a) elastic deformation versus time and (b) strain rate versus time, with a smooth appearance of plasticity in between a plasticity onset threshold $U_{y}$ and saturation $U_{Y}$. Plasticity starts at $U_{y}=0.75\, U_{Y}$, with the linear interpolation of eq. \[h\_interpolate\]. Same legend as previous figure.[]{data-label="fig:CreepSmooth"}](CreepFunctionhUmFigureU.eps "fig:")]{} [(a)]{} [![Response to constant applied stress: (a) elastic deformation versus time and (b) strain rate versus time, with a smooth appearance of plasticity in between a plasticity onset threshold $U_{y}$ and saturation $U_{Y}$. Plasticity starts at $U_{y}=0.75\, U_{Y}$, with the linear interpolation of eq. \[h\_interpolate\]. Same legend as previous figure.[]{data-label="fig:CreepSmooth"}](CreepFunctionhUmFigureStrainRate.eps "fig:")]{} [(b)]{}
A creep experiment in a rheometer applies a constant stress $\sigma_\mathrm{app}$. It determines the effective viscosity $\eta_\mathrm{eff}$ from the steady shear rate: $$\eta_\mathrm{eff}=\lim_{t\rightarrow \infty} \;
\frac{\sigma_\mathrm{app}}{\dot\varepsilon(t)}.$$
The rheological response is found from eq. (\[eq:stresselasticviscous\]) with $\sigma=\sigma_\mathrm{app}$, and from eq. (\[eq-evol-u-scalar\]). The elastic loading and the strain rate are plotted on Fig. (\[fig:Creep\]). The elastic deformation saturates to $\sigma_\mathrm{app}/\mu$ when it is below the threshold $U_{Y}$, and that it saturates to $U_{Y}$ when above the threshold, over a characteristic time $\lambda=\eta/\mu$.
At long times, the strain rate tends towards vanishing values below yield stress (the flow stops), and tends to finite values above the yield: $\dot\varepsilon(t\rightarrow
\infty)=(\sigma_\mathrm{app}-\mu U_{Y})/\eta$. We thus deduce the effective viscosity: $$\begin{aligned}
\eta_\mathrm{eff}&=& \infty \quad \mathrm{when} \quad
\sigma_\mathrm{app}\leq \mu U_{Y},\\
\eta_\mathrm{eff}&=&\frac{\eta}{1-\frac{\mu
U_{Y}}{\sigma_\mathrm{app}}}\quad \mathrm{when} \quad \sigma_\mathrm{app}>\mu
U_{Y}.\end{aligned}$$
Taking a smooth plastic transition (eq. \[eq-evol-u-smooth\]) does not changes the overall features, except that the deceleration times below yield are no longer superimposed, see fig. (\[fig:CreepSmooth\]).
Oscillating shear
-----------------
![Long time periodic response to an oscillatory shear, for three amplitudes: $\varepsilon/U_Y=0.75,
1.1$ and $1.8$. (a) in the case where the plasticity appears abruptly at $U_{Y}$, (b) gradual transition in between $U_{y}$ and $U_{Y}$, with $U_{y}/U_{Y}=0.5$ (eq. \[h\_interpolate\]).[]{data-label="fig:GprimeGsecond_Heaviside.eps"}](GprimeGsecond_Heaviside.eps "fig:")(a)\
![Long time periodic response to an oscillatory shear, for three amplitudes: $\varepsilon/U_Y=0.75,
1.1$ and $1.8$. (a) in the case where the plasticity appears abruptly at $U_{Y}$, (b) gradual transition in between $U_{y}$ and $U_{Y}$, with $U_{y}/U_{Y}=0.5$ (eq. \[h\_interpolate\]).[]{data-label="fig:GprimeGsecond_Heaviside.eps"}](GprimeGsecond_Um.eps "fig:")(b)\
\[fig:GprimeGsecond\_Square.eps\]
A Couette apparatus is another typical rheometry experiment [@mas96; @lau02; @rou05; @hoh05]. It measures the stress $\sigma(t)$ on the walls while imposing an applied sinusoidal shear strain of pulsation $\omega=2\pi/T$: $$\scalardeformation=\gamma \sin(\omega t).$$
### Hysteresis cycle and non-linear response {#hyster}
To test the effect of hysteresis of the model we calculate the response to such a periodic oscillatory shear bounded by amplitudes $\gamma$ and $-\gamma$. In the slow shear limit, the frequency does not play any explicit role. We can thus keep it fixed without loss of generality.
We integrate eq. (\[eq-evol-u-scalar\]). The periodic elastic deformation *vs* strain curve is plotted on Fig. (\[fig:GprimeGsecond\_Heaviside.eps\]) top. The stress response is linear in strain below the threshold, and saturates above in plastic regime, exhibiting a strong hysteresis.
Reversing the sign of the loading instantly stops any plasticity and the reponse becomes purely elastic. Multiple loading does not increase the slope of the loading part, nor the value of saturation yield; the foam is described as perfectly plastic. Such features are observed in experiments on other amorphous solids [@aub99; @cou03] (as opposed to strain-hardening features of crystalline metals [@cha87]).
Integration of eq. (\[def\_h\]) describes a smooth variation of deformation, see Fig. (\[fig:GprimeGsecond\_Square.eps\]) bottom.
### Storage and loss moduli: predictions
In complex notation the stress response $\sigma^*$ is linked to the strain $\scalardeformation^*$ by $\sigma^*=(G'+iG'')\scalardeformation^*$. Here $G'$ is the storage modulus and $G''$ the loss modulus of the material, defined as the in-phase and out phase part of the response (first term in a Fourier series, see non linear models [@hyu02; @sim03; @miy06]).
When increasing the amplitude $\gamma$ of the imposed sinusoidal shear strain, the response is first linear until the amplitude at which $G'$ and $G''$ start to vary. In both the linear and non-linear regimes, the storage and loss moduli are calculated as: $$\begin{aligned}
G'&=&-\frac{1}{\gamma^2}\frac{1}{\pi \omega}
\int_0^T \sigma(t)
d\dot{\scalardeformation},
\nonumber\\
G''&=&\quad\frac{1}{\gamma^2}\frac{1}{\pi}
\int_0^T
\sigma(t)d\scalardeformation,\label{eq:G''integral}\end{aligned}$$ $G'$ is proportional to the area enclosed by the $(\sigma(t),\dot{\scalardeformation}(t))$ curve, while $G''$ is proportional to the area enclosed by the $(\sigma(t), \scalardeformation(t))$ curve. When plasticity occurs, the cycle has a non-vanishing area in the $(\sigma(t),
\scalardeformation(t))$ diagram, meaning a non-vanishing loss modulus $G''$.
In the present model $\sigma(t)$ depends on the current elastic deformation $U(t)$ and shear rate $\dot{\scalardeformation}(t)$ (eq. \[eq:stresselasticviscous\]). For the case of the abrupt elastic/plastic transition, the analytical integration of areas is simple and provides the following solutions for the moduli. Using eqs. (\[eq:G”integral\]) we obtain, when $\gamma \ll U_Y $: $$\begin{aligned}
G'&\simeq &\mu
\nonumber\\
G''&=&\eta,
\label{eq:G-low-amplitude}\end{aligned}$$ [which is the usual linear visco-elastic regime. Note that our model predicts frequency independant moduli, for a fixed small amplitude $\gamma$. ]{} At large amplitudes, when $\gamma \gg U_Y $: $$\begin{aligned}
G'&\simeq &\mu
\frac{4}{\pi}\left(\frac{U_Y}{\gamma}\right)^{3/2},
\nonumber\\
G''&=&\mu\frac{4U_Y}{\pi\gamma}+\eta. \label{eq:Gsecondhigh}\end{aligned}$$
These asymptotic dependencies in $\gamma^{-3/2}$ and $\gamma^{-1}$, respectively, are obtained analytically and are robust with respect to the model. The analytical expression of $G'$ and $G''$ over the whole range of $\gamma$, but with $\eta=0$ is explicitly presented in refs. [@lab04; @hoh06].
For a smooth yield function $h$, predictions are obtained numerically and plotted on Figs. (\[fig:GprimeGsecondfigure\_general.eps\]-\[fig:GprimeGsecondRouyer\]).
![ Storage and loss moduli [*versus*]{} strain amplitude for a monodisperse emulsion. Symbols: experimental $G'$ (circles) and $G''$ (triangles) in a close-packed emulsion (Fig. 1 of ref. [@mas95], fraction of the continuous phase 20%, droplet size 0.53 $\mu$m, oscillation pulsation $\omega=1$ rad $s^{-1}$). Lines: models for $G'$ (solid line), and for $G''$ with an abrupt transition (dashed line), with viscosity (dash-dotted line), with viscosity and a smooth yield function $h=(U/U_Y)^2$, $U_{y}=0$ (solid line). Model parameters: shear modulus $\mu=1.7\; 10^3$ Pa, yield deformation $U_Y=0.045$, viscosity $\eta=30$ Pa.s. []{data-label="fig:GprimeGsecondfigure_general.eps"}](GprimeGsecond_Mason2.eps){width="8.5cm"}
![ Same as Fig. (\[fig:GprimeGsecondfigure\_general.eps\]) for a polydisperse foam [@sai99]. Liquid fraction 5%, bubble size 40 to 70 $\mu$m, $\omega=1$ rad $s^{-1}$. Lines: models for $G'$ (solid line), and for $G''$ with an abrupt transition (dashed line), with viscosity and a smooth yield function $h=(U/U_Y)^2$, $U_{y}=0$ (solid line). Model parameters: $\mu=100$ Pa, $U_Y=0.2$, $\eta=15$ Pa.s. []{data-label="fig:GprimeGsecondJalmes"}](GprimeGsecondJalmes95_SVM_figure.eps){width="8.5cm"}
(8,8)(0,0) (-0.5,0)[![ Same as Fig. (\[fig:GprimeGsecondfigure\_general.eps\]) for a monodisperse foam [@rou05]. Liquid fraction 8%, bubble size 21 $\mu$m, $\omega=1$ rad $s^{-1}$. Data of $G'$ and $G''$ are normalised by $\mu$ and $\gamma$ by $U_Y$, while $\eta \omega / \mu=0.08$. Same legend as figure \[fig:GprimeGsecondJalmes\]. [Inset: Strain-rate frequency superposition: same data plotted as a function of pulsation $\omega$, for a given maximum strain rate $\dot\gamma_{0}=\omega \gamma$ [@wys06].]{} []{data-label="fig:GprimeGsecondRouyer"}](GprimeGsecond_Rouyer.eps "fig:"){width="9cm"}]{} (4.7,4.5)[![ Same as Fig. (\[fig:GprimeGsecondfigure\_general.eps\]) for a monodisperse foam [@rou05]. Liquid fraction 8%, bubble size 21 $\mu$m, $\omega=1$ rad $s^{-1}$. Data of $G'$ and $G''$ are normalised by $\mu$ and $\gamma$ by $U_Y$, while $\eta \omega / \mu=0.08$. Same legend as figure \[fig:GprimeGsecondJalmes\]. [Inset: Strain-rate frequency superposition: same data plotted as a function of pulsation $\omega$, for a given maximum strain rate $\dot\gamma_{0}=\omega \gamma$ [@wys06].]{} []{data-label="fig:GprimeGsecondRouyer"}](GprimeGsecond_Wyss.eps "fig:"){width="3.5cm"}]{} (6,4.5)[$\omega/(\dot\gamma_{0}/U_{Y})$]{}
### Comparison with experiments on emulsions and foams
Rheometry measurements of monodisperse emulsions [@mas95] (Fig. \[fig:GprimeGsecondfigure\_general.eps\]) and polydisperse foams [@sai99; @rou05] (Figs. \[fig:GprimeGsecondJalmes\] and \[fig:GprimeGsecondRouyer\]) directly yield, without hypotheses, the values of the material parameters required by the model. The shear modulus $\mu$ is read from the value of $G'$ at low amplitude. The viscosity $\eta$ is read from the value of $G''$ at low amplitude (or the value of the minimum, in Fig. \[fig:GprimeGsecondfigure\_general.eps\], where the two data points at $\gamma<10^{-3}$ have too large error bars to be taken into account, according to T. Mason, private communication). The yield deformation $U_Y$ is read from the intersection of low amplitude plateau of $G'$ and its large amplitude $-3/2$ exponent power-law.
A purely elasto-plastic model is enough to predict $G'$ correctly, over the whole range of amplitude, including the $-3/2$ exponent power-law. This simplest model also describes correctly the large amplitude trend for $G''$.
The low amplitude value of $G''$ can be modelled by including a viscosity (dash-dotted line of figure \[fig:GprimeGsecondfigure\_general.eps\]), which confirms that viscosity is relevant even in such slowly sheared models. This procedure is at the expense of a slight over-prediction at large amplitudes. This latter aspect suggests a possible shear-thinning, that is a decrease of the viscosity $\eta$ with the shear rate, similarly to the observed reduction of the drag of foams in motion in channels [@prin83].
The agreement between the data and the model, without adjustable parameter, is good. It is still improved, even for $G''$ at intermediate amplitudes, if we account for the fact that the first T1s appear gradually at a value $U_{y}$ lower than $U_Y$ (solid lines on figures for $G''$). Even the value of $U_{y}$ itself is not very important, and in order to avoid introducing a free parameter we use here $U_{y}=0$ and a smooth yield function.
[These result suggest that data can be rescaled with the yield deformation $U_{Y}$, and it suggest a rescaling when plotting data as a function of frequency, following the strain-rate frequency superposition method (SRFS [@wys06]). This method consider measurements with a fixed maximum strain rate $\dot\gamma_{0}=\omega \gamma$, it is therefore equivalent to vary frequency or oscillation amplitude. The natural rescaling for pulsation that appears is $\omega/b(\dot\gamma_{0})=\omega/(\dot\gamma_{0}/U_{Y})=U_{Y}/\gamma$, see inset of Fig (\[fig:GprimeGsecondRouyer\]), and our model predicts the global shapes of the moduli curves as observed in [@wys06]. The characteristic frequency $b(\dot\gamma_{0})$ is here linear in $\dot\gamma_{0}$, also in agreement with the trend observed in [@wys06] for large enough strain rates. ]{}
Discussion {#disc}
==========
Discussion of the predictions
-----------------------------
Independently from us, Höhler et al. solve a purely elasto-plastic model [@hoh06; @lab04]. Since they neglect the viscosity, they can eliminate $U$ and replace it by $\sigma/\mu$. The agreement of their model with data of Fig. \[fig:GprimeGsecondRouyer\], as well as other experimental data, is good for $G'$ over the whole range of amplitude; and also for $G''$ but only at large amplitude, where the dissipation comes from the relaxation after T1s rather than from $\eta$.
Our models describes better: (i) $G"$ at low amplitude using viscosity; and (ii) $G''$ at intermediate amplitudes, if we account for the fact that the first T1s appear at a value $U_{y}$ lower than $U_Y$.
The predicted curves are robust with respect to $U_{y}$. This implies that we do not need to fit it; but that, conversely, we are not yet able to deduce $U_{y}$ from $G''$ data. If we had a direct experimental measurement of $U_{y}$, we could inject it in the model to predict $G''$ at intermediate amplitudes, near $U_Y$, but the resulting predictions would be very similar to the present ones.
While $U_{y}$ corresponds to the onset of isolated plastic events, at a deformation $U_Y$ the plastic events have a macroscopic effect: they catch the total strain; the foam flows without increasing its deformation any longer.
The effective viscosity $\eta_\mathrm{eff}$ diverges to infinite values when $\sigma_\mathrm{app}/\mu \rightarrow U_{Y}^+$. This means that the foam comes close to its yield deformation, in the fluid sense. This change in behaviour was shown by [@cru02], and modelled by a granular model with a velocity dependant friction. Here, the dynamics is entirely driven by a constitutive equation with shear-rate independent parameters. It is the transient elastic loading that drives a transient flow, which stops when the stress is not strong enough.
The trends of figure (\[fig:Creep\]) agree with experimental data on various material reported in [@cru02], at least with granular materials and emulsion. The present model does not predict the apparent shear-thinning behaviour observed with their experimental data with foams [@cru02], where an increase of shear rate with time is found [@cou02]. [The low amplitude predictions for the visco-elastic regime (see Eq. \[eq:G-low-amplitude\]) are frequency independant moduli. The limitation of the present model is thus that it does not fully capture the slight increase of the loss moduli at low frequency, and the $\omega^{1/2}$ trend at large frequencies (see for instance the case of foams [@hoh05]). ]{}
Elastic, plastic, viscous model
-------------------------------
A complete model for an elastic, plastic, viscous foam requires to recognise the role of three physical variables $U$, $\dot{\scalardeformation}$, $\P$. There is a relation between them (eq. \[def\_P\]). Unless specific approximations apply, a foam’s representative volume element (RVE) is characterised by two independent variables: we suggest to select the local elastic deformation $U$, and the local shear rate $\dot{\scalardeformation}$, which are intuitive and physically relevant. Both of course depends on the sample’s past history, but this history plays no explicit role. Both are always defined, whether in elastic, plastic or viscous regime [@aub03]. Two recent works [@wys06; @mar06z] find that $G'$ and $G''$ depend on the strain amplitude and on the strain rate (rather than on the frequency), in the same spirit as our phase diagram (fig. \[diagramme\_vep\]).
Each volume element can thus be plotted as a point in a phase diagram (Fig. \[diagramme\_vep\]); that is, the $(\dot{\scalardeformation},U)$ plane [@pic05]. In a heterogeneous flow, different volume elements of the same foam are plotted as different points. A volume element’s evolution is a trajectory on this plane. [*Simple*]{} materials correspond to the axes of the plane: pure elastic and pure plastic regimes on the vertical axis, pure viscous regime (Navier-Stokes) on the horizontal axis.
The importance of $U$ is the most original feature of the present model: $U$ cannot be entirely determined by $\dot{\scalardeformation}$ since the latter can change sign; $U$ cannot be entirely determined by $\sigma$ if the viscous contribution is not negligible.
The yield function describing the occurrence of plasticity can be linked to the traditional hardening modulus, used for the description of plastic materials [@sim98]. It is defined as $K=d\sigma/d\varepsilon_{p}$, while the elastic modulus is $\mu=d\sigma/U$. In the present model, the hardening modulus is dependant on the elastic deformation: $K=1/h(U)-1$. It therefore vanishes when $U$ tends to its saturation value $U_{Y}$: at this point the material does not harden any more.
A deformation beyond $U_Y$ is not accessible when starting from rest (Fig. \[diagramme\_vep\]). But the foam could initially be prepared (for instance artificially [@eli99]) in a configuration very far from equilibrium. Under a steady shear rate $\dot{\scalardeformation}$, the deformation $U$ always tends towards $U_Y (\dot{\scalardeformation})$, whether from below or from above.
In a flowing foam, there is always a viscous dissipation. Its contribution becomes dominant in front of the plastic dissipation if $\dot{\scalardeformation} > \mu U / \eta$. This is compatible with the slow shear criterion, $\dot{\scalardeformation} \ll \tau_{\rm relax} ^{-1}$ (ie on a second Weissenberg number that is here $\mathrm{Wi}_{m}=\dot{\scalardeformation} \tau_{\rm relax}\ll 1$), if there is a scale separation between the microscopic relaxation time $\tau_{\rm relax}$ towards local equilibrium and the large scale time $\eta/\mu$. The dimensionless ratio $ \mu\tau_{\rm relax} / \eta$ of microscopic to macroscopic times is analogous to the parameter $I$ of granular materials [@jop06].
Perspectives {#sec:Perspectives}
------------
This model could in principle be generalised to higher velocity gradients [@sar06]. This would require a high flow velocity varying over a small scale, and $\tau_{\rm relax} $ could play an explicit role. The deviation from equilibrium, of order $\dot{\scalardeformation} \tau_{\rm relax} $, would become significant: for instance, under a steady shear the limit value of $U$ could become larger than $U_Y$.
A rheometer such as a Couette apparatus can measure $\sigma_{12}=\sigma$ (tangential force per unit wall surface) and $\varepsilon_{12}=\varepsilon/2$ (components of the symmetrized deformation gradient), in a coordinate system aligned with walls. [For comparison with tensorial data it is especially important to bear in mind that there is a [*factor $1/2$*]{}: the threshold $U_{Y}$ on oscillation amplitude $\varepsilon$ as measured by a Couette rheometer, corresponds to a threshold $U_{Y}/2$ on the tensorial deformation $\varepsilon_{12}$. ]{} The present scalar approach can be generalised to take into account such an influence of the orientation of material deformation, as well as spatial variations [@mar06; @rau06testexp].
The present paper is a contribution to a lively debate. Can a foam be described as a continuous material? We tend here to answer “yes", in the same spirit as many recent papers which describe or predict rheological properties at large scale [@asi03; @dollet_local; @mar06; @jan06; @lab04; @hoh06; @sar06]. [Statistical descriptions of fluctuations and their correlations [@sol97; @fal98; @pic05; @cou05; @cru02] are then useful in describing even more complex rheological behaviour such as shear banding [@kab03] or growing correlation lengths cale near the glass transition [@Berthier2005]. ]{} Interestingly, even in granular materials, where very large scale fluctuations are known to occur, a recent paper emphasises the dominant role of the “continuous material" description based on averages [@jop06].
Acknowledgments {#acknowledgments .unnumbered}
===============
We thank E. Janiaud for critical reading of the manuscript, R. Höhler for comparison of our calculations before publication, F. Rouyer for providing experimental data, S. Ataei Talebi, I. Cheddadi, B. Dollet, C. Quilliet, C. Raufaste, and P. Saramito for discussions, T. Mason and A. Saint-Jalmes for comments on their experimental data.
Appendix: Elastic-plastic transition {#sec:appendix .unnumbered}
====================================
Transient response from rest
----------------------------
We assume (in this section only) that [*the deformation rate $\dot{\scalardeformation}$ keeps a constant sign*]{}. Under this essential assumption, we can calculate analytically the transient response during a shearing experiment. That is, the relation $U(\scalardeformation)$ between applied strain $\scalardeformation =
\int \dot{\scalardeformation}\; {\rm d}t$ and elastic deformation $U$.
The yield function $h$ is defined to interpolate between $h(U)= 0$ for $0<U<U_{y}$, and $h(U_Y)=1$. By direct integration, eq. (\[eq-evol-u-smooth\]) yields: $$\scalardeformation = \int_0^U \; \frac{{\mathrm d}U}{1-h(U)}.
\label{gamma_(U)}$$ Here, without further loss of generality, we have also assumed (but it is easy to relax) that $\scalardeformation=U=0$ at the start of the experiment, and that $\dot{\scalardeformation}\geq 0$, so that $U \geq0$ too. Eq. (\[gamma\_(U)\]) yields the function $\scalardeformation(U)$, which can be inverted to obtain $U(\scalardeformation)$. These functions can be measured on experiments and compared with predictions derived from direct measurements of $h(U)$.
Whatever the function $h(U)$, eq. (\[gamma\_(U)\]) implies that $U \approx \scalardeformation$ as long as $U< U_{y}$: applied and elastic deformation are equal in the elastic regime. At the onset of plasticity (or topological changes), $U>U_{y}$, they differ. When $U$ gets close to $U_Y$ the r.h.s. of eq. (\[gamma\_(U)\]) diverges. Thus, when $\scalardeformation$ increases arbitrarily, $U$ tends asymptotically towards the saturation value $U_Y$.
Examples of yield functions $h$
-------------------------------
-------------------------------------------------------------------------------------------
Yield function $h $ Elastic response $U(\scalardeformation)/U_Y$
----------------------------------------- -------------------------------------------------
$ {\cal H}( U - U_Y)$ $\scalardeformation- {\cal
H}(\scalardeformation/U_Y-1)\scalardeformation$
finite $U_{y}$ (eq. \[h\_interpolate\]) Eq. (\[finite\_Um\])
$( U /U_Y)^0$ $0$
$( U /U_Y)^1$ $1-\exp (-\scalardeformation/U_Y)$
$( U /U_Y)^2$ $\tanh (\scalardeformation/U_Y)$
$\sin^2( U /U_Y)$ $\arctan (\scalardeformation/U_Y)$
-------------------------------------------------------------------------------------------
: Elastic deformation for different examples of yield function $h$, for a non-deformed initial condition $U(0)=0$ and with $\dot\scalardeformation$ of constant sign.[]{data-label="tab:hfunctions"}
Table \[tab:hfunctions\] proposes a few examples of yield functions $h$, and some are plotted on figure \[fig:HandFromrest\].
Eq. (\[eq-evol-u-scalar\]) is only a particular case of the more general eq. (\[eq-evol-u-smooth\]), with $h$ being the discontinuous Heaviside function: $$h(U)={\cal H}( U - U_Y).
\label{eq:hHeaviside}$$ Eq. (\[gamma\_(U)\]) thus includes the case of the abrupt transition.
![ Responses from rest for some examples of yield functions in Table I. (a) $h(U)$; (b) $U/U_Y$ versus $\scalardeformation/U_Y$: since they are very similar, for clarity only some of them are plotted. Thick solid line: abrupt transition, $h(U)={\cal H}( U - U_Y)$. Thin line: finite $U_{y}$, here $U_{y}=0.75\; U_Y$, and linear interpolation $h(U)=({ U -U_{y}})/({U_Y
-U_{y}}){\cal H}( U - U_{y})$ (eq. \[h\_interpolate\]). Dashes: vanishing $U_{y}$ and quadratic interpolation, $h=(U/U_Y)^2$. Dots: vanishing $U_{y}$ and linear interpolation, $h=(U/U_Y)$. []{data-label="fig:HandFromrest"}](H.eps "fig:") (a)\
![ Responses from rest for some examples of yield functions in Table I. (a) $h(U)$; (b) $U/U_Y$ versus $\scalardeformation/U_Y$: since they are very similar, for clarity only some of them are plotted. Thick solid line: abrupt transition, $h(U)={\cal H}( U - U_Y)$. Thin line: finite $U_{y}$, here $U_{y}=0.75\; U_Y$, and linear interpolation $h(U)=({ U -U_{y}})/({U_Y
-U_{y}}){\cal H}( U - U_{y})$ (eq. \[h\_interpolate\]). Dashes: vanishing $U_{y}$ and quadratic interpolation, $h=(U/U_Y)^2$. Dots: vanishing $U_{y}$ and linear interpolation, $h=(U/U_Y)$. []{data-label="fig:HandFromrest"}](Fromrest.eps "fig:") (b)
An example of a yield function with finite $U_{y}$ is a piece-wise linear function: $$\begin{aligned}
U \leq U_{y} \quad & \quad h(U) = 0, \nonumber \\
U \geq U_{y} \quad & \quad h(U) = \frac{ U -U_{y}}{U_Y -U_{y}}.
\label{h_interpolate}\end{aligned}$$ and eq. (\[gamma\_(U)\]) yields directly: $$\begin{aligned}
\label{finite_Um}
\scalardeformation \leq U_{y} \quad \quad
&U(\scalardeformation) & = \scalardeformation
, \nonumber \\
\scalardeformation\geq U_{y} \quad \quad
& U(\scalardeformation) & = \frac{\scriptstyle U_Y}{\scriptstyle U_{Y}-U_{y}}
\nonumber \\
& & - \left(\frac{ \scriptstyle U_Y}{\scriptstyle U_{Y}-U_{y}}-\scriptstyle U_{y}\right)\; e^{-\left(
\frac{ \scalardeformation - U_{y} }{U_Y -U_{y}}
\right)}
.\end{aligned}$$
We can interpolate between abrupt and smooth transitions, using the family of model power-law yield functions: $$h(U)=\left(\frac{ U }{U_Y}\right)^n.
\label{eq:hpowerlaw}$$ For instance, the quadratic expression $h(U)=(U/U_Y)^2$ yields $U(\scalardeformation)=
U_Y\tanh(\scalardeformation/U_Y)$. With these functions, plasticity appears more or less gradually, as soon as $ U >0$. That is, $U_{y}=0$. The limit $n\rightarrow \infty$ is the Heaviside function (eq. \[eq:hHeaviside\]).
More generally, the yield function can be thought as the convolution of the Heaviside function ${\cal H}$ and a distribution of yield values $p_Y$: $$h(U)=\int p_Y(U_Y) {\cal H}(\vert U\vert-U_Y) dU_{Y}.
\label{eq:hconvolution}$$ For instance, if the distribution of yield values $p$ is a Dirac peak at $U_Y$, it results in a Heaviside yield function $h$ (eq. \[eq:hHeaviside\]).
Robustness with respect to the choice of $h$
--------------------------------------------
Some functions $U(\scalardeformation)$ from table \[tab:hfunctions\] are plotted on figure \[fig:HandFromrest\]b. Strikingly, they do not depend much on the actual expression of $h(U)$. In fact, only the expression of $h$ near $U_Y$ matters; the relation between $\scalardeformation$ and $U$ is robust. The elastic deformation $U$ is close to the imposed strain $\scalardeformation$ at low applied strain, and tends to a saturation value at large applied strain.
The only important feature of $h$ is its derivative $h'$ just below the yield point: $$h' = \left( \frac{dh(U)}{dU} \right)_{U \to U_Y^-}.$$ It determines how the fraction in the r.h.s. of eq. (\[gamma\_(U)\]) diverges. Thus $U(\scalardeformation)$ is not the same if $h'$ is zero or infinite, or even not defined as in eq. . If it is infinite, $U$ reaches the saturation value at a finite value of applied deformation.
Conversely, if $h'$ is finite, as in most examples of Table I, the behaviour is universal. In eq. (\[gamma\_(U)\]), the fraction diverges as $(U_Y-U) h'$. Thus, whatever the value of $h'$, $\scalardeformation(U)$ diverges logarithmically and $U(\scalardeformation)$ tends exponentially towards $U_Y$.
[53]{} natexlab\#1[\#1]{}bibnamefont \#1[\#1]{}bibfnamefont \#1[\#1]{}citenamefont \#1[\#1]{}url \#1[`#1`]{}urlprefix\[2\][\#2]{} \[2\]\[\][[\#2](#2)]{}
, ** (, ).
, ** (, , ).
, ** (, , ).
, ** (, ).
, , , ** (, ).
, , , ** (, ).
, ****, ().
, ** (, , ).
, ****, ().
, ** (, ).
, , , ****, ().
, , , , ****, ().
, ****, ().
, ****, ().
, , , , ****, ().
, ****, ().
, , , , .
, , , ****, ().
, , , , ****, ().
, , , , , , .
, Ph.D. thesis, ().
, , , .
, .
, , , ****, ().
, , , , ****, ().
, , , , , ****, ().
, ** (, ).
, ** (, ).
, , , ****, ().
, ****, ().
, ****, ().
, ****, ().
, , , .
, , , ****, ().
, ****, ().
, ().
, , , ****, ().
, , , ****, ().
, , , , ****, ().
, , , , , ****, ().
, , , , ****, ().
, , , ****, ().
, , , ****, ().
, ****, ().
, , , , ****, ().
, , , , , , , , ****, ().
, Ph.D. thesis, ().
, ** (, ).
, , , , , ****, ().
, , , ****, ().
, .
, .
, , , , , , , , ****, ().
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The strong Coulomb interactions and the small exciton radii in two-dimensional metal dichalcogenides can result in very fast capture of electrons and holes of excitons by mid-gap defects from Auger processes. In the Auger processes considered here, an exciton is annihilated at a defect site with the capture of the electron (or the hole) by the defect and the hole (or the electron) is scattered to a high energy. In the case of excitons, the probability of finding an electron and a hole near each other is enhanced many folds compared to the case of free uncorrelated electrons and holes. Consequently, the rate of carrier capture by defects from Auger scattering for excitons in metal dichalcogenides can be 100-1000 times larger than for uncorrelated electrons and holes for carrier densities in the $10^{11}$-$10^{12}$ cm$^{-2}$ range. We calculate the capture times of electrons and holes by defects and show that the capture times can be in the sub-picosecond to a few picoseconds range. The capture rates exhibit linear as well as quadratic dependence on the exciton density. These fast time scales agree well with the recent experimental observations [@Shi13; @Lagarde14; @Korn11; @Wangb14], and point to the importance of controlling defects in metal dichalcogenides for optoelectronic applications.'
author:
- 'Haining Wang, Jared H. Strait, Changjian Zhang, Weimin Chan, Christina Manolatou, Sandip Tiwari, Farhan Rana'
title: Fast Exciton Annihilation by Capture of Electrons or Holes by Defects via Auger Scattering in Monolayer Metal Dichalcogenides
---
Introduction
============
Many body interactions play an important role in determining the electronic and optoelectronic properties of two-dimensional (2D) transition metal dichalcogenides (TMDs). The exciton binding energies in 2D chalcogenides are almost an order of magnitude larger compared to other bulk semiconductors [@Fai10; @Xu13; @Changjian14; @timothy; @Chernikov14]. The strong Coulomb interactions and small exciton radii in 2D-TMDs result in large optical oscillator strengths [@Changjian14; @Konabe14; @Berg14] and short radiative lifetimes [@Wanga14]. In this paper we show that the same factors also result in very fast capture of electrons and holes of excitons by defects from Auger processes leading to fast non-radiative recombination rates. The basic idea can be understood as follows. Consider the Auger process in which a hole (in the valence band) scatters off an electron (in the conduction band) and is captured by a mid-gap defect level and the electron (in the conduction band) takes the energy released in the hole capture process. In the case of uncorrelated electrons and holes, the rate for this process is proportional to the product of the hole density $p$ and the probability of finding an electron near the hole, which is proportional to the electron density $n$. But in the case of tightly bound excitons, an electron is present near the hole with a very high probability proportional to $|\phi(\vec{r}=0)|^{2}$, where $\phi(\vec{r})$ is related to the exciton wavefunction (see the discussion below). Therefore, the rate for a hole (or an electron) in a tightly bound exciton to get captured by a defect is proportional to the exciton density times $|\phi(\vec{r}=0)|^{2}$. Generally speaking, Auger rates in semiconductors are considered to be important only at large carrier densities [@Landsberg92]. But given the small exciton radii in 2D-TMDs (in the 7-10$\AA$ range), $|\phi(\vec{r}=0)|^{2}$, which is inversely proportional to the square of the exciton radius, can be extremely large and, consequently, Auger capture rates in 2D-TMDs can be very fast. Compared to the rates for direct electron-hole recombination via interband Auger scattering (exciton-exciton annihilation), which can be limited by the orthogonality of the conduction and valence band Bloch states, the rates for the capture of electrons and holes of excitons by defects can be very fast when the defect states have a good overlap with the conduction or valence band Bloch states.
Quantum efficiencies of TMD light emitters and detectors that have been reported are extremely poor; in the .0001-.01 range [@Lopez13; @Ross14; @Yin12; @Steiner13; @Pablo14]. Similar quantum efficiencies for TMDs have been observed in photoluminescence experiments [@Fai10; @Feng10; @Wangb14]. Therefore, most of the electrons and holes injected electrically or optically in TMDs recombine non-radiatively. Given that the average radiative lifetimes of excitons in TMDs are in the range of hundreds of picoseconds to a few nanoseconds [@Wanga14], the non-radiative recombination or capture times in TMDs are expected to be of the order of a few picoseconds. Several experimental results on the ultrafast carrier dynamics in photoexcited monolayer MoS$_{2}$ do indeed point to non-radiative recombinaton and/or capture times in the few picoseconds range [@Shi13; @Korn11; @Lagarde14; @Wangb14]. The mechanisms by which electrons and holes recombine non-radiatively and/or are captured by defects, and the associated time scales, remain to be clarified. The results in this paper show that electrons and holes of excitons in TMDs can get captured by defects on very short times scales that are in the sub-picosecond to a few picoseconds range resulting in exciton annihilation. The capture rates exhibit linear as well as quadratic dependence on the exciton density. The quadratic dependence of the exciton annihilation rate on the exciton density is generally considered to be an exclusive characteristic of exciton-exciton annihilation processes via interband Auger scattering. Although the discussion in this paper focuses on monolayer MoS$_{2}$, the analysis and the results presented here are expected to be relevant to all 2D-TMDs, and are expected to be useful in designing metal dichalcogenide optoelectronic devices as well as in helping to understand and interpret experimental data [@Shi13; @Lagarde14; @Korn11; @Wangb14].
Theoretical Model
=================
![Two basic Auger processes for the capture of an electron (a) or a hole (b) of an exciton by a defect state are depicted [@Landsberg92; @Landsberg80].[]{data-label="fig:auger_1"}](./auger_1){width=".3\textwidth"}
Introduction
------------
The two basic Auger processes for the capture of an electron (a) or a hole (b) of an exciton by a defect state are depicted in Fig.\[fig:auger\_1\]. Proper partitioning of the Hamiltonian is important in order to compute the rates of these processes. We discuss the terms in the Hamiltonian describing various processes below.
The Non-interacting Hamiltonian {#sec:nonintH}
-------------------------------
The crystal structure of a monolayer of group-VI dichalcogenides $MX_{2}$ (e.g. $M$=Mo,W and $X$=S,Se) consist of $X$-$M$-$X$ layers, and within each layer the $M$ atoms (or the $X$ atoms) form a 2D hexagonal lattice. Each $M$ atoms is surrounded by 6 nearest neighbor $X$ atoms in a trigonal prismatic geometry with $D_{3h}^{1}$ symmetry. The valence band maxima and conduction band minima occur at the $K$ and $K'$ points in the Brillouin zone. Most of the weight in the conduction and valence band Bloch states near the $K$ and $K'$ points resides on the d-orbitals of $M$ atoms [@Falko13; @yao12; @timothy]. The spin up and down valence bands are split near the $K$ and $K'$ points by 0.1-0.2 eV due to the spin-orbit-coupling [@Falko13; @yao12; @timothy; @Liu14]. In comparison, the spin-orbit-coupling effects in the conduction band are much smaller [@Liu14]. Assuming only d-orbitals for the conduction and valence band states, and including spin-orbit coupling, one obtains the following simple spin-dependent tight-binding Hamiltonian (in matrix form) near the $K$($K'$) points [@yao12], $$\left[
\begin{array}{cc}
\Delta/2 & \hbar v k_{-} \\
\hbar v k_{+} & -\Delta/2 + \lambda \tau \sigma
\end{array} \right] \label{eq:H1}$$ Here, $\Delta$ is related to the material bandgap, $\sigma=\pm1$ stands for the electron spin, $\tau=\pm1$ stands for the $K$ and $K'$ valleys, $2\lambda$ is the splitting of the valence band due to spin-orbit coupling, $k_{\pm}=\tau k_{x}\pm ik_{y}$, and the velocity parameter $v$ is related to the coupling between the orbitals on neighboring $M$ atoms. From density functional theories [@Lam12; @Falko13], $v\approx 5-6 \times 10^5 $ m/s. The wavevectors are measured from the $K$($K'$) points. The d-orbital basis used in writing the above Hamiltonian are $| d_{z^{2}} \rangle$ and $(| d_{x^{2}-y^{2}}\rangle + i\tau | d_{xy}\rangle)/\sqrt{2}$ [@yao12]. We will use the symbol $s$ for the combined valley ($\tau$) and spin ($\sigma$) degrees of freedom. Defining $\Delta_{s}$ as $\Delta - \lambda \tau \sigma$, the energies and eigenvectors of the conduction and valence bands are [@yao12; @Efimkin13], $$E_{{c \atop v},s}(\vec{k}) = \frac{\lambda \tau \sigma}{2} + \gamma \sqrt{ (\Delta_{s}/2)^{2} + (\hbar v k)^{2}}$$ $$| v_{{c \atop v},\vec{k},s}\rangle = \left[ \begin{array}{c} \cos(\theta_{\gamma,\vec{k},s}/2) e^{-i\tau\phi_{\vec{k}}/2} \\ \tau \gamma \sin(\theta_{\gamma,\vec{k},s}/2) e^{i\tau\phi_{\vec{k}}/2} \end{array} \right] \label{eq:wf1}$$ Here, $\gamma =1$ (or $-1$) stands for the conduction (or the valence) band, $\phi_{\vec{k}}$ is the phase of the wavevector $\vec{k}$, and, $$\cos(\theta_{\gamma,\vec{k},s}) = \gamma \frac{\Delta_{s}}{2 \sqrt{ (\Delta_{s}/2)^{2} + (\hbar v k)^{2}}} \label{eq:cs}$$ Near the conduction band minima and valence band maxima, the band energy dispersion is parabolic with well-defined effective masses, $m_{e}$ and $m_{h}$, for electrons and holes, respectively.
The Hamiltonian describing electron states in the conduction band, valence band, and a mid-gap defect state is, $$\begin{aligned}
H_{o} & = & \sum_{\vec{k},s}E_{c,s}(\vec{k}) c^{\dagger}_{\vec{k},s} c_{\vec{k},s} + \sum_{\vec{k},s}E_{v,s}(\vec{k}) b^{\dagger}_{\vec{k},s} b_{\vec{k},s} \nonumber \\
& & + \sum_{\sigma}E_{d} d^{\dagger}_{\sigma}d_{\sigma}\end{aligned}$$ Here, $c_{\vec{k},s}$, $b_{\vec{k},s}$, and $d_{\sigma}$ are the destruction operators for the conduction band, valence band, and defect states, respectively. The bandgap is $E_{g_{s,s'}} = E_{c,s}(\vec{k}=0) - E_{v,s'}(\vec{k}=0)$. Since only the smallest bandgap will be relevant in the discussion that follows, we will drop the spin/valley indices from $E_{g_{s,s'}}$ for simplicity.
Electron-Hole Interaction and Exciton States {#sec:Heh}
--------------------------------------------
The Coulomb interaction between the electrons and holes can be included by adding the following term to the Hamiltonian, $$H_{eh} = \frac{1}{A} \sum_{\vec{k},\vec{k}',\vec{q},s,s'} V(\vec{q}) F_{s,s'}(\vec{k},\vec{k}',\vec{q}) c^{\dagger}_{\vec{k}+\vec{q},s} b^{\dagger}_{\vec{k}'-\vec{q},s'} b_{\vec{k}',s'} c_{\vec{k},s}$$ $V(\vec{q})$ is the 2D Coulomb potential and equals $e^{2}/2\epsilon_{o} \epsilon(\vec{q})q$. The wavevector-dependent dielectric constant $\epsilon(\vec{q})$ for monolayer MoS$_{2}$ is given by Zhang et al. [@Changjian14] and Berkelbach et al. [@timothy]. $F_{s,s'}(\vec{k},\vec{k}',\vec{q})$ is [@Efimkin13], $$F_{s,s'}(\vec{k},\vec{k}',\vec{q}) = \langle v_{c,\vec{k}+\vec{q},s}| v_{c,\vec{k},s} \rangle \, \langle v_{v,\vec{k}'-\vec{q},s'} | v_{v,\vec{k}',s'} \rangle \label{eq:F1}$$ Near the conduction band minima, where $\hbar v k << \Delta_{s}$, $\cos(\theta_{\gamma,\vec{k},s}) \approx 1$ and $\sin(\theta_{\gamma,\vec{k},s}) << 1$. Similarly, near the valence band maxima, $\sin(\theta_{\gamma,\vec{k},s}) \approx 1$ and $\cos(\theta_{\gamma,\vec{k},s}) << 1$. Therefore, for wavevectors near the band extrema one can make the approximation [@Efimkin13], $$F_{s,s'}(\vec{k},\vec{k}',\vec{q}) = e^{i(\tau\phi_{\vec{k}+\vec{q}} - \tau\phi_{\vec{k}} + \tau'\phi_{\vec{k}'} - \tau'\phi_{\vec{k}'-\vec{q}})/2} \label{eq:F2}$$ Exciton states are approximate eigenstates of the Hamiltonian $H_{o} + H_{eh}$. Assuming that the ground state of the semiconductor is $|\psi_{o}\rangle$, which consists of a filled valence band and an empty conduction band, an exciton state with in-plane momentum $\vec{Q}$ can be constructed from the ground state as follows [@Changjian14; @Efimkin13], $$|\psi_{s,s',\alpha}(\vec{Q})\rangle = \frac{1}{\sqrt{A}}\sum_{\vec{k}} \psi_{\alpha,\vec{Q}}(\vec{k}) c_{\vec{k} + \frac{m_{e}}{m_{ex}}\vec{Q},s}^{\dagger} b_{\vec{k}-\frac{m_{h}}{m_{ex}}\vec{Q},s'}|\psi_{o} \rangle \label{eq:exciton}$$ The exciton wavefunction is $\psi_{\alpha,\vec{Q}}(\vec{k})$. The electron and hole effective masses are $m_{e}$ and $m_{h}$, respectively. The exciton mass is $m_{ex} = m_{e} + m_{h}$, and the reduced electron-hole mass is $m_{r}$. If one writes the exciton wavefunction as, $$\psi_{\alpha,\vec{Q}}(\vec{k}) = \phi_{\alpha}(\vec{k}) e^{i(\tau\phi_{\vec{k}+ (m_{e}/m_{ex})\vec{Q}} + \tau'\phi_{\vec{k}'-(m_{h}/m_{ex})\vec{Q}})/2} \label{eq:psiphi}$$ then the exciton wavefunction $\phi_{\alpha}(\vec{k})$ satisfies the standard exciton eigenvalue equation [@Changjian14; @timothy], $$\begin{aligned}
& & \left[\bar{E}_{c}(\vec{k}) - \bar{E}_{v}(\vec{k}) \right]\phi_{\alpha}(\vec{k}) - \frac{1}{A} \sum_{\vec{q}} V(\vec{q}) \phi_{\alpha}(\vec{k}-\vec{q}) \nonumber \\
& & = E_{\alpha}(\vec{Q}) \phi_{\alpha}(\vec{k}) \label{eq:eigen}\end{aligned}$$ with an eigenvalue $E_{\alpha}(\vec{Q})$ given by, $$E_{\alpha}(\vec{Q}) = E_{g} - E_{\alpha} + \frac{\hbar^{2} Q^{2}}{2 m_{ex}}$$ where, $E_{\alpha}$ is the exciton binding energy. The energy $E_{\alpha}(\vec{Q})$ is measured with respect to the energy of the ground state $|\psi_{o}\rangle$. [*Note that the phase factors cancel out and do not appear in the exciton eigenvalue equation*]{}. The exciton wavefunctions are orthonormal and complete in the sense [@Kira12], $$\int \frac{d^{2}\vec{k}}{(2\pi)^{2}} \phi^{*}_{\alpha}(\vec{k}) \phi_{\beta}(\vec{k}) = \delta_{\alpha,\beta} \label{eq:orth}$$ $$\sum_{\alpha} \phi_{\alpha}(\vec{k}) \phi^{*}_{\alpha}(\vec{k}') = (2\pi)^{2} \delta^{2}(\vec{k}-\vec{k}') \label{eq:comp}$$ The sum over $\alpha$ above includes all the discrete bound exciton states as well as the continuum of ionized exciton states. Finally, the probability of finding an electron and a hole at a distance $\vec{r}$ in the exciton state $| \psi_{\alpha,\vec{Q}}(\vec{k}) \rangle$ can be computed by destroying an electron and a hole using the real-space field destruction operators and then taking the overlap of the resulting state with the ground state $|\psi_{o}\rangle$. The result is $|\phi_{\alpha}(\vec{r})|^{2}$ where $\phi_{\alpha}(\vec{r})$ is the Fourier transform of $\phi_{\alpha}(\vec{k})$. Note that $\phi_{\alpha}(\vec{r})$ is not the Fourier transform of $\psi_{\alpha,\vec{Q}}(\vec{k})$, which also includes extra phase factors (see (\[eq:psiphi\]).
Exciton Basis {#sec:ex_basis}
-------------
In what follows, we will use the exciton basis. The exciton creation operator $B^{\dagger}_{s,s',\alpha}(\vec{Q})$ can be defined as, $$B^{\dagger}_{s,s',\alpha}(\vec{Q}) = \frac{1}{\sqrt{A}}\sum_{\vec{k}} \psi_{\alpha,\vec{Q}}(\vec{k}) c_{\vec{k} + \frac{m_{e}}{m_{ex}}\vec{Q},s}^{\dagger} b_{\vec{k}-\frac{m_{h}}{m_{ex}}\vec{Q},s'}$$ Using the completeness and the orthogonality of the exciton wavefunctions given in (\[eq:comp\]) and (\[eq:orth\]), we get, $$c_{\vec{k},s}^{\dagger} b_{\vec{k}',s'} = \frac{1}{\sqrt{A}}\sum_{\alpha} \psi^{*}_{\alpha,\vec{Q}}(\vec{k}_{r}) B^{\dagger}_{s,s',\alpha}(\vec{Q})$$ Here, $\vec{k}_{r}$ and $\vec{Q}$ equal $(m_{h}/m_{ex})\vec{k} + (m_{e}/m_{ex})\vec{k}'$ and $\vec{k}-\vec{k}'$ on the left hand side, respectively. Products of electron and hole creation and destruction operators can thus be expressed in terms of the exciton operators.
![The computed orbitals of the defect states in a MoS$_{2}$ monolayer corresponding to a sulfur vacancy are shown (from Noh et al. [@Noh14]). (a) and (b) show $A_{1}$ state and the two degenerate $E$ states, respectively.[]{data-label="fig:s_vac"}](./s_vac_b){width=".45\textwidth"}
Defect States
-------------
TMDs ($MX_{2}$), and in particular Monolayer MoS$_{2}$, are known to have several different kinds of point defects, such as $M$ and $X$ vacancies and interstitials, impurity atoms, in addition to grain boundaries and dislocations [@Sofo04; @Komsa12; @Seifert13; @Kong13; @Noh14; @Guinea14; @Robertson13; @Hao13; @VanDerZande13]. The goal in this Section is not to give a detailed description of different defect states in TMDs, something well beyond the scope of this paper, but to capture the essential physics in a way that would enable us to obtain capture rates for electrons and holes and present the main ideas associated with the capture processes.
Since the Bloch states form a complete set, the wavefunction $\psi_{d}(\vec{r})$ of the electron in the defect state can be expanded in terms of the Bloch states from all the bands [@Landsberg92]. In most cases of practical interest, only Bloch states in the vicinity of certain points, $\vec{K}_{s}$, in the Brillouin zone, such as $\Gamma$, $M$, $K$ and $K'$ in the case of 2D-TMDs, need to be included in the expansion and therefore one may write, $$\psi_{d}(\vec{r}) = \frac{1}{\sqrt{A}} \sum_{n,\vec{k},s} c_{n,s}(\vec{k}) \frac{e^{i(\vec{K}_{s} + \vec{k}).\vec{r}}}{\sqrt{A}} u_{n,\vec{k},s}(\vec{r}) \label{eq:defect1}$$ In the expression above, $u_{n,\vec{k},s}(\vec{r})$ are the periodic parts of the Bloch functions. The sum over $n$ runs over all the energy bands. Whereas shallow defect levels can usually be described well by limiting the summation above to a single band, deep mid-gap defect levels generally have contributions from multiple bands [@Landsberg92; @Landsberg80]. The above expression can usually be cast in much simpler forms for specific defect states.
As an example, we consider the case of the deep point defect in MoS$_{2}$ due to a sulfur atom vacancy. A sulfur atom vacancy is a common defect in MoS$_{2}$ monolayers and can have a small formation energy [@Kong13; @Robertson13; @Noh14]. The three states within the bandgap associated with a sulfur vacancy have been obtained previously using ab-initio techniques [@Kong13; @Robertson13; @Noh14]. These defect states consist of: (i) a single $A_{1}$ state, made up of mostly the $d_{xz}$ and $d_{yz}$ orbitals of the Mo atoms adjacent to the missing S atom, with an energy few tenths of an eV above the valence band maxima, and (ii) two degenerate $E$ states, made up of mostly the $d_{z^{2}}$, $d_{x^{2}-y^{2}}$, and $d_{xy}$ orbitals of the Mo atoms adjacent to the missing S atom, with an energy 1.4-1.6 eV above the valence band maxima. All the defect states are spin-degenerate and correspond to the one ($A_{1}$) and two dimensional ($E$) representations of the trigonal symmetry group $C_{3v}$. The computed orbitals of these states are shown in Fig.\[fig:s\_vac\] (from Noh et al. [@Noh14]). A defect state can be an efficient center for non-radiative recombination due to Auger scattering only if it has good overlaps with the Bloch states of both the conduction and the valence bands. The $E$ states fit this criterion. The $E$ states can be described well by limiting the summation in the expression above to the Bloch states of the conduction and the valence band extrema at the $K$ and $K'$ points. Since all the orbitals forming the $E$ states have weights almost entirely on the Mo atoms adjacent to the missing S atom, one may write $c_{{c \atop v},s}(\vec{k}) \approx \chi_{d}(\vec{k}) e^{i\gamma\tau\phi_{\vec{k}}/2} b_{{c \atop v},s}$. Since $e^{i\gamma\tau\phi_{\vec{k}}/2} u_{{c \atop v},\vec{k},s}(\vec{r})$ does not vary much with $\vec{k}$ near the band extrema, the sum in (\[eq:defect1\]) can be rearranged to give, $$\psi_{d}(\vec{r}) = \chi_{d}(\vec{r}) \sum_{{n=c,v \atop s}} b_{n,s} e^{i\vec{K}_{s}.\vec{r}} e^{i\gamma\tau\phi_{\underline{\vec{k}}}/2} u_{n,\underline{\vec{k}},s}(\vec{r}) \label{eq:defect2}$$ Here, the line under $\vec{k}$ means that any wavevector near the band extrema can be chosen. The function $\chi_{d}(\vec{r})$ is expected to be localized at the defect, becoming very small at the second nearest Mo atom near the defect site.
Hamiltonian for the Capture of Holes and Electrons {#sec:Hc}
--------------------------------------------------
Consider process (b) in Fig.\[fig:auger\_1\] in which a hole scatters off an electron and is captured by a defect and the electron is scattered to a higher energy. The relevant term in the Coulomb interaction Hamiltonian that describes the hole capture process in Fig.\[fig:auger\_1\](b) can be written as, $$H_{hc} = \frac{1}{A} \sum_{\vec{k},\vec{k}',\vec{q},s,s'} V(\vec{q}) M_{s,s'}(\vec{k},\vec{k}',\vec{q}) c^{\dagger}_{\vec{k}+\vec{q},s} b^{\dagger}_{\vec{k}',s'} d_{\sigma'} c_{\vec{k},s} + h.c.$$ The overlap factor $M_{s,s'}(\vec{k},\vec{k}',\vec{q})$ equals, $$\begin{aligned}
& & M_{s,s'}(\vec{k},\vec{k}',\vec{q}) = \langle v_{c,\vec{k}+\vec{q},s} | v_{c,\vec{k},s} \rangle \nonumber \\
& & \times \sum_{n=c,v} b_{n,s'} e^{i\gamma\tau'\phi_{\underline{\vec{k}}}/2} \langle v_{v,\vec{k}',s'} | v_{n,\underline{\vec{k}},s'} \rangle \nonumber \\
& & \times \frac{1}{\sqrt{A}} \int d^{2}\vec{r} \, \chi_{d}(\vec{r}) \, e^{-i(\vec{k}' + \vec{q}).\vec{r}} \nonumber \\
& & \approx e^{i(\tau\phi_{\vec{k}+\vec{q}} - \tau\phi_{\vec{k}} - \tau'\phi_{\vec{k}'})/2} \frac{b_{v,s'}}{\sqrt{A}} \chi_{d}(\vec{k}'+\vec{q})\end{aligned}$$ Similarly, the electron capture process (Fig.\[fig:auger\_1\](a)) is described by the Hamiltonian, $$H_{ec} = \frac{1}{A} \sum_{\vec{k},\vec{k}',\vec{q},s,s'} V(\vec{q}) L_{s,s'}(\vec{k},\vec{k}',\vec{q}) d^{\dagger}_{\sigma} b^{\dagger}_{\vec{k}',s'} b_{\vec{k}'+\vec{q},s'} c_{\vec{k},s} + h.c.$$ where overlap factor $L_{s,s'}(\vec{k},\vec{k}',\vec{q})$ equals, $$L_{s,s'}(\vec{k},\vec{k}',\vec{q}) \approx e^{i(\tau'\phi_{\vec{k}'+\vec{q}} - \tau'\phi_{\vec{k}'} - \tau \phi_{\vec{k}})/2} \frac{b^{*}_{c,s}}{\sqrt{A}} \chi^{*}_{d}(\vec{k}+\vec{q})$$ The potential of the defect does not appear in the Hamiltonian above. The reason for this is that it has already been taken into account in defining the non-interacting Hamiltonian, and its eigenstates, in Section (\[sec:nonintH\]).
Electron and Hole Capture Rates for Excitons
============================================
We assume an initial state described by the density operator $\rho_{i}$ in which the exciton occupation $n_{s,s',\alpha}(\vec{Q})$, defect occupation $f_{d}$, and conduction and valence band occupations are given by, $$\begin{aligned}
&& \langle d^{\dagger}_{\sigma} d_{\sigma'} \rangle = f_{d} \, \delta_{\sigma,\sigma'} \nonumber \\ \label{eq:ex_avg}
&& \langle c^{\dagger}_{\vec{k},s} c_{\vec{k}',s'}\rangle = f_{c,s}(\vec{k}) \delta_{s,s'} \delta_{\vec{k},\vec{k}'} \nonumber \\
&& \langle b^{\dagger}_{\vec{k},s} b_{\vec{k}',s'} \rangle = f_{v,s}(\vec{k}) \delta_{s,s'} \delta_{\vec{k},\vec{k}'} \nonumber \\
&& \langle B^{\dagger}_{s,s',\alpha}(\vec{Q}) B_{s,s',\alpha}(\vec{Q}) \rangle = n_{s,s',\alpha}(\vec{Q}) + \nonumber \\
&& \frac{1}{A}\sum_{\vec{k}} |\phi_{\alpha}(\vec{k})|^{2} f_{c,s}(\vec{k} + \frac{m_{e}}{m_{ex}} \vec{Q}) \left[ 1 - f_{v,s'}(\vec{k} - \frac{m_{h}}{m_{ex}} \vec{Q}) \right] \nonumber \\ \end{aligned}$$ The angled brackets stand for ensemble averaging with respect to the the density operator $\rho_{i}$. Since the excitons are not exact bosons, the value of $\langle B^{\dagger}_{s,s',\alpha}(\vec{Q}) B_{s,s',\alpha}(\vec{Q}) \rangle$ is not just equal to the exciton occupation $n_{s,s',\alpha}(\vec{Q})$. Using the cluster expansion to evaluate $\langle B^{\dagger}_{s,s',\alpha}(\vec{Q}) B_{s,s',\alpha}(\vec{Q}) \rangle$ results in the additional Hartree-Fock term shown above [@Kira06; @Koch06]. The same extra term also shows up in the luminescence spectra of excitons [@Kira12], and, as discussed below, this term results in a quadratic dependence of the capture rate on the exciton density at large exciton densities.
We assume that the electron and hole densities for different spins/valleys (including both free carriers and bound excitons) are $n_{s}$ and $p_{s'}$, respectively, and the defect density is $n_{d}$. The initial ensemble consists of states that are approximate eigenstates of $H_{o} + H_{eh}$ but not of $H_{o} + H_{eh} + H_{hc} + H_{ec}$. Therefore, we consider $H_{hc}$ and $H_{ec}$ as perturbations.
Electron Capture Rate
---------------------
We first consider process (a) in Fig.\[fig:auger\_1\] in which the electron is captured by a defect. The average electron capture rate $R_{ec}$ (units: per unit area per second) can be calculated from the first order perturbation theory using the exciton basis described in Section \[sec:ex\_basis\] and the average values given in (\[eq:ex\_avg\]). The details of the calculations are given in the Appendix. The final result is, $$\begin{aligned}
& & R_{ec} \approx \frac{2\pi}{\hbar} n_{d}(1-f_{d}) \sum_{s,s',\alpha} D_{v,s'}(q_{\alpha}) |\chi_{d}(q_{\alpha})|^{2} |b_{c,s}|^{2} \nonumber \\
& & \times \left| \frac{1}{A} \sum_{\vec{k}_{r}} V(q_{\alpha}\hat{x}-\vec{k}_{r}) \phi_{\alpha}(\vec{k}_{r}) \right|^{2} \left[ n_{s,s',\alpha} \right. \nonumber \\
& & \left. + \frac{1}{A^{2}} \sum_{\vec{k},\vec{Q}} |\phi_{\alpha}(\vec{k})|^{2} f_{c,s}(\vec{k} + \frac{m_{e}}{m_{ex}} \vec{Q}) \left[ 1 - f_{v,s'}(\vec{k} - \frac{m_{h}}{m_{ex}} \vec{Q}) \right] \right] \nonumber \label{eq:Rec_0} \\ \end{aligned}$$ Here, $D_{v,s'}$ is the valence band density of states (per valley per spin) evaluated at the energy of the scattered hole whose wavevector is $q_{\alpha}$. $q_{\alpha}$ is approximately given by the relation, $E_{v,s'}(0) - E_{v,s'}(q_{\alpha}) = E_{g} - E_{\alpha} - E_{d}$. Note that none of the phase factors appear in the above result. The exciton density $n_{s,s',\alpha}$ is, $$n_{s,s',\alpha} = \int \frac{d^{2}\vec{Q}}{(2\pi)^{2}} \, n_{s,s',\alpha}(\vec{Q})$$ If $q_{\alpha} >> k_{r}$ for all values of $k_{r}$ for which $\phi_{\alpha}(\vec{k}_{r})$ is significant, then the above expression reduces to, $$\begin{aligned}
R_{ec} & = & \frac{2\pi}{\hbar} n_{d} \left( 1-f_{d} \right) \sum_{s,s',\alpha} D_{v,s'}(q_{\alpha}) |V(q_{\alpha})|^{2} |\chi_{d}(q_{\alpha})|^{2} \nonumber \\
& & \times |b_{c,s}|^{2} \left[ |\phi_{\alpha}(\vec{r}=0)|^{2} n_{s,s',\alpha} + G_{\alpha} n_{s}p_{s'} \right] \label{eq:Rec_1}\end{aligned}$$ Expression for $G_{\alpha}$ is given in the Appendix. $G_{\alpha}$ is significant for only the lowest few exciton states.
Hole Capture Rate
-----------------
The rate for process (b) in Fig.\[fig:auger\_1\] in which the hole is captured by a defect can be calculated in the same way. The result is, $$\begin{aligned}
& & R_{hc} \approx \frac{2\pi}{\hbar} n_{d}f_{d} \sum_{s,s',\alpha} D_{c,s}(q_{\alpha}) |\chi_{d}(q_{\alpha})|^{2} |b_{v,s'}|^{2} \nonumber \\
& & \times \left| \frac{1}{A} \sum_{\vec{k}_{r}} V(q_{\alpha}\hat{x}-\vec{k}_{r}) \phi_{\alpha}(\vec{k}_{r}) \right|^{2} \left[ n_{s,s',\alpha} \right. \nonumber \\
& & \left. + \frac{1}{A^{2}} \sum_{\vec{k},\vec{Q}} |\phi_{\alpha}(\vec{k})|^{2} f_{c,s}(\vec{k} + \frac{m_{e}}{m_{ex}} \vec{Q}) \left[ 1 - f_{v,s'}(\vec{k} - \frac{m_{h}}{m_{ex}} \vec{Q}) \right] \right] \nonumber \\ \label{eq:Rhc_0}\end{aligned}$$ where now $q_{\alpha}$ is approximately given by the relation, $E_{c,s}(q_{\alpha}) - E_{c,s}(0) = E_{d} - E_{\alpha}$. And, as before, if $q_{\alpha} >> k_{r}$ for all values of $k_{r}$ for which $\phi_{\alpha}(\vec{k}_{r})$ is significant, then the above expression reduces to, $$\begin{aligned}
R_{hc} & = & \frac{2 \pi}{\hbar} n_{d} f_{d} \sum_{s,s',\alpha} D_{c,s}(q_{\alpha}) |V(q_{\alpha})|^{2} |\chi_{d}(q_{\alpha})|^{2} \nonumber \\
& & \times |b_{v,s'}|^{2} \left[ |\phi_{\alpha}(\vec{r}=0)|^{2} n_{s,s',\alpha} + G_{\alpha} n_{s}p_{s'} \right] \label{eq:Rhc_1}\end{aligned}$$
Coulomb Correlations and Enhancement of the Auger Capture Rates {#sec:corr}
---------------------------------------------------------------
Equation (\[eq:Rec\_1\]) for the electron capture rate can also be written as, $$\begin{aligned}
R_{ec} & = & \frac{2\pi}{\hbar} n_{d} \left( 1-f_{d} \right) \sum_{s,s',\alpha} D_{v,s'}(q_{\alpha}) |V(q_{\alpha})|^{2} |\chi_{d}(q_{\alpha})|^{2} \nonumber \\
& & \times |b_{c,s}|^{2} n_{s}p_{s'} \left[ G_{\alpha} + g_{s,s',\alpha}(\vec{r}=0) \right] \label{eq:Rec_2}\end{aligned}$$ where, $g_{s,s',\alpha}(\vec{r}=0) = |\phi_{\alpha}(\vec{r}=0)|^{2} n_{s,s',\alpha}/(n_{s}p_{s'})$. The quantity inside the square brackets in (\[eq:Rec\_2\]), $G_{\alpha} + g_{s,s',\alpha}(\vec{r}=0)$, describes the enhancement in the probability of finding an electron and a hole close to each other as a result of the attractive Coulomb interactions. It is interesting to compare the electron capture rate in (\[eq:Rec\_2\]) with the result obtained assuming no electron-hole attractive interaction (i.e. $H_{eh}=0$), $$\begin{aligned}
R_{ec} & = & \frac{2\pi}{\hbar} n_{d} \left( 1-f_{d} \right) \nonumber \\
& \times & \sum_{s,s'} D_{v,s'}(q_{o}) |V(q_{o})|^{2} |\chi_{d}(q_{o})|^{2} |b_{c,s}|^{2} n_{s} p_{s'} \label{eq:Rec_3}\end{aligned}$$ where $q_{o}$ is approximately given by the relation, $E_{v,s'}(0) - E_{v,s'}(q_{o}) = E_{g} - E_{d}$. It can be seen that the capture rate in (\[eq:Rec\_2\]) is larger by the same enhancement factor. Assuming all the electrons and holes are in the lowest ($\alpha =1$) bound exciton state, values of $D_{v,s'}$ and $|b_{c,s}|$ are independent of the valley/spin indices, and the exciton density is $n_{ex} = \sum_{s,s'} n_{s,s',\alpha=1}$, the comparison between (\[eq:Rec\_2\]) and (\[eq:Rec\_3\]) shows that the enhancement of the electron capture rate in the case of excitons is roughly proportional to $G_{\alpha=1} + |\phi_{\alpha=1}(\vec{r}=0)|^{2}/n_{ex}$. Given that the radius of the lowest exciton state in monolayer MoS$_{2}$ is in the 7-10 $\AA$ range [@Changjian14], the enhancement, assuming an exciton density of $10^{12}$ cm$^{-2}$, is in the 72-138 range, and in the 644-1308 range if the exciton density is assumed to be $10^{11}$ cm$^{-2}$. Therefore, the correlations in the positions of the electrons and the holes as a result of the attractive Coulomb interaction make electrons and holes in tightly bound excitons in TMDs far more susceptible to capture by defects compared to uncorrelated free carriers. Interestingly, even when the exciton density $n_{s,s',\alpha}$ is zero the capture rate in (\[eq:Rec\_2\]) is enhanced by the factors $G_{\alpha}$ compared to the rate in (\[eq:Rec\_3\]) for uncorrelated electrons and holes. Therefore, Coulomb correlations in the positions of electrons and holes due to the attractive interaction between them enhances the Auger scattering rates even at the Hartree-Fock level.
Numerical Results and Discussion
================================
Carrier Capture Times at Low Exciton Densities {#sec:num1}
----------------------------------------------
For numerical computations, we consider monolayer MoS$_{2}$ on a quartz substrate, as is the case in many experiments. We first assume that the exciton density is small enough ($\le$$10^{12}$ cm$^{-2}$) to allow one to ignore phase-space filling effects [@Changjian14]. We use the wavevector-dependent dielectric constant $\epsilon(\vec{q})$ for monolayer MoS$_{2}$ on quartz given by Zhang et al. [@Changjian14]. The defect state wavefunction is given in (\[eq:defect2\]). The values of $|b_{c,s}|^{2}$ and $|b_{v,s}|^{2}$ are assumed to be independent of the valley/spin indices. This is a good approximation for many important cases. For example, in the case of the sulfur vacancy in MoS$_{2}$ discussed earlier, the $E$ states have a total weight of $\sim$0.25 on the $d_{z^{2}}$ orbitals of the Mo atoms adjacent to the missing sulfur atom [@Yong14]. Since the conduction band Bloch states of both $K$ and $K'$ valleys are made up of mostly the $d_{z^{2}}$ orbitals of Mo atoms, $|b_{c,s}|^{2}$ is the same for both the valleys. We approximate the envelope, $\chi_{d}(\vec{r})$, of the defect state wavefunction in (\[eq:defect2\]) by a Gaussian, $\chi_{d}(\vec{r}) = \sqrt{2/(\pi a_{d}^{2})}e^{-r^{2}/a_{d}^{2}}$, where $a_{d} \approx 3$ $\AA$ (see Fig.\[fig:s\_vac\]). Note that the in-plane S-Mo bound length in MoS$_{2}$ is $\sim$1.83 $\AA$. Fig.\[fig:capture1\] plots the computed capture times of electrons ($\tau_{ec}$) and holes ($\tau_{hc}$) of excitons assuming that all the excitons are in the lowest state ($\alpha=1$). In the low exciton density limit considered here these capture times are independent of the exciton density. The defect density $n_{d}$ is assumed to be $2\times 10^{11}$ cm$^{-2}$. The capture times for electrons and holes shown in Fig.\[fig:capture1\] have been normalized by multiplying them by $|b_{c}|^{2}$ and $|b_{v}|^{2}$, respectively, given the uncertainty in the exact values of these parameters. In the calculation of the electron capture times the defect state is assumed to be empty ($f_{d}=0$), and in the calculation of the hole capture times the defect state is assumed to be full ($f_{d}=1$).
![The capture times of electrons ($\tau_{ec}$) and holes ($\tau_{hc}$) of excitons by defects in monolayer MoS$_{2}$ on quartz are plotted as a function of the defect energy within the bandgap. The exciton binding energy is $E_{\alpha=1}$ is 0.4 eV and the material bandgap is 2.3 eV [@Changjian14]. The plotted capture times for electrons and holes have been normalized by multiplying them by $|b_{c}|^{2}$ and $|b_{v}|^{2}$, respectively. The defect density $n_{d}$ is $2\times 10^{11}$ cm$^{-2}$.[]{data-label="fig:capture1"}](./capture1){width=".45\textwidth"}
The curves shown in Fig.\[fig:capture1\] can provide results in different situations. For example, in the case of the $E$ states associated with a sulfur vacancy, if $|b_{c}|^{2}$ is assumed to be $\sim$0.25 [@Yong14], then the electron capture time curve in Fig.\[fig:capture1\] would need to be multiplied by 4 in order to get the actual electron capture times. If the $E$ state energy is assumed to $\sim$1.5 eV above the valence band edge [@Noh14], then the electron capture time comes out to be $\sim$2.4 ps. Since the capture times decrease inversely with the defect density $n_{d}$, the capture times shown in Fig.\[fig:capture1\] can be interpolated for different values of the defect density. For example, a defect density of $8\times 10^{11}$ cm$^{-2}$ would result in an electron capture time of 0.6 ps for the $E$ state of a sulfur vacancy (under the same assumptions as stated above).
Fig.\[fig:capture1\] shows that shallower traps have much shorter capture times than deeper traps. This can be understood as follows. Energy conservation requires that the scattered electron (hole), in a hole (electron) capture process, takes away most of the energy. The deeper the trap the more the final energy of the scattered particle. Also, momentum conservation requires that the momentum of the scattered particle be provided by the relevant Fourier component of the defect state wavefunction. Therefore, the deeper the trap the larger the momentum transfer. Since in Fourier space the defect state wavefunction is $\chi_{d}(\vec{q}) = \sqrt{2\pi a_{d}^{2}} e^{-q^{2}a_{d}^{2}/4}$, larger momentum transfers result in smaller capture rates. Note that this result is largely independent of the exact assumed form of the defect state wavefunction. In addition, the Coulomb potential $V(\vec{q})$ also decreases for larger momentum transfers. Although the final density of states available to the scattered particle increases with the particle energy (for non-parabolic energy band dispersions in 2D), this increase is not enough to offset the reduction in the capture rates due to the factors mentioned above.
Since the energy width of the valence and conduction bands in MoS$_{2}$ are less than 1.2 eV and 0.6 eV [@Lam12; @Louie1], respectively, the limited horizontal extents of the curves in Fig.\[fig:capture1\] ensure that the electron (hole) scattered to a high energy in the hole (electron) capture process is scattered within the same band consistent with the assumptions made in this work. It is, however, possible for the scattered particle to go into a different band. For example, slightly away from the $K$ ($K'$) points, the next higher conduction band has Bloch states with a large weight on the $d_{z^{2}}$ orbitals of Mo atoms and these Bloch states will have large overlap with the Bloch states near the conduction band bottom [@Guinea13]. It should also be noted that the weights $|b_{c}|^{2}$ and $|b_{v}|^{2}$ for defects could be very small or zero. For example, in the case of sulfur vacancy $A_{1}$ states both $|b_{c}|^{2}$ and $|b_{v}|^{2}$ are expected to be very small [@Robertson13; @Noh14; @Yong14].
![The inverse capture time ($\tau^{-1}_{ec}$) for the electron of an exciton in monolayer MoS$_{2}$ on quartz is plotted as a function of the exciton density. The plotted capture time has been normalized by multiplying it by $|b_{c}|^{2}$. The defect density $n_{d}$ is $2\times 10^{11}$ cm$^{-2}$ and the defect energy $E_{d}$ is assumed to be 1.5 eV above the valence band edge. The inverse capture time increases with the exciton density $n_{ex}$ roughly as, $\tau^{-1}_{ec} \sim A + Bn_{ex}$ ($A$ and $B$ are constants).[]{data-label="fig:capture2"}](./capture2){width=".45\textwidth"}
Carrier Capture Times at High Exciton Densities {#sec:num2}
-----------------------------------------------
At large exciton densities (typically larger than $10^{12}$ cm$^{-2}$, but smaller than $10^{13}$ cm$^{-2}$, for 2D-TMDs [@Changjian14]), phase-space filling effects cannot be ignored in the description of the exciton states. We use the formalism developed by Kira and Koch [@Kira12; @Kira06]. When phase-space filling is taken into account, exciton eigenvalue equation in the relative co-ordinates becomes non-Hermitian (see the Appendix) and its solutions are expressed in terms of the [*left*]{} and the [*right*]{} eigenfunctions, $\phi^{L}_{\alpha,s,s'}(\vec{k},\vec{Q})$ and $\phi^{R}_{\alpha,s,s'}(\vec{k},\vec{Q})$, respectively. These eigenfunctions are a also a function of the center of mass momentum $\vec{Q}$, and are related as follows [@Kira12; @Kira06], $$\begin{aligned}
& & \phi^{R}_{s,s',\alpha}(\vec{k},\vec{Q}) = \phi^{L}_{s,s',\alpha}(\vec{k},\vec{Q}) \left[ f_{v,s'}(\vec{k} - \frac{m_{h}}{m_{ex}} \vec{Q}) \right. \nonumber \\
& & \left. - f_{c,s}(\vec{k} + \frac{m_{e}}{m_{ex}} \vec{Q}) \right] \end{aligned}$$ and obey the orthogonality relation, $$\int \frac{d^{2}\vec{k}}{(2\pi)^{2}} [\phi^{L}_{s,s',\alpha}(\vec{k},\vec{Q})]^{*} \phi^{R}_{s,s',\beta}(\vec{k},\vec{Q}) = \delta_{\alpha,\beta} \label{eq:orth_2}$$ In terms of these eigenfunctions, the expression for the electron capture rate becomes, $$\begin{aligned}
& & R_{ec} \approx \frac{2\pi}{\hbar} n_{d}(1-f_{d}) \frac{1}{A} \sum_{s,s',\alpha,\vec{Q}} D_{v,s'}(q_{\alpha}) |\chi_{d}(q_{\alpha})|^{2} |b_{c,s}|^{2} \nonumber \\
& & \times \left| \frac{1}{A} \sum_{\vec{k}_{r}} V(q_{\alpha}\hat{x}-\vec{k}_{r}) \phi^{R}_{s,s',\alpha}(\vec{k}_{r},\vec{Q}) \right|^{2} \left[ n_{s,s',\alpha}(\vec{Q}) \right. \nonumber \\
& & \left. + \frac{1}{A} \sum_{\vec{k}} |\phi^{L}_{s,s',\alpha}(\vec{k},\vec{Q})|^{2} f_{c,s}(\vec{k} + \frac{m_{e}}{m_{ex}} \vec{Q}) \right. \nonumber \\
& & \left. \times \left[ 1 - f_{v,s'}(\vec{k} - \frac{m_{h}}{m_{ex}} \vec{Q}) \right] \right] \nonumber \label{eq:Rec_4} \\ \end{aligned}$$ The expression for the capture rate of holes in the high exciton density case follows similarly from (\[eq:Rhc\_0\]). When all electrons and holes exist as excitons, self-consistency requires that the distribution functions are given by [@Kira12], $$\begin{aligned}
& & f_{c,s}(\vec{k}) = \frac{1}{A} \sum_{s',\alpha,\vec{Q}} [\phi^{L}_{s,s',\alpha}(\vec{k},\vec{Q})]^{*} \phi^{R}_{s,s',\alpha}(\vec{k},\vec{Q}) n_{s,s',\alpha}(\vec{Q}) \nonumber \\
& & 1 - f_{v,s'}(\vec{k}) = \frac{1}{A} \sum_{s,\alpha,\vec{Q}} [\phi^{L}_{s,s',\alpha}(\vec{k},\vec{Q})]^{*} \phi^{R}_{s,s',\alpha}(\vec{k},\vec{Q}) n_{s,s',\alpha}(\vec{Q}) \nonumber \label{eq:dist} \\\end{aligned}$$ Equations (\[eq:Rec\_4\]) and (\[eq:dist\]) show that the capture rate $R_{ec}$ has terms that go linearly as well as quadratically with the exciton density. The quadratic dependence comes from the Hartree-Fock term in the evaluation of $\langle B^{\dagger}_{s,s',\alpha}(\vec{Q}) B_{s,s',\alpha}(\vec{Q}) \rangle$ (see Equation (\[eq:ex\_avg\])). It can be understood as coming from the Auger scattering between the electron of one exciton and the hole of another exciton. Recall from the discussion in Section \[sec:corr\] that even at the Hartree-Fock level Auger scattering between electrons and holes is enhanced due to the Coulomb correlations compared to uncorrelated electrons and holes.
For numerical computations, we again consider monolayer MoS$_{2}$ on a quartz substrate, as in Section \[sec:num1\]. We solve the exciton eigenvalue equation for different exciton densities and obtain the exciton radii and the exciton binding energies [@Changjian14]. For simplicity, we consider the case when all the electrons and holes are in the lowest ($\alpha =1$) bound exciton state. Fig.\[fig:capture2\] plots the inverse capture time ($\tau^{-1}_{ec}$) of the electron of an exciton in monolayer MoS$_{2}$ on quartz as a function of the exciton density. The plotted capture time has been normalized by multiplying it by $|b_{c}|^{2}$. The defect density $n_{d}$ is $2\times 10^{11}$ cm$^{-2}$ and the defect energy $E_{d}$ is assumed to be 1.5 eV above the valence band edge. The inverse capture time increases with the exciton density $n_{ex}$ roughly as, $\tau^{-1}_{ec} \sim A + Bn_{ex}$ ($A$ and $B$ are constants), indicating that the capture rate $R_{ec}$ has both linear and quadratic dependence on the exciton density ($R_{ec} \sim A n_{ex} + B n_{ex}^{2}$). The term quadratic in the exciton density in $R_{ec}$ becomes significant at exciton densities higher than $\sim$10$^{12}$ cm$^{-2}$. When interpreting experimental data, this quadratic increase of the carrier capture rate with the exciton density can make exciton annihilation via carrier capture by defects indistinguishable from direct electron-hole recombination via interband Auger scattering (exciton-exciton annihilation), the rate of which is also expected to go quadratically with the exciton density.
Comments and Conclusion
=======================
The results presented in this paper show that the capture times for electrons and holes of excitons in TMDs can be very short - from less than a picosecond to a few picoseconds. These numbers agree well with the recently reported experimental results on the ultrafast carrier dynamics in photoexcited monolayer MoS$_{2}$ where fast relaxation times in the few picoseconds range were observed [@Shi13; @Korn11; @Lagarde14; @Wangb14]. In addition, the results in Fig.\[fig:capture1\] and Fig.\[fig:capture2\] are largely independent of the carrier temperature which is also consistent with the experimental observations [@Lagarde14; @Wangb14].
The expressions given in this work could overestimate (underestimate) the capture rates (times). The reasons are as follows. The magnitude of the intraband overlap integrals for Bloch states were assumed to equal unity in Section \[sec:Hc\] and only phase differences were taken into account. At energies much different from the band edge energies, the Bloch states are different from the band edge Bloch states, and consequently the magnitude of the overlap integrals are smaller than unity. For example, the two-band $k.p$ model in Section \[sec:nonintH\] shows that at wavevector $\vec{k}$ the conduction (valence) band Bloch states have contributions from the valence (conduction) band Bloch states at $\vec{k}=0$ with a weight given by $0.5 - 0.5 (\Delta_{s}/2)/\sqrt{(\Delta_{s}/2)^2 + (\hbar v k)^{2}}$. This implies a 15% weight at energies in the band that are $\sim$0.5 eV away from the band edge. In addition, both the conduction and valence band Bloch states are expected to get contributions from other lower and higher bands at large wavevectors [@Falko13]. However, we don’t expect the essential physics to change significantly or the rates to change by more than a factor of unity when these sources of error are removed. We should also point out that the rates for carrier capture by defects in 2D-TMDs can vary from sample to sample as the nature of defects is expected to depend on the method of sample preparation.
Acknowledgments
===============
The authors would like to acknowledge helpful discussions with Paul L. McEuen and Michael G. Spencer, and support from CCMR under NSF grant number DMR-1120296, AFOSR-MURI under grant number FA9550-09-1-0705, and ONR under grant number N00014-12-1-0072.
Appendices
==========
Details on the Electron Capture Rate
------------------------------------
In this Section, we derive the expression for the electron capture rate given in (\[eq:Rec\_0\]). The derivation of the hole capture rate is similar. We assume an initial state described by the density matrix $\rho_{i}$ in which the exciton occupation is $n_{s,s',\alpha}(\vec{Q})$, the defect density is $n_{d}$, the defect occupation is $f_{d}$, and the electron and hole densities (including both free carriers and bound excitons) are $n_{s}$ and $p_{s'}$, respectively. The average values of various operators are as given in (\[eq:ex\_avg\]). The rate of change of the total electron density is, $$\dot{n} = \frac{dn}{dt} = \frac{d}{dt} \left( \frac{1}{A}\sum_{s,\vec{k}} c^{\dagger}_{\vec{k},s} c_{\vec{k},s} \right)$$ Defining the interaction representation for the time development of operators as, $$O^{I}(t) = e^{\frac{i}{\hbar}(H_{o}+H_{eh}) t} O e^{-\frac{i}{\hbar}(H_{o}+H_{eh}) t}$$ the rate $R_{ec}$ for the electron capture by the defect can be found by picking the appropriate term from the expression obtained using the first order perturbation theory, $$\begin{aligned}
\langle \frac{dn}{dt} \rangle = \lim_{\eta \rightarrow 0} \frac{i}{\hbar}An_{d}\int_{-\infty}^{t} dt' \, e^{\eta t'} \, {\rm Tr}\left\{ \rho_{i} \left[ H_{ec}^{I}(t') , \dot{n}^{I}(t) \right] \right\} \nonumber \\\end{aligned}$$ Since the exciton states are approximate eigenstates of the Hamiltonian $H_{o} + H_{eh}$ we have, $$e^{\frac{i}{\hbar}(H_{o}+H_{eh}) t} B_{s,s',\alpha}(\vec{Q}) e^{-\frac{i}{\hbar}(H_{o}+H_{eh}) t} \approx B_{s,s',\alpha}(\vec{Q}) e^{-i\frac{E_{\alpha}(\vec{Q})}{\hbar}t}$$ It is therefore convenient to express the conduction and valence band creation and destruction operators appearing in $H_{ec}$ using the exciton basis described in Section \[sec:ex\_basis\]. We also point out here that the ensemble average of a product of operators of the form, $$\langle e^{\frac{i}{\hbar}(H_{o}+H_{eh}) t} c^{\dagger}_{\vec{k}_{1},s_{1}} b_{\vec{k}'_{1},s'_{1}} b^{\dagger}_{\vec{k}'_{2},s'_{2}} c_{\vec{k}_{2},s_{2}} e^{-\frac{i}{\hbar}(H_{o}+H_{eh}) t} \rangle$$ needs to be evaluated using the cluster expansion and keeping the correlation terms as well as the Hartree-Fock term [@Kira06; @Koch06]. The final result is, $$\begin{aligned}
& & R_{ec} = \frac{2\pi}{\hbar} n_{d}(1-f_{d}) \frac{1}{A^{4}} \mathop{\sum_{s,s',\vec{k}_{r},\vec{k}'_{r}}}_{\vec{Q},\vec{q},\alpha} |b_{c,s}|^{2} V^{*}(\vec{q}-\vec{k}'_{r}) V(\vec{q}-\vec{k}_{r}) \nonumber \\
& & \times |\chi_{d}(\vec{q} + (m_{e}/m_{ex})\vec{Q})|^{2} \phi^{*}_{\alpha}(\vec{k}'_{r}) \phi_{\alpha}(\vec{k}_{r}) \left[ n_{s,s',\alpha}(\vec{Q}) \right. \nonumber \\
& & \left. + \frac{1}{A} \sum_{\vec{k}} |\phi_{\alpha}(\vec{k})|^{2} f_{c,s}(\vec{k} + \frac{m_{e}}{m_{ex}} \vec{Q}) \left[ 1 - f_{v,s'}(\vec{k} - \frac{m_{h}}{m_{ex}} \vec{Q}) \right] \right] \nonumber \\
& & \times \delta \left( E_{g} - E_{\alpha} + \frac{\hbar^{2}Q^{2}}{2m_{ex}} - E_{d} \right. \nonumber \\
& & \left. - E_{v,s'}(0) + E_{v,s'}(\vec{q} - \frac{m_{h}}{m_{ex}}\vec{Q}) \right)\end{aligned}$$ Note that all the phase factors have canceled out. The exciton center of mass kinetic energy, $\hbar^{2}Q^{2}/2m_{ex}$, is expected to be much smaller than the energy difference $E_{g} - E_{\alpha} - E_{d}$. The former is expected to be in the few tens of meV range and the latter in the hundreds of meV range. The energy conserving delta function then enforces $q$ to the value determined by the condition $E_{v,s'}(0) - E_{v,s'}(q_{\alpha}) = E_{g} - E_{\alpha} - E_{d}$. Once the magnitude of $\vec{q}$ has been fixed in this way, it is easy to see that $R_{ec}$ does not depend on the angle of $\vec{q}$. So one may assume $q \approx q_{\alpha} \hat{x}$ and obtain, $$\begin{aligned}
& & R_{ec} \approx \frac{2\pi}{\hbar} n_{d}(1-f_{d}) \sum_{s,s',\alpha} D_{v,s'}(q_{\alpha}) |\chi_{d}(q_{\alpha})|^{2} |b_{c,s}|^{2} \nonumber \\
& & \times \left| \frac{1}{A} \sum_{\vec{k}_{r}} V(q_{\alpha}\hat{x}-\vec{k}_{r}) \phi_{\alpha}(\vec{k}_{r}) \right|^{2} \left[ n_{s,s',\alpha} \right. \nonumber \\
& & \left. + \frac{1}{A^{2}} \sum_{\vec{k},\vec{Q}} |\phi_{\alpha}(\vec{k})|^{2} f_{c,s}(\vec{k} + \frac{m_{e}}{m_{ex}} \vec{Q}) \left[ 1 - f_{v,s'}(\vec{k} - \frac{m_{h}}{m_{ex}} \vec{Q}) \right] \right] \nonumber \\ \end{aligned}$$ Here, $D_{v,s'}$ is the valence band density of states (per valley per spin) evaluated at the energy of the scattered hole whose wavevector is $q_{\alpha}$. The exciton density $n_{s,s',\alpha}$ is, $$n_{s,s',\alpha} = \int \frac{d^{2}\vec{Q}}{(2\pi)^{2}} \, n_{s,s',\alpha}(\vec{Q})$$ If $q_{\alpha} >> k_{r}$ for all values of $k_{r}$ for which $\phi_{\alpha}(\vec{k}_{r})$ is significant, then the above expression reduces to, $$\begin{aligned}
R_{ec} & = & \frac{2\pi}{\hbar} n_{d} \left( 1-f_{d} \right) \sum_{s,s',\alpha} D_{v,s'}(q_{\alpha}) |V(q_{\alpha})|^{2} |\chi_{d}(q_{\alpha})|^{2} \nonumber \\
& & \times |b_{c,s}|^{2} \left[ |\phi_{\alpha}(\vec{r}=0)|^{2} n_{s,s',\alpha} + G_{\alpha} n_{s}p_{s'} \right] \label{eq:Rec_app}\end{aligned}$$ Equation (\[eq:Rec\_app\]) contains the exciton density $n_{s,s',\alpha}$ as well as the electron and hole densities (including both free carriers and bound excitons) $n_{s}$ and $p_{s'}$, respectively. The latter appear as a result of the Hartree-Fock term in the cluster expansion [@Kira06; @Koch06]. $G_{\alpha}$ is, $$\begin{aligned}
&& G_{\alpha} = \frac{|\phi_{\alpha}(\vec{r}=0)|^{2}}{n_{s}p_{s'} \, A^{2}} \sum_{\vec{k},\vec{Q}} |\phi_{\alpha}(\vec{k})|^{2} f_{c,s}(\vec{k} + \frac{m_{e}}{m_{ex}} \vec{Q}) \nonumber \\
&& \times \left[ 1 - f_{v,s'}(\vec{k} - \frac{m_{h}}{m_{ex}} \vec{Q}) \right]\end{aligned}$$ $G_{\alpha}$ is expected to be significant for only the lowest few exciton states.
If all the electrons and holes are assumed to be in the lowest ($\alpha =1$) bound exciton state then self-consistency requires that the distribution functions are given by [@Kira12], $$\begin{aligned}
f_{c,s}(\vec{k}) & = & |\phi_{\alpha=1}(\vec{k})|^{2} n_{s} \nonumber \\
1 - f_{v,s'}(\vec{k}) & = & |\phi_{\alpha=1}(\vec{k})|^{2} p_{s'} \label{eq:dist_2}\end{aligned}$$ Here, $n_{s} = \sum_{s'} n_{s,s',\alpha=1}$ and $p_{s'} = \sum_{s} n_{s,s',\alpha=1}$. One then obtains, $$\begin{aligned}
&& G_{1} = |\phi_{\alpha=1}(\vec{r}=0)|^{2} \frac{1}{A^{2}} \sum_{\vec{k},\vec{Q}} |\phi_{\alpha=1}(\vec{k} + \frac{m_{e}}{m_{ex}} \vec{Q})|^{2} \nonumber \\
&& \times |\phi_{\alpha=1}(\vec{k})|^{2} |\phi_{\alpha=1}(\vec{k} - \frac{m_{h}}{m_{ex}} \vec{Q})|^{2}\end{aligned}$$ Assuming the standard 2D exciton wavefunction [@Changjian14], $G_{1}$ equals $128/(5\pi) \approx 8.15$.
Description of Excitons States in the High Exciton Density Limit
----------------------------------------------------------------
In the high exciton density case, phase filling effects cannot be ignored in the description of the exciton states [@Changjian14; @Kira12]. The exciton wavefunctions, $\phi^{L}_{\alpha,s,s'}(\vec{k},\vec{Q})$ and $\phi^{R}_{\alpha,s,s'}(\vec{k},\vec{Q})$ satisfy the eigenvalue equations [@Kira12], $$\begin{aligned}
&& \left[ E_{c,s}(\vec{k}+(m_{e}/m_{ex})\vec{Q}) - E_{v,s'}(\vec{k}-(m_{h}/m_{ex})\vec{Q}) \right] \nonumber \\
&& \times \phi^{L}_{\alpha,s,s'}(\vec{k},\vec{Q}) - \frac{1}{A}\sum_{\vec{k}'} V(\vec{k}-\vec{k}') \phi^{L}_{\alpha,s,s'}(\vec{k}',\vec{Q}) \nonumber \\
&& \times \left[ f_{v,s'}(\vec{k}' - \frac{m_{h}}{m_{ex}} \vec{Q}) - f_{c,s}(\vec{k}' + \frac{m_{e}}{m_{ex}} \vec{Q}) \right] \nonumber \\
&& = E_{s,s',\alpha}(\vec{Q}) \phi^{L}_{s,s',\alpha}(\vec{k},\vec{Q})\end{aligned}$$ $$\begin{aligned}
&& \left[ E_{c,s}(\vec{k}+(m_{e}/m_{ex})\vec{Q}) - E_{v,s'}(\vec{k}-(m_{h}/m_{ex})\vec{Q}) \right] \nonumber \\
&& \times \phi^{R}_{\alpha,s,s'}(\vec{k},\vec{Q}) - \left[ f_{v,s'}(\vec{k} - \frac{m_{h}}{m_{ex}} \vec{Q}) \right. \nonumber \\
&& \left. - f_{c,s}(\vec{k} + \frac{m_{e}}{m_{ex}} \vec{Q}) \right] \frac{1}{A}\sum_{\vec{k}'} V(\vec{k}-\vec{k}') \phi^{R}_{\alpha,s,s'}(\vec{k}',\vec{Q}) \nonumber \\
&& = E_{s,s',\alpha}(\vec{Q}) \phi^{R}_{s,s',\alpha}(\vec{k},\vec{Q})\end{aligned}$$ The exciton wavefunctions satisfy the orthogonality and completeness relations, $$\int \frac{d^{2}\vec{k}}{(2\pi)^{2}} [\phi^{L}_{s,s',\alpha}(\vec{k},\vec{Q})]^{*} \phi^{R}_{s,s',\beta}(\vec{k},\vec{Q}) = \delta_{\alpha,\beta} \label{eq:orth_app}$$ $$\sum_{\alpha} \phi^{L}_{s,s',\alpha}(\vec{k},\vec{Q}) [\phi^{R}_{s,s',\alpha}(\vec{k}',\vec{Q})]^{*} = (2\pi)^{2} \delta^{2}(\vec{k}-\vec{k}') \label{eq:comp_app}$$ We also define, $$\begin{aligned}
\psi^{L/R}_{s,s',\alpha}(\vec{k},\vec{Q}) & = & \phi^{L/R}_{s,s',\alpha}(\vec{k},\vec{Q}) \nonumber \\
& \times & e^{i(\tau\phi_{\vec{k}+ (m_{e}/m_{ex})\vec{Q}} + \tau'\phi_{\vec{k}'-(m_{h}/m_{ex})\vec{Q}})/2}\end{aligned}$$ The exciton creation operator $B^{\dagger}_{s,s',\alpha}(\vec{Q})$ is defined as, $$B^{\dagger}_{s,s',\alpha}(\vec{Q}) = \frac{1}{\sqrt{A}}\sum_{\vec{k}} \psi^{L}_{s,s',\alpha}(\vec{k},\vec{Q}) c_{\vec{k} + \frac{m_{e}}{m_{ex}}\vec{Q},s}^{\dagger} b_{\vec{k}-\frac{m_{h}}{m_{ex}}\vec{Q},s'}$$ Using the completeness and the orthogonality of the exciton wavefunctions given in (\[eq:comp\_app\]) and (\[eq:orth\_app\]), we get, $$c_{\vec{k},s}^{\dagger} b_{\vec{k}',s'} = \frac{1}{\sqrt{A}} \sum_{\alpha} \psi^{R}_{s,s',\alpha}(\vec{k}_{r},\vec{Q}) B^{\dagger}_{s,s',\alpha}(\vec{Q})$$ where, $\vec{k}_{r}$ and $\vec{Q}$ equal $(m_{h}/m_{ex})\vec{k} + (m_{e}/m_{ex})\vec{k}'$ and $\vec{k}-\vec{k}'$ on the left hand side, respectively.
[99]{}
K. F. Mak, C. Lee, J. Hone, J. Shan, and T. F. Heinz, Phys. Rev. Lett. 105, 136805 (2010). J. S. Ross, S. Wu, H. Yu, N. J. Ghimire, A. M. Jones, G. Aivazian, J. Yan, D. G. Mandrus, D. Xiao, W. Yao, and X. Xu, Nat. Comm. 4, 1474 (2013). C. Zhang, H. Wang, W. Chan, C. Manolatou, F. Rana, Phys. Rev. B, 89, 205436 (2014). T. C. Berkelbach, M. S. Hybertsen, and D. R. Reichman, Phys. Rev. B 88, 045318 (2013). A. Chernikov, T. C. Berkelbach, H. M. Hill, A. Rigosi, Y. Li, O. B. Aslan, D. R. Reichman, M. S. Hybertsen, T. F. Heinz, Phys. rev. Lett., 113, 076802 (2014). D. K. Efimkin, Yu. E. Lozovik, Phys. Rev. B, 87, 245416 (2013). S. Konabe, S. Okada, Phys. Rev. B, 90, 15304 (2014). G. Berghauser, E. Malic, Phys. Rev. B, 89, 125309 (2014).
G. Liu, W. Shan, Y. Yao, W. Yao, D. Xiao, Phys. Rev., B, 88, 085433 (2014). H. Wang, C. Zhang, W. Chan, C. Manolatou, S. Tiwari, F. Rana, arXiv:1409.3996 (2014). P. T. Landsberg, “Recombination in Semiconductors”, Cambridge University Press, Cambridge, UK (1992). D J Robbins, P T Landsberg, Journal of Physics C, 13, 2425 (1980). H. Shi, R. Yan, Rusen, S. Bertolazzi, J. Brivio, B. Gao, A. Kis, D. Jena, H. Xing, Huili, L. Huang, ACS Nano, 7, 1072 (2013). D. Lagarde, L. Bouet, X. Marie, C. R. Zhu, B. L. Liu, T. Amand, P H. Tan, B. Urbaszek, Phys. Rev. Lett., 112, 047401 (2014). T. Korn, S. Heydrich, M. Hirmer, J. Schmutzler, C. Schüller, App. Phys. Lett., 99, 102109 (2011). D. Xiao, Gui-Bin Liu, W. Feng, X. Xu, and W. Yao, Phys. Rev. Lett. 108, 196802 (2012).
Private communication with Yong-Sung Kim [@Noh14]. E. Cappelluti, R. Roldan, J. A. Silva-Guillen, P. Ordejon, F. Guinea, Phys. Rev., B, 88, 075409 (2013). M. Kira, S. W. Koch, Prog. Quant. Electron., 30, 155 (2006). S. W. Koch, M. Kira, G. Khitrova, H. M. Gibbs, Nature Materials, 5, 523 (2006). T. Cheiwchanchamnangij and W. R. L. Lambrecht, Phys. Rev. B 85, 205302 (2012). D. Y. Qiu, F. H. da Jornada, and S. G. Louie, Phys. Rev. Lett. 111, 216805 (2013). H. Wang, C. Zhang, F. Rana,arXiv:1409.4518 (2014). M. Kira, S. W. Koch, “Semiconductor Quantum Optics”, Cambridge University Press, NY (2012). O. Lopez-Sanchez, D. Lembke, M. Kayci, A. Radenovic, and A. Kis, Nature Nanotechnology 8, 497 (2013). J. S. Ross, P. Klement, A. M. Jones, N. J. Ghimire, J. Yan, D. G. Mandrus, T. Taniguchi, K. Watanabe, K. Kitamura, W. Yao, et al., Nature Nanotechnology 9, 268 (2014). Z. Yin, H. Li, H. Li, L. Jiang, Y. Shi, Y. Sun, G. Lu, Q. Zhang, X. Chen, and H. Zhang, ACS Nano 6, 74 (2012). R. S. Sundaram, M. Engel, A. Lombardo, R. Krupke, A. C. Ferrari, P. Avouris, and M. Steiner, Nano Letters 13, 1416 (2013). B. W. H. Baugher, H. O. H. Churchill, Y. Yang, and P. Jarillo-Herrero, Nature Nanotechnology 9, 262 (2014). A. Splendiani, L. Sun, Y. Zhang, T. Li, J. Kim, C.-Y. Chim, G. Galli, and F. Wang, Nano Letters, 10, 1271 (2010). A. Kormanyos, V. Zolyomi, N. D. Drummond, P. Rakyta, G. Burkard and V. I. Falko, Phys. Rev. B 88, 045416 (2013). A. M. van der Zande, P. Y. Huang, D. A. Chenet, T. C. Berkelbach, Y. You, G.-H. Lee, T. F. Heinz, D. R. Reichman, D. A. Muller, and J. C. Hone, Nature Materials, 12, 554 (2013).
J. D. Fuhr, A. Saul, and J. O. Sofo, Phys. Rev. Lett. 92, 026802 (2004). H. P. Komsa, J. Kotakoski, S. Kurasch, O. Lehtinen, U. Kaiser, and A. V. Krasheninnikov, Phys. Rev. Lett., 109, 035503 (2012). A. N. Enyashin, M. Bar-Sadan, L. Houben, and G. Seifert, The Journal of Physical Chemistry C, 117, 10842 (2013). W. Zhou, X. Zou, S. Najmaei, Z. Liu, Y. Shi, J. Kong, J. Lou, P. M. Ajayan, B. I. Yakobson, J. C. Idrobo, Nano Lett., 13, 2615 (2013). J. Noh, H. Kim, Y. Kim, Phys. Rev., B, 89, 205417 (2014). D. Liu, Y. Guo, L. Fang, J. Robertson, Appl. Phys. Lett., 103, 183113 (2013). S. Yuan, R. Roldan, M. I. Katsnelson, and F. Guinea, Phys. Rev. B 90, 041402 (2014). H. Qiu, T. Xu, Z. Wang, W. Ren, H. Nan, Z. Ni, Q. Chen, S. Yuan, F. Miao, F. Song, et al., Nature Communications, 4, 2642 (2013).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In high-energy particle physics, workflow management systems are primarily used as tailored solutions in dedicated areas such as Monte Carlo production. However, physicists performing data analyses are usually required to steer their individual workflows manually which is time-consuming and often leads to undocumented relations between particular workloads. We present a generic analysis design pattern that copes with the sophisticated demands of end-to-end HEP analyses and provides a `make`-like execution system. It is based on the open-source pipelining package Luigi which was developed at Spotify and enables the definition of arbitrary workloads, so-called Tasks, and the dependencies between them in a lightweight and scalable structure. Further features are multi-user support, automated dependency resolution and error handling, central scheduling, and status visualization in the web. In addition to already built-in features for remote jobs and file systems like Hadoop and HDFS, we added support for WLCG infrastructure such as LSF and CREAM job submission, as well as remote file access through the Grid File Access Library. Furthermore, we implemented automated resubmission functionality, software sandboxing, and a command line interface with auto-completion for a convenient working environment. For the implementation of a $t\bar{t}H$ cross section measurement, we created a generic Python interface that provides programmatic access to all external information such as datasets, physics processes, statistical models, and additional files and values. In summary, the setup enables the execution of the entire analysis in a parallelized and distributed fashion with a single command.'
address: 'III. Physics Institute A, RWTH Aachen University, Germany'
author:
- 'M Erdmann, B Fischer, R Fischer, M Rieger'
title: 'Design and Execution of `make`-like, distributed Analyses based on Spotify’s Pipelining Package Luigi'
---
Motivation {#sec:motivation}
==========
The management of scientific workflows presents a complex challenge in today’s physics working environments in the context of mastering a specific research question. Current high-energy physics (HEP) analyses are designed to function on a large *scale* as they often require a significant amount of resources in terms of computing time, memory consumption, and disk space. A logical solution to resource limitation is parallelization on large-scale computing centers. Main challenges are, e.g., thousands of datasets to be processed, application of multivariate techniques like deep learning, and the diversity of applied algorithms such as final state event reconstruction. As a consequence, the *complexity* of an analysis, i.e., the degree of granularity and inhomogeneity of workloads, is increased. The relation between scale and complexity and the impact on analysis conception are indicated in figure \[fig:scalecomplexity\].
![Scale and complexity as specification measures for physics analyses and their impact on the choice of structural conception.[]{data-label="fig:scalecomplexity"}](figures/scalecomplexity){width="50.00000%"}
Physicists are usually required to steer their individual workloads manually, i.e., start (remote) processes, monitor their execution, handle potential failures and resubmission, and eventually initiate dependent workloads once the preceding ones finished successfully. In addition, intermediate or final output data must be often retrieved for further validation. On a large scale, this *management* task is not only time-consuming for the operating physicist, but, in contrast to automated approaches, also represents a risk for errors, e.g., the loss of information on the interplay between particular workloads. In fact, in certain cases the reproducibility of physics results might not be guaranteed.
We present a design pattern for physics analyses conception that copes with the challenges posed by scale and complexity. Development and testing took place alongside a $t\bar{t}H$ cross section measurement analysis. Therefore, both its usability and suitability could be demonstrated in a thorough context. Its core is based on the pipelining package Luigi [@luigi], which provides guidance on structuring arbitrary workloads (section \[sec:luigi\]). Scalability on HEP infrastructure is ensured by introducing common interfaces to remote computing facilities (section \[sec:submission\]) and distributed storage systems (section \[sec:storage\]). Furthermore, the portability of software and computing environments via sandboxing is discussed (section \[sec:sandboxes\]).
Luigi {#sec:luigi}
=====
Luigi is a Python module that helps users to “build complex pipelines of batch jobs, handle dependency resolution, and create visualizations to help manage multiple workflows” [@luigi2]. The execution model is based on targets and follows a `make`-like approach as it only computes what is really necessary in order to produce the output of a requested workload [@make]. While its initial development started at Spotify, it was made open-source in 2012, and is now a community-driven project with numerous contributors.
Building Blocks
---------------
Conceptually, Luigi’s core functionality is divided into five distinct components represented by Python classes: `Task`, `Target`, `Parameter`, `Worker`, and `Scheduler`. Their implementation is both lightweight and extensible, enabling users to model arbitrary workflows. Tasks are representations of atomic workload units that constitute a workflow. They can require one or more other tasks to denote a directional dependency. The common interface between dependent tasks is accomplished via targets, i.e., containers for arbitrary data such as file paths, database entries, or meta information. Tasks define their output as a collection of targets that should be created at run time as part of their actual payload. Parameters can alter the default behavior of tasks, effectively resulting in task *classes* to be considered as *templates*. They are registered at task definition while their actual value is assigned per task *instance* and can be changed at execution time by the user. It is often useful to pass parameter values to required tasks and to encode them in output target information, e.g. as path fragment of a file target. A task is executed in the context of a worker, which holds stateful information such as execution status, run time, and error data. Workers place tasks in dedicated subprocesses which results in an inherent way of local parallelization. In addition, workers can communicate with a central scheduler, which not only visualizes worker information and task dependencies in a modern web interface, but also provides a global task lock. This mechanism prevents situations where, e.g., two users try to run the same task and produce identical output in the same location, which could potentially lead to file corruption. A code example showing the definition of a simple task class including relevant building blocks and an exemplary dependency tree are presented in figure \[fig:task\]. An architectural overview scheme of the components is shown in figure \[fig:arch\].
Further features include automatic failure handling, command-line interface generation per task, task templates that support map-reduce jobs (e.g. for Hadoop, Pig, or Cascading), and file system abstractions (e.g. for HDFS).
[0.54]{} ![a) Code example showing the definition of a task class (`Reconstruction`) with parameters (`dataset`), required dependencies (`Selection`), produced output targets (`LocalTarget`), and a generic payload (`run`). b) Exemplary task dependency tree. The uppermost task is triggered, all tasks below constitute dependencies.[]{data-label="fig:task"}](figures/taskcode "fig:"){width="\textwidth"}
[0.45]{} ![a) Code example showing the definition of a task class (`Reconstruction`) with parameters (`dataset`), required dependencies (`Selection`), produced output targets (`LocalTarget`), and a generic payload (`run`). b) Exemplary task dependency tree. The uppermost task is triggered, all tasks below constitute dependencies.[]{data-label="fig:task"}](figures/tree "fig:"){width="\textwidth"}
Execution Model
---------------
Luigi’s execution model follows a `make`-like approach. It is initiated by invoking a task with appropriate parameters via the command-line interface or programmatically within Python. A task is considered complete when all of its output targets exist. Therefore, the most abstract target definition only contains a single `exists()` method which evaluates to `True` if the target exists, or `False` otherwise. This rule is initially used to built up a dependency tree. Starting from the triggered task, all branches of the tree are traversed recursively until a completed task is reached. Consequently, all tasks collected up to this point will be run using a configurable number of workers, spawned from the process that started the execution. As a result, Luigi computes only what is really necessary to produce the output of the triggered task. This paradigm can be classified as output-driven, in contrast to other, workflow-centered approaches where users start workflows as a whole. Benefits entail automatic output bookkeeping via target existence, transparent and deterministic reproducibility of results, and hence, decoupling of data and algorithms in collaborative working environments.
Remote Computing Resources {#sec:submission}
==========================
Current high-energy physics data analyses are often comprised of hundreds of thousands computing jobs. This usually exceeds local resources and makes interfaces to large-scale computing facilities inevitable. However, typical working environments require users to change existing and/or write additional code to incorporate such interfaces for remote job submission and processing.
We created a mechanism for employing remote computing resources on top of Luigi’s task model that follows two paradigms:
1\. Remote computing capabilities of a task should only require minimal, rather descriptive code changes.
2\. The decision on the actual run location is not hard-coded, but can be made at execution time.
Technically, this is achieved via *mixin* inheritance of an additional task class that provides submission and status retrieval capabilities. The rest of the code remains unchanged, although one can optionally define additional requirements that must be met before submission is commenced. When multiple mixin classes with submission capabilities are given, the particular behavior is steered by a task parameter. In addition, a user can specify the rules how one or multiple tasks translate into one or multiple jobs. The remote status is retrieved by a local representation of the corresponding task via polling which can be resumed in case of disruptions. As an example, we created an implementation for using resources of the Worldwide LHC Computing Grid (WLCG) which constitutes an appropriate use-case in the context of high-energy physics infrastructure. The mechanism also implements several mandatory features that add actual value to the daily working experience. They include automatic resubmission for failed jobs, the possibility to define pilot jobs, early stopping criteria with status prediction, and hooks for publishing status information to common job dashboards.
Distributed Data Storage {#sec:storage}
========================
The typical disk space consumption of a large-scale physics analyses may amount to multiple tens of gigabytes. Also, when dealing with remote computing resources and a high degree of parallelization, data should be read from and written to distributed, high-throughput storage systems.
We created a Luigi file system implementation and corresponding target classes that cope with most of the storage back-ends that are deployed in the current high-energy physics landscape. It is based on Python bindings of the Grid File Access Library (GFAL) [@gfal], which uses plugins to provide support for various back-ends, such as dCache, SRM, GridFTP, XRootD, Amazon S3, WebDAV, and Dropbox [@srm; @dcache; @gridftp; @xrootd]. The implemented interface resembles the local file target implementation, extended by convenience methods for handling file transfers. User authentication and session management is controlled via environment variables or can be configured programmatically. Mandatory features include batch transfers, automatic retries for robustness against network and connection disruptions, transfer validation, and local caching to reduce overhead due to redundant requests.
Environment Sandboxing {#sec:sandboxes}
======================
Besides portability of specific software environments across remote computing resources, the reproducibility of results over time poses a crucial challenge to the design of physics analyses. Dependencies on software and data of the host environment should be avoided as they typically are subject to change as part of maintenance and security measures. Furthermore, a particular workload might require software in an environment that is not compatible with other workloads. Possible solutions are virtual machines or virtual environments, e.g. via Linux containers, that are to be retained as *images* alongside analysis code [@boettiger]. Virtualization tools like Docker and Vagrant are promising candidates for managing images on a long-term basis [@docker; @docker3; @docker2; @vagrant].
We invented a generalized approach that combines the exchangeability of software environments with Luigi’s task and execution model. So-called *sandboxes* are configured on task level and allow for the definition of rules to either force the switch to a specific sandbox or to identify a fallback sandbox based on runtime conditions. They are created on demand, host the execution of one ore more tasks, and are terminated automatically once all tasks are completed successfully. By construction, sandboxes cannot be nested but are rather spawned consecutively by the executing worker process. This way, each task in a dependency tree can define its own, independent sandbox, which, when left unchanged, leads to reproducible and stable results over time.
![Architectural overview showing the interaction between components on local and remote resources. The use of a central scheduler is optional.[]{data-label="fig:arch"}](figures/arch){width="90.00000%"}
Conclusions {#sec:conclusions}
===========
The presented tools and concepts for generic analyses conception constitute a novel approach for coping with the increasing demands of modern high-energy physics data analysis. The Luigi pipelining package is a viable solution to address the complexity of structuring and executing workloads in a `make`-like fashion. Scalability in the scope of high-energy physics infrastructure is added in a non-intrusive way by interfacing common job submission systems and remote data storage on arbitrary locations. In addition, a customizable sandboxing mechanism ensures the integrity of software and computing environments, and therefore the reproducibility of physics results. Possible implementations are based on virtual environments via Linux containers (Docker) and virtual machines (Vagrant). As the described approach does not introduce constraints on the software or data structures to be used, it is considered a toolbox providing an *analysis design pattern* rather than a *framework*.
A resulting workflow represents a well-defined formulation of the interplay between particular tasks which often exists only in the “physicist’s head”. Since all knowledge about the analysis structure is preserved, loss of information is avoided, e.g. in situations when a team or allocations of duties are subject to change. Furthermore, targets define a flexible but clear interface between tasks which helps to enhance the exchange between individual physicists, small teams or larger groups. In a broader context, the presented project provides the means to extend the concept of collaboration beyond the sharing of code. Eventually, the resulting increase of transparency and reproducibility paves the way for *analysis preservation*.
\
While started as a private project alongside a $t\bar{t}H$ cross section measurement, the development of all concepts described in this article is being published in the *luigi analysis workflow* project (<https://github.com/riga/law>).
Acknowledgments {#acknowledgments .unnumbered}
===============
This project is supported by the Ministerium für Wissenschaft und Forschung, Nordrhein-Westfalen, the Bundesministerium für Bildung und Forschung (BMBF), the Helmholz Alliance Physics at the Terascale, and the Deutsche Forschungsgemeinschaft (DFG).
References {#references .unnumbered}
==========
[9]{} The Luigi Authors, 2011 - 2017, <https://github.com/spotify/luigi>
Spotify AB, 2012, *Luigi is now open source: build complex pipelines of tasks*, <https://developer.spotify.com/news-stories/2012/09/24/hello-world>
Free Software Foundation, 2016, *GNU Make*, <https://www.gnu.org/software/make>
The GFAL2 Authors, *Grid File Access Library*, <https://dmc.web.cern.ch/projects/gfal-2>
Donno F *et al.*, *Storage resource manager version 2.2: design, implementation, and testing experience*, JPCS, 2008 **119(6)** 062028
Millar P *et al.*, *dCache, agile adoption of storage technology*, Computing in High Energy and Nuclear Physics, 2012, New York City (USA), <https://bib-pubdb2.desy.de/record/140491>
Kourtellis N *et al.*, *Data transfers in the grid: workload analysis of globus GridFTP*, Proceedings of the 2008 International Workshop on Data-aware Distributed Computing, 2008, 29-38, doi:10.1145/1383519.1383523
Dorigo A *et al.*, *XROOTD - A highly scalable architecture for data access*, WSEAS Transactions on Computers, 2005
Boettiger C, *An introduction to Docker for reproducible research*, ACM SIGOPS Operating Systems Review, Special Issue on Repeatability and Sharing of Experimental Artifacts, 2015 **49(1)**, 71-79, doi:10.1145/2723872.2723882, [arXiv:1410.0846v1](https://arxiv.org/abs/1410.0846)
Docker Inc., 2016 - 2017, <http://www.docker.com>
Merkel D, *Docker: lightweight Linux containers for consistent development and deployment*, Linux J., 2014 **239(2)**, [http://dl.acm.org/
citation.cfm?id=2600241](http://dl.acm.org/
citation.cfm?id=2600241)
Matthias K and Kane S P, *Docker: Up and Running: Shipping Reliable Containers in Production*, O’Reilly Media, 2015, ISBN:978-1491917572
Hashimoto M, *Vagrant: Up and Running: Create and Manage Virtualized Development Environments*, O’Reilly Media, 2013, ISBN:978-1449335830
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In this paper, throughput and energy efficiency of secure wireless transmission of delay sensitive data generated by random sources is studied. A fading broadcast model in which the transmitter sends confidential and common messages to two receivers is considered. It is assumed that the common and confidential data, generated from Markovian sources, is stored in buffers prior to transmission, and the transmitter operates under constraints on buffer/delay violation probability. Under such statistical quality of service (QoS) constraints, effective capacity of time-varying wireless transmissions and effective bandwidth of Markovian sources are employed to determine the throughput. In particular, secrecy capacity is used to describe the service rate of buffers containing confidential messages. Moreover, energy per bit is used as the energy efficiency metric and energy efficiency is studied in the low signal-to-noise (SNR) regime. Specifically, minimum energy per bit required for the reliable communication of common and confidential messages is determined and wideband slope expressions are identified. The impact of buffer/delay constraints, correlation between channels, source characteristics/burstiness, channel knowledge at the transmitter, power allocation, and secrecy requirements on the throughput and energy efficiency of common and confidential message transmissions is identified.'
author:
- 'Mustafa Ozmen and M. Cenk Gursoy [^1]'
title: 'Secure Transmission of Delay-Sensitive Data over Wireless Fading Channels'
---
Common and confidential messages, effective bandwidth, effective capacity, energy efficiency, fading broadcast channel, Markovian arrivals, secrecy capacity, statistical buffer/delay QoS constraints, throughput.
Introduction {#sec:intro}
============
Inherent broadcast nature of wireless transmissions results in their susceptibility to eavesdropping, which makes security one of the critical considerations in wireless networks. One way to address the security problem is by exploiting the attributes of the wireless physical layer. From an information-theoretic perspective, Wyner in [@wyner] laid the theoretical foundations of physical-layer security by introducing the wiretap channel wherein the eavesdropper receives a degraded version of the signal received by the legitimate user. In this model, secure communication becomes possible without any shared secret key. As a performance metric, secrecy capacity is defined as the supremum of the achievable communication rates from the transmitter to the legitimate user while the wiretapper is kept ignorant of the information being sent. An extension to a more general wiretap channel model is considered in [@csiszar] where the secrecy capacity is derived for nondegraded broadcast channels with confidential messages.
In addition to security considerations, energy efficiency in wireless systems has also been studied intensively in order to address energy costs and environmental concerns [@survey-Feng]. The importance of energy-efficient operation has further increased with the unprecedented growth both in mobile data traffic and in the number of mobile devices and networks in recent years. Hence, it is imperative to understand the fundamental performance limits in terms of throughput and energy efficiency in order to utilize the energy/power resources that are scarce especially in mobile scenarios.
Moreover, it is also important to note that while establishing secure links and utilizing the limited resources are critical, yet another concern is that certain quality-of-service (QoS) guarantees need to be provided particularly for delay-sensitive data traffic. For instance, in a number of applications such as voice over IP (VoIP), multimedia streaming, interactive video, and online gaming, constraints on delay, packet loss, or buffer overflow probabilities are imposed in the system design so that the end-users experience satisfactory performance levels. In this setting, effective QoS provisioning for delay-sensitive data traffic depends on the accuracy of the source traffic models. For instance, voice traffic can be modeled as an ON/OFF Markov process, and for variable bit-rate video traffic, autoregressive, Markovian, or Markov-modulated processes can be used [@survey-VBRvideotraffic].
With the above-mentioned motivations, our primary goal in this paper is to identify the throughput and energy efficiency of secure wireless transmissions in the presence of statistical QoS requirements of delay-sensitive data traffic generated by random sources.
Literature Overview
-------------------
As noted above, addressing security considerations is essential in wireless communication networks due to the ease in eavesdropping of wireless transmissions. With this motivation, information-theoretic security has been extensively investigated. For instance, in [@liang-broad] and [@boundsecrecy] wiretap channels with fading have been studied whereas authors in [@mimowiretap1] and [@mimowiretap2] incorporated the multiple antenna settings to wiretap channels. Furthermore, the energy efficiency of secure and reliable communication schemes have been addressed in several recent studies. The work in [@securelowsnr] addressed secure communication in the low signal-to-noise ratio (SNR) regime and identified the minimum energy *per secret bit* and the wideband slope (which are two key performance metrics in the low SNR regime [@verdu]). Motivated similarly by energy efficiency requirements, Comaniciu and Poor in [@eesecrecy1] investigated the security-energy tradeoff from an information theoretic perspective. Zhang *et al.* in [@eesecrecy2] studied three-node MIMO wiretap channels in order to design an energy efficient precoder. Ng *et al.* in [@eesecrecy3] considered secure OFDMA systems and addressed the energy efficient resource allocation problem. Kalantari *et al.* in [@eesecrecy4] investigated the power control in wiretap interference channels where users either work together or act as selfish nodes. Similar to our motivation, Chen and Lei in [@eesecrecy5] took energy efficiency, security and QoS guarantees into account jointly and worked on maximizing the secrecy energy efficiency while having constraints on delay. In [@zhu1] and [@zhu2], Zhu *et al.* investigated the cross layer scheduling of OFDMA networks with both open and private data transmissions. In [@fixedsecrecy], two medium-access protocols were proposed and the mean service rate, the source’s data queue and the secret keys queue was analyzed. Shafie and Al-Dhahir [@shafie] proposed a network scheme that consists of a source node and a destination in the presence of buffer aided relay node and an eavesdropper, while taking the data burstiness of source and energy recycling process at the relay into account. In [@shafie2], secure and stable throughput region is investigated by employing beamforming based cooperative jamming that depends on the channel side information available at the transmitter. In [@mao], authors assumed that only the distribution of eavesdropper is known at the transmitter and studied the problem that maximizes the long-term data admission rate while having constraints on the secrecy outage and stability of the data queue. Khalil *et al.* in [@khalil] derived upper and lower bounds on the secrecy capacity of the flat fading channel with limitations on delay. For more details regarding the advances in this rich field of physical-layer security in wireless communications, we refer to surveys and overviews provided in [@liang-foundtrend]–[@Zou].
As a theory to address the delay and other deterministic service guarantees, network calculus has been introduced by Cruz in early 1990s [@cruz_part1], [@cruz_part2]. Thereafter, Chang in [@chang] introduced the theory of effective bandwidth of a time-varying source as a stochastic version of the network calculus [@Changbook], [@ChangZajic]. Effective bandwidth theory identified the performance and resource requirements in the presence of statistical QoS constraints which are imposed as limitations on buffer/delay violation probabilities. Effective bandwidths of various source models have been studied in the literature. For instance, Elwalid and Mitra in [@elwalid] investigated the effective bandwidth of Markovian traffic sources while Markov fluids, Markov-modulated Poisson sources, and general stationary sources were addressed in [@ebw] and [@costasweber].
In addition to time varying source characteristics, the channel characteristics vary with time in wireless communications. Hence, in wireless models, accurate characterization of the throughput rely on the understanding of the queueing system operating with time-varying service and arrival rates. Wu and Negi defined the effective capacity [@dapeng] as the maximum constant arrival rate that can be supported by time-varying transmission rates. Effective capacity is essentially described as a dual concept to effective bandwidth by applying the theory of effective bandwidth to a model with a time-varying channel capacity and regarding the channel service process as a random source with negative rate. Effective capacity has been employed in determining the performance of wireless systems under QoS constraints (see e.g., [@gursoy-Twireless09]–[@itjournal] and references therein). We have analyzed secrecy effective capacity and optimal power control in [@deli-secrecy], and energy efficiency under queueing and secrecy constraints in [@asilomar-ozmen], considering only constant-arrival rates.
Contributions
-------------
As noted above, wireless physical-layer security has recently been intensively studied. On the other hand, energy costs of security, the tradeoff between energy efficiency and secrecy, and analysis of secure wireless transmissions in the presence of random arrivals and delay/buffer constraints have been addressed up to a lesser degree. Indeed, to the best of our knowledge, most recent studies addressed average delay and stability requirements in this context, and not considered statistical queueing constraints, such as buffer overflow and delay violation limitations, which are frequently imposed especially in real-time applications. Motivated by these, we study the secure communication of delay-sensitive data traffic generated from Markovian sources (e.g., discrete-time Markov, Markov fluid, discrete-time and continuous-time Markov modulated Poisson sources) and investigate the fundamental performance limits of secure throughput and energy efficiency under statistical buffer/delay violation constraints. In particular, we can list the contributions of this paper as follows:
- Considering two-state (ON/OFF) Markovian source models, throughput expressions for common and confidential messages in terms of source statistics, effective capacity of wireless transmissions of common and confidential messages, and QoS exponent $\theta$ are provided.
- Energy efficiency metrics, namely the minimum energy per bit and wideband slope, are identified for discrete-time Markov, Markov fluid, and Markov-modulated Poisson arrival models again in terms of important system, channel, and source parameters.
- The effect of source randomness, channel correlation, secrecy requirements, buffer/delay QoS constraints on the performance metrics are identified for both common and confidential messages from both analytical characterizations and numerical results.
- Throughput and energy efficiency metrics are obtained when the transmitter knows the channel statistics but not the realizations of the channel fading, and therefore sends the confidential data at a fixed rate.
Channel Model {#sec:channelmodel}
=============
![Two-receiver broadcast channel model.[]{data-label="fig:chmodel"}](chmodel.eps){width="45.00000%"}
As depicted in Figure \[fig:chmodel\], we consider a fading broadcast channel in which a transmitter sends common and confidential messages to two receivers. Messages are stored in buffers before being transmitted. Specifically, confidential messages intended for receiver 1 and receiver 2 are kept in buffers labeled 1 and 2, respectively, as shown in Fig. \[fig:chmodel\], and common messages are stored in buffer 0. Since delay-sensitive data traffic is considered, statistical queueing constraints are imposed in order to limit buffer overflows and delay violations. We assume flat-fading between the transmitter and receivers. The channel input-output relation can be expressed as $$\begin{aligned}
\begin{array}{ll}
y_j = h_jx + n_j \text{ for } j = 1,2
\end{array} \label{eq:i-o}\end{aligned}$$ where $x$ is the channel input and $y_j$ is the output at the $j^{\text{th}}$ receiver for $j \in \{1,2\}$. Input signal includes both confidential and common messages. Average transmitted signal energy is ${\mathbb{E}}\{|x|^2\} = {\mathcal{E}}$. Moreover, in (\[eq:i-o\]), $h_i$ denotes the fading coefficient in the channel between the transmitter and receiver $j$. Finally, $n_j$ denotes the zero-mean, circularly-symmetric, complex Gaussian background noise at receiver $j$ with variance ${\mathbb{E}}\{|n_j|^2\} = N_{0}$. Hence, the *input signal-to-noise ratio* (SNR) is $$\begin{gathered}
{{\text{\footnotesize{SNR}}}}= \frac{{\mathbb{E}}\{|x|^2\}}{{\mathbb{E}}\{|n_j|^2\}} = \frac{{\mathcal{E}}}{N_{0}} \mbox{ } \quad j=1,2. \label{eq:snr}\end{gathered}$$ While fading coefficients can have arbitrary distributions with finite energies, we assume that block-fading is experienced. Hence, the realizations of the fading coefficients stay fixed for a block of symbols and change independently for the next block.
Preliminaries {#sec:prelim}
=============
In this section, we describe the preliminary concepts employed in our subsequent analysis of secure throughput and energy efficiency in the presence of Markov arrivals and statistical queueing constraints.
Effective Bandwidth of Markovian Arrivals {#subsec:markovarrivals}
-----------------------------------------
Effective bandwidth characterizes the minimum constant transmission rate required to support the given random data arrival process while satisfying statistical queueing constraints on buffer overflows and delay violations. In this paper, we assume that the data to be sent is generated from Markovian sources and, as noted above, is initially stored in a buffer before transmission. Statistical constraints are imposed on the buffer length. In particular, we assume that the buffer overflow probability satisfies $$\label{eq:theta}
\lim_{q \to \infty} \frac{\log \Pr\{Q \ge q\}}{q} = -\theta
$$ where $Q$ denotes the stationary queue length, and $\theta$ is the decay rate of the tail distribution of the queue length. The above limiting formula implies that for large $q_{\max}$, we have $\Pr\{Q \ge q_{\max}\} \approx e^{-\theta q_{\max}}$. A closer approximation is [@dapeng] $$\begin{gathered}
\Pr\{Q \ge q\} \approx \varsigma e^{-\theta q} \label{eq:overflowprob-rev}\end{gathered}$$ where $\varsigma = \Pr\{Q > 0\}$ is the probability of non-empty buffer. Hence, for a sufficiently large threshold, the buffer overflow probability decays exponentially with rate controlled by the QoS exponent $\theta$. Note that as $\theta$ increases, stricter queueing or QoS limitations are imposed.
Conversely, for a given buffer threshold $q$ and overflow probability limit $\epsilon = \Pr\{Q \ge q\}$, the desired value of $\theta$ can be determined as $$\begin{gathered}
\theta = \frac{1}{q}\log_e \frac{\varsigma}{\epsilon}.\end{gathered}$$
In the given setting, the delay violation probability is also characterized to decay exponentially and is approximated by [@Du-Zhang] $$\begin{gathered}
\Pr\{D \ge d\} \approx \varsigma e^{-\theta a^*(\theta) d} \label{eq:delayviolation}\end{gathered}$$ where $D$ is the queueing delay in the buffer at steady state, $d$ is the delay threshold, and $a^*(\theta)$ is the effective bandwidth of the arrival process, described below.
Let $\{a(k), k=1,2,\ldots\}$ be a sequence of nonnegative random variables, describing the random arrival rates. Also let the time-accumulated arrival process be denoted by $A(t) = \sum_{k=1}^{t} a(k)$. Then, the effective bandwidth is given by the asymptotic logarithmic moment generating function of $A(t)$ [@chang], i.e., $$\begin{gathered}
a(\theta)= \lim_{ t \rightarrow \infty}\frac{1}{\theta t}\log {\mathbb{E}}\left\{e^{\theta A(t)}\right\}.\end{gathered}$$
We consider four types of Markovian sources, namely discrete-time Markov source, Markov fluid source, discrete-time Markov-modulated Poisson source and continuous-time Markov-modulated Poisson source. We mainly concentrate on a simple two-state (ON-OFF) model. For these sources, we briefly describe below the effective bandwidth, which characterizes the minimum constant transmission (or service) rate required to support the given time-varying data arrivals while the buffer overflow probability satisfies (\[eq:theta\]).
### Discrete Markov Source {#subsubsec:discreteMarkov}
Data arrival process from this source is modeled as a discrete-time Markov chain. As noted above, we consider a two-state Markov chain in which $r$ bits arrive (i.e., the arrival rate is $r$ bits/block) in the ON state while there are no arrivals in the OFF state. The transition probability matrix $\mathbf{J}$ for this two-state source is $$\begin{aligned}
\mathbf{J}=\left[
\begin{matrix}
p_{11} & p_{12} \\
p_{21} & p_{22}
\end{matrix}\right] \label{eq:probmatrixtwostate}\end{aligned}$$ where $p_{11}$ denotes the probability of staying in the OFF state and $p_{22}$ denotes the probability of staying in the ON state. The probabilities of transitioning from one state to a different one are therefore denoted by $p_{21} = 1 - p_{22}$ and $p_{12} = 1 - p_{11}$. Given the transition probability matrix, the effective bandwidth is formulated as [@chang] $$\begin{aligned}
a(\theta, r) = \frac{1}{\theta} \log_e\!\!\left(\!\!\tfrac{p_{11}+p_{22} e^{r\theta}+\sqrt{ (p_{11}+p_{22}e^{r\theta})^2 - 4(p_{11}+p_{22}-1)e^{r\theta} } }{2}\right). \label{eq:2discreteEBW}
$$
### Markov Fluid Source
Data arrival process from a Markov fluid source is modeled as a continuous-time Markov chain with a generating matrix ${\mathbf{G}}$. The generating matrix for the two-state case is in the form of $$\begin{aligned}
{\mathbf{G}}=\left[
\begin{matrix}
-\alpha & \alpha \\
\beta & -\beta
\end{matrix}\right] \label{eq:generating}\end{aligned}$$ where $\alpha$ and $\beta$ are the transition rates from one state to another. When the arrival rates for the two-state model are $r$ and $0$ and hence we basically have ON and OFF states, the effective bandwidth is given by $$\begin{aligned}
a(\theta)=\frac{1}{2\theta}\left[\theta r -(\alpha+\beta)+\sqrt{(\theta r -(\alpha+\beta))^2+4\alpha\theta r} \right]. \label{eq:2fluidEBW}\end{aligned}$$
### Discrete-Time Markov-Modulated Poisson Source
When data arrivals are modeled as a Poisson process with intensity that is determined by a discrete Markov chain, the source is described as a discrete-time Markov modulated-Poisson process (MMPP). Essentially, discrete-time MMPP is similar to discrete Markov processes but with the difference that the instantaneous arrival rate in each Markov state is Poisson distributed rather than being fixed. Hence, MMPP source has more uncertainty or burstiness comparatively. We again assume that the MMPP source has two states (namely ON and OFF) with different Poisson arrival intensities. In particular, when the source is in the ON state, the Poisson intensity is $r$, while the intensity is zero and hence there are no arrivals in the OFF state. For the Markov chain, we use the same transition probability matrix $\mathbf{J}$ in . Under these assumptions, the effective bandwidth is given in at the top of the next page.
$$\begin{aligned}
a^*(\theta, r) = \frac{1}{\theta} \log_e\!\!\left(\!\!\tfrac{p_{11}+p_{22} e^{r(e^\theta-1)}+\sqrt{ (p_{11}+p_{22}e^{r(e^\theta-1)})^2 - 4(p_{11}+p_{22}-1)e^{r(e^\theta-1)} } }{2}\right) \label{eq:discMMPPEB}
$$
### Continuous-Time Markov-Modulated Poisson Source
In this case, the data arrival rate is again Poisson distributed but with intensity that varies according to a continuous-time Markov chain. We similarly consider a two-state ON-OFF model and assume that the Poisson arrival intensity is $r$ in the ON state whereas there is no arrival in the OFF state. Employing the same generating matrix ${\mathbf{G}}$ as in , the effective bandwidth can be written as $$\begin{aligned}
\hspace{-0.3cm}a(\theta)=&\frac{1}{2\theta}\left[\left(e^\theta-1\right) r -(\alpha+\beta)\right] \nonumber
\\
&+\frac{1}{2\theta}\sqrt{\big[\left(e^\theta-1\right) r -(\alpha+\beta)\big]^2+4\alpha\left(e^\theta-1\right) r} . \label{eq:2MMPPEBW}\end{aligned}$$
Effective Capacity of Wireless Transmissions
--------------------------------------------
Effective capacity provides the maximum constant arrival rate that a given time-varying service process can support while the buffer overflow probability decays exponentially as described in [@dapeng]. Let $\{R[k], k=1,2,\ldots\}$ denote the discrete-time stationary and ergodic stochastic service process and $S[t]\triangleq
\sum_{k=1}^{t}R[k]$ be the time-accumulated process. Then, the effective capacity is given by [@dapeng] $$C_E({{\text{\footnotesize{SNR}}}},\theta)=-\lim_{t\rightarrow\infty}\frac{1}{\theta
t}\log_e{\mathbb{E}\{e^{-\theta S[t]}\}}.$$ We assume that the fading coefficients $\{h_i\}$ change independently from one block to another. Under this assumption, the effective capacity simplifies to $$\label{ec}
C_{E}({{\text{\footnotesize{SNR}}}},\theta)=-\frac{1}{\theta}\log_e\mathbb{E}\{e^{-\theta
R}\},$$ where $R$ is the instantaneous service (or equivalently data transmission) rate. For instance, the maximum service rate in a single-user fading Gaussian channel is given by the instantaneous channel capacity expressed as $$\begin{gathered}
R = \log_2(1 + {{\text{\footnotesize{SNR}}}}z)\end{gathered}$$ where $z = |h|^2$ denotes the fading power and $h$ is the fading coefficient of the channel. The service rates for the transmission of common and confidential messages in the fading broadcast channel addressed in this paper are described below in Section \[subsec:secrecy\].
Throughput and Energy Efficiency Metrics {#subsec:metrics}
----------------------------------------
In this section, we formulate the throughput and energy efficiency metrics for wireless links in the presence of random source arrivals, statistical queueing constraints, and time-varying transmission rates. Specifically, we consider two-state Markovian arrival models (described in Section \[subsec:markovarrivals\]) in which the average arrival rates are $r$ and $0$ in the ON and OFF states, respectively[^2]. Stationary distribution of the Markov chains is denoted by $\boldsymbol \pi = [\pi_1, \pi_2]$ where $\pi_1$ and $\pi_2$ are the probabilities of the OFF and ON states, respectively. Therefore, the source average arrival rate is simply $$\begin{gathered}
r_{{\text{avg}}} =\pi_2r = P_{{\text{ON}}}r \label{eq:avgarrival}\end{gathered}$$ which is equal to the average departure rate when the queue is in steady state [@ChangZajic]. Then, we seek to determine the maximum average arrival rate $r_{\text{avg}}^*$ that can be supported by the fading channel described in Section \[sec:channelmodel\] while satisfying the statistical QoS requirements given in the form in (\[eq:theta\]). As shown in [@ChangZajic Theorem 2.1], if the effective bandwidth of the arrival process is equal to the effective capacity of the service process, i.e., $$\begin{aligned}
a(\theta, r) = C_E({{\text{\footnotesize{SNR}}}}, \theta), \label{eq:equalityforQoS}\end{aligned}$$ then, (\[eq:theta\]) is satisfied, i.e., buffer overflow probability decays exponentially fast with rate controlled by the QoS exponent $\theta$. Hence, we can determine from (\[eq:equalityforQoS\]) the ON-state maximum arrival rate $r^*({{\text{\footnotesize{SNR}}}}, \theta)$ that can be supported by the wireless channel for given ${{\text{\footnotesize{SNR}}}}$ and QoS exponent $\theta$. Then, the maximum average arrival rate (and hence the throughput) is $$\begin{gathered}
r_{\text{avg}}^*({{\text{\footnotesize{SNR}}}}, \theta) = r^*({{\text{\footnotesize{SNR}}}}, \theta) P_{{\text{ON}}}. \label{eq:avgarrivalrate}\end{gathered}$$
In this paper, we employ energy per bit as the performance metric of energy efficiency. In our setup, we define energy per bit as $$\begin{gathered}
\frac{E_b}{N_0} = \frac{{{\text{\footnotesize{SNR}}}}}{ r_{\text{avg}}^*({{\text{\footnotesize{SNR}}}},\theta)}.\end{gathered}$$
In our analysis, following the approach in [@verdu], we study the minimum energy per bit and the wideband slope, which is defined as the slope of the spectral efficiency curve at zero spectral efficiency. The minimum energy per bit $\frac{E_b}{N_0}_{{\text{min}}}$ under QoS constraints can be obtained from [@verdu] $$\label{eq:ebnomin-ra}
\frac{E_b}{N_0}_{{\text{min}}}=\lim_{{{\text{\footnotesize{SNR}}}}\rightarrow0}\frac{{{\text{\footnotesize{SNR}}}}}{r_{\text{avg}}^*({{\text{\footnotesize{SNR}}}},\theta)}=\frac{1}{ \dot{r}_{\text{avg}}^*(0) }.$$ At $\frac{E_b}{N_0}_{{\text{min}}}$, the slope $\mathcal {S}_0$ of the spectral efficiency versus $E_b/N_0$ (in dB) curve can be found from [@verdu] $$\label{eq:widebandslope-ra}
\mathcal{S}_0=-\frac{2\big(\dot{r}_{\text{avg}}^*(0)\big)^2} { \ddot{r}_{\text{avg}}^*(0)}\log_e{2}$$ where $\dot{r}_{\text{avg}}^*(0)$ and $\ddot{r}_{\text{avg}}^*(0)$ are the first and second derivatives, respectively, of the function $r_{\text{avg}}^*({{\text{\footnotesize{SNR}}}},\theta)$ with respect to ${{\text{\footnotesize{SNR}}}}$ at ${{\text{\footnotesize{SNR}}}}= 0$. $\frac{E_b}{N_0}_{{\text{min}}}$ and $\mathcal{S}_0$ provide a linear approximation of the spectral efficiency curve at low spectral efficiencies.
Instantaneous Secrecy Capacity of Confidential Messages and Capacity of Common Message Transmissions {#subsec:secrecy}
----------------------------------------------------------------------------------------------------
In this section, we describe the secrecy capacity in detail in a general case in which the transmitter sends both common and confidential messages[^3] to two receivers, and, with that, we identify the service rates of our queueing model. Confidential and common messages are sent simultaneously and it is assumed that common message is decoded at the receiver in the presence of the interference from the confidential message transmission. Confidential messages of two receivers are sent necessarily using time-division duplexing depending on the channel strengths. More specifically, confidential message is only sent to the receiver with the higher received SNR.
Secrecy capacity quantifies the maximum achievable rates of secure communication. For instance, it is well-known that the secrecy capacity of confidential message transmission with the signal-to-noise ratio denoted by SNR in the presence of an eavesdropper is given by $$\begin{aligned}
R({{\text{\footnotesize{SNR}}}}) = \left[\log_2(1 + {{\text{\footnotesize{SNR}}}}z_m)-\log_2(1 + {{\text{\footnotesize{SNR}}}}z_e)\right]^+ \label{eq:secrecyrate}.\end{aligned}$$ Note that the above formula of secrecy capacity is a generic one with $z_m$ and $z_e$ denoting the magnitude squares of the fading coefficients of channels of the intended user and eavesdropper, respectively. When the transmitter sends separate confidential messages to each user as we have assumed and described in Section \[sec:channelmodel\], the unintended user can be regarded as an eavesdropper.
Having two confidential messages and one common message to send, transmitter allocates its power for the transmission of these messages. We assume that when confidential message intended for receiver $i$ is being sent, $\delta_i$ portion of the power is used for confidential message transmission while $(1-\delta_i)$ portion of the power is used for common message transmission. Additionally, we define the regions for time-division duplexing of confidential messages as $$\begin{aligned}
\Gamma_1&=\left\{(z_1, z_2) \in \mathbb{R}^{2+} : z_1 \geq z_2\right\}, \nonumber
\\
\Gamma_2&=\left\{(z_1, z_2) \in \mathbb{R}^{2+} : z_1 < z_2\right\}. \nonumber\end{aligned}$$ For instance, when we have $(z_1, z_2) \in \Gamma_1$, only confidential message intended for receiver $1$ is transmitted along with the common message[^4]. As previously stated, the common message is decoded in the presence of interference from confidential message transmissions. Both users can decode the common message when it is sent at a rate they both can decode, implying that the common message is sent at the minimum rate that both channels can support. Hence, the instantaneous transmission rate of the common message becomes $$\begin{aligned}
R_0({{\text{\footnotesize{SNR}}}})=& \log_2\left(1+\frac{(1-\delta_1){{\text{\footnotesize{SNR}}}}z_2}{1+\delta_1{{\text{\footnotesize{SNR}}}}z_2}\right) {\mathbf{1}}\left\{\Gamma_1\right\} \nonumber
\\
&+\log_2\left(1+\frac{(1-\delta_2){{\text{\footnotesize{SNR}}}}z_1}{1+\delta_2{{\text{\footnotesize{SNR}}}}z_1}\right) {\mathbf{1}}\left\{\Gamma_2\right\}. \label{eq:rate0}\end{aligned}$$ After subtracting the common message from the received signal, the receiver with the better channel can decode its confidential message without any interference from the common message. Therefore, we can express the instantaneous transmission rate of confidential messages intended for receivers 1 and 2, respectively, as $$\begin{aligned}
R_1({{\text{\footnotesize{SNR}}}})= \log_2\left(\frac{1+\delta_1{{\text{\footnotesize{SNR}}}}z_1}{1+\delta_1{{\text{\footnotesize{SNR}}}}z_2}\right) {\mathbf{1}}\left\{\Gamma_1\right\} \label{eq:rate1}\end{aligned}$$ $$\begin{aligned}
R_2({{\text{\footnotesize{SNR}}}})= \log_2\left(\frac{1+\delta_2{{\text{\footnotesize{SNR}}}}z_2}{1+\delta_2{{\text{\footnotesize{SNR}}}}z_1}\right) {\mathbf{1}}\left\{\Gamma_2\right\} \label{eq:rate2}\end{aligned}$$ where ${\mathbf{1}}\{\cdot\}$ denotes the indicator function[^5].
Throughput of Secure Transmissions with Random Data Arrivals Under QoS Constraints
==================================================================================
In this section, we investigate the throughput of the transmission of confidential and common messages, considering different random source types introduced in Section \[subsec:markovarrivals\]. In order to highlight the impact of random arrivals, we also address the case of a source with a constant arrival rate. For each source type, we characterize the maximum *average* arrival rate as the maximum throughput. Thus, we determine the throughput by deriving the maximum average arrival rate in terms of SNR for both constant-rate arrivals and the four Markovian arrival models.
We note that our initial analysis considers perfect channel side information (CSI) at the transmitter. Hence, we assume that the transmitter knows the realizations of $z_1$ and $z_2$. This is an accurate assumption, for instance, in a cellular scenario in which the base station knows the channel conditions and the users are not malicious but still the confidential messages are to be kept private from the unintended user. We address the case of no CSI subsequently in Section \[section:noCSI\].
### Constant-Rate Source {#subsubsec:dMarkov}
Throughput in the case of constant-rate arrival is given by the effective capacity. For each message, the effective capacity is given by $$\begin{aligned}
C_{Ei}({{\text{\footnotesize{SNR}}}},\theta_i)=-\frac{1}{\theta_i}\log_e\mathbb{E}\{e^{-\theta_i
R_i({{\text{\footnotesize{SNR}}}})}\} \text{ for } i=0,1,2. \label{eq:sc}\end{aligned}$$ Note that for $i = 1 \text{ and } 2$, we have the maximum constant arrival rates of the confidential messages at the transmitter, which are intended for receivers 1 and 2, respectively. For $i = 0$, we have the maximum constant arrival rate of the common message at the transmitter. Note further that the QoS constraint $\theta_i$ of different messages can in general be different. We also define the function ${\text{g}}_i({{\text{\footnotesize{SNR}}}})$ as $$\begin{aligned}
{\text{g}}_i({{\text{\footnotesize{SNR}}}}) = {\mathbb{E}}\left\{e^{-\theta_i R_i({{\text{\footnotesize{SNR}}}})}\right\} = e^{-\theta_i C_{Ei}({{\text{\footnotesize{SNR}}}},\theta_i)}. \label{eq:gdef}\end{aligned}$$ Note that with this definition, we have $$\begin{aligned}
C_{Ei}({{\text{\footnotesize{SNR}}}},\theta_i) = -\frac{1}{\theta_i}\log_e {\text{g}}_i({{\text{\footnotesize{SNR}}}}).\end{aligned}$$ As it will be seen in subsequent subsections, maximum average arrival rates for random sources can also be concisely expressed using the function ${\text{g}}_i({{\text{\footnotesize{SNR}}}})$.
### Discrete Markov Source {#subsubsec:dMarkov}
In this case, we assume that (confidential and/or common) message arrivals to the buffers at the transmitter are according to a discrete-time Markov chain. In the case of ON-OFF discrete Markov source, introducing effective bandwidth expression in (\[eq:2discreteEBW\]) into , and solving for $r$, we can obtain the maximum arrival rate $r^*({{\text{\footnotesize{SNR}}}}, \theta)$ and then express the maximum average arrival rate as a function of the effective capacity $C_E$ as $$\begin{aligned}
r_{{\text{avg}}i}^*({{\text{\footnotesize{SNR}}}}, \theta_i)\!&=\!\frac{P_{{\text{ON}}}}{\theta_i}\!\log_e\!\left(\!\frac{e^{2\theta_i C_{Ei}({\text{\scriptsize{SNR}}}, \theta_i)}\! -\! p_{11}e^{\theta_i C_{Ei}({\text{\scriptsize{SNR}}}, \theta_i)}} {1\!-\!p_{11}\!-\!p_{22}\!+\!p_{22}e^{\theta_i C_{Ei}({\text{\scriptsize{SNR}}}, \theta_i)}}\!\right)\! \nonumber
\\
\!&=\!\frac{P_{\text{ON}}}{\theta_i}\!\log_e\!\left(\!\frac{1 - p_{11}{\text{g}}_i({\text{\scriptsize{SNR}}}) } {(1\!-\!p_{11}\!-\!p_{22}){\text{g}}_i^2({\text{\scriptsize{SNR}}})\! +\!p_{22}{\text{g}}_i({\text{\scriptsize{SNR}}}) }\!\right) \label{eq:2discreteravg}\end{aligned}$$ for $i = 0,1,2$, where ${\text{g}}_i({{\text{\footnotesize{SNR}}}})$ is defined in (\[eq:gdef\]).
Note that the probability of the ON state is given by $
P_{{\text{ON}}}=\frac{1-p_{11}}{2-p_{11}-p_{22}}.
$ If we use the assumption $p_{11}=1-s$ and $p_{22}=s$ (and hence $P_{{\text{ON}}}=s$), the expression for average arrival rate can be simplified further as $$\begin{gathered}
r_{{\text{avg}}i}^*({{\text{\footnotesize{SNR}}}}, \theta_i)=\frac{s}{\theta_i}\log_e\left(\frac{e^{\theta_i C_{Ei}({\text{\scriptsize{SNR}}}, \theta_i)} - (1-s)} {s}\right). \label{eq:ravgdiscreteq}\end{gathered}$$
![Maximum average arrival rate of the confidential message of the first user $r_{{\text{avg}},1}^*$ vs. average signal-to-noise ratio ${{\text{\footnotesize{SNR}}}}$ when $\theta_1 = 1$ and $\delta_1=0.5$.[]{data-label="fig:ravg1SNRdisc"}](ravg1SNR_disc_v2.eps){width="45.00000%"}
It can be easily verified that $r_{{\text{avg}}i}^*$ is a monotonic function of $s$, i.e., as $s$ (and hence ON-state probability $P_{{\text{ON}}}=s$) increases, the maximum average arrival rate increases. We see this effect in Fig. \[fig:ravg1SNRdisc\] where we plot the relationship between maximum average arrival rate of the confidential message of the first user vs. average ${{\text{\footnotesize{SNR}}}}$ curves for different values of $s$ and correlation coefficient $\rho$. We consider a Rayleigh fading environment and assume that the fading powers $z_1$ and $z_2$ are exponentially distributed with unit means, i.e., ${\mathbb{E}}\{z_1\} = {\mathbb{E}}\{z_2\} = 1$, and correlation coefficient $\rho = \frac{\text{cov}(z_1, z_2)}{\sqrt{\text{var}(z_1)\text{var}(z_1)}}$. Numerical evaluation verifies that as $s$ increases, maximum average arrival rate increases for given ${{\text{\footnotesize{SNR}}}}$ and $\rho$. Hence, as the source becomes less bursty, throughput improves. Also, the correlation between the channels of the legitimate user and eavesdropper has an impact on the throughput. Higher correlation values lead to diminished secrecy capacity, which results in smaller throughput values.
![Buffer overflow probability $\Pr\{Q>q\}$ vs. buffer threshold $q$ for both confidential and common messages when $\theta_1=0.5$, $\theta_2=2$, $\theta_0=1$ , $\delta_1=\delta_2=0.7$, $p_{11} = p_{22} = 0.8$ and SNR = $1$. []{data-label="fig:bufferperfCSI"}](buffer_perfCSI.eps){width="45.00000%"}
We have also performed buffer simulations to further verify our theoretical analysis. Initially, we set the values of the QoS exponent $\theta_i$, SNR, source state transition probabilities $p_{11}$ and $p_{22}$ of the ON/OFF discrete Markov source, and determined the maximum average arrival rate the system can support using the theoretical characterizations in this section. We also calculated the corresponding maximum data arrival rate $r_i$ in the ON state. Then, we initiated the simulation by generating the random data arrivals according to the Markov source model, and generating the Gaussian fading coefficients for the service rates. In this process, we have kept track of the buffer length over $10^7$ runs. We have compared the simulated buffer lengths with different thresholds to determine how frequently a threshold is exceeded and identify the overflow probabilities. In Fig. \[fig:bufferperfCSI\], we plot the buffer overflow probability (in logarithmic scale) vs. buffer threshold $q$. We obtain excellent results from these simulations. Specifically, we determined the simulated QoS exponent values $\theta_{\text{sim}}$ from the slopes of the buffer overflow probability curves in the figure[^6]. The simulated $\theta_{\text{sim}}$ values were obtained as $2.0171, 0.9433, 0.5018$ when the corresponding theoretical $\theta$ values were $2, 1, 0.5$, respectively. Hence, if we originally set $\theta = 2$ and design the system accordingly, the buffer overflow probability decays with QoS exponent $\theta_{\text{sim}} = 2.0171$, matching the prediction very well.
### Markov Fluid Source {#subsubsec:fMarkov}
Similarly as in the case of discrete Markov source, for the ON-OFF Markov fluid source, incorporating (\[eq:2fluidEBW\]) into , we determine the maximum average arrival rate as $$\begin{aligned}
r_{{\text{avg}}i}^*({{\text{\footnotesize{SNR}}}}, \theta_i)&=P_{{\text{ON}}} \frac{\theta_i C_{Ei}({\text{\scriptsize{SNR}}}, \theta_i) +\alpha+\beta} {\theta_i C_{Ei}({\text{\scriptsize{SNR}}}, \theta_i) +\alpha} \, C_{Ei}({{\text{\footnotesize{SNR}}}}, \theta_i) \nonumber
\\
&=-\frac{P_{{\text{ON}}}}{\theta_i} \frac{\alpha+\beta-\log_e{\text{g}}_i({{\text{\footnotesize{SNR}}}})} {\alpha -\log_e{\text{g}}_i({{\text{\footnotesize{SNR}}}})} \, \log_e{\text{g}}_i({{\text{\footnotesize{SNR}}}}) \label{eq:2fluidravg}\end{aligned}$$ for $i = 0,1,2$. Note that the probability of ON state is given as $
P_{{\text{ON}}}=\frac{\alpha}{\alpha+\beta}.
$
![Maximum average arrival rate of the confidential message of the second user $r_{{\text{avg}},2}^*$ vs. average signal-to-noise ratio ${{\text{\footnotesize{SNR}}}}$ when $\theta_2 = 1$, $\rho=0.05$ and $\delta_2=0.5$.[]{data-label="fig:ravg2SNRfluid"}](ravg2SNR_fluid.eps){width="45.00000%"}
In Fig. \[fig:ravg2SNRfluid\], we plot the maximum average arrival rate of the confidential message of the second user as a function of average ${{\text{\footnotesize{SNR}}}}$ while considering different channels and Markov fluid sources. Specifically, we assume different pairs of the source state transition rates $\alpha$ and $\beta$ and different expected channel gains ${\mathbb{E}}\{z_2\}=\gamma$. As in Fig. \[fig:ravg1SNRdisc\], we still assume that $z_1$ and $z_2$ are exponentially distributed, and ${\mathbb{E}}\{z_1\} =1$. It is observed that increasing $\alpha$ and decreasing $\beta$ simultaneously increase the ON-state probability $P_{{\text{ON}}}$ and reduce the burstiness of the source, and as a result, throughput increases. Furthermore, better channel conditions for the legitimate user lead to improved throughput due to increase in secrecy capacity.
### Discrete-Time Markov Modulated Poisson Source {#subsubsec:dMMPP}
In order to express the maximum average arrival rate in terms of $C_E$, we again insert the effective bandwidth expression in (\[eq:discMMPPEB\]) into and obtain
$$\begin{gathered}
\hspace{-.4cm}r_{{\text{avg}}i}^*({{\text{\footnotesize{SNR}}}}, \theta_i)\!=\!\frac{P_{{\text{ON}}}} {\left(e^{\theta_i}-1\right)}\!\log_e\!\left(\!\frac{1 - p_{11}{\text{g}}_i({\text{\scriptsize{SNR}}}) } {(1\!-\!p_{11}\!-\!p_{22}){\text{g}}_i^2({\text{\scriptsize{SNR}}})\! +\!p_{22}{\text{g}}_i({\text{\scriptsize{SNR}}}) }\!\right). \label{eq:ravgdMMPP}\end{gathered}$$
### Continuous-Time Markov Modulated Poisson Source
We find the following maximum average arrival rate $r_{\text{avg}}^*$ by incorporating into (\[eq:equalityforQoS\]): $$\begin{gathered}
r_{{\text{avg}}i}^*({{\text{\footnotesize{SNR}}}}, \theta_i)=-\frac{P_{{\text{ON}}}}{\left(e^{\theta_i}-1\right)}\frac{\alpha+\beta-\log_e{\text{g}}_i({{\text{\footnotesize{SNR}}}})} {\alpha -\log_e{\text{g}}_i({{\text{\footnotesize{SNR}}}})} \, \log_e{\text{g}}_i({{\text{\footnotesize{SNR}}}}). \label{eq:2MMPPr}\end{gathered}$$
Energy Efficiency Of Secure Transmissions with Random Data Arrivals Under QoS Constraints
=========================================================================================
In this section, we investigate the energy efficiency of the transmission of confidential and common messages for various source types discussed previously. Using the throughput formulas we have obtained, we analyze the energy efficiency and derive closed-form expressions of the minimum energy per bit and wideband slope.
Minimum Energy per Bit
----------------------
The minimum energy per bit in characterizes the minimum energy needed to send one bit reliably over the wireless fading channel under statistical queueing constraints. Lower minimum energy per bit levels indicate higher energy efficiency. First, we formulate the minimum energy per bit for the confidential messages as $$\begin{aligned}
\frac{E_b}{N_0}_{{\text{min}},i}=\lim_{{{\text{\footnotesize{SNR}}}}\rightarrow0} \frac{\delta_i \Pr(\Gamma_i) {{\text{\footnotesize{SNR}}}}}{r_{{\text{avg}}i}^*({{\text{\footnotesize{SNR}}}},\theta_i)} = \frac{\delta_i \Pr(\Gamma_i) }{ \dot{r}_{{\text{avg}}i}^*(0) }\label{eq:ebnomin_conf}\end{aligned}$$ for $i=1,2$. Similarly for the common message, the minimum energy per bit becomes $$\begin{aligned}
\frac{E_b}{N_0}_{{\text{min}},0}&=\lim_{{{\text{\footnotesize{SNR}}}}\rightarrow0} \frac{\left[(1-\delta_1) \Pr(\Gamma_1)+(1-\delta_2) \Pr(\Gamma_2) \right]{{\text{\footnotesize{SNR}}}}}{r_{{\text{avg}}0}^*({{\text{\footnotesize{SNR}}}},\theta_0)} \nonumber
\\&= \frac{(1-\delta_1) \Pr(\Gamma_1)+(1-\delta_2) \Pr(\Gamma_2) }{ \dot{r}_{{\text{avg}}0}^*(0) }.\label{eq:ebnomin_common}\end{aligned}$$
Below, we initially characterize the minimum energy per bit for the case of constant-rate arrivals, and subsequently show that the same minimum energy per bit levels are achieved when discrete-time Markov and Markov ON-OFF sources are considered.
\[prop:ebnocons\] When the data arrival rate is constant, the minimum energy per bit expressions for the confidential message transmissions to receivers 1 and 2 under QoS constraints are given, respectively, by $$\begin{aligned}
{\frac{E_b}{N_0}}_{\text{min},1}= \frac{\Pr(\Gamma_1)\log_e2}{{\mathbb{E}}_{\Gamma_1}\!\!\left\{z_1- z_2\right\}} \label{eq:EbN01}\end{aligned}$$ $$\begin{aligned}
{\frac{E_b}{N_0}}_{\text{min},2}= \frac{\Pr(\Gamma_2)\log_e2}{{\mathbb{E}}_{\Gamma_2}\left\{ z_2-z_1\right\}} \label{eq:EbN02},\end{aligned}$$ and the minimum energy per bit for the common message transmission under QoS constraints is given by $$\begin{aligned}
{\frac{E_b}{N_0}}_{\text{min},0}= \frac{[(1-\delta_1)\Pr(\Gamma_1)+(1-\delta_2)\Pr(\Gamma_2)]\log_e2} {(1-\delta_1){\mathbb{E}}_{\Gamma_1}\!\!\left\{ z_2 \right\}+(1-\delta_2){\mathbb{E}}_{\Gamma_2}\!\!\left\{z_1 \right\}} \label{eq:EbN00}\end{aligned}$$ where $\Pr(\Gamma_1)=\Pr(z_1 < z_2)$, $\Pr(\Gamma_2)=\Pr(z_1 < z_2)$, and $\delta_i$ is fraction of the power used for the transmission of the confidential message to receiver $i$. Moreover, ${\mathbb{E}}_{\Gamma_1}$ denotes the expectation in region $\Gamma_1$ while ${\mathbb{E}}_{\Gamma_2}$ is similarly defined in the complement region $\Gamma_2$.
*Proof:* See Appendix \[subsec:ebnocons\].
When $z_1$ and $z_2$ are independent and exponentially distributed with ${\mathbb{E}}\{z_1\}=1$ and ${\mathbb{E}}\{z_2\}=\gamma$, we have $\Pr(\Gamma_1)=\frac{1}{\gamma+1}$ and $\Pr(\Gamma_2)=\frac{\gamma}{\gamma+1}$, and we can get closed-form expressions for the minimum energy per bit formulations as follows: $$\begin{gathered}
{\frac{E_b}{N_0}}_{\text{min},1}=\log_e2, \,\,\,\,\, {\frac{E_b}{N_0}}_{\text{min},2}=\frac{\log_e2}{\gamma} \label{eq:ebnomin-loge2}
\\
{\frac{E_b}{N_0}}_{\text{min},0}=\frac{\gamma+1}{\gamma}\log_e2.\end{gathered}$$
Interestingly, for both ON-OFF discrete-time Markov and Markov fluid sources, minimum energy per bit expressions are the same as those attained in the presence of constant-rate sources.
\[prop:ebno\_discrete\] When data arrivals are modeled as ON-OFF discrete-time Markov or Markov fluid processes, the minimum energy per bit expressions for confidential and common message transmissions under QoS constraints remains the same as those for the constant arrival rate model and hence are given by , , and , respectively.
*Proof:* See Appendix \[subsec:ebno\_discrete\].
Heretofore, we have seen that the minimum bit energy expressions do not depend on either the queueing constraints or the source randomness. More specifically, minimum bit energy of confidential/common message transmissions are the same regardless of the value of the QoS exponent $\theta$ and whether data arrives at a constant rate or according to an ON-OFF Markov process. However, this is not the case when we consider more bursty Markov-modulated Poisson arrivals, as shown in the result below.
\[prop:ebno\_MMPP\] When the source arrivals are modeled as ON-OFF discrete-time or continuous-time MMPPs, the minimum energy per bit expressions for confidential and common message transmissions under QoS constraints are given, respectively, by $$\begin{aligned}
{\frac{E_b}{N_0}}_{\text{min},1}= \frac{(e^{\theta_1}-1)\Pr(\Gamma_1)\log_e2}{\theta_1{\mathbb{E}}_{\Gamma_1}\!\!\left\{z_1- z_2\right\}} \label{eq:EbN01MMPP}\end{aligned}$$ $$\begin{aligned}
{\frac{E_b}{N_0}}_{\text{min},2}= \frac{(e^{\theta_2}-1)\Pr(\Gamma_2)\log_e2}{\theta_2{\mathbb{E}}_{\Gamma_2}\left\{ z_2-z_1\right\}} \label{eq:EbN02MMPP}\end{aligned}$$ $$\begin{aligned}
{\frac{E_b}{N_0}}_{\text{min},0}= \frac{(e^{\theta_0}\!-1)[(1-\delta_1)\Pr(\Gamma_1)+(1-\delta_2)\Pr(\Gamma_2)]\log_e2} {\theta_0\left[(1-\delta_1){\mathbb{E}}_{\Gamma_1}\!\!\left\{ z_2 \right\}+(1-\delta_2){\mathbb{E}}_{\Gamma_2}\!\!\left\{z_1 \right\}\right]}. \label{eq:EbN00MMPP}\end{aligned}$$
*Proof:* See Appendix \[subsec:ebno\_MMPP\].
For MMPP sources, minimum energy per bit now depends on the QoS exponent through the term $\frac{e^\theta-1}{\theta}$. Since $\frac{e^\theta-1}{\theta} > 1$ for $\theta > 0$ and increases with increasing $\theta$, a higher energy per bit is required for MMPP sources (compared to constant-rate and ON-OFF Markov sources) and energy cost grows as the QoS constraints become more stringent. Interestingly, energy per bit expressions still do not depend on the specific parameters of the random arrival model (such as transition probabilities/rates of the Markov chain and intensity of the Poisson arrivals).
As also noted before, Proposition \[prop:ebno\_discrete\] shows that the minimum energy per bit for discrete-time Markov and Markov fluid sources are the same as for the constant-rate source. The primary intuitive reasoning behind this result is that the minimum energy per bit is an asymptotic performance metric achieved as ${{\text{\footnotesize{SNR}}}}\to 0$, and the impact of source burstiness significantly diminishes at these asymptotically low ${{\text{\footnotesize{SNR}}}}$ levels for discrete-time Markov and Markov fluid sources. Specifically, as ${{\text{\footnotesize{SNR}}}}$ diminishes, the fixed arrival rate (in the ON-state of the Markov models) that can be supported by the wireless channel decreases as well, resulting in less and less impact on buffer overflows and delay violations.
On the other hand, if the arrival process is MMPP, the intensity of the Poisson process is reduced with decreasing ${{\text{\footnotesize{SNR}}}}$. However, the arrival process is still a Poisson process but with a smaller intensity, meaning that there is still a probability, however small, for the instantaneous arrival rate in the ON state to be large since the arrival rate depends on the realization of a Poisson distributed random variable. Hence, MMPP source is more bursty in the low-${{\text{\footnotesize{SNR}}}}$ regime than discrete-time Markov and Markov fluid sources, and this is reflected in the larger minimum energy per bit values as shown in the results of Proposition \[prop:ebno\_MMPP\].
Wideband Slope
--------------
Minimum energy per bit ${\frac{E_b}{N_0}}_{\text{min}}$ is the ultimate performance limit of energy-efficient operation. At the same time, it is an asymptotic performance metric achieved in the limit as ${{\text{\footnotesize{SNR}}}}$ vanishes. In this subsection, we complement the ${\frac{E_b}{N_0}}_{\text{min}}\!\!\!\!-$analysis by characterizing the wideband slope of confidential and common message transmissions for different source models. Unlike the minimum energy per bit, wideband slope is distinct for each source and depends on the source statistics. In this subsection, we also provide numerical results to demonstrate the effectiveness of the linear approximation of the throughput in the low-${{\text{\footnotesize{SNR}}}}$ regime in terms of ${\frac{E_b}{N_0}}_{\text{min}}$ and wideband slope $\mathcal{S}_0$, and to identify the impact of secrecy requirements, source randomness, QoS constraints, and channel correlation on energy efficiency.
### Constant-Rate Sources
\[prop:widebandslope\] For constant-rate arrivals, the wideband slope expressions for common and confidential message transmissions under QoS constraint are given by $$\begin{aligned}
\mathcal{S}_{0,i}\!=\! \frac{2\left({\mathbb{E}}\left\{ \dot{f}_i(0)\right\}\right)^2} { \frac{\theta_i}{\log_e\!2}\text{var}\!\left(\dot{f}_i(0)\right) \!- {\mathbb{E}}\!\left\{ \ddot{f}_i(0)\right\}} \label{eq:S0i}\end{aligned}$$ for $i=0,1,2$ where we have defined $f_i({{\text{\footnotesize{SNR}}}})=R_i({{\text{\footnotesize{SNR}}}}) \log_e2$ with $R_i({{\text{\footnotesize{SNR}}}})$ being the instantaneous rate of confidential or common message given in (\[eq:rate0\])–(\[eq:rate2\]), and the first and second derivatives of $f_i({{\text{\footnotesize{SNR}}}})$ at ${{\text{\footnotesize{SNR}}}}= 0$ are given by $$\begin{aligned}
\dot{f}_1(0)&= \delta_1 \left(z_1 - z_2\right) {\mathbf{1}}\!\left\{z_1\!\geq \!z_2\right\}, \nonumber
\\
\dot{f}_2(0)&= \delta_2 \left( z_2-z_1 \right) {\mathbf{1}}\!\left\{z_1\!< \!z_2\right\}, \nonumber
\\
\dot{f}_0(0)&= (1-\delta_1) z_2{\mathbf{1}}\!\left\{z_1\!\geq \!z_2\right\}+ (1-\delta_2) z_1{\mathbf{1}}\!\left\{z_1\!< \!z_2\right\}, \nonumber
\\
\ddot{f}_1(0)&= -\delta_1^2 \left[z_1^2 - z_2^2\right] {\mathbf{1}}\!\left\{z_1\!\geq \!z_2\right\}, \nonumber
\\
\ddot{f}_2(0)&= -\delta_2^2 \left[ z_2^2-z_1^2 \right]{\mathbf{1}}\!\left\{z_1\!< \!z_2\right\}, \nonumber
\\
\ddot{f}_0(0)&= -(1-\delta_1^2)z_2^2{\mathbf{1}}\!\left\{z_1\!\geq \!z_2\right\}-(1-\delta_2^2)z_1^2 {\mathbf{1}}\!\left\{z_1\!< \!z_2\right\}. \label{eq:f0}\end{aligned}$$
*Proof:* See Appendix \[subsec:widebandslope\].
Above, $\mathcal{S}_{0,0}$ is the wideband slope for common message transmission while $\mathcal{S}_{0,1}$ and $\mathcal{S}_{0,2}$ denote the wideband slope of confidential message transmissions to receivers 1 and 2, respectively.
For independent and exponentially distributed $z_1$ and $z_2$ with ${\mathbb{E}}\{z_1\}=1$ and ${\mathbb{E}}\{z_2\}=\gamma$, the wideband slope expressions simplify to $$\begin{aligned}
\mathcal{S}_{0,1}\!=\! \frac{2} { \frac{\theta_1}{\log_e\!2}\left(1+2\gamma\right) +4\gamma+2} \label{eq:S01}
\\
\mathcal{S}_{0,2}\!=\! \frac{2} { \frac{\theta_2}{\log_e\!2}\left(1+\frac{2}{\gamma}\right) +\frac{4}{\gamma}+2}. \label{eq:S02}\end{aligned}$$ If we further assume that $\delta_1=\delta_2=\delta$, then the wideband slope for common message becomes $$\begin{aligned}
\mathcal{S}_{0,0}\!=\! \frac{2} { \frac{\theta_0}{\log_e\!2}+ \frac{1-\delta^2}{(1-\delta)^2}}. \label{eq:S00}\end{aligned}$$
### Discrete-Time Markov Sources
Next, we consider ON-OFF discrete-time Markov sources with transition probabilities denoted by $p_{ij}$ for $i,j \in \{1,2\}$.
\[prop:wbsdisc\] The wideband slope expressions for confidential and common message transmissions under QoS constraint are given by $$\begin{aligned}
\mathcal{S}_{0,i}\!=\! \frac{2\left({\mathbb{E}}\left\{ \dot{f}_i(0)\right\}\right)^2} {\eta\frac{\theta_i}{\log_e\!2}\!\left({\mathbb{E}}\!\left\{ \dot{f}_i(0)\right\}\right)^2\!\!\! + \! \frac{\theta_i}{\log_e\!2}\text{var}\!\left(\dot{f}_i(0)\right) \!- {\mathbb{E}}\!\left\{ \ddot{f}_i(0)\right\}} \label{eq:S0idisc}\end{aligned}$$ for $i=0,1,2$, where $\dot{f}_i(0)$ and $\ddot{f}_i(0)$ are given in (\[eq:f0\]). Additionally, $\eta$ above is defined as $$\eta = \frac{(1-p_{22})(p_{11}+p_{22})}{(1-p_{11})(2-p_{11}-p_{22})}. \label{eq:eta}$$
*Proof:* See Appendix \[subsec:wbsdisc\]
Again, for independent and exponentially distributed $z_1$ and $z_2$ with ${\mathbb{E}}\{z_1\}=1$ and ${\mathbb{E}}\{z_2\}=\gamma$, the wideband slope expressions are given as $$\begin{aligned}
\mathcal{S}_{0,1}\!=\! \frac{2} { \frac{\theta_1\eta}{\log_e\!2}+\frac{\theta_1}{\log_e\!2}\left(1+2\gamma\right) +4\gamma+2}, \label{eq:S01disc}
\\
\mathcal{S}_{0,2}\!=\! \frac{2} { \frac{\theta_2\eta}{\log_e\!2}+\frac{\theta_2}{\log_e\!2}\left(1+\frac{2}{\gamma}\right) +\frac{4}{\gamma}+2}. \label{eq:S02disc}\end{aligned}$$ If we further assume that $\delta_1=\delta_2=\delta$, then the wideband slope for common message becomes $$\begin{aligned}
\mathcal{S}_{0,0}\!=\! \frac{2} { \frac{\theta_0\eta}{\log_e\!2}+\frac{\theta_0}{\log_e\!2}+ \frac{1-\delta^2}{(1-\delta)^2}}. \label{eq:S00disc}\end{aligned}$$ When compared with the corresponding wideband slope expressions in – for the constant-rate source, we notice that wideband slope formulas above in (\[eq:S01disc\])–(\[eq:S00disc\]) for the discrete Markov source differ only due to the presence of the term $\frac{\theta \eta}{\log_e2}$, which reflects essentially the source randomness with the parameter $\eta$. This additional term leads to smaller wideband slopes, indicating the detrimental impact of source randomness on energy efficiency. Note also that when $p_{11} = 0$ and $p_{22} = 1$, discrete Markov essentially becomes a constant-rate source and we have $\eta = 0$.
![Maximum average arrival rate of first user’s confidential message $r_{{\text{avg}}1}^*$ vs. energy per bit $\frac{E_b}{N_0}$ in dB when $\theta_1 = 1$, $\rho=0.05$ and $\delta_1=0.5$.[]{data-label="fig:ravg1EbN0disc"}](ravg1EbNo_disc_v2.eps){width="45.00000%"}
In Fig. \[fig:ravg1EbN0disc\], the maximum average arrival rate of the confidential message for the first user vs. energy per bit is plotted. We consider an ON-OFF discrete Markov source with $p_{11}=1-s$ and $p_{22}=s$ (and hence $P_{{\text{ON}}}=s$). We assume $\theta=1$ and $\delta_1=0.5$. The channel power gains $z_1$ and $z_2$ are exponentially distributed with ${\mathbb{E}}\{z_1\}=1$, ${\mathbb{E}}\{z_2\}=\gamma$ and correlation coefficient $\rho=0.05$. As predicted, the minimum energy per bit does not depend on source burstiness or the second user channel statistics, i.e., $\gamma$. There is a slight increase in the minimum energy per bit values achieved in the cases of secrecy as compared to no secrecy. The main reason for this is the correlation in the channel conditions of the two users. Without any correlation, the minimum energy per bit becomes equal to $-1.59$ dB. As a result of similar minimum energy per bit values, wideband slope becomes a critical performance indicator in the low-SNR regime. We notice that wideband slope diminishes when secrecy requirements are imposed and also when source burstiness increases with diminishing ON-state probability $P_{{\text{ON}}}=s$. We also observe that, as the second user (or equivalently eavesdropper) channel conditions improve, i.e., as $\gamma$ increases, we have smaller wideband slopes.
![Maximum average arrival rate of second user’s confidential message $r_{{\text{avg}}2}^*$ vs. energy per bit $\frac{E_b}{N_0}$ in dB when $\theta_2 = 1$, $\rho=0.05$ and $\delta_2=0.5$.[]{data-label="fig:ravg2EbN0disc"}](ravg2EbNo_disc_v2.eps){width="45.00000%"}
In Fig. \[fig:ravg2EbN0disc\], the maximum average arrival rate of the confidential messages for the second user vs. energy per bit is plotted. Similarly as before, we set $\theta=1$, $\rho=0.05$ and $\delta_2=0.5$. Again, the minimum energy per bit does not depend on source burstiness. On the other hand, we observe that wideband slope increases as source becomes less bursty, i.e., as $q$ increases. Also, better channel conditions for the legitimate user (i.e., larger $\gamma$) increases the energy efficiency.
![Maximum average arrival rate of common message $r_{{\text{avg}}0}^*$ vs. energy per bit $\frac{E_b}{N_0}$ in dB when $\theta_0 = 1$, $\rho=0.05$ and $\delta_1=\delta_2=0.5$.[]{data-label="fig:ravg0EbN0disc"}](ravg0EbNo_disc_v2.eps){width="45.00000%"}
We illustrate the spectral efficiency curve for the common message in Fig. \[fig:ravg0EbN0disc\], assuming the parameter setting $\theta = 1$, $\rho=0.05$ and $\delta_1=\delta_2=0.5$. We again verify that source characteristics do not play a role in the value of the minimum energy per bit. Better channel conditions for the second user improve the overall energy efficiency of the transmission of the common message by improving the minimum energy per bit. We also notice that wideband slope is the same when we alter the channel conditions. However, source burstiness has a negative impact on the wideband slope, thus, on the energy efficiency as well.
### Markov Fluid Sources
In the following, we characterize the wideband slope in the case of ON-OFF Markov fluid arrivals with transition rates $\alpha$ and $\beta$.
\[prop:wbsfluid\] The wideband slope expressions for confidential and common message transmissions under QoS constraint are given by $$\begin{aligned}
\mathcal{S}_{0,i}\!=\! \frac{2\left({\mathbb{E}}\left\{ \dot{f}_i(0)\right\}\right)^2} {\zeta\frac{\theta_i}{\log_e\!2}\!\left({\mathbb{E}}\!\left\{ \dot{f}_i(0)\right\}\right)^2\!\!\! + \! \frac{\theta_i}{\log_e\!2}\text{var}\!\left(\dot{f}_i(0)\right) \!+ {\mathbb{E}}\!\left\{ \ddot{f}_i(0)\right\}} \label{eq:S0ifluid}\end{aligned}$$ for $i=0,1,2$, where $\dot{f}_i(0)$ and $\ddot{f}_i(0)$ are defined in . Note further that $\zeta$ is defined as $$\zeta=\frac{2\beta}{\alpha(\alpha+\beta)}. \label{eq:zeta}$$
*Proof:* See Appendix \[subsec:wbsfluid\].
Similarly as for the previous arrival models, we can simplify the wideband expressions for the confidential message transmissions to the following when we have independent and exponentially distributed $z_1$ and $z_2$ with ${\mathbb{E}}\{z_1\}=1$ and ${\mathbb{E}}\{z_2\}=\gamma$: $$\begin{aligned}
\mathcal{S}_{0,1}\!=\! \frac{2} { \frac{\theta_1\zeta}{\log_e\!2}+\frac{\theta_1}{\log_e\!2}\left(1+2\gamma\right) +4\gamma+2}, \label{eq:S01fluid}
\\
\mathcal{S}_{0,2}\!=\! \frac{2} { \frac{\theta_2\zeta}{\log_e\!2}+\frac{\theta_2}{\log_e\!2}\left(1+\frac{2}{\gamma}\right) +\frac{4}{\gamma}+2}. \label{eq:S02fluid}\end{aligned}$$ If we further assume that $\delta_1=\delta_2=\delta$, then the wideband slope for common message becomes $$\begin{aligned}
\mathcal{S}_{0,0}\!=\! \frac{2} { \frac{\theta_0\zeta}{\log_e\!2}+\frac{\theta_0}{\log_e\!2}+ \frac{1-\delta^2}{(1-\delta)^2}}. \label{eq:S00fluid}\end{aligned}$$ The common theme in the above expressions and the ones corresponding to other source types (i.e., expressions in (\[eq:S01\])–(\[eq:S00\]) and (\[eq:S01disc\])–(\[eq:S00disc\])) is that wideband slope expressions depend on three critical factors: QoS exponent $\theta$, source burstiness parameter ($\zeta$ in the case of Markov fluid source and $\eta$ in the case discrete Markov source, which both become zero when the arrival rate is constant), and channel statistics through ${\mathbb{E}}\{z_2\} = \gamma$. For instance, wideband slopes diminish as $\theta$ increases and more stringent buffer/delay constraints are imposed.
We depict, in Fig. \[fig:ravg1EbN0fluid\], the maximum average arrival rate of the confidential message for the first user vs. energy per bit for Markov fluid sources with different values of $\alpha$ and $\beta$. We assume $\theta = 1$, $\gamma=1$ and $\delta_1=0.5$. In the case of no secrecy, the minimum energy per bit is equal to $-1.59$ dB and it remains unchanged under different source characteristics. With secrecy, source burstiness again does not impact the minimum energy per bit. However, as channel correlation increases, the energy efficiency degrades due to higher minimum energy per bit. Additionally, the source characteristics have significant impact on the wideband slope e.g. wideband slope decreases as source becomes more bursty (i.e., as we change the state transition rates from $\alpha=9$ and $\beta=1$ to $\alpha=1$ and $\beta=9$).
![Maximum average arrival rate of first user’s confidential message $r_{{\text{avg}}1}^*$ vs. energy per bit $\frac{E_b}{N_0}$ when $\theta_1 = 1$, $\gamma=1$ and $\delta_1=0.5$.[]{data-label="fig:ravg1EbN0fluid"}](ravg1EbNo_fluid.eps){width="45.00000%"}
In Fig. \[fig:ravg0EbN0fluid\], the maximum average arrival rate of the common messages for the first user vs. energy per bit for the source with Markov fluid characteristics is illustrated. We assume $\theta = 1$, $\gamma=1$ and $\delta_1=\delta_2=0.5$. The minimum energy per bit only changes for the different values of correlation. Interestingly, in this case the energy efficiency gets better with increasing correlation. The intuition behind this is that the common message throughput is limited by the worst of the channels of the first and second users. As correlation increases, the discrepancy between the conditions of the channels is reduced, improving the throughput. Additionally, wideband slope is higher for less bursty systems.
![Maximum average arrival rate of common message $r_{{\text{avg}}0}^*$ vs. energy per bit $\frac{E_b}{N_0}$ when $\theta_0 = 1$, $\gamma=1$ and $\delta_1=\delta_2=0.5$.[]{data-label="fig:ravg0EbN0fluid"}](ravg0EbNo_fluid.eps){width="45.00000%"}
### Discrete-Time MMPP Sources
Next, we address ON-OFF discrete-time MMPP sources.
\[prop:wbsMMPPdisc\] The wideband slope expressions for confidential and common message transmissions under QoS constraint are given by $$\begin{aligned}
\mathcal{S}_{0,i}\!=\! \frac{\frac{2\theta_i}{e^{\theta_i}-1}\left({\mathbb{E}}\left\{ \dot{f}_i(0)\right\}\right)^2} {\eta\frac{\theta_i}{\log_e\!2}\!\left({\mathbb{E}}\!\left\{ \dot{f}_i(0)\right\}\right)^2\!\!\! + \! \frac{\theta_i}{\log_e\!2}\text{var}\!\left(\dot{f}_i(0)\right) \!+ {\mathbb{E}}\!\left\{ \ddot{f}_i(0)\right\}}. \label{eq:S0i}\end{aligned}$$ for $i=0,1,2$ where $\dot{f}_i(0)$ and $\ddot{f}_i(0)$ are defined in and $\eta$ is defined in .
We omit the proof as it is rather straightforward due to the relationship between the throughputs of the discrete Markov source and the discrete MMPP source.
![Maximum average arrival rate of common message $r_{{\text{avg}}0}^*$ vs. energy per bit $\frac{E_b}{N_0}$ in dB when $\rho=0.8$, $\gamma=1$, $p_{11}=0.1$ and $p_{22}=0.9$.[]{data-label="fig:ravg0EbN0MMPPdisc"}](ravg0EbNo_MMPPdisc.eps){width="45.00000%"}
In Fig. \[fig:ravg0EbN0MMPPdisc\], we illustrate the maximum average arrival rate of the common message vs. energy per bit when the source is ON-OFF discrete-time MMPP. We set $\rho=0.8$, $\gamma=1$, $p_{11}=0.1$ and $p_{22}=0.9$, and study the impact of different values of $\theta_0$ and $\delta_i$. For the MMPP source, the minimum energy per bit depends on the QoS exponent $\theta_0$ and it will improve when $\theta_0$ decreases, indicating less stringent queueing constraints. Power allocation has no impact on the minimum energy per bit. However, with more power allocated to the common message, the wideband slope becomes higher.
### Continuous-Time MMPP Sources
Finally, we consider continuous-time MMPP sources.
\[prop:wbsMMPPfluid\] The wideband slope expressions for confidential and common message transmissions under QoS constraint are given by $$\begin{aligned}
\mathcal{S}_{0,i}\!=\! \frac{\frac{2\theta_i}{e^{\theta_i}-1}\left({\mathbb{E}}\left\{ \dot{f}_i(0)\right\}\right)^2} {\zeta\frac{\theta_i}{\log_e\!2}\!\left({\mathbb{E}}\!\left\{ \dot{f}_i(0)\right\}\right)^2\!\!\! + \! \frac{\theta_i}{\log_e\!2}\text{var}\!\left(\dot{f}_i(0)\right) \!+ {\mathbb{E}}\!\left\{ \ddot{f}_i(0)\right\}} \label{eq:S0i}\end{aligned}$$ for $i=0,1,2$ where $\dot{f}_i(0)$ and $\ddot{f}_i(0)$ are defined in and $\zeta$ is defined in .
We again omit the proof due to the fact that the result readily follows from the relationship between the throughputs of Markov fluid and fluid MMPP sources.
![Maximum average arrival rate of common message $r_{{\text{avg}}0}^*$ vs. energy per bit $\frac{E_b}{N_0}$ in dB when $\rho=0$, $\gamma=1$, $\alpha=9$ and $\beta=1$.[]{data-label="fig:ravg0EbN0MMPPfluid"}](ravg0EbNo_MMPPfluid.eps){width="45.00000%"}
In Fig. \[fig:ravg0EbN0MMPPfluid\], we illustrate the maximum average arrival rate of the common message vs. energy per bit when the source is ON-OFF continuous-time MMPP. We assume $\rho=0$, $\gamma=1$, $\alpha=9$ and $\beta=1$. We again study the impact of $\theta_0$ and $\delta_i$ on energy efficiency similarly as in Fig. \[fig:ravg0EbN0MMPPdisc\] but the main difference is that there is no correlation in Fig. \[fig:ravg0EbN0MMPPfluid\]. As in the previous discussion, the minimum energy per bit improves with lower $\theta_0$ values, and increasing the power allocation on common message transmission increases the wideband slope, and thus improves the energy efficiency. Interestingly, when compared with Fig. \[fig:ravg0EbN0MMPPdisc\], we notice that having no correlation between the channels of the two users hurts the energy efficiency of the common message transmission as it increases the minimum energy per bit significantly.
Throughput and Energy Efficiency with no Channel Knowledge at the Transmitter {#section:noCSI}
=============================================================================
In this section, we depart from the perfect transmitter CSI assumption of the previous sections and consider a scenario in which the transmitter has no CSI. Specifically, we assume that the transmitter does not know the realizations of the channel fading coefficients, which is relevant in cases in which the eavesdropper is passive and malicious. This also represents a worst-case scenario due to the fact that even the legitimate channel is not known. Treating the eavesdropper as malicious, we address a special case of the previously treated system model. In particular, we do not consider common message transmission and assume that the transmitter just intends to send confidential messages to receiver 1 while keeping them private from receiver 2 (which is regarded as the eavesdropper).
Not knowing the realizations of the channel fading coefficients $h_1$ and $h_2$, the transmitter sends the data at the fixed rate of $\lambda$ bits/s/Hz. As before, instantaneous secrecy capacity $R({{\text{\footnotesize{SNR}}}}) =\left[\log_2(1 + {{\text{\footnotesize{SNR}}}}z_1)-\log_2(1 + {{\text{\footnotesize{SNR}}}}z_2)\right]^+$ quantifies the maximum achievable rates of secure communication where $z_i = |h_i|^2$. Hence, if $\lambda \le R({{\text{\footnotesize{SNR}}}})$, then reliable and secure communication is attained and therefore the transmitted message is decoded correctly while eavesdropper is being kept ignorant of the message. If, on the other hand, $\lambda > R({{\text{\footnotesize{SNR}}}})$, secrecy outage occurs. Under these assumptions, the wireless link can be modeled as a two-state discrete-time Markov chain. Specifically, the channel is assumed to be in the ON state if $\lambda \le R({{\text{\footnotesize{SNR}}}})$, while the channel is in the OFF state when $\lambda > R({{\text{\footnotesize{SNR}}}})$. The steady-state probability for the ON state can be easily obtained as $$\begin{aligned}
{{\text{P}}}\{\Gamma\}&={{\text{P}}}\{R({{\text{\footnotesize{SNR}}}})>\lambda\}={{\text{P}}}\left\{z_1 > 2^\lambda z_2+ \frac{2^\lambda-1}{{{\text{\footnotesize{SNR}}}}}\right\}
\\
&=\int_0^\infty \int_{2^\lambda z_2+ \frac{2^\lambda-1}{{\text{\scriptsize{SNR}}}}}^\infty p(z_1, z_2)dz_1 dz_2 \label{eq:probgamma}\end{aligned}$$ where we define $\Gamma=\left\{(z_1, z_2) \in \mathbb{R}^+ : \lambda < R({{\text{\footnotesize{SNR}}}})\right\}$.
Effective Capacity with no Channel Knowledge at the Transmitter
---------------------------------------------------------------
In [@Changbook Chap. 7, Example 7.2.7], it is shown for Markov modulated processes that $$\begin{gathered}
\label{eq:theta-envelope}
\frac{\Lambda(\theta)}{\theta} = \frac{1}{\theta} \log_e {\mathbb{E}}\{ \rho\big(\phi(\theta){{\mathbf{M}}}\big)\}.\end{gathered}$$ Above, ${{\mathbf{M}}}$ is the transition matrix of the underlying Markov process, and $\phi(\theta)$ is a diagonal matrix whose components are the moment generating functions of the processes in the Markov states. We assume that the fading coefficients $\{h_i\}$ change independently from one block to another. Under this assumption, the effective capacity can be obtained as $$\begin{aligned}
\label{eq:impcsiCE}
C_E({{\text{\footnotesize{SNR}}}},\theta) = -\frac{\Lambda(-\theta)}{\theta} = -\frac{1}{\theta}\log_e\left[1-{{\text{P}}}\{\Gamma_1\}\left(1-e^{-\theta \lambda}\right)\right]\end{aligned}$$ where ${{\text{P}}}\{\Gamma\}$ is the channel ON-state probability given in .
Energy Efficiency with Discrete Markov Sources
----------------------------------------------
First, we consider ON-OFF discrete Markov sources. We also assume that channel fading powers $z_1$ and $z_2$ are independent exponentially distributed with means 1 and $\gamma$, respectively. In the following result, we characterize the considered energy efficiency metrics under these assumptions.
\[prop:impcsi\] The minimum energy per bit and wideband slope achieved with fixed-rate secure transmissions in the presence of an eavesdropper with ON-OFF discrete Markov data arrivals and statistical QoS constraints are given by $$\begin{aligned}
\frac{E_b}{N_0}_{{\text{min}}}=e(\gamma+1)\log_e2, \quad \text{ and} \label{eq:impcsiebno}
\\
\mathcal{S}_{0}\!=\frac{1}{\frac{\theta (\eta-1)}{2\log_e2}+\frac{\theta e (\gamma+1) }{2\log_e2}+ e\gamma+\frac{e (\gamma+1)}{2} }, \label{eq:impcsiS0}\end{aligned}$$ respectively, with $\eta$ defined in .
*Proof:* See Appendix \[subsec:impcsi\].
As in the perfect CSI case, the minimum energy per bit in does not depend on the QoS exponent $\theta$ and source statistics while the wideband slope in depends on both. Specifically, wideband slope decreases with stricter QoS limitations (i.e., with increasing $\theta$) and increased source burstiness (i.e., with larger $\eta$).
It is also interesting to compare the minimum energy per bit expressions achieved with perfect CSI and no CSI. Recall from (\[eq:ebnomin-loge2\]) that with perfect CSI, the minimum energy per bit for the confidential message transmission to receiver 1 assuming exponentially distributed fading powers with ${\mathbb{E}}\{z_1\} = 1$ and ${\mathbb{E}}\{z_2\} = \gamma$ is $$\begin{gathered}
\frac{E_b}{N_0}_{{\text{min}}} = \log_e2.\end{gathered}$$ Comparing this with (\[eq:impcsiebno\]), we immediately identify the additional energy cost per bit of not having channel knowledge at the transmitter as $\left[e(\gamma+1)-1\right] \log_e2$. Hence, the characterization in Proposition \[prop:impcsi\] nicely quantifies the energy cost of not having transmitter CSI in secure wireless transmissions.
![Buffer overflow probability $\Pr\{Q>q\}$ vs. buffer threshold $q$ for different values of $\theta$. $p_{11} = p_{22} = 0.8$, SNR = $0.05$ []{data-label="fig:bufferqueue"}](buffersim.eps){width="45.00000%"}
Following the same methodology as described in the discussion of Fig. \[fig:bufferperfCSI\], we have again performed simulations in the case of no transmitter CSI. In Fig. \[fig:bufferqueue\], we plot the buffer overflow probability vs. buffer threshold $q$. We again have very good agreement with theoretical predictions. In particular, the simulated $\theta_{\text{sim}}$ values were obtained as $1.9306, 1.0657, 0.5109$ when the corresponding theoretical $\theta$ values were $2, 1, 0.5$, respectively.
![Maximum average arrival rate $ r_{{\text{avg}}}^*$ vs. energy per bit $\frac{E_b}{N_0}$ with various source statistics when $\theta = 0.5$.[]{data-label="fig:impcsidisc"}](impcsiebno_disc_p11p22.eps){width="45.00000%"}
As also noted above, Proposition \[prop:impcsi\] shows that while the minimum energy per bit does not depend on the source statistics and QoS exponent $\theta$, the wideband slope depends on both and decreases as burstiness parameter $\eta$ increases. We see these clearly in Fig. \[fig:impcsidisc\], where we plot the maximum average arrival rate vs. energy per bit for discrete Markov sources with varying statistics. As predicted, the minimum energy per bit stays same at $5.76$ dB, which is more than 7 dB larger than the minimum energy per bit of $-1.59$ dB achieved in the case of perfect CSI. We also observe that source with smaller $p_{11}$ and greater $p_{22}$ (while keeping $p_{11} + p_{22} = 1$) has a smaller $\eta$ value and correspondingly larger wideband slope. Hence, lower source burstiness benefits the energy efficiency.
Energy Efficiency with Markov Fluid Sources
-------------------------------------------
In this section, we consider ON-OFF Markov fluid sources and similarly as in the previous section identify the energy efficiency metrics.
\[prop:impcsifluid\] The minimum energy per bit and wideband slope achieved with fixed-rate secure transmissions in the presence of an eavesdropper with ON-OFF Markov fluid data arrivals and statistical QoS constraints are given by $$\begin{aligned}
\frac{E_b}{N_0}_{{\text{min}}}=e(\gamma+1)\log_e2, \quad \text{ and} \label{eq:impcsiebnofluid}
\\
\mathcal{S}_{0}\!=\frac{1}{\frac{\theta (\zeta-1)}{2\log_e2}+\frac{\theta e (\gamma+1) }{2\log_e2}+ e\gamma+\frac{e (\gamma+1)}{2} }, \label{eq:impcsiS0fluid}\end{aligned}$$ respectively, where $\zeta$ is defined in
*Proof:* See Appendix \[subsec:impcsifluid\].
![Maximum average arrival rate $ r_{{\text{avg}}}^*$ vs. energy per bit $\frac{E_b}{N_0}$ with various source statistics when $\theta = 0.5$.[]{data-label="fig:impcsifluid"}](impcsiebno_fluid_a_v2.eps){width="45.00000%"}
In Fig. \[fig:impcsifluid\], we depict the maximum average arrival rate vs. energy per bit curves for Markov fluid sources. We change the fixed the rate parameter $a$ (introduced in (\[eq:lambdataylor\])) and compare the curves. As expected, the optimal selection of $a=1$ returns the smallest value for the minimum energy per bit. However, as ${{\text{\footnotesize{SNR}}}}$ increases, higher performance is achieved when operating with $a=0.8$, indicating that $a = 1$ is optimal only at sufficiently small values of ${{\text{\footnotesize{SNR}}}}$.
Conclusion {#sec:Conc}
==========
In this paper, we have analyzed the throughput and energy efficiency of secure broadcast transmissions of two confidential messages and multicast transmission of a common message to two users under statistical delay/buffer QoS constraints. Considering Markovian data arrivals to the buffers, we have identified the maximum average arrival rates (or equivalently the throughput) and have determined the minimum energy per secret bit for confidential message transmission and minimum energy per bit for common message transmission. We have also obtained a linear approximation of the maximum average arrival rates in terms of $\frac{E_b}{N_0}$ by identifying the wideband slope.
The key observations and results are the following. Secure throughput is shown to decrease as the source becomes more bursty or the channel correlation increases. Throughput also diminishes when more stringent QoS requirements (indicated by higher values of the QoS exponent $\theta$) are imposed. In terms of energy efficiency, correlation works in favor of common message transmission while it works against the confidential message transmission. We have seen that in general, security requirements, source burstiness, and QoS constraints increase energy requirements. This is due to the facts that security considerations increase minimum energy per bit, and QoS constraints and source burstiness, while not having an impact on the minimum energy per bit in the cases of discrete Markov and Markov fluid sources, reduce the wideband slope. We have also noted that the more bursty MMPP sources require minimum energy per bit values that depends on the QoS exponent $\theta$ and increases with increasing $\theta$. Hence, in this case, energy requirements grow significantly as buffer/delay constraints become stricter.
Finally, assuming no instantaneous channel knowledge at the transmitter and fixed-rate transmissions, we have identified the throughput and energy efficiency expressions for both discrete discrete Markov and Markov fluid sources. Via these characterizations, we have identified the additional energy costs due to not knowing the channel.
Proof of Proposition \[prop:ebnocons\] {#subsec:ebnocons}
--------------------------------------
First, we define minimum energy per bit for the confidential messages as $$\begin{aligned}
\frac{E_b}{N_0}_{{\text{min}},i}= \frac{\delta_i \Pr(\Gamma_i) }{ \dot{r}_{{\text{avg}}i}^*(0) }\label{eq:ebnomin_conf2}\end{aligned}$$ where $i=1,2$. Similarly for the common message, the minimum energy per bit becomes $$\begin{aligned}
\frac{E_b}{N_0}_{{\text{min}},0}= \frac{(1-\delta_1) \Pr(\Gamma_1)+(1-\delta_2) \Pr(\Gamma_2) }{ \dot{r}_{{\text{avg}}0}^*(0) }.\label{eq:ebnomin_common2}\end{aligned}$$ As the arrival rate is constant, we can use effective capacity as the throughput formula. Therefore, we can exchange $\dot{r}_{\text{avg}}^*(0)$ with $\dot C_{E}(0)$ in the minimum energy per bit equation. For the proofs, we primarily focus on the ${\text{g}}({{\text{\footnotesize{SNR}}}})$ function that is defined in .
Now, the first derivative of $C_E({{\text{\footnotesize{SNR}}}})$ with respect to SNR is easily seen to be given by $$\begin{gathered}
\dot C_{Ei}({{\text{\footnotesize{SNR}}}}) = -\frac{1}{\theta_i }\frac{\dot {\text{g}}_i({{\text{\footnotesize{SNR}}}})}{{\text{g}}_i({{\text{\footnotesize{SNR}}}})} \label{eq:dotcei0}\end{gathered}$$ where $\dot{{\text{g}}}_i({{\text{\footnotesize{SNR}}}})$ denote the first derivative of the function ${\text{g}}_i({{\text{\footnotesize{SNR}}}})$ with respect to ${{\text{\footnotesize{SNR}}}}$. It can be readily seen that ${\text{g}}_i(0)=1$. If we use $f_i({{\text{\footnotesize{SNR}}}})$ as the instantaneous service rate in nats (i.e. $R_i({{\text{\footnotesize{SNR}}}})=f_i({{\text{\footnotesize{SNR}}}})\log_e2$), then we have the relation $$\begin{gathered}
\dot{{\text{g}}}_i(0)=-\frac{\theta_i}{\log_e2}{\mathbb{E}}\left\{\dot{f}_i(0)\right\} \label{eq:dotgi0}\end{gathered}$$ where the first derivative expressions $\dot{f}_i(0)$ for $i =0,1,2$ are given by $$\begin{aligned}
\label{eq:dotfis}
\dot{f}_1(0)&= \delta_1 \left(z_1 - z_2\right) {\mathbf{1}}\!\left\{z_1\!\geq \!z_2\right\}, \nonumber
\\
\dot{f}_2(0)&= \delta_2 \left( z_2-z_1 \right) {\mathbf{1}}\!\left\{z_1\!< \!z_2\right\}, \nonumber
\\
\dot{f}_0(0)&= (1-\delta_1) z_2{\mathbf{1}}\!\left\{z_1\!\geq \!z_2\right\}+ (1-\delta_2) z_1{\mathbf{1}}\!\left\{z_1\!< \!z_2\right\}.\end{aligned}$$ By inserting $\dot{f}_i(0)$ formulations above to , and then $\dot{{\text{g}}}_i(0)$ to consecutively, we obtain the minimum energy per bit expressions for confidential and common messages in - using and .
Proof of Proposition \[prop:ebno\_discrete\] {#subsec:ebno_discrete}
--------------------------------------------
First, we prove the result for the discrete Markov source. We need to obtain the first derivative of $r_{{\text{avg}},i}^*({{\text{\footnotesize{SNR}}}})$. Let us rewrite the maximum average arrival rate in as $$\begin{aligned}
r_{{\text{avg}},i}^*({{\text{\footnotesize{SNR}}}},\theta_i)=\frac{P_{{\text{ON}}}}{\theta_i}&\Big[\log_e(1-p_{11}{\text{g}}_i({{\text{\footnotesize{SNR}}}})) -\log_e({\text{g}}_i({{\text{\footnotesize{SNR}}}}))\Big. \nonumber
\\
&\Big.-\log_e\big((1-p_{11}-p_{22}){\text{g}}_i({{\text{\footnotesize{SNR}}}})+p_{22}\big)\Big]\end{aligned}$$ where ${\text{g}}_i({{\text{\footnotesize{SNR}}}})$ is defined in . Taking the first derivative with respect to ${{\text{\footnotesize{SNR}}}}$, we obtain $$\begin{aligned}
\dot{r}_{{\text{avg}},i}^*({{\text{\footnotesize{SNR}}}},\theta_i)=\frac{P_{{\text{ON}}}}{\theta_i}& \Bigg[\frac{-p_{11}\dot{{\text{g}}_i}({{\text{\footnotesize{SNR}}}})}{1-p_{11}{\text{g}}_i({{\text{\footnotesize{SNR}}}})} -\frac{\dot{{\text{g}}_i}({{\text{\footnotesize{SNR}}}})}{{\text{g}}_i({{\text{\footnotesize{SNR}}}})}\Big. \nonumber
\\
&\Big.-\frac{(1-p_{11}-p_{22})\dot{{\text{g}}_i}({{\text{\footnotesize{SNR}}}})}{(1-p_{11}-p_{22}){\text{g}}_i({{\text{\footnotesize{SNR}}}})+p_{22}}\Bigg]. \label{eq:dotravg_disc}\end{aligned}$$ When we let ${{\text{\footnotesize{SNR}}}}\to 0$, the first derivative expression becomes $$\begin{aligned}
\dot{r}_{{\text{avg}},i}^*(0)&=\frac{\dot{{\text{g}}}_i(0)}{\theta_i} P_{{\text{ON}}}\Bigg[-\frac{p_{11}}{1-p_{11}} -1-\frac{1-p_{11}-p_{22}}{1-p_{11}} \Bigg]
\\
&=-\frac{\dot {\text{g}}_i(0)}{\theta}=\frac{\dot{f}_i(0)}{\log_e2}\label{eq:dotravg0_disc}\end{aligned}$$ where $P_{{\text{ON}}}=\tfrac{1-p_{11}}{2-p_{11}-p_{22}}$. Note that ${\text{g}}(0)=1$. Plugging the result in and into and , we immediately obtain (\[eq:EbN01\]) - (\[eq:EbN00\]). Now, we show the proof for the Markov fluid source. We evaluate the derivative of $r_{{\text{avg}},i}^*({{\text{\footnotesize{SNR}}}})$ in with respect to ${{\text{\footnotesize{SNR}}}}$ and obtain given at the top of the next page.
$$\begin{aligned}
\dot r_{{\text{avg}},i}^*({{\text{\footnotesize{SNR}}}}, \theta_i)=-\frac{P_{{\text{ON}}}}{\theta_i} \left\{\log_e{\text{g}}_i({{\text{\footnotesize{SNR}}}})\tfrac{\operatorname{d}\!}{\operatorname{d}\!{\text{\scriptsize{SNR}}}} \Bigg[\frac{\alpha+\beta-\log_e{\text{g}}_i({\text{\scriptsize{SNR}}})}{\alpha-\log_e{\text{g}}_i({\text{\scriptsize{SNR}}})}\Bigg]+ \frac{\alpha+\beta-\log_e{\text{g}}_i({\text{\scriptsize{SNR}}})}{\alpha-\log_e{\text{g}}_i({\text{\scriptsize{SNR}}})} \frac{\dot{{\text{g}}_i}({{\text{\footnotesize{SNR}}}})}{{\text{g}}_i({{\text{\footnotesize{SNR}}}})} \right\} \label{eq:dotravg_fluid}\end{aligned}$$
When we let ${{\text{\footnotesize{SNR}}}}\to 0$, the first derivative expression simplifies to $$\begin{aligned}
\dot{r}_{{\text{avg}},i}^*(0)&=-\frac{P_{{\text{ON}}}}{\theta_i} \frac{\alpha+\beta}{\alpha} \dot{{\text{g}}}_i(0)=\frac{\dot{f}_i(0)}{\log_e2} \label{eq:dotravg0_fluid}\end{aligned}$$ where $P_{{\text{ON}}}=\tfrac{\alpha}{\alpha+\beta}$. Note that ${\text{g}}(0)=1$. Plugging the result in and into and , we immediately obtain (\[eq:EbN01\]) - (\[eq:EbN00\]).
Proof of Proposition \[prop:ebno\_MMPP\] {#subsec:ebno_MMPP}
----------------------------------------
The proof is straightforward as we note that the maximum average arrival rate $r_{{\text{avg}},i}^*({{\text{\footnotesize{SNR}}}})$ of discrete-time MMPP source in is the scaled version of that of the discrete Markov source in . The scaling factor is $\frac{\theta_i}{e^\theta_i-1}$. The same assertion can be made for the relationship between the maximum average arrival rates of continuous-time MMPP in and Markov fluid source in . Therefore, the minimum energy per bit expressions for discrete-time and continuous-time MMPP sources can be obtained by scaling the formulations in - with $\frac{e^\theta_i-1}{\theta_i}$.
Proof of Proposition \[prop:widebandslope\] {#subsec:widebandslope}
-------------------------------------------
Let us recall that the wideband slope is given by $$\label{eq:widebandsloperavg}
\mathcal{S}_0=-\frac{2\big(\dot{r}_{\text{avg}}^*(0)\big)^2} { \ddot{r}_{\text{avg}}^*(0)}\log_e{2}.$$ When the arrival rate is constant, we can exchange $r_{{\text{avg}},i}^*({{\text{\footnotesize{SNR}}}})$ with $C_{Ei}({{\text{\footnotesize{SNR}}}})$. For the wideband slope, in addition to the first derivative of the throughput, we also need to obtain the second derivative of the throughput. Second derivatives of the effective capacity at ${{\text{\footnotesize{SNR}}}}=0$ can be computed as $$\begin{gathered}
\ddot C_{Ei}({{\text{\footnotesize{SNR}}}}) = -\frac{1}{\theta_i }\left[\frac{\ddot {\text{g}}_i({{\text{\footnotesize{SNR}}}})}{{\text{g}}_i({{\text{\footnotesize{SNR}}}})}-\left(\frac{\dot {\text{g}}_i({{\text{\footnotesize{SNR}}}})}{{\text{g}}_i({{\text{\footnotesize{SNR}}}})}\right)^2 \right]. \label{eq:ddotcei0}\end{gathered}$$ To simplify this equation, we derive the second derivative of ${\text{g}}_i({{\text{\footnotesize{SNR}}}})$ at ${{\text{\footnotesize{SNR}}}}=0$ as $$\begin{gathered}
\ddot{{\text{g}}}_i(0)=-\frac{\theta_i}{\log_e2} {\mathbb{E}}\left\{\ddot{f}_i(0)\right\}+ \left(\frac{\theta_i}{\log_e2}{\mathbb{E}}\left\{\ddot{f}_i(0)\right\}\right)^2, \label{eq:ddotgi0}\end{gathered}$$ where the second derivative expressions $\ddot{f}_i(0)$ for $i=0,1,2$ are given by $$\begin{aligned}
\label{eq:ddotfis}
\ddot{f}_1(0)&= -\delta_1^2 \left[z_1^2 - z_2^2\right] {\mathbf{1}}\!\left\{z_1\!\geq \!z_2\right\}, \nonumber
\\
\ddot{f}_2(0)&= -\delta_2^2 \left[ z_2^2-z_1^2 \right]{\mathbf{1}}\!\left\{z_1\!< \!z_2\right\}, \nonumber
\\
\ddot{f}_0(0)&= -(1-\delta_1^2)z_2^2{\mathbf{1}}\!\left\{z_1\!\geq \!z_2\right\}-(1-\delta_2^2)z_1^2 {\mathbf{1}}\!\left\{z_1\!< \!z_2\right\}.\end{aligned}$$
We insert $\dot{f}_i(0)$ in and $\ddot{f}_i(0)$ in onto $\dot{{\text{g}}}_i(0)$ in and $\ddot{{\text{g}}}_i(0)$ in in order to obtain first and second derivative expressions of the effective capacity. By incorporating the and on we obtain wideband slope expression in .
Proof of Proposition \[prop:wbsdisc\] {#subsec:wbsdisc}
-------------------------------------
In order to find the wideband slope, we need to determine the second derivative of the maximum average arrival rate with respect to ${{\text{\footnotesize{SNR}}}}$. As ${{\text{\footnotesize{SNR}}}}\rightarrow0$ the second derivative expression is given by $$\begin{aligned}
\ddot{r}_{{\text{avg}},i}^*(0,\theta_i)=&\frac{\ddot{{\text{g}}_i}(0)}{\theta_i} P_{{\text{ON}}}\Bigg[-\frac{p_{11}}{1-p_{11}} -1-\frac{(1-p_{11}-p_{22})}{1-p_{11}} \Bigg] \nonumber
\\
&+\frac{\left[\dot{{\text{g}}_i}(0)\right]^2}{\theta_i}P_{{\text{ON}}}\Bigg[-\frac{p_{11}^2}{(1-p_{11})^2} +1\!+\frac{(1-p_{11}-p_{22})^2}{(1-p_{11})^2} \Bigg]\nonumber
\\
=&-\frac{\ddot{{\text{g}}_i}(0)}{\theta_i}+(1-\eta)\frac{\left[\dot{{\text{g}}_i}(0)\right]^2}{\theta_i} \label{eq:ddotravg0_disc}\end{aligned}$$ where $\eta$ is defined in (\[eq:eta\]). The fact that $g_i(0) = 1$ is taken into account in (\[eq:ddotravg0\_disc\]). Finally, inserting and into , the wideband slope expression in is readily obtained.
Proof of Proposition \[prop:wbsfluid\] {#subsec:wbsfluid}
--------------------------------------
In order to find the wideband slope, we need to determine the second derivative of the maximum average arrival rate with respect to ${{\text{\footnotesize{SNR}}}}$. When ${{\text{\footnotesize{SNR}}}}\rightarrow0$, the second derivative expression is given by $$\begin{aligned}
\ddot{r}_{{\text{avg}},i}^*(0,\theta_i)=&-\frac{P_{{\text{ON}}}}{\theta_i} \left\{\frac{2\beta}{\alpha^2}\dot{{\text{g}}_i}(0)+\frac{\alpha+\beta}{\alpha}\left(\ddot{{\text{g}}_i}(0)-\dot{{\text{g}}_i}(0)\right)\right\}\nonumber
\\
=&-\frac{\ddot{{\text{g}}_i}(0)}{\theta_i}+(1-\zeta)\frac{\left[\dot{{\text{g}}_i}(0)\right]^2}{\theta_i} \label{eq:ddotravg0_fluid}\end{aligned}$$ where $\zeta$ is defined in (\[eq:zeta\]) and we again use the fact that ${\text{g}}_i(0) = 1$. Finally, inserting and into , we obtain the wideband slope expression in .
Proof of Proposition \[prop:impcsi\] {#subsec:impcsi}
------------------------------------
First, we define $${\text{g}}({{\text{\footnotesize{SNR}}}})=1-{{\text{P}}}\{\Gamma_1\}\left(1-e^{-\theta \lambda}\right). \label{ravgdiscV2}$$ The maximum average arrival rate of the ON-OFF discrete Markov source can be rewritten as $$\begin{aligned}
r_{\text{avg}}^*({{\text{\footnotesize{SNR}}}})=\frac{P_{{\text{ON}}}}{\theta}\left[\log_e\left(\frac{1-p_{11}{\text{g}}({{\text{\footnotesize{SNR}}}})}{(1\!-\!p_{11}\!-\!p_{22}){\text{g}}^2({{\text{\footnotesize{SNR}}}})+p_{22}{\text{g}}({{\text{\footnotesize{SNR}}}})}\right)\right].\end{aligned}$$
In order to find the minimum energy per bit and wideband slope, we need to determine the first and second derivatives of the maximum average arrival rate with respect to ${{\text{\footnotesize{SNR}}}}$. Initially, we take the first derivative of maximum average arrival rate and let ${{\text{\footnotesize{SNR}}}}\to 0$ as follows: $$\begin{aligned}
\dot{r}_{\text{avg}}^*(0)&=\frac{\dot{{\text{g}}}(0)}{\theta} P_{{\text{ON}}}\Bigg[-\frac{p_{11}}{1-p_{11}} -1-\frac{1-p_{11}-p_{22}}{1-p_{11}} \Bigg]
\\
&=-\frac{\dot {\text{g}}(0)}{\theta}. \label{eq:dotravg0_disc2}\end{aligned}$$ For this, we also need to characterize the first derivative of ${\text{g}}({{\text{\footnotesize{SNR}}}})$. We start with the Taylor series expansion of the fixed rate $\lambda$ in the low-SNR regime: $$\lambda=\frac{a}{\log_e2}{{\text{\footnotesize{SNR}}}}+ \frac{b}{\log_e2}{{\text{\footnotesize{SNR}}}}^2 + o({{\text{\footnotesize{SNR}}}}^2) \label{eq:lambdataylor}.$$ Now, the first derivative of ${\text{g}}({{\text{\footnotesize{SNR}}}})$ is given by $$\label{eq:dotgsnrimpcsi}
\dot{{\text{g}}}({{\text{\footnotesize{SNR}}}})=-\frac{\partial}{\partial{{\text{\footnotesize{SNR}}}}}{{\text{P}}}\{\Gamma_1\}\left(1-e^{-\theta \lambda}\right)+{{\text{P}}}\{\Gamma_1\}\frac{\partial e^{-\theta \lambda}}{\partial{{\text{\footnotesize{SNR}}}}}.$$ As ${{\text{\footnotesize{SNR}}}}\rightarrow0$, we have $\lambda\rightarrow0$. Therefore at ${{\text{\footnotesize{SNR}}}}=0$, we have $$\label{eq:dotg0impcsi}
\dot{{\text{g}}}(0)=\lim_{{{\text{\footnotesize{SNR}}}}\to0}{{\text{P}}}\{\Gamma_1\} (-\theta)e^{-\theta \lambda} \frac{\partial \lambda}{\partial{{\text{\footnotesize{SNR}}}}}.$$
To proceed we need to obtain the probability expression ${{\text{P}}}\{\Gamma_1\}$. For independent and exponentially distributed $z_1$ and $z_2$ with unit mean, we can obtain $$\begin{aligned}
{{\text{P}}}\{\Gamma_1\} &= \int_0^\infty e^{-z_2}\int_{2^\lambda z_2+ \frac{2^\lambda-1}{{\text{\scriptsize{SNR}}}}}^\infty e^{-z_1}dz_1 dz_2
\\
&=e^{-\frac{2^\lambda-1}{{\text{\scriptsize{SNR}}}}}\int_0^\infty e^{-(2^\lambda+1)z_2} dz_2
\\
&=e^{-\frac{2^\lambda-1}{{\text{\scriptsize{SNR}}}}}\frac{1}{2^\lambda+1}.\end{aligned}$$ Now, we can simplify the expression in as $$\dot{{\text{g}}}(0)=-\frac{e^{-a}}{2}\theta \frac{a}{\log_e2}, \label{eq:dotg0impcsiv2}$$ and inserting this expression into , we obtain the minimum energy per bit as $$\frac{E_b}{N_0}_{{\text{min}}}=-\frac{\theta}{\dot{{\text{g}}}(0)}=\frac{2\log_e2}{ae^{-a}}.$$ Finally, we want to determine the smallest possible minimum energy per bit expression. It can be easily seen that the smallest value for the minimum energy per bit is obtained when $a=1$, leading to the minimum energy per bit expression in .
In order to find the wideband slope, we first determine the second derivative of the maximum average arrival rate with respect to ${{\text{\footnotesize{SNR}}}}$ and then evaluate it at ${{\text{\footnotesize{SNR}}}}= 0$. The resulting equation is given above in .
$$\begin{aligned}
\ddot{r}_{\text{avg}}^*(0)=&\frac{\ddot{{\text{g}}}(0)}{\theta} P_{{\text{ON}}}\Bigg[-\frac{p_{11}}{1-p_{11}} -1-\frac{(1-p_{11}-p_{22})}{1-p_{11}} \Bigg] +\frac{\left[\dot{{\text{g}}}(0)\right]^2}{\theta}P_{{\text{ON}}}\Bigg[-\frac{p_{11}^2}{(1-p_{11})^2} +1\!+\frac{(1-p_{11}-p_{22})^2}{(1-p_{11})^2} \Bigg]\nonumber
\\
=&-\frac{\ddot{{\text{g}}}(0)}{\theta}+(1-\eta)\frac{\left[\dot{{\text{g}}}(0)\right]^2}{\theta} \label{eq:ddotravg0_disc2}.\end{aligned}$$
Note that, $\eta$ is defined in . The first derivative of ${\text{g}}({{\text{\footnotesize{SNR}}}})$ at ${{\text{\footnotesize{SNR}}}}=0$ is given by , and the second derivative is $$\begin{aligned}
\ddot{{\text{g}}}(0)=&\lim_{{{\text{\footnotesize{SNR}}}}\to0}2\frac{\partial {{\text{P}}}\{\Gamma_1\}}{\partial{{\text{\footnotesize{SNR}}}}}\frac{\partial e^{-\theta \lambda}}{\partial{{\text{\footnotesize{SNR}}}}}+{{\text{P}}}\{\Gamma_1\}\frac{\partial e^{-\theta \lambda}}{\partial{{\text{\footnotesize{SNR}}}}^2}
\\
=&2\left(-\frac{1}{4}e^{-a}(a^2+a+2b)(-\theta)\frac{a}{\log_e2}\right)\nonumber
\\
&+\frac{e^{-a}}{2}\left(-\theta\frac{2b}{\log_e2}+\theta^2\frac{a^2}{(\log_e2)^2}\right)
\\
=&\frac{e^-a}{2} \left[\theta \frac{a^3+a^2}{\log_e2}+\theta^2 \frac{a^2}{(\log_e2)^2}+\theta \frac{2b}{\log_e2}(a-1) \right]. \label{eq:ddotg0impcsi}\end{aligned}$$
The wideband slope expression can be determined inserting the first and second derivative expressions in and into : $$\begin{aligned}
\mathcal{S}_{0}&=\frac{2 \frac{\left(\dot{{\text{g}}}(0)\right)^2}{\theta^2} }{ \frac{\ddot{{\text{g}}}(0) }{\theta}+\frac{\eta-1}{\theta} \left(\dot{{\text{g}}}(0)\right)^2} \log_e2
\\
&=\frac{1}{\frac{\theta (\eta-1)}{2\log_e2}+\frac{\theta}{\log_e2 e^{-a}}+\frac{a+1}{e^{-a}}+\frac{2b(a-1)}{a^2 e^{-a}} }. \label{eq:widebandslopeinab}\end{aligned}$$ Since the wideband slope is defined as the slope at the minimum energy per bit, we set $a=1$. Note that with this choice, parameter $b$ vanishes as $2b(a-1)\to 0$ in . Thus, we obtain the formulation in .
Proof of Proposition \[prop:impcsifluid\] {#subsec:impcsifluid}
-----------------------------------------
The maximum average arrival rate of Markov fluid source can be rewritten as $$\begin{aligned}
r_{\text{avg}}^*({{\text{\footnotesize{SNR}}}})=-\frac{P_{{\text{ON}}}}{\theta}\left[1+\frac{\beta}{\alpha-\log_e({\text{g}}({{\text{\footnotesize{SNR}}}}))}\right]\log_e({\text{g}}({{\text{\footnotesize{SNR}}}})). \label{eq:ravgfluidv2}\end{aligned}$$
By taking the first derivative of the expression in and letting ${{\text{\footnotesize{SNR}}}}\to 0$, we obtain the following: $$\begin{aligned}
\dot{r}_{\text{avg}}^*(0)&=-\frac{P_{{\text{ON}}}}{\theta}\left[1+\frac{\beta}{\alpha}\right]\dot{{\text{g}}}(0)=-\frac{\dot {\text{g}}(0)}{\theta}. \label{eq:dotravg0_fluid2}\end{aligned}$$
By combining with as $a\to 1$, and inserting into , we obtain the minimum energy per bit given in .
Next, we take the second derivative of the maximum average arrival rate with respect to ${{\text{\footnotesize{SNR}}}}$ and then evaluate it at ${{\text{\footnotesize{SNR}}}}= 0$ as $$\begin{aligned}
\ddot{r}_{\text{avg}}^*(0)=&-\frac{P_{{\text{ON}}}}{\theta}\left[\left(\frac{2\beta}{\alpha^2}-1-\frac{\beta}{\alpha}\right)\left(\dot{{\text{g}}}(0)\right)^2 + \left(1+\frac{\beta}{\alpha}\right) \ddot{{\text{g}}}(0) \right]
\\
=&-\frac{\ddot{{\text{g}}}(0)}{\theta}+(1-\zeta)\frac{\left[\dot{{\text{g}}}(0)\right]^2}{\theta} \label{eq:ddotravg0_fluid2}.\end{aligned}$$ Note that, $\zeta$ is defined in . We derive the wideband slope expression by using , and . Again, since the wideband slope is defined at the minimum energy per bit, we set $a=1$ and obtain the formulation in .
[99]{} A. D. Wyner, “The wire-tap channel," *Bell Syst. Tech. J.*, vol. 54, pp. 1355–1367, Oct. 1975
I. Csiszár and J. Körner, “Broadcast channels with confidential messages," *IEEE Trans. Inform. Theory*, vol. 3, pp. 339–348, May 1978.
D. Feng, C. Jiang, G. Lim, L. J. Cimini, Jr., G. Feng, and G. Y. Li, “A survey of energy-efficient wireless communications," *IEEE Commun. Surveys & Tutorials*, vol. 15, no. 1, pp. 167–178, 2013.
S. Tanwir and H. Perros, “A survey of VBR video traffic models," *IEEE Commun. Surveys & Tutorials*, vol. 15, no. 4, pp. 1778–1802, 2013.
P. K. Gopala, L. Lai, and H. El Gamal, “On the secrecy capacity of fading channels," *IEEE Trans. Inf. Theory*, vol. 54, no. 10, pp. 4687– 4698, Oct. 2008.
Y. Liang, L. Lai, H. V. Poor, S. Shamai (Shitz), “A broadcast approach for fading wiretap channels," *IEEE Trans. Inform. Theory*, vol. 60, no. 2, pp. 842–858, Feb. 2014.
H. Jeon, N. Kim, J. Choi, H. Lee, J. Ha, “Bounds on secrecy capacity over correlated ergodic fading channels at high SNR," *IEEE Trans. Inform. Theory*, vol. 57, no. 4, pp. 1975-1983, Apr. 2011.
P. Baracca, N. Laurenti, S. Tomasin, “Physical layer authentication over MIMO fading wiretap channels," *IEEE Trans. Wireless Commun.*, vol. 11, no. 7, pp. 2564-2573, July. 2012.
J. Li, A. P. Petropulu, “Ergodic secrecy rate for multiple-antenna wiretap channels with Rician fading," *IEEE Trans. Inform. For. and Sec.*, vol. 6, no.3, pp. 861-867, Sept. 2011.
S. Verdú, “Spectral efficiency in the wideband regime," *IEEE Trans. Inform. Theory*, vol.48, no.6 pp.1319-1343. Jun.2002.
M. C. Gursoy “Secure communication in the low-SNR regime," *IEEE Trans. Commun.*, vol. 60, no. 4, pp. 1114-1123, Apr. 2012.
C. Comaniciu, H. V. Poor, “On energy-secrecy trade-offs for Gaussian wiretap channels," *IEEE Trans. Inform. For. and Sec.*, vol. 8, no.2, pp. 314-323, Feb. 2013.
H. Zhang, Y. Huang, S. Li and L. Yang, “Energy-efficient precoder design for MIMO wiretap channels," *IEEE Commun. Letters*, vol. 18, no.9, pp. 1559-1562, Sept. 2014.
D. W. K. Ng, E. S. Lo and R. Schober, “Energy-efficient resource allocation for secure OFDMA systems," *IEEE Trans. Vehic. Tech.*, vol. 61, no.6, pp. 2572 - 2585, July 2012.
A. Kalantari, S. Maleki, G. Zheng, S. Chatzinotas and B. Ottersen, “Joint power control in wiretap interference channels," *IEEE Trans. Wireless Commun.*, vol. 14, no.7, pp. 3810 - 3823, July 2015.
X. Chen, L. Lei, “Energy-efficient optimization for physical layer security in multi-antenna downlink networks with QoS guarantee," *IEEE Commun. Letters*, vol. 17, no.4, pp. 637-640, Apr. 2013.
X. Zhu, B. Yang, C. Chen, L. Xue, X. Guan and F. Wu, “Cross-layer scheduling for OFDMA-based cognitive radio systems with delay and security constraints,” *IEEE Trans. Veh. Tech.*, vol. 64, no. 12, pp. 5919-5934, Dec. 2015.
X. Zhu, B. Yang and X. Guan, “Cross-layer scheduling with secrecy demands in delay-aware OFDMA network,” IEEE Wireless Communications and Networking Conference (WCNC), Shanghai, 2013, pp. 1339-1344.
A. El Shafie and N. Al-Dhahir, “On secure communications over a wiretap channel with fixed-rate transmission: Protocol design and queueing analysis,” *IEEE Wireless Commun. Letters*, vol. 4, no. 4, pp. 453-456, Aug. 2015.
A. El Shafie and N. Al-Dhahir, “Secure communications in the presence of a buffer-aided wireless-powered relay with self-energy recycling,” *IEEE Wireless Communications Letters*, vol. 5, no. 1, pp. 32-35, Feb. 2016.
A. El Shafie, T. Q. Duong and N. Al-Dhahir, “QoS-based design for security enhancement in TDMA-based wireless networks,” IEEE Globecom Workshops, Washington D.C., 2016
Z. Mao, C. E. Koksal and N. B. Shroff, “Achieving full secrecy rate with low packet delays: An optimal control approach,” *IEEE Journal on Selected Areas in Commun.*, vol. 31, no. 9, pp. 1944-1956, Sept. 2013.
K. Khalil, O. O. Koyluoglu, H. E. Gamal and M. Youssef, “Opportunistic secrecy with a strict delay constraint,” *IEEE Trans. on Commun.*, vol. 61, no. 11, pp. 4700-4709, Nov. 2013.
Y. Liang, H. V. Poor, and S. Shamai (Shitz), “Information theoretic security," *Found. Trends Commun. Inform. Theory*, vol. 5, no. 4–5, pp. 355-–580, 2009.
R. Liu and W. Trappe, *Securing Wireless Communications at the Physical Layer*, Springer-Verlag, 2010.
M. Block and J. Barros, *Physical-Layer Security: From Information Theory to Security Engineering*, Cambridge University Press, 2011.
R. Bassily, *et al.,* “Cooperative security at the physical layer: A summary of recent advances," *IEEE Signal Process. Mag.,* vol. 30, no. 5, pp. 16-28, Sept. 2013.
Y.-W. P. Hong, P.-C. Lan, and C.-C. J. Kuo, “Enhancing physical-layer secrecy in multiantenna wireless systems: An overview of signal processing approaches," *IEEE Signal Process. Mag.,* vol. 30, no. 5, pp. 29-40, Sept. 2013.
A. Yener and S. Ulukus, “Wireless physical-layer security: Lessons learned from information theory," *Proc. of the IEEE,* vol. 103, no. 10, pp. 1814-1825, Oct. 2015.
Y. Zou, J. Zhu, X. Wang, and L. Hanzo, “A survey on wireless security: Technical challenges, recent advances, and future trends," *Proc. of the IEEE*, vol. 104, no. 9, pp. 1727–1765, Sep. 2016.
R. L. Cruz, “A calculus for network delay, part I: Network elements in isolation," *IEEE Trans. Inf. Theory*, vol. 37, pp. 114–-131, Jan. 1991.
R. L. Cruz, “A calculus for network delay, part II: Network analysis," *IEEE Trans. Inf. Theory*, vol. 37 pp. 132–-141, Jan. 1991.
C.-S. Chang, “Stability, queue length, and delay of deterministic and stochastic queuing networks," *IEEE Trans. Auto. Control*, vol. 39, no. 5, pp. 913-931, May 1994.
C.-S. Chang, *Performance Guarantees in Communication Networks*, Springer-Verlag, 2000.
C.-S. Chang and T. Zajic, “Effective bandwidths of departure processes from queues with time varying capacities," Proc. of the IEEE INFOCOM, 1995.
A. I. Elwalid and D. Mitra, “Effective bandwidth of general Markovian traffic sources and admission control of high speed networks," *IEEE/ACM Trans. on Networking.*, vol. 1, no. 3, pp. 329-343, Jun. 1993.
G. Kesidis, J.Walrand, and C.-S. Chang, “Effective bandwidths for multiclass Markov fluids and other ATM sources," *IEEE/ACM Trans. on Networking.*, vol. 1, no.4, pp. 424-428, Aug. 1993.
C. Courcoubetis and R. Weber, “Effective bandwidth for stationary sources," *Prob. Eng. Inf. Sci.*, 9, pp. 285-296, 1995.
D. Wu and R. Negi “Effective capacity: a wireless link model for support of quality of service," *IEEE Trans. Wireless Commun.*, vol.2,no. 4, pp.630-643. July 2003.
Q. Du and X. Zhang, “Statistical QoS provisioning for wireless unicast/multicast of multi-layer video streams," *IEEE J. Sel. Areas Commun.*, vol. 28, no. 3, pp. 420–433, Apr. 2010.
M. C. Gursoy, D. Qiao, and S. Velipasalar “Analysis of energy efficiency in fading channels under QoS constraints," *IEEE Trans. Wireless Commun.*, vol. 8, no. 8, pp. 4252–4263, Aug. 2009.
L. Musavian, and Q. Ni, “Effective capacity maximization with statistical delay and effective energy efficiency requirements," *IEEE Trans. Wireless Commun.* vol. 14, no. 7, pp. 3824 - 3835, July 2015.
C. She, C. Yang and L. Liu, “Energy-efficient resource allocation for MIMO-OFDM systems serving random sources with statistical QoS requirement", IEEE Trans. Commun., vol. 63, no. 11, pp. 4125-4141, Nov. 2015.
M. Ozmen and M. C. Gursoy, “Wireless throughput and energy efficiency with random arrivals and statistical queueing constraints," *IEEE Trans. Inform. Theory*, vol. 62, no. 3, pp. 1375-1395, March 2016. D. Qiao, M. C. Gursoy, and S. Velipasalar “Secure wireless communication and optimal power control under statistical queueing constraints," *IEEE Trans. Inform. Forensics and Security*, vol. 6, no. 3, pp. 628–639, Sep. 2011.
M. Ozmen and M. C. Gursoy, “Throughput and energy efficiency under queueing and secrecy constraints," Proc. of the Asilomar Conference on Signals, Systems and Computers, Nov. 2012
[^1]: The authors are with the Department of Electrical Engineering and Computer Science, Syracuse University, Syracuse, NY, 13244 (e-mail: [email protected], [email protected]).
[^2]: For discrete Markov and Markov fluid sources, we have a constant arrival rate of $r$ in the ON state, while the arrival rate is Poisson distributed with intensity (or equivalently average value) $r$ for MMPP sources.
[^3]: Here, we consider standard information-theoretic arguments regarding the definition of messages and how they are encoded and transmitted over fading channels (see e.g., [@Gopala], [@liang-broad]).
[^4]: We note that the event of $z_1=z_2$ occurs with zero probability if the fading powers $z_1$ and $z_2$ have continuous distributions, as frequently assumed in the statistical modeling of the wireless fading channel in the literature. However, in the case of discrete fading distributions, this event is in general a non-zero probability event. In such a case, the secrecy capacity is zero, and hence no confidential message transmission can be performed. All the power can be allocated to the transmission of the common message by setting $\delta_1=0$.
[^5]: The secrecy rate expressions in (\[eq:rate1\]) and (\[eq:rate2\]) are derived from the generic expression in (\[eq:secrecyrate\]) For instance, in (\[eq:rate1\]), $z_1$ and $z_2$ correspond to $z_m$ and $z_e$, respectively, and the signal-to-noise ratio is $\delta_1 {{\text{\footnotesize{SNR}}}}$. Additionally, the indicator function essentially represents the operation $[\cdot]^+$, ensuring that the secrecy rate is zero if $z_1 < z_2$.
[^6]: Note from (\[eq:overflowprob-rev\]) that the overflow probability is expected to behave in logarithmic scale as $\log \Pr\{Q \ge q\} \approx -\theta q + \log \varsigma$. Hence, the slope of the logarithmic overflow probability vs. buffer threshold $q$ curve is proportional to $-\theta$.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Using a generalized parton model approach including spin and intrinsic parton motion effects, and assuming the validity of factorization for large $p_T$ jet production in hadronic collisions, we study the azimuthal distribution around the jet axis of leading pions, produced in the jet fragmentation process. We identify the observable leading-twist azimuthal asymmetries, which are generated by all the physically allowed combinations of transverse momentum dependent (TMD) parton distribution and fragmentation functions. In particular, we show how one can isolate the Collins and Sivers contributions, and suggest a test of the process dependence of the Sivers function by considering the effect of color-gauge invariant initial and final state interactions.'
author:
- 'U. D’Alesio, F. Murgia C. Pisano[^1]'
title: Azimuthal distributions of pions inside a jet in hadronic collisions
---
\[1999/12/01 v1.4c Il Nuovo Cimento\]
Introduction
============
Transverse single-spin and azimuthal asymmetries in high-energy hadronic reactions have raised a lot of interest in the last years (see e.g. Ref. [@D'Alesio:2007jt] and references therein). In particular, the huge spin asymmetries measured in the inclusive forward production of pions in high-energy proton-proton collisions, at moderately large transverse momentum, cannot be explained in the realm of leading-twist (LT) perturbative QCD (pQCD), based on the usual collinear factorization theorems. Out of the theoretical approaches proposed in order to account for these measurements, in the following we will adopt the so-called transverse momentum dependent (TMD) formalism, which takes into account spin and intrinsic parton motion effects assuming a pQCD factorization scheme. In this framework, single-spin and azimuthal asymmetries are generated by TMD polarized partonic distribution and fragmentation functions, among which the most relevant from a phenomenological point of view are the Sivers distribution [@Sivers:1989cc] and, for transversely polarized quarks, the Boer-Mulders distribution [@Boer:1997nt] and the Collins fragmentation function [@Collins:1992kk] (similar functions can be defined for linearly polarized gluons, see e.g. Ref. [@Anselmino:2005sh]).
Azimuthal asymmetries in the distribution of pions inside a large transverse momentum jet are quite interesting observables [@D'Alesio:2010am] and are presently under active investigation at the Relativistic Heavy Ion Collider (RHIC). In contrast to inclusive pion production, where several underlying competing mechanisms (mainly the Sivers and Collins ones) cannot be separated, and in close analogy with the semi-inclusive deep inelastic scattering (SIDIS) case, one could discriminate among different effects by taking suitable moments of these asymmetries. In principle, quark and gluon originating jets can also be distinguished, at least in some kinematic regimes. Furthermore, information on the size of TMD distributions and fragmentation functions can be gained, in a kinematic region in which they are still poorly known. This would be very helpful in clarifying the role played by the quark(gluon) Sivers distribution and by the Collins(-like) fragmentation function in the sizable single spin asymmetries observed at RHIC for single inclusive pion production. A similar analysis where transverse partonic motion was considered only in the fragmentation process, aimed at a study of the universality of the Collins function for quarks, was presented in Ref. [@Yuan:2007nd]. Our approach, named generalized parton model (GPM), has in principle a richer structure in the observable azimuthal asymmetries, because intrinsic motion is taken into account in the initial hadrons as well. Preliminary data from the STAR collaboration at RHIC seem to support our model, since they report on a Sivers asymmetry for neutral pions larger than zero [@Poljak:2010tm; @Poljak:2011vu] and compatible with our predictions. However, since factorization has not been proven in this case, but is rather taken as a reasonable phenomenological assumption, the validity of the scheme and the universality of the TMD distributions involved require a severe scrutiny by further comparison with experiments.
\[sec-results\] Theoretical framework
=====================================
We consider the process $p^{\uparrow}p\to{\rm jet}+\pi+X$, with one of the protons in a transverse spin state described by the four-vector $S$. We work in the $p\,p$ c.m. frame, where the polarized proton moves along the $+\hat{\bm{Z}}_{\rm cm}$ direction, and define $(XZ)_{\rm cm}$ as the production plane containing the colliding beams and the observed jet, with $(\bm{p}_{\rm j})_{X_{\rm cm}}>0$. In this frame $S = (0, \cos\phi_{S},\sin\phi_{S},0) $ and $p_{\rm j} =
p_{{\rm j}\,T}(\cosh \eta_{\rm j},1,0,\sinh \eta_{\rm j})$, where $\eta_{\rm j} = -\log[\tan(\theta_{\rm j}/2)]$ is the jet (pseudo)rapidity. We denote by $z$ and $\bm{k}_{\perp\pi}$ respectively the light-cone momentum fraction and the transverse momentum of the observed pion inside the jet w.r.t. the jet (fragmenting parton) direction of motion. Calculations have been performed within the GPM framework at leading-order (LO) in pQCD utilizing the helicity formalism. More details can be found in Ref. [@D'Alesio:2010am].
The final expression for the single-transverse polarized cross section has the following general structure: $$\begin{aligned}
2{\rm d}\sigma(\phi_{S},\phi_\pi^H) &\sim & {\rm d}\sigma_0
+{\rm d}\Delta\sigma_0\sin\phi_{S}+
{\rm d}\sigma_1\cos\phi_\pi^H+ {\rm d}\sigma_2\cos2\phi_\pi^H+
{\rm d}\Delta\sigma_{1}^{-}\sin(\phi_{S}-\phi_\pi^H)
\nonumber\\
&+& {\rm d}\Delta\sigma_{1}^{+}\sin(\phi_{S}+\phi_\pi^H)
+{\rm d}\Delta\sigma_{2}^{-}\sin(\phi_{S}-2\phi_\pi^H)+
{\rm d}\Delta\sigma_{2}^{+}\sin(\phi_{S}+2\phi_\pi^H)\,,
\label{d-sig-phi-SA}\end{aligned}$$ where $\phi_\pi^H$ is the azimuthal angle of the pion three-momentum around the jet axis, as measured in the fragmenting parton helicity frame [@D'Alesio:2010am]. In close analogy with the SIDIS case, in order to single out the different contributions of interest, we can define appropriate azimuthal moments, $$\begin{aligned}
A_N^{W(\phi_{S},\phi_\pi^H)}(\bm{p}_{\rm j},z,k_{\perp\pi})
&=&
2\,\frac{\int{\rm d}\phi_{S}{\rm d}\phi_\pi^H\,
W(\phi_{S},\phi_\pi^H)\,[{\rm d}\sigma(\phi_{S},\phi_\pi^H)-
{\rm d}\sigma(\phi_{S}+\pi,\phi_\pi^H)]}
{\int{\rm d}\phi_{S}{\rm d}\phi_\pi^H\,
[{\rm d}\sigma(\phi_{S},\phi_\pi^H)+
{\rm d}\sigma(\phi_{S}+\pi,\phi_\pi^H)]}\,,
\label{gen-mom}\end{aligned}$$ where $W(\phi_{S},\phi_\pi^H)$ is some appropriate circular function of $\phi_{S}$ and $\phi_\pi^H$.
Phenomenology
=============
In this section we present and discuss some phenomenological implications of our approach in kinematic configurations accessible at RHIC by the STAR and PHENIX experiments. We consider both central ($\eta_{\rm j}=0$) and forward ($\eta_{\rm j}=3.3$) (pseudo)rapidity configurations at a c.m. energy $\sqrt{s} =$ 200 GeV (different c.m. energies, namely $\sqrt{s}=62.4$ and 500 GeV, are also studied in [@D'Alesio:2010am]).
For numerical calculations all TMD distribution and fragmentation functions are taken in the simplified form where the functional dependence on the parton light-cone momentum fraction and on transverse motion are completely factorized, assuming a Gaussian-like flavour-independent shape for the transverse momentum component. Concerning the parameterizations of the transversity and quark Sivers distributions, and of the Collins functions, we consider two sets: SIDIS 1 [@Anselmino:2005ea; @Anselmino:2007fs] and SIDIS 2 [@Anselmino:2008sga; @Anselmino:2008jk]. Furthermore, for the usual collinear distributions, we adopt the LO unpolarized set GRV98 [@Gluck:1998xa]. For fragmentation functions, we adopt two well-known LO sets among those available in the literature, the set by Kretzer [@Kretzer:2000yf] and the DSS one [@deFlorian:2007aj]. Our choice is dictated by the subsequent use of the two available parametrization sets for the Sivers and Collins functions in our scheme, that have been extracted in the past years by adopting these sets of fragmentation functions. Since the range of the jet transverse momentum (the hard scale) covered is significant, we take into account proper evolution with scale. On the other hand, the transverse momentum component of all TMD functions is kept fixed with no evolution with scale. In all cases considered, $\bm{k}_{\perp\pi}$ is integrated over and, since we are interested in leading particles inside the jet, we present results obtained integrating the light-cone momentum fraction of the observed hadron, $z$, in the range $z\geq 0.3$.
We have considered first, for $\pi^+$ production only, a scenario in which the effects of all TMD functions are over-maximized. By this we mean that all TMD functions are maximized in size by imposing natural positivity bounds. The transversity distribution has been fixed at the initial scale by saturating the Soffer bound and then we let it evolve. Moreover, the relative signs of all active partonic contributions are chosen so that they sum up additively. In this way we set an upper bound on the absolute value of any of the effects playing a potential role in the azimuthal asymmetries. Therefore, all effects that are negligible or even marginal in this scenario may be directly discarded in subsequent refined phenomenological analyses. See Ref. [@D'Alesio:2010am] for a more detailed discussion.
As a second step in our study we consider, for both neutral and charged pions, only the surviving effects, involving TMD functions for which parameterizations are available from independent fits to other spin and azimuthal asymmetries data in SIDIS, Drell-Yan (DY), and $e^+e^-$ processes.
![ The estimated quark Collins asymmetry for the $p^\uparrow p\to {\rm jet}+\pi + X$ process, obtained adopting the parameterizations SIDIS 1 and SIDIS 2 respectively, at $\sqrt{s}=200$ GeV in the forward rapidity region. The dotted black vertical line delimits the region $x_F \approx 0.3$. \[asy-an-coll-par200\] ](asy_coll_par_SIDIS1.eps "fig:"){width="40.00000%"} ![ The estimated quark Collins asymmetry for the $p^\uparrow p\to {\rm jet}+\pi + X$ process, obtained adopting the parameterizations SIDIS 1 and SIDIS 2 respectively, at $\sqrt{s}=200$ GeV in the forward rapidity region. The dotted black vertical line delimits the region $x_F \approx 0.3$. \[asy-an-coll-par200\] ](asy_coll_par_SIDIS2.eps "fig:"){width="40.00000%"}
![The estimated quark and gluon contributions to the Sivers asymmetry for the $p^\uparrow p\to {\rm jet}+\pi + X$ process, obtained adopting the parametrization sets SIDIS 1 and SIDIS 2, at forward rapidity and $\sqrt{s}=200$ GeV. The dotted black vertical line delimits the region $x_F\approx 0.3$. \[asy-an-siv-par200\] ](asy_siv_par_pip.eps "fig:"){width="35.00000%"} ![The estimated quark and gluon contributions to the Sivers asymmetry for the $p^\uparrow p\to {\rm jet}+\pi + X$ process, obtained adopting the parametrization sets SIDIS 1 and SIDIS 2, at forward rapidity and $\sqrt{s}=200$ GeV. The dotted black vertical line delimits the region $x_F\approx 0.3$. \[asy-an-siv-par200\] ](asy_siv_par_pi0.eps "fig:"){width="35.00000%"} ![The estimated quark and gluon contributions to the Sivers asymmetry for the $p^\uparrow p\to {\rm jet}+\pi + X$ process, obtained adopting the parametrization sets SIDIS 1 and SIDIS 2, at forward rapidity and $\sqrt{s}=200$ GeV. The dotted black vertical line delimits the region $x_F\approx 0.3$. \[asy-an-siv-par200\] ](asy_siv_par_pim.eps "fig:"){width="35.00000%"}
In Fig. \[asy-an-coll-par200\] we present, in the forward rapidity region, the quark generated asymmetry $A_N^{\sin(\phi_{S}-\phi_\pi^H)}$, which comes mainly from the convolution between the transversity distribution and the Collins fragmentation function. The plots are obtained adopting the parameterizations SIDIS 1 and SIDIS 2. The Collins asymmetry for neutral pions results to be almost vanishing, in agreement with preliminary RHIC data [@Poljak:2010tm; @Poljak:2011vu]. The dotted black vertical line delimits the region $x_F \approx 0.3$, with $x_F=2 p_{{\rm j}\,L}/\sqrt{s}$, beyond which the SIDIS parameterizations for transversity are extrapolated outside the Bjorken $x$ region covered by SIDIS data and are therefore plagued by large uncertainties. For this reason, the two parameterizations for charged pions give comparable results only below this limit (notice the difference in scale between the two panels). A measurement of such asymmetries would be then very important and helpful in clarifying the large $x$ behaviour of the quark transversity distribution. Furthermore, it turns out that in the central rapidity region the quark Collins asymmetries are practically negligible.
In Fig. \[asy-an-siv-par200\] we show, for both neutral and charged pions, the quark and gluon contributions to the Sivers asymmetry $A_N^{\sin\phi_{S}}$, which cannot be disentangled, in the forward rapidity region as a function of $p_{{\rm j}\,T}$. The quark term is obtained adopting the SIDIS 1 and SIDIS 2 parameterizations. The almost unknown gluon Sivers function is tentatively taken positive and saturated to an updated version of the bound obtained in Ref. [@Anselmino:2006yq] by considering PHENIX data for the $\pi^0$ transverse single spin asymmetry at mid-rapidity. Similarly to the case of the Collins asymmetry, the two parameterizations give comparable results only in the $p_{{\rm j}\,T}$ region where the behaviour of the quark Sivers distribution is reasonable well constrained by SIDIS data (see again the dotted black vertical line).
Test of the process dependence of the Sivers function
=====================================================
In the GPM approach one applies the TMD distribution and fragmentation functions extracted from SIDIS, considering them to be universal. Very recently [@D'Alesio:2011mc] the azimuthal distribution of leading pions inside jets has been studied allowing for process dependence of the (quark) Sivers function. This is referred to as the colour gauge invariant (CGI) GPM [@Gamberg:2010tj]. Namely, we have taken into account the effects of initial (ISIs) and final state interactions (FSIs) between the active parton and spectator remnants in the different scattering sub-processes. Since the details of such interactions depend on the particular partonic reaction considered, they render the Sivers function non-universal (see [@Gamberg:2010tj] and references therein). The oft-discussed case is the difference between the FSIs in SIDIS and the ISIs in DY scattering which leads to the prediction of an opposite relative sign of the Sivers functions in the two processes. This is considered to be a crucial test of our understanding of single spin asymmetries in QCD and still has to be confirmed by experiments.
When applying similar reasoning to hadron production in $p\,p$ collisions, typically the Sivers function has a more complicated colour factor structure since both ISIs and FSIs contribute. However, in the forward rapidity region the process under study is dominated by only one channel, $qg\to qg$, with the final quark identified with the observed jet. As a consequence, one finds that the predictions for the Sivers asymmetries obtained with and without inclusion of color gauge factors are comparable in size but with [*opposite signs*]{} [@D'Alesio:2011mc], similarly to the DY case. Therefore the experimental observation of a sizable asymmetry could easily discriminate among the two approaches and test the universality properties of the Sivers function in hadronic reactions. Our results are shown in Fig. \[fig1\], where we plot $A_N^{\sin\phi_
{S}}$ as a function of the jet transverse momentum $p_{{\rm j}T}$ at $\eta_{\rm j}=3.3$, for the RHIC energy $\sqrt{s}=500$ GeV, integrated over $\bm{k}_{\perp\pi}$ and $z$ ($z\ge 0.3$).
![The Sivers asymmetry $A_N^{\sin\phi_{S}}$ for the $p^\uparrow p\to {\rm jet} +\pi +X$ process as a function of $p_{{\rm j}T}$, at fixed jet rapidity $\eta_{\rm j}=3.3$, for the RHIC energy $\sqrt{s}=500$ GeV.[]{data-label="fig1"}](asy_jet_pip_500.eps "fig:"){width="35.00000%"} ![The Sivers asymmetry $A_N^{\sin\phi_{S}}$ for the $p^\uparrow p\to {\rm jet} +\pi +X$ process as a function of $p_{{\rm j}T}$, at fixed jet rapidity $\eta_{\rm j}=3.3$, for the RHIC energy $\sqrt{s}=500$ GeV.[]{data-label="fig1"}](asy_jet_pi0_500.eps "fig:"){width="35.00000%"} ![The Sivers asymmetry $A_N^{\sin\phi_{S}}$ for the $p^\uparrow p\to {\rm jet} +\pi +X$ process as a function of $p_{{\rm j}T}$, at fixed jet rapidity $\eta_{\rm j}=3.3$, for the RHIC energy $\sqrt{s}=500$ GeV.[]{data-label="fig1"}](asy_jet_pim_500.eps "fig:"){width="35.00000%"}
The predictions labeled SIDIS 1 and SIDIS 2 are similar in the intermediate $p_{{\rm j}T}\le 5.5$ GeV region (corresponding to $x_F<0.3$). This region is then optimal to test directly the process dependence of the Sivers function. Note that for $\sqrt s = 200$ GeV the behavior of our estimates would be similar to that shown in Fig. \[fig1\], gaining almost a factor of 2 in size. However the range of $p_{{\rm j}T}$ covered would be narrower ($p_{{\rm j}T}\le 6.5$ GeV) and with $x_F \le 0.3$ now corresponding to $p_{{\rm j}T}\le 2.5$ GeV.
Finally, as a natural extension of this analysis, one can consider single inclusive jet asymmetry in $p\,p$ scattering, which is described solely by the Sivers function. The results obtained for $A_N^{\sin\phi_{S}}$ (not shown) look almost indistinguishable from the case of neutral pion-jet production (central panel of Fig. \[fig1\]).
\[sec-conclusions\] Conclusions
===============================
We have presented a study of the azimuthal asymmetries measurable in the distribution of leading pions inside a large-$p_T$ jet produced in single-transverse polarized proton proton collisions for kinematic configurations accessible at RHIC. To this end, we have adopted a generalized TMD parton model approach with inclusion of spin and intrinsic parton motion effects both in the distribution and in the fragmentation sectors. In contrast to inclusive pion production, where the Sivers and Collins mechanisms cannot be separated, the leading-twist azimuthal asymmetries discussed above would allow one to single out the different effects by taking suitable moments of the asymmetries. This will give us the opportunity of testing the factorization and universality assumptions, and of gaining information on the size and *sign* of the various TMD functions in a kinematic region not covered by SIDIS data.
C.P. is supported by Regione Autonoma della Sardegna under grant PO Sardegna FSE 2007-2013, L.R. 7/2007. U.D. and F.M. acknowledge partial support by Italian MIUR under PRIN 2008, and by the European Community (FP7 grant agreement No. 227431).
[20]{}
.
; .
. . . . . . these proceedings, arXiv:1111.0755 (2011).
. . . .
. . .
.
.
.
[^1]: Speaker. Talk given at Third International Workshop on Transverse Polarization Phenomena in Hard Scattering (Transversity 2011), Veli Lošinj, Croatia, 29 August - 2 September 2011.
| {
"pile_set_name": "ArXiv"
} |
---
author:
- |
Jun Xiang$^1$, Guoshuai Zhang$^1$, Jianhua Hou$^1$, Sang Nong$^2$, Rui Huang$^3$\
$^1$ South-Central University for Nationalities, Wuhan, China\
$^2$ Huazhong University of Science and Technology, Wuhan, China\
$^3$ The Chinese University of Hong Kong, Shenzhen, China\
[email protected], [email protected], [email protected], [email protected],
bibliography:
- 'main.bib'
title: Multiple Target Tracking by Learning Feature Representation and Distance Metric Jointly
---
| {
"pile_set_name": "ArXiv"
} |
The behaviour of a trapped Bose-Einstein condensate under rotation has attracted much attention in the recent years. References [@FS; @GRPG; @Feder; @Wilkin; @Rokhsar; @Ben; @BP; @WG; @LF; @KMP; @JKMR; @JK; @NU] have dealt with this problem theoretically both in the Thomas-Fermi limit of strong interactions, as well as in the limit of weak interactions. Borrowed from the field of nuclear physics, the terminology “yrast state" refers to the state with the lowest energy for a given angular momentum $L$ that the system has. Of equal importance to the yrast state are the low-lying excitations, which determine the thermodynamic behaviour of the system, and they are also crucial for the stability of the yrast states under external perturbations.
In the present study we consider the hamiltonian $H= H_0 + V$. Here $$\begin{aligned}
H_0 = \sum_i \left[ - \frac {\hbar^{2}} {2M} {{\mbox{\boldmath $\nabla$}}}_{i}^{2} +
\frac 1 2 \, M \omega^{2} r_{i}^{2} \right]
\label{h0}\end{aligned}$$ includes the kinetic energy of the particles and the potential energy due to the trapping potential, where $M$ is the atom mass, and $\omega$ is the frequency of the harmonic confining potential, which we assume to be isotropic. Also the interaction between the particles $V$ is assumed to be short-ranged, $$\begin{aligned}
V = \frac 1 2 U_{0} \sum_{i \neq j} \delta({\bf r}_{i} - {\bf r}_{j}),
\label{int}\end{aligned}$$ where $U_0 = 4 \pi \hbar^2 a/M$ is the matrix element for atom-atom collisions, with $a$ being the scattering length. We consider the limit of weak interactions, where the typical interaction energy is much less than the typical oscillator quantum of energy, $$\begin{aligned}
n U_0 \ll \hbar \omega,
\label{conditionwi}\end{aligned}$$ with $n$ being the density of atoms. The above condition allows us to work in the subspace of single-particle states with no radial excitations, $$\Phi_m({\bf r}) = \frac 1 {(m! \pi a_0^3)^{1/2}}
\left( \frac {\rho}{a_0} \right)^{|m|} e^{i m \phi}
e^{-(\rho^{2}+z^2)/2 a_0^2}.
\label{phim}$$ Here $\rho, z$, and $\phi$ are cylindrical polar coordinates, and $a_0 = (\hbar/M \omega)^{1/2}$ is the oscillator length. These states are degenerate in the absence of interactions (while states with radial excitations lie higher by an energy of order $\hbar \omega$). The whole problem thus reduces to incorporating the interactions between the atoms, which lift the degeneracy of the states.
One of us has determined in Ref.[@Ben] the yrast line of a weakly-interacting Bose-Einstein condensate for an effective attraction between the atoms. In the same reference the yrast line has also been determined for the case of an effective repulsion between the atoms and for $L \ll N$. In Ref.[@KMP] we have given a more detailed description of the yrast line within the mean-field Gross-Pitaevskii approximation for a wide range of values of the ratio $L/N$.
If one studies the limit $L \ll N$, as shown in Ref.[@Ben], in a state with a $2^{\lambda}$-pole excitation, $$\begin{aligned}
| \lambda \rangle = | (m=0)^{N-1}, (m=\lambda)^{1} \rangle,
\label{statel}\end{aligned}$$ where $m$ is the state with angular momentum $m \hbar$, the excitation (interaction) energy is $$\begin{aligned}
\epsilon_{\lambda} = \langle \lambda | V | \lambda \rangle -
\langle 0 | V | 0 \rangle =
- N v_0 \left( 1 - \frac 1 {2^{\lambda - 1}} \right),
\label{elambda}\end{aligned}$$ where $| 0 \rangle = | (m=0)^N \rangle$ is the ground state, and $v_0 = U_0/a_0^3$. The above equation implies that the highest gain in the interaction energy per unit of angular momentum comes from the $\lambda = 2$ or 3 excitations, and therefore the quadrupole and octupole excitations are expected to carry the angular momentum for $L
\ll N$. As a result, in this limit there is a quasi-degeneracy between the low-lying states, as we discuss below in detail. By saying quasi-degenerate states, we mean that the energy separation between them is of order $v_0$, and not of order $N v_0$.
Bertsch and Papenbrock [@BP] have examined the ground state of a weakly-interacting Bose-Einstein condensate under rotation numerically by diagonalizing the hamiltonian $H$ in the subspace of degenerate states (\[phim\]) for a given angular momentum. In Ref.[@NU] Nakajima and Ueda have performed similar numerical calculations in the limit where the angular momentum per particle $L/N$ is much less than 1, and have found that the quasi-degenerate states which we mentioned in the previous paragraph lie above the yrast by an energy, which in the asymptotic limit $N \rightarrow \infty$ is given by $1.59 N_3 (N_3-1) v_0/4$, where $N_3$ is the number of octupole excitations.
In this study we give an analytical derivation of this result with use of a diagrammatic perturbation-theory approach. The starting point of our analysis is the fact that the quadrupole and octupole $\lambda =2,3$ excitations are dominant for $L \ll N$ [@Ben; @KMP], and we therefore assume that in this limit the angular momentum is carried by $\lambda =2$ and $\lambda =3$ excitations only. In addition, the condensate is dominated by atoms which do not have any angular momentum. Our approach therefore consists of considering a condensate with $N_0$ atoms in the state $m=0$, $N_2$ atoms in the state with $m=2$, and $N_3$ in the state with $m=3$, and then treating the other states perturbatively by keeping the appropriate diagrams and using perturbation theory to get the correction to the energy. The “bare" interaction energy with particles in the $m=0$, 2 and 3 states is given by (see Fig.1 and Table I) $$\begin{aligned}
{\cal E}^{(0)} = \frac 1 2 N_0 (N_0 - 1)&+&\frac 3 {16} N_2 (N_2 -1)
+ \frac 5 {32} N_3 (N_3 -1) +
\nonumber \\
&+&\frac 1 2 N_0 N_2 + \frac 1 4
N_0 N_3 + \frac 5 8 N_2 N_3,
\label{bare}\end{aligned}$$ where the interaction energy is measured in units of $v_0$. The diagrams shown in Fig.2 lower the energy, as second-order perturbation theory implies. For example, for the process shown in Fig.2(a) where two particles with $m=2$ scatter to a state with $m=0$ and $m=4$, and then back to the initial state, the matrix element for each vertex is $$\begin{aligned}
{\cal M}=\frac {\sqrt 6} {16} \sqrt {N_2 (N_2-1) (N_0+1) (N_4+1)} \,\, v_{0},
\label{mee}\end{aligned}$$ where $N_m$ is the number of atoms with angular momentum $m \hbar$. In Eq.(\[mee\]) we have included the factor $N_4+1$ for clarity, although as we mentioned above $N_4$ is assumed to be equal to zero. The result of Eq.(\[mee\]) is most easily derived by writing the interaction energy, Eq.(\[int\]), as $$\begin{aligned}
V = \frac 1 2 U_0 \sum_{i,j,k,l}
I_{i,j,k,l} \, a_i^\dagger a_j^\dagger a_k a_l,
\label{intersq}\end{aligned}$$ where $a_i$ and $a_i^\dagger$ are annihilation and creation operators respectively, and $$\begin{aligned}
I_{i,j,k,l} &=&
\int \Phi_i^*({\bf r}) \Phi_j^*({\bf r}) \Phi_k({\bf r})
\Phi_l({\bf r}) \, d{\bf r} \phantom{XXX}
\nonumber \\ \phantom{XXX}
&=& \frac {(i+j)!} {2^{(i+j)}
\sqrt{i!\, j! \, k! \, l!}} \int |\Phi_0({\bf r})|^4 d{\bf r}
\label{integral}\end{aligned}$$ when $i+j=k+l$ and zero otherwise. In the case we consider here only the integrals $I_{0,4,2,2}$ and $I_{4,0,2,2}$ give a non-zero result reflecting the symmetrization of the bosonic wavefunction.
As Eq.(\[elambda\]) implies, the difference between the energy in the intermediate and the initial states is $\Delta E= \epsilon_4 +
\epsilon_0 - 2 \epsilon_2 = -N v_{0}/8$. Therefore, according to perturbation theory, the correction to the energy is $$\begin{aligned}
\frac {|{\cal M}|^{2}} {\Delta E} = - \frac 3 {16} N_2 (N_2-1) v_{0}
\label{mnde}\end{aligned}$$ in the limit $L \ll N$. Table I gives the correction to the energy for the processes shown in Fig.2. Adding these terms to ${\cal E}^{(0)}$, we see that the corrected interaction energy ${\cal E}^{(1)}$ is, in units of $v_0$ and to leading order in $N^{-1}$, $$\begin{aligned}
{\cal E}^{(1)}&=&\frac 1 2 N_0 (N_0 - 1) +
\frac 5 {34} N_3 (N_3 -1) + \frac 1 2 N_0 N_2 +
\nonumber \\ &+&\frac 1 4
N_0 N_3 - \frac 1 4 N_2 N_3 - \frac 3 4 N_3 - \frac 1 2 N_2.
\label{barec}\end{aligned}$$ To conserve particle number and angular momentum, we have the following constraints $$\begin{aligned}
N_0 + N_2 + N_3 = N \,\,\, {\rm and} \,\,\, 2 N_2 + 3 N_3 =L.
\label{cond}\end{aligned}$$ Using them, we express ${\cal E}^{(1)}$ in terms of $N, L$, and $N_3$, thus finding $$\begin{aligned}
{\cal E}^{(1)} = \frac 1 4 N (2 N - L -2) + \frac {27} {68} N_3 (N_3 -1).
\label{barecf}\end{aligned}$$ The number 27/68 coincides with the numerical result 1.59/4 reported in Ref.[@NU]. Here $N_3$ can take all the non-negative integer values that are consistent with the constraints of Eq.(\[cond\]). Thus the number of excited states described by Eqs.(\[cond\]) and (\[barecf\]) is equal to the integer part of $L/6$. These statements are exact asymptotically, i.e., for $N \rightarrow \infty$, since there are processes which couple the quadrupole $m=2$ excitations with the octupole $m=3$ excitations and they contribute terms of order $1/N$ to the excitation energy. Concerning the yrast state, a consequence of Eq.(\[barecf\]) is that its energy is given by the expression $$\begin{aligned}
{\cal E}_0 = \frac N 4 (2 N - L -2),
\label{2syra}\end{aligned}$$ in agreement with Refs.[@BP; @KMP; @JK]. Finally a result of Eq.(\[barecf\]) is that the yrast state for $L \ll N$ is dominated by quadrupole excitations, i.e., to leading order, $$\begin{aligned}
\frac {N_2} N = \frac 1 2 \frac L N .
\label{2syr}\end{aligned}$$
In summary we have developed an effective theory which describes the ground state and the low-lying excited states of a weakly-interacting Bose-Einstein condensate under rotation, in the limit where the angular momentum is much less than the number of atoms. This study demonstrates that there are low-lying excited states which differ by an energy of order $v_0$ (and not $N v_0$), and we have found agreement with a previous numerical study [@NU] of the same problem. We should point out that for all the values of $L/N$ we have examined numerically, except the case $L/N \ll 1$, we found that the low-lying excited states are separated from the yrast state by an energy of order $N v_0$, and in that respect the limit $L \ll N$ seems to be unique.
We would like to thank C. Pethick, A. Jackson and M. Koskinen for useful discussions. G.M.K. would like to thank the Foundation of Research and Technology, Hellas (FORTH) for its hospitality. S.M.R. would like to acknowledge financial support from the “Bayerische Staatsministerium für Wissenschaft, Forschung und Kunst".
[The six diagrams contributing to the bare interaction energy ${\cal E}^{(0)}$, Eq.(\[bare\]). The straight lines denote atoms with angular momentum given by the numbers written above the lines; dashed lines denote the interaction.]{}
\[FIG1\]
[The five (additional) diagrams contributing to the interaction energy ${\cal E}^{(1)}$, Eq.(\[barecf\]).]{}
\[FIG2\]
A. A. Svidzinsky and A. L. Fetter, Phys. Rev. Lett. [**84**]{}, 5919 (2000).
J. J. García-Ripoll and V. M. Pérez-García Phys. Rev. A [**60**]{}, 4864 (1999).
D. L. Feder, C. W. Clark, and B. I. Schneider, Phys. Rev. A [**61**]{}, 011601(R) (1999); Phys. Rev. Lett. [**82**]{}, 4956 (1999).
N. K. Wilkin, J. M. F. Gunn, and R. A. Smith, Phys. Rev. Lett. [**80**]{}, 2265 (1998).
D. A. Butts, and D. S. Rokhsar, Nature [**397**]{}, 327 (1999).
B. Mottelson, Phys. Rev. Lett. [**83**]{}, 2695 (1999).
G. F. Bertsch, and T. Papenbrock, Phys. Rev. Lett. [**83**]{}, 5412 (1999).
N. K. Wilkin and J. M. F. Gunn, Phys. Rev. Lett. [**84**]{}, 6 (2000).
M. Linn, and A. L. Fetter, Phys. Rev. A [**60**]{}, 4910 (1999).
G. M. Kavoulakis, B. Mottelson, and C. J. Pethick, Phys. Rev. A [**62**]{}, 63605 (2000).
A. D. Jackson, G. M. Kavoulakis, B. Mottelson, and S. M. Reimann, e-print cond-mat/0004309.
A. D. Jackson and G. M. Kavoulakis, Phys. Rev. Lett. [**85**]{}, 2854 (2000).
T. Nakajima and M. Ueda, e-print cond-mat/0007004.
[Diagram]{} [Energy]{}
------------- ----------------------
1(a) $N_0(N_0-1)/2$
1(b) $3 N_2(N_2-1)/16$
1(c) $5 N_3(N_3-1)/32$
1(d) $N_0 N_2/2$
1(e) $N_0 N_3/4$
1(f) $5 N_2 N_3/8$
2(a) $-3 N_2 (N_2-1)/16$
2(b) $-5 N_3 (N_3-1)/544$
2(c) $-N_2 N_3/8$
2(d) $-N_2/2$
2(e) $-3 N_3 (N_2+1)/4$
: The contribution of the diagrams shown in Figs.1 and 2 to the interaction energy (in units of $v_0$).[]{data-label="table"}
0.5pc
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We propose a robust dynamic walking controller consisting of a dynamic locomotion planner, a reinforcement learning process for robustness, and a novel whole-body locomotion controller (WBLC). Previous approaches specify either the position or the timing of steps, however, the proposed locomotion planner simultaneously computes both of these parameters as locomotion outputs. Our locomotion strategy relies on devising a reinforcement learning (RL) approach for robust walking. The learned policy generates multi step walking patterns, and the process is quick enough to be suitable for real-time controls. For learning, we devise an RL strategy that uses a phase space planner (PSP) and a linear inverted pendulum model to make the problem tractable and very fast. Then, the learned policy is used to provide goal-based commands to the WBLC, which calculates the torque commands to be executed in full-humanoid robots. The WBLC combines multiple prioritized tasks and calculates the associated reaction forces based on practical inequality constraints. The novel formulation includes efficient calculation of the time derivatives of various Jacobians. This provides high-fidelity dynamic control of fast motions. More specifically, we compute the time derivative of the Jacobian for various tasks and the Jacobian of the centroidal momentum task by utilizing Lie group operators and operational space dynamics respectively. The integration of RL-PSP and the WBLC provides highly robust, versatile, and practical locomotion including steering while walking and handling push disturbances of up to 520 N during an interval of 0.1 sec. Theoretical and numerical results are tested through a 3D physics-based simulation of the humanoid robot Valkyrie.'
author:
- 'Donghyun Kim, Jaemin Lee, and Luis Sentis, [^1][^2][^3]'
bibliography:
- 'tro2017.bib'
title: Robust Dynamic Locomotion via Reinforcement Learning and Novel Whole Body Controller
---
[Shell : Bare Demo of IEEEtran.cls for IEEE Journals]{}
Biped Locomotion, Reinforcement Learning, Whole Body Control
Introduction
============
We explore the performance of a reinforcement learning (RL) process and a new whole-body impedance and force controller for robust dynamic locomotion on a full biped humanoid dynamical model. Full-bodied 3D humanoid dynamic walking based on inverted pendulum (IP) dynamics and RL has not been studied to date. To enable the RL process to run efficiently, we found that utilizing phase-space planning (PSP) [@Zhao:2012de] provides a space of practical parameters that enables the transition function to operate in the reduced inverted pendulum manifold. A key advantage of using an IP model is that it generalizes locomotion for many types of systems with a light dependency with their concrete kinematic structure. Using an IP model not only reduces the search space but also enables the same procedure to be used across different types of full-bodied bipedal humanoid robots. A closely related work to ours, [@MacAlpine:2012vp] utilizes a trajectory parametrization of the dynamic locomotion process [@Graf:2009uz], specifically customized for a particular robot, the NAO. As such, our study can be viewed as a generalization of this type of work to more general models typically used in the dynamic walking communities [@dynamic-walking-2017] – e.g. inverted pendulums. In addition, the previous line of work on NAO robots, has remained quiet with respect to quantifying robustness. In the work presented here, we address robustness as a main thrust performing detailed simulations with respect to unplanned and large disturbances at moderate walking speeds.
Since IP dynamics are in a reduced manifold, we propose a new type of whole-body controllers that are highly robust and effective to transfer the IP-based locomotion process into the full humanoid model. In particular we incorporate new efficiently-computed feedforward terms, momentum and balancing tasks, and more accurate contact models while maintaining the key capability of task prioritization. In addition we completely reformulate the control structures for whole-body control with respect to previous work of ours on this area of whole-body control. We accommodate for the new models and also achieve high computational efficiency. As such, we build on our long time history of devising whole-body controllers, this time around making significant algorithmic changes. We believe these transformations constitute a quantum leap with respect to whole-body controllers with dynamic locomotion capabilities.
The combination of RL-IP-PSP for locomotion pattern generation achieves significant robustness by training a neural network through an actor-critic process with many possible center-of-mass states, representing potential disturbances, then learning successful step timing and foot position policies for recovery. Utilizing step timing and foot positions is not typically explored because simultaneously varying both of these parameters results in nonlinear system dynamics. [@Herdt:2010bh] proposed a model predictive control (MPC) method for synthesizing walking patterns based on desired foot positions, kinematic limits, and given step timing. [@Kryczka:2015ck] formulated a nonlinear optimization problem to solve two walking steps ahead of time to reduce computational cost. [@Khadiv:2016hm] linearized an optimization problem by searching for a solution one step ahead of time. In contrast, instead of relying on runtime optimizations, we train a control policy offline using an IP model, and use it afterwards for real-time control of full humanoid robots being physically disturbed. Therefore, our learned locomotion planning generator can plan hundreds of steps ahead of time in an instant, compared to the stepping time scale. As such, speed is a key characteristic of the proposed planning and control framework compared to the state of the art.
Devising a new robust dynamic locomotion generator is insufficient to be directly used in full humanoid robots. Therefore building on our expertise in this area, we devise a new type of whole-body controller (WBC) which we call whole-body locomotion controller (WBLC) that focuses on speed, unilateral contact constraints, and speedy prioritized task control. The proposed WBC algorithm enables to efficiently compute projection-based hierarchical task controllers [@Sentis:COl04Q1j] and at the same time incorporate contact inequality constraints which are represented by a quadratic programming (QP) process[@koolen2013summary; @feng2015optimization; @kuindersma2014efficiently], hierarchical quadratic programming (HQP) [@saab2013dynamic]. While QP based controllers have been very successful for field application their computational cost is considerably higher than that of projection-based methods. In contrast, projection-based methods have not incorporated before inequality constraints such as unilateral contact and friction constraints. Our proposed WBC algorithm combines for first time QP and projection-based methods.
![[**Type of humanoid platform our controller explores.**]{} The left image shows NASA’s Valkyrie humanoid robot, with 135.9 $\si[per-mode=symbol]{\kilo\gram}$ weight and 1.83 $\si[per-mode=symbol]{\meter}$ height. The right image shows our dynamic simulation of Valkyrie using the physics based simulator SrLib.[]{data-label="fig:valkyrie_figure"}](A_valkyrie_figure){width="0.8\columnwidth"}
In our study, we introduce a centroidal angular momentum (CAM) tasks to improve agile locomotion performance and we examine its effect when used in the whole-body control hierarchy. For example, when a robot quickly shakes or rotates its body, we reduced undesirable arm movements that result from angular momentum compensation by introducing an arm motion task with higher priority than the CAM task.
In the proposed WBLC, we devise new projection-based recursive structures that incorporate unilateral contact and friction constraints, yielding the desired reduction on the computational cost compared to other QP based algorithms. However, we don’t only achieve computational efficiency by combining QP and projection-based methods. We do achieve it by important improvements on the computation of the projection-based operations themselves.
Indeed, in conventional projection-based methods, the computational cost of some operations is considerably high. For instance, one well-known WBC algorithm that uses joint acceleration, [@siciliano1991general], includes costly terms such as the time derivative of a null space projection matrix. In addition, many WBC algorithms contain computations for Coriolis/centrifugal and gravitational forces projected on in operational task space [@sentis2005synthesis; @mansard2009unified], which are costly to calculate specially as the number of control tasks increases. Our WBLC algorithms eliminates this problems. An analytic solution of the time derivatives of Jacobians is devised by employing Lie group operators, and an implementation using the Rigid Body Dynamic Library. We also eliminate the need to compute Coriolis/centrifugal terms for every task priority.
![[**PIPM and CoM phase plots.**]{} (a) PIPM 3D position moving on a variable height surface. (b) Overlapped PIPM phase plots corresponding CoM paths during dynamic walking. In the sagittal plane, we can see multiple parabolas connected to each other corresponding to various walking steps. Parabolas in the frontal plane produce limit cycles.[]{data-label="fig:psp_concept"}](A_phase_planner_concept){width="\columnwidth"}
Overall, the main contributions of our study are as follows. First, devising a novel learning framework for robust dynamic locomotion under push disturbances achieving virtually instantaneous re-planning of an order of magnitude more steps that the state-of-the-art. Second, we devise an elegant method to introduce steering capabilities to phase-space planning for dynamically moving in all directions. Third, we devise a new whole-body locomotion controllers, which yields the benefits of QP based computation of reaction forces and projection-based prioritized task control. Due to many optimizations, we believe that this controller is one of the fastest WBC’s that fulfills both prioritization and practical inequality constraints. Lastly, we integrate all of these algorithms into a comprehensive software and conduct thorough testing on robust dynamic locomotion under large push disturbances on physics-based simulations of Valkyrie.
The paper is organized as follows. In Section \[sec:rl\_psp\], we describe the RL-IP-PSP process after briefly reviewing related work in Section \[sec:related\_work\]. We then introduce the formulation of WBLC in Section \[sec:wblc\] and the methods to efficiently obtain the time derivative of Jacobians for motion control and CAM tasks. In Section \[sec:result\], we study the effects of our framework in agile task such as shaking the robot’s body, walking while steering, and recovering from large pushes while walking. We do all of this using a model of NASA’s Valkyrie robot and the srLib physics based engine[^4](Fig.\[fig:valkyrie\_figure\]). A more exhaustive review of previous work can be found in Appendix \[sec:related\_work\].
Reinforcement Learning based\
Phase Space Planner {#sec:rl_psp}
=============================
We devise an RL process around a PSP framework, the latter significantly enhancing the learning efficiency by exploiting the inherent directional walking constraints of PSP. PSP generates effective step switching information using simplified models such as the prismatic inverted pendulum model (PIPM). In Fig. \[fig:psp\_concept\], we show phase plots across multiple walking steps of the CoM sagittal and lateral phase portraits based on PIPM dynamics. In the sagittal plane, the path consists of connected parabolas, while in the frontal plane, the walking path follows semi-periodic parabolas in a closed cycle. For convenience, we will use $x$ for the sagittal plane and $y$ for the frontal plane.
Phase Space Planner {#sec:psp_explain}
-------------------
Leading step planning generators, such as Divergent Component of Motion [@Englsberger:2015jp], find CoM paths given step positions and their timing as input information. The ZMP Preview Control method [@Kajita:2003iq] has different mechanics but the generated output can be interpreted as finding the CoM path given step position and timing input information. In contrast, PSP finds the step switching time and lateral foot positions, given sagittal foot positions and apex velocities (Fig. \[fig:psp\_exp\]). The apex states are those at the instant when the sagittal CoM velocity is at its minimum; equivalently, they can be considered as states where the sagittal CoM position is zero in a local frame attached to the stance foot, i.e. below the CoM sagittal position.
In Fig. \[fig:psp\_exp\](a), we can see that the current robot’s CoM state and the next desired apex state uniquely define a switching state [[{width="0.8em"}]{}]{}, a switching time, and an apex time. These timings are used to find the next lateral foot position, $p_{y,2}$. Note that the resultant locomotion trajectory is straight forward if $\dot{x}_{apex}$ is a positive number and $\dot{y}_{apex} =0$. In contrast, the proposed algorithm, applies a simple and elegant modification that allows to dynamically steer the biped in any direction of walking (see Fig. \[fig:psp\_exp\](b)). When we need to turn walking direction, we re-initialize the orientation of the local frame $\{b\}$ to the new direction and project the current state to the new frame. The original PSP algorithm devised locomotion trajectories via numerical integration. However, for algorithmic speed purposes, the methods presented here assume that the CoM height is linear allowing us to exploit an analytical solution (see Appendix. \[sec:append\_psp\]).
Considering a one step ahead plan, an initial CoM state, and desired future states $[p_x,~\dot{x}_{apex},~\dot{y}_{apex}]^{\top}$ PSP finds the next step position and timing, $\begin{bmatrix} p_y, &t_{switch}\end{bmatrix}^{\top}$. Notice that the walking direction is indicated using apex velocities, $[\dot{x}_{apex}, ~\dot{y}_{apex}]^{\top}$. We will now see that our formulation of PSP makes the RL problem more efficient by reducing the dimensionality of the learned state variables.
The Reinforcement Learning Problem {#sec:rl_process}
----------------------------------
As mentioned before, a central part of our walking methodology is to achieve robustness via reinforcement learning. The technique we use is the Actor-Critic with Eligibility Traces method. We summarize this process in Algorithm \[code:rl\_explain\] which is an adaptation [@sutton2011reinforcement].
Initialize policy weights $\mathbf{\theta}$ and state-value weights $\mathbf{w}$
We define $\mathbf{s}$, as CoM apex states, $\mathbf{s} \triangleq \begin{bmatrix} y_{apex}, &\dot{x}_{apex}, &\dot{y}_{apex} \end{bmatrix}^{\top}$. Notice that $\mathbf{s}$ does not include the variable $x_{apex}$ because it is assumed to be always zero in the local frame. We define actions, $\mathbf{a} \triangleq \begin{bmatrix} p_x, &\dot{x}_{apex}, &\dot{y}_{apex} \end{bmatrix}^{\top}$, as input parameters to the PSP process. A transition function, $T(\mathbf{s},~\mathbf{a})$, computes the next apex state, $\mathbf{s}'$, and the instantaneous reward value. In Fig. \[fig:transition\_fn\], we show the transition function, consisting of two stages: 1) finding step timing and position values via PSP, and 2) computing the next apex state via an analytic solution of the linear inverted pendulum model (LIPM). The first stage is described in Appendix \[sec:append\_psp\] and allows to find $t_{switch}$, $t_{apex}$, and $p_y$ from the current apex state and chosen action. The second stage, finds the next apex state using the analytic solution of the CoM dynamics (see Eq. ). In Algorithm \[code:Phase\_Space\] the process of finding the switching times and the next apex states is explained in detail.
The next item, $\hat{v}(\mathbf{s}, \bm{w})$, corresponds to the value function - similar to the cost-to-go function in Dynamic Programming. We store its learned values using a radial basis function (RBF) neural network [@cualinmultidimensional]. The network uses a three-dimensional input vector consisting of the CoM apex state. $$\label{eq:state_range}
\begin{split}
& \cdot ~ -0.14 \leq y_{apex} \leq 0.2 ~(\si[per-mode=symbol]{\meter}),\\[1mm]
& \cdot ~ \ \ 0.03 \leq \dot{x}_{apex} \leq 0.61 ~(\si[per-mode=symbol]{\meter\per\second}), \\[1mm]
& \cdot ~ -0.55 \leq \dot{y}_{apex} \leq 0.55 ~(\si[per-mode=symbol]{\meter\per\second}).
\end{split}$$ The hidden layer consists of a bias term and $18\times 30 \times 56$ Gaussian functions with centers on a grid with 2$\si[per-mode=symbol]{\centi\meter}$ spacing along each input dimension. The policy function also consists of an RBF neural network but a little different from the value function because of actions are chosen based on an stochastic evaluation.
![[**Transition Function.**]{} The transition function relies on two models: PSP and LIPM. Given an apex state and a considered action, PSP computes step timing and location information that serve as inputs to the LIPM model. Then, LIPM solves for the next apex state based on the given state and the provided inputs.[]{data-label="fig:transition_fn"}](A_transition_fn){width="\columnwidth"}
Fig. \[fig:rb\_policy\] shows that outputs of the RBF network are means and standard deviations of truncated normal distributions, $\pi(\mathbf{a}|\mathbf{s}, \bm{\theta})$. The range of the distributions are selected by considering the desired walking speed and step length limits as follows: $$\begin{split}
& \cdot ~ 0.1 \leq p_x \leq 0.5 ~ (\si[per-mode=symbol]{\meter}),\\[1mm]
& \cdot ~ 0.03 \leq \dot{x}_{apex} \leq 0.37 ~(\si[per-mode=symbol]{\meter\per\second}),\\[1mm]
& \cdot ~ -0.25 \leq \dot{y}_{apex} \leq 0.25 ~(\si[per-mode=symbol]{\meter\per\second}).
\end{split}$$ The network’s outputs are linearly weighted by $\bm{\theta}$; thus, the purpose of RL is to find the weights $\bm{\theta}$ given candidate actions that minimized the desired cost.
The instantaneous reward is defined by the forward velocity error and lateral step size error: $$R = -(\dot{x}_{apex}^{nom} - \dot{x}_{apex})^2 - 15 \times (p_y^{nom} - p_y)^2 -(\dot{y}_{apex})^2.$$ The set target for the learning process is to achieve recovery behaviors that maintain a straight forward direction, $\dot{y}_{apex} = 0$ while keeping a nominal lateral directional step size. We choose $\dot{x}_{apex}^{nom} = 0.2\si[per-mode=symbol]{\meter\per\second}$ and $p_{y}^{nom} = 0.3\si[per-mode=symbol]{\meter}$. The reward comes from the transition function described before, given current apex and action states selected from the truncated distributions.
If the next predicted apex state incurs a terminal condition, the transition function gives a negative reward of $-5.0$, and the process terminates and starts a new iteration. The set of safe conditions (i.e. opposite to the terminal conditions) is the intersection of the following predicates: $$\label{eq:terminal_cond}
\begin{split}
&\cdot ~ t_{apex} > 0.12 \ (\si[per-mode=symbol]{\second}),\\[1mm]
&\cdot ~ t_{switch} > 0.12 \ (\si[per-mode=symbol]{\second}),\\[1mm]
&\cdot ~ 0.1 < p_{y} < 0.5 \ (\si[per-mode=symbol]{\meter}),
\end{split}$$ which reflect the ability of the robot to swing its legs and the lateral step length. Notice that we do not include a predicate about the sagittal step length because it is already bounded by the allowable action range. The learning process ends when the variance of the learned policy becomes small enough ($<0.07$ units in our case). The usual number of iterations required to complete the learning process is about 30,000, and the process usually takes about 1 $\si[per-mode=symbol]{\minute}$ to compute on a dual-core, 3.0 GHz, Intel i7 processor thanks to the speed of our analytic PSP method.
![ [**Radial Basis Function Neural Network for Walking Policy Representation.**]{} Outputs of the neural network are means and standard deviations of each action value. The truncated normal distributions defined by the outputs are used to stochastically pick actions.[]{data-label="fig:rb_policy"}](A_radial_basis_exp){width="\columnwidth"}
Evaluation of Learned Policy
----------------------------
Fig. \[fig:rl\_check\] shows that the performance of the RL-based planner increases as the number of iterations increases. By watching the posture of a robot at different CoM states, we choose the nominal apex state to be $\begin{bmatrix} y_{apex}, &\dot{x}_{apex}, &\dot{y}_{apex} \end{bmatrix}=[~0.056, ~0.2, ~0~]^{\top}$. We proceed by simulating push disturbances to the CoM based on various external forces and directions. We use mean values of the final learned policy as desired actions rather than randomly picking actions from the normal distributions. The results are shown in Fig. \[fig:rl\_check\] showing the learned policy obtained after many iterations and their enhancements on the walking patterns. In this figure, initially our simulated robot stands with the right foot on the ground and we simulate push disturbances to the left, right, and forward directions of its body. For example, the [[{width="0.8em"}]{}]{} post impulse apex state, $[~0.05, ~0.39, ~0.33~]^{\top}$, is the result of an impulse applied to the left-forward direction of the robot’s body. Red lines are interrupted within a few walking step indicating that the initial policies fail to find proper actions. In contrast, pink lines correspond to the final learned policy which achieves infinite walking steps without falling given the initial push disturbances.
Whole Body Locomotion Control {#sec:wblc}
=============================
We devise a new whole-body locomotion control algorithm, dubbed WBLC, that specifies tasks using a hierarchy of accelerations and uses quadratic programming to determine contact forces. Fig. \[fig:wblc\] describes the overall process for computing the torque commands. The details are described below.
Acceleration-based Formula with Hierarchy
-----------------------------------------
Task level controllers are computed in operational space as acceleration commands and converted to joint accelerations using differential forward kinematics, $$\begin{split}
\dot{\mathbf{x}}_{1} &= \bm{J}_{1} \dot{\mathbf{q}},\\
\ddot{\mathbf{x}}_{1} &= \bm{J}_{1}\ddot{\mathbf{q}} + \dot{\bm{J}}_{1}\dot{\mathbf{q}},
\end{split}$$ where $\mathbf{x} \in \mathbb{R}^{n}$ and $\mathbf{q} \in \mathbb{R}^{m}$ represent the task’s operational coordinate and the joint positions, respectively, and $\bm{J}$ is the corresponding Jacobian matrix. Then, the joint acceleration for a desired task acceleration, $\ddot{\mathbf{x}}_{1}^{d}$ can be resolved as $$\label{eq:ddot_q_first}
\ddot{\mathbf{q}}_{1} = \overline{\bm{J}}_{1} \left( \ddot{\mathbf{x}}_{1}^{d} - \dot{\bm{J}}_{1} \dot{\mathbf{q}} \right) = \overline{\bm{J}}_{1} \ddot{\mathbf{e}}_{1}^{d},$$ where $\overline{\bm{J}}_1$ indicates the dynamically consistent inverse of $\bm{J}_1$, i.e. $$\overline{\bm{J}}_{1} = \bm{A}^{-1} \bm{J}_{1}^{\top} \left( \bm{J}_{1} \bm{A}^{-1}\bm{J}_{1}^{\top} \right)^{-1},$$ where $\bm{A} \in \mathbb{R}^{m \times m}$ indicates the mass/inertia matrix of the rigid body model of the robot. If we consider now the mapping of two operational tasks $\ddot{\mathbf{x}}_{1}^{d}$ and $\ddot{\mathbf{x}}_{2}^{d}$, we propose the following task hierarchy mapping $$\ddot{\mathbf{q}} = \overline{\bm{J}}_{1} \ddot{\mathbf{e}}_{1}^{d} + \overline{\bm{J}_{2|1}} \left( \ddot{\mathbf{e}}_{2}^{d} - \bm{J}_{2} \ddot{\mathbf{q}}_{1} \right),
\label{eq:two_task}$$ where $\overline{\bm{J}_{2|1}}\triangleq \overline{\left( \bm{J}_{2} \bm{N}_{1} \right)}$ represents the Jacobian associated with the second task, $\bm{J}_2$, projected into the null space of the first task, $\bm{N}_1=\bm{I} - \overline{\bm{J}}_{1} \bm{J}_{1}$, which by definition is orthogonal to the Jacobian associated with the first task, $\bm{J}_{1}$. The Equation (\[eq:two\_task\]) can be extended to the general $n$ task case, using the following hierarchy $$\ddot{\mathbf{q}}_{[task]} = \overline{\bm{J}}_{1} \ddot{\mathbf{e}}_{1}^{d} + \sum_{k=2}^{n} \ddot{\mathbf{q}}_{k},\quad (n\geq 2)
\label{eq:n_tasks}$$ with $$\begin{split}
&\ddot{\mathbf{q}}_{k} = \overline{\bm{J}}_{k|prec(k)} \left( \ddot{\mathbf{e}}_{k}^{d} - \bm{J}_{k} \sum_{i=1}^{k-1} \ddot{\mathbf{q}}_{i} \right),\\
&\bm{J}_{k|prec(k)} = \bm{J}_{k} \bm{N}_{prec(k)}, \\
&\bm{N}_{prec(k)} = \prod_{s=1}^{k-1} \bm{N}_{s|s-1} \quad (k\geq 2, \quad \bm{N}_{1|0} = \bm{N}_1), \\
&\bm{N}_{s|s-1} = \bm{I} - \overline{\bm{J}}_{s|prec(s)} \bm{J}_{s|prec(s)}
\quad (s \geq 2) \textrm{.}
\end{split}$$ This task hierarchy is similar, albeit not identical to [@siciliano1991general]. Compared to it, our proposed method is more concise, resulting in less computations for similar control specifications. In particular we do not require the computation of time derivatives of prioritized Jacobians. Details on the similarities and differences between these two works are discussed in Appendix \[append\_b\].
![[**Block Diagram of the Proposed Whole-Body Locomotion Controller.**]{} In WBLC, motion commands are compounded as joint accelerations based on null-space projection methods. The CM task specification is usd to compute reaction forces via QP optimization including unilateral contacts and friction cone constraints. The computed joint acceleration and reaction forces are used to sove for torque commands, which are the final output of WBLC.[]{data-label="fig:wblc"}](A_block_diagram){width="1.0\columnwidth"}
Optimizing Reaction Forces of Underactuated Robots
--------------------------------------------------
Based on the desired joint acceleration given in (Eq. ), WBLC finds torque commands via the following equation: $$\label{eq:multi_dyn}
\bm{A}(\mathbf{q}) \ddot{\mathbf{q}}^{d} + \mathbf{b}(\mathbf{q},\dot{\mathbf{q}}) + \mathbf{g} (\mathbf{q}) + \bm{J}_{r}^{\top}\mathbf{F}_{r} = \bm{U}^{\top} \bm{\tau},$$ where $\mathbf{q} \in \mathbb{R}^{n+6}$ and $\mathbf{b}(\mathbf{q},\dot{\mathbf{q}})$ and $\mathbf{g}(\mathbf{q})$ are the joint space coriolis/centrifugal and gravity terms, respectively. $\mathbf{F}_{r}$ and $\bm{J}_{r}$ represent the reaction forces and the corresponding contact Jacobian. $\bm{\tau} \in \mathbb{R}^{n}$ and $\bm{U}^{\top} \in \mathbb{R}^{(n+6) \times (n)}$ represent the actuator torque commands and the selection matrix mapping actuated torque to the floating base dynamics. Note that $\ddot{\mathbf{q}}^d$ is chosen as, $$\label{eq:qqd}
\ddot{\mathbf{q}}^d = \ddot{\mathbf{q}}_{[task]} + \bm{N}_{n|prec(n)} \ddot{\mathbf{q}}_{res},$$ with $$\bm{N}_{n|prec(n)} = \bm{N}_{prec(n)} \bm{N}_{n|n-1}$$ where $\ddot{\mathbf{q}}_{res}$ is a residual joint acceleration command.
To find the reaction forces $\mathbf{F}_r$, we specify a centroidal momentum (CM) operational task. A CM task consists of linear and angular momenta portions. The linear part, corresponds to the robot’s CoM behavior, $\mathbf{F}_{cm,lin}$, and is typically used for locomotion planning. On the other hand, the angular behavior, the so-called CAM, $\mathbf{F}_{cm,ang}$, is typically set to zero value. Setting the angular task to zero creates conflict with other tasks, such as body rotational tasks. We circumvent this problem by projecting angular behavior as a lower priority task than body rotational tasks as we will soon see. In addition, sometimes it is not possible to simultaneously fulfill linear and angular momentum specifications. For that reason, we specific CoM behavior as a hard constraint while relaxing angular behavior, i.e. $$\begin{split}
\min_{\mathbf{F}_r}\quad & \mathbf{F}_{r}^{\top} \bm{Q} \mathbf{F}_{r} + \| \mathbf{F}_{cm,ang}^d - \bm{W}_{ang} \mathbf{F}_{r} \|^{2} \\[1.5mm]
\textrm{Subject to.} \quad& \mu |\mathbf{F}_{r,z} | \geq |\mathbf{F}_{r,x}| \\
& \mu |\mathbf{F}_{r,z} | \geq |\mathbf{F}_{r,y}| \\
& \mathbf{F}_{cm,lin}^{d} - \bm{W}_{lin}\mathbf{F}_{r} = \mathbf{0}
\end{split}$$ where $\mathbf{F}_{cm,lin}^{d}$ and $\mathbf{F}_{cm,ang}^{d}$ are the desired linear and angular parts of the CM, $\mu$ represents a friction coefficient related to the contact surfaces, $\bm{Q}$ is a weighting matrix, and $\bm{W}_{ang}$ and $\bm{W}_{lin}$ are mappings from reaction forces to angular and linear momenta behaviors. Based on the results of this optimization, $\mathbf{F}_{r}$, the desired value of the CAM task can be calculated as follows: $$\label{eq:cm-definition}
\bm{I}_{cm} \ddot{\mathbf{x}}_{cm}^{d} = \mathbf{F}_{cm}^{d} = \left[ \begin{array} {cc} \mathbf{F}_{cm,lin}^{d} & \bm{W}_{ang} \mathbf{F}_{r} \end{array} \right]^{\top},$$ where $\bm{I}_{cm}$ is a spatial inertial term. Notice that the term $\bm{W}_{ang} \mathbf{F}_{r}$ might be different than $\mathbf{F}_{cm,ang}^{d}$ since the desired angular behavior might violate friction cone constraints. From the above equation, we extract the desired CM acceleration command $\ddot{\mathbf{x}}_{cm}^{d}$ for usage in the controller hierarchy. More concretely, $\ddot{\mathbf{x}}_{cm}^{d} = \left ( \ddot{\mathbf{x}}_{CoM}^{d} \, \mathbf{\alpha}_{ang}^{d} \right)$, where the first term within the parenthesis is the desired CoM acceleration and the second term is the desired angular acceleration. Both of these commands are used separately in the hierarchy defined in Eq. , to produce the joint acceleration command $\ddot{\mathbf{q}}_{[task]}$ which in turn yields $\ddot{\mathbf{q}}^d$ via Eq. . Plugging this last term into Eq. we obtain $$\bm{A}\left( \ddot{\mathbf{q}}_{[task]} + \bm{N}_{n|prec(n)} \ddot{\mathbf{q}}_{res} \right) +\mathbf{b} + \mathbf{g} + \bm{J}_{r}^{\top}\mathbf{F}_{r} = \bm{U}^{\top} \bm{\tau},$$ which can be written in matrix form as $$\label{eq:final_step}
\left[\begin{array}{cc}
\bm{U}^{\top} & -\bm{A}\bm{N}_{n|prec(n)}
\end{array}\right]
\left[ \begin{array}{c}
\bm{\tau} \\ \ddot{\mathbf{q}}_{res}
\end{array}\right] = \bm{\tau}_{[task]},$$ where we have defined the term $$\label{eq:task_torque}
\bm{\tau}_{[task]} \triangleq \bm{A}\ddot{\mathbf{q}}_{[task]} + \mathbf{b} + \mathbf{g} + \bm{J}_r^{\top}\mathbf{F}_r.$$ We now have an underdetermined matrix system which can be solved via pseudo inversion as $$\left[ \begin{array}{c}
\bm{\tau} \\ \ddot{\mathbf{q}}_{res}
\end{array}\right] = \left[\begin{array}{cc}
\bm{U}^{\top} & -\bm{A}\bm{N}_{n|prec(n)}
\end{array}\right]^{+} \bm{\tau}_{[task]}
\label{eq:final_cmd}$$ where $(.)^{+}$ represents the Moore-Penrose pseudo inverse operation.
Time Derivative of Jacobian {#sec:jdot}
===========================
![[**Tracking performance comparison with and without the term $\dot{\bm{J}}\dot{\mathbf{q}}$.**]{} A three-DoF planar manipulator is used to control its end effector to follow a vertical line (red lines) with 2 Hz frequency (blue dashed lines are the end-effector path). The tracking results demonstrte that the (a) controller, which accounts for $\dot{\bm{J}}\dot{\mathbf{q}}$, outperforms the (b) controller.[]{data-label="fig:jdot_linkage"}](A_jdot_linkage){width="\columnwidth"}
The ability to compute efficiently the time derivative of Jacobian operators for fast operational space control has been overlooked. However it plays an important role on robustifying fast movements. Fig.\[fig:jdot\_linkage\] shows that the tracking performance of a simple serial manipulator is significantly enhanced by using the term $\dot{\bm{J}}\dot{\mathbf{q}}$ via Operational Space Control (OPC), where $J$ is the Jacobian of the end effector and $\dot q$ is the vector of joint velocities. Notice that $\dot J$ is used in our WBLC in Equation . The commanded task is to follow a vertical line defined by the function, $\mathbf{x}^{d} = [~0.62, ~0.23 \sin(4\pi t)~]^{\top}$. The OPC input is $$\ddot{\mathbf{x}} =
\begin{bmatrix}
0 \\
-36.32 \sin(4\pi t)
\end{bmatrix} + K_p (\mathbf{x}^{d} - \mathbf{x}) + K_v(\dot{\mathbf{x}}^{d} - \dot{\mathbf{x}}).$$ We will use Lie group theory to compute the derivatives of point task Jacobians [@kimlie]. We implement this functionality using the Rigid Body Dynamics Library [^5], which is a popular open source dynamics toolbox. In addition, we also devise a new method to compute the time derivative of the CM Jacobian which cannot be computed using Lie group theory.
Time Derivative of Point Jacobian {#sec:time_der_pt_jacobian}
---------------------------------
Lie group operators provide convenient analytic derivations for Jacobian computations. The $SE(3)$ orientation and position representation of a rigid body in three-dimensional space consists of orientation matrix ($\bm{R}$) and a position vector ($\mathbf{p}$). It can also be represented via the $4\times 4$ homogeneous transformation, $$\bm{T}_{g,i} = \left[
\begin{array}{cc}
\bm{R}_{g,i} & \mathbf{p}_{g,i} \\
0 & 1
\end{array}
\right],$$ where, $\bm{R}_{g,i}$ and $\mathbf{p}_{g,i}$ represent the orientation and position of the $i^{th}$ frame in global coordinates respectively (See Fig. \[fig:openchain\]). The velocity representation in $se(3)$ consists of the 6-dimensional vector, $[\mathbf{w}, \mathbf{v} ]^T$, and yields the $4\times 4$ homogeneous equality, $$\bm{V}_i \triangleq \left[
\begin{array}{cc}
[\mathbf{w}_i]^{\times} & \mathbf{v}_i \\
0 & 0
\end{array}
\right],$$ where $$[\mathbf{w}_i]^{\times} \triangleq \begin{bmatrix}
0 & -w_{i,3} & w_{i,2} \\
w_{i,3} & 0 & -w_{i,1} \\
-w_{i,2} & w_{i,1} & 0
\end{bmatrix}$$ Here, $w_{i,1}, w_{i,2}, w_{i,3}$ are relative angular velocities along the three Cartesian axes, and $\mathbf{v}_i$ is the linear velocity. It can be shown that $\bm{V}_i =\bm{T}_{g, i}^{-1} \dot{\bm{T}}_{g,i}$, and corresponds to the generalized velocity seen from the $i^{th}$ frame. The velocity in the global frame associated with $\bm{V}_i$ can be obtained via adjoint derivations,
$$\begin{split}
{\rm Ad}_{{T}_{g,i}} \left( \bm{V}_i\right) &= \bm{T}_{g,i} \bm{V}_i \bm{T}_{g,i}^{-1} \\
&= \bm{T}_{g,i} \bm{T}_{g,i}^{-1} \dot{\bm{T}}_{g,i} \bm{T}_{g,i}^{-1}\\
& = \dot{\bm{T}}_{g,i} \bm{T}_{g,i}^{-1}.
\end{split}$$
The adjoint mapping operator is defined as $$\label{eq:adjoint_mtx}
{\rm Ad}_{T_{i,j}} \triangleq \begin{bmatrix}
\bm{R}_{i,j} & 0 \\
[\mathbf{p}_{i,j}]^{\times} \bm{R}_{i,j} & \bm{R}_{i,j}
\end{bmatrix},$$ where $\bm{R}_{i,j}$ and $\mathbf{p}_{i,j}$ are relative rotations and positions between points $i$ and $j$. The generalized velocity of point $p$ in local coordinates (see Fig. \[fig:openchain\]) can be represented as $$\begin{split}
\mathbf{V}_p & = {\rm Ad}_{{T}_{p,n}} \mathbf{V}_n \\
&= {\rm Ad}_{{T}_{p,n-1}}\mathbf{V}_{n-1} + {\rm Ad}_{{T}_{p,n}}\mathbf{S}_n \dot{\mathbf{q}}_n \\
& \qquad \vdots \\
&= {\rm Ad}_{{T}_{p,0}}\mathbf{V}_{0} + \sum_{i=1}^{n} {\rm Ad}_{{T}_{p,i}}\mathbf{S}_{i} \dot{\mathbf{q}}_{i}.
\end{split}$$
![[**Multi-DoF Openchain.**]{} The openchain consists of $n$ joints. At the end of the chain, the end-effector is attached to link $n$.[]{data-label="fig:openchain"}](A_linkage_explain){width="0.8\columnwidth"}
{width="\linewidth"}
Because $\{0\}$ frame is the global frame (an inertial frame), $\mathbf{V}_{0}$ is equal to zero. On the other hand $\mathbf{S}_{i}$ maps joint velocities to $\mathit{R}^{6}$, e.g. $\mathbf{S}_i$ is $[~0, 0, 1, 0, 0, 0~]^T$ means $\dot q_i$ is a revolute joint rotating along the $z$ local axis. The first three positions of $\mathbf{S}_{i}$ represent rotational axes while the last three positions represent prismatic axes. It can be shown that the Jacobian of a point $p$, is equal to $$\bm{J}_p =
\begin{bmatrix}
{\rm Ad}_{{T}_{p,1}} \mathbf{S}_1
& Ad_{{T}_{p,2}}\mathbf{S}_2
&\cdots & {\rm Ad}_{{T}_{p,n}}\mathbf{S}_n
\end{bmatrix}$$ Furthermore, let’s break down the adjoint operators into the following chain operation $${\rm Ad}_{{T}_{p,i}} \mathbf{S}_i = {\rm Ad}_{{T}_{p, p'}} {\rm Ad}_{{T}_{p', n}} {\rm Ad}_{{T}_{n, i}} \mathbf{S}_i.$$ Here $p'$ is a virtual point representing the position of $p$ but with local orientation with respect to frame $n$. As such it represents just a position offset. In this case $\bm{T}_{p', n}$ is constant and the $i$-th column of the time derivative of the Jacobian can be resolved as $$\begin{split}
\dot{\bm{J}}_{p,i}
& = \dot{\overbrace{\left\{{\rm Ad}_{{T}_{p,i}}\mathbf{S}_i \right\}}} \\
& = \dot{\overbrace{\left\{{\rm Ad}_{{T}_{p, p'}}\right\}}} {\rm Ad}_{{T}_{p', n}} {\rm Ad}_{{T}_{n, i}} \mathbf{S}_i \\
& \quad + {\rm Ad}_{{T}_{p, p'}} {\rm Ad}_{{T}_{p', n}} \dot{\overbrace{\left\{ {\rm Ad}_{{T}_{n, i}} \mathbf{S}_{i} \right\}}} \\
&= {\rm Ad}_{{T}_{p,p'}} {\rm ad}_{{V}_{p,p'}} {\rm Ad}_{{T}_{p', n}} {\rm Ad}_{{T}_{n, i}} \mathbf{S}_i \\
&\quad + {\rm Ad}_{{T}_{p,p'}} {\rm Ad}_{{T}_{p',n}} \left\{{\rm Ad}_{{T}_{n,i}} {\rm ad}_{{V}_{n,i}} \mathbf{S}_i + {\rm Ad}_{{T}_{n,i}} (\dot{\mathbf{S}}_i ) \right\}
\end{split}$$
Here we have used $\dot{{\rm Ad}_{{T}}} = {\rm Ad}_{{T}} ({\rm
ad}_{\bm{V}})$ and since $V = \bm{T}^{-1}\dot{\bm{T}}$, then we define ${\rm ad}_{\bm{T}^{-1}\dot{\bm{T}}} \triangleq \bm{\bm{T}^{-1}\dot{\bm{T}}} - \bm{\dot{\bm{T}}\bm{T}^{-1}}$.
Time Derivative of the Centroidal Momentum Jacobian
---------------------------------------------------
The previous equations for the time derivative of point Jacobians are not applicable to the CM Jacobian. The latter can be obtained from the CM task definition of Eq. linear part is simply the weighted sum of time derivatives of each link’s CoM Jacobian. However, the angular part is not straightforward. Instead of finding $\dot{\bm{J}}_{cm}$, we can find the multiplication of $\dot{\bm{J}}_{cm}$ and the joint velocities, $\dot{\mathbf{q}}$, via operational space dynamics: $$\bm{\mathit{\Lambda}}_{cm}(\mathbf{q})\ddot{\mathbf{x}} + \bm{\mu}_{cm}(\mathbf{q}, \dot{\mathbf{q}}) + \mathbf{p}_{cm}(\mathbf{q}) = \mathbf{F}_r,$$ Here, $\bm{\mathit{\Lambda}}_{cm}$, $\bm{\mu}_{cm}$, and $\mathbf{p}_{cm}$ are an inertia matrix, coriolis and centrifugal force, and gravitational force of the CM operational task, respectively. Since there is no coriolis and centrifugal effects in CM space, $\bm{\mu}_{cm}$ is zero. Thus, $\dot{\bm{J}}_{cm}\dot{\mathbf{q}}$ must be equal to $\bm{J}_{cm}\bm{A}^{-1}\mathbf{b}$: $$\dot{\bm{J}}_{cm}\dot{\mathbf{q}} =
\bm{J}_{cm}\bm{A}^{-1}\mathbf{b}.$$ All terms in $\bm{J}_{cm}\bm{A}^{-1}\mathbf{b}$ are easily computable usingoff-the-shelf dynamics libraries.
Results {#sec:result}
=======
To verify the performance of the proposed methods, we conduct three demonstrations: 1) body orientation control while changing the task hierarchy, 2) dynamic locomotion with directional change, and 3) push-recovery from various directions while walking. Toward these investigations, we implement our algorithms on a simulation of the Valkyrie humanoid robot, and test it using the physics based simulation SrLib. Because our focus is on locomotion, we fix the finger and wrist joints, bringing the total number of joints to 28. To incorporate floating body dynamics, prismatic and ball joints are introduced to connect Valkyrie’s pelvis to a fixed frame.
In the simulation environment, we use a friction coefficient between the ground and the robot’s feet of 0.8. On the other hand the friction cone constraints used in WBLC are set to a value of 0.65 to be conservative. In case our contact control solver fails to find proper reaction forces, we allow for solutions that violate friction constraints by relaxing the friction coefficient to a value of 1.75. The resulting control solution implies that slip occurs but only for very short times (in general less than 0.005 $\si[per-mode=symbol]{\second}$). This simple technique doesn’t incur an increase in computational complexity while greatly enhancing the robustness of WBLC with respect to external disturbances.
Body Orientation Control with Various Task Hierarchies
------------------------------------------------------
Body shaking behavior is a difficult skill that we use here to study the dynamic performance of WBLC tasks. In traditional humanoid control methods, CoM and CAM tasks are controlled within the same priority level. We propose to split them via WBLC into different hierarchy levels. To demonstrate this feature, we define the following six tasks:
- $\ddot{\mathbf{x}}_{1} \in \mathbb{R}^{3}$: Linear CoM
- $\ddot{\mathbf{x}}_{2} \in \mathbb{R}^{3}$: Centroidal Angular Momentum (CAM)
- $\ddot{\mathbf{x}}_{3} \in \mathbb{R}^{3}$: Body Orientation
- $\ddot{\mathbf{x}}_{4} \in \mathbb{R}^{22}$: Partial Joint Posture (all joints except shoulder pitch, shoulder roll, and knee pitch)
- $\ddot{\mathbf{x}}_{5} \in \mathbb{R}^{3}$: Pelvis Orientation
- $\ddot{\mathbf{x}}_{6} \in \mathbb{R}^{28}$: Full Joint Posture
Most tasks above are self-explanatory. We introduce a partial joint posture consisting on keeping the initial joint positions for all robot joints except for the shoulder pitch and roll, and the knee pitch. This task is used for the sole purpose of testing performance when multiple tasks conflict. In particular, the partial joint posture conflicts with the CAM task within the above task set and viceversa. For our test, we use two hierarchies: $$\begin{split}
\mathbb{H}_{1} &= \left\{ \ddot{\mathbf{x}}_{1} \rightarrow \ddot{\mathbf{x}}_{2} \rightarrow \ddot{\mathbf{x}}_{3} \rightarrow \ddot{\mathbf{x}}_{4} \rightarrow \ddot{\mathbf{x}}_{5} \rightarrow \ddot{\mathbf{x}}_{6} \right\} \\
\mathbb{H}_{2} &= \left\{ \ddot{\mathbf{x}}_{1} \rightarrow \ddot{\mathbf{x}}_{3} \rightarrow \ddot{\mathbf{x}}_{4} \rightarrow \ddot{\mathbf{x}}_{5} \rightarrow \ddot{\mathbf{x}}_{2} \rightarrow \ddot{\mathbf{x}}_{6} \right\}
\end{split}$$ The second hierarchy, $\mathbb{H}_{2}$, is more appropriate than the first one, $\mathbb{H}_{1}$ to achieve accurate control of the body shaking (orientation) task. This is accomplished by assigning higher priority to the body orientation task and moving backwards the CAM task. As shown in Fig. \[fig:shaking\_body\] (a), changing the hierarchy levels cause different whole body motions. Fig. \[fig:shaking\_body\] (b) shows the body orientation task error for the two task hierarchies. The body orientation performance from $\mathbb{H}_{2}$ is better than that of $\mathbb{H}_{1}$. This can be seen in the interval from 4.0 $\si[per-mode=symbol]{\second}$ to 4.5 $\si[per-mode=symbol]{\second}$ in Fig. \[fig:shaking\_body\] (b). In addition, the different hierarchies cause not only different movements but also different torque profiles. As shown in Fig. \[fig:shaking\_body\] (c), higher torques are needed for $\mathbb{H}_{2}$ than for $\mathbb{H}_{1}$.
Dynamic Walking with Directional Change
---------------------------------------
Walking can be broken down into three phases: double contact, right foot contact, and left foot contact. To represent these phases we define the following task hierarchy in WBLC:
- $\ddot{\mathbf{x}}_{1} \in \mathbb{R}^{3}$: Linear CoM position
- $\ddot{\mathbf{x}}_{2} \in \mathbb{R}^{3}$: Pelvis Orientation
- $\ddot{\mathbf{x}}_{3} \in \mathbb{R}^{3}$: Body Orientation
- $\ddot{\mathbf{x}}_{4} \in \mathbb{R}^{3}$: (for the single contact phases) Foot Orientation
- $\ddot{\mathbf{x}}_{5} \in \mathbb{R}^{3}$: (for the single contact phases) Foot Position
- $\ddot{\mathbf{x}}_{6} \in \mathbb{R}^{6}$: Neck and Torso Joint Posture
- $\ddot{\mathbf{x}}_{7} \in \mathbb{R}^{3}$: Centroidal Angular Momentum
- $\ddot{\mathbf{x}}_{8} \in \mathbb{R}^{10}$: Arms Joint Posture
To produce swing foot trajectories, we define third degree B-splines, which guarantee acceleration continuity. The orientation coordinates for the robot’s body, pelvis, and feet are described using quaternions. For each step, these orientation tasks are commanded to smoothly switch from the current frame to next one. Given initial CoM states, our locomotion planner computes foot positions and their timing while satisfying the desired walking directional changes. In our test shown in Fig. \[fig:turn\], Valkyrie takes first 12 steps while continuously changing its walking direction by $18.8^{\si[per-mode=symbol]{\degree}}$ per step. After that, Valkyrie takes 5 forward steps with no directional change. Then, Valkyrie takes another 12 steps while changing direction by, $-18.8^{\si[per-mode=symbol]{\degree}}$ per step. The user only specifies the walking directions while RL-PSP automatically finds the foot positions and their timing using the learned policy. The learned policy consists only on switching states and step locations. The desired position, velocity, and acceleration of the CoM are computed with the analytic equation of the LIP model at runtime.
![[**Continuous walking directional change**]{} Valkyrie shows a complex dynamic walking pattern involving changing walking direction. (a) shows a top view of Valkyrie and its walking path. (b) Shos how the robot’s CoM state mapped to the next local frame. Local frames rotate with the desired walking direction. For each step, the stance foot becomes the origin of the local frame, and the orientation of the frame is aligned with the desired walking direction. The previous switching CoM state is projected to the current local frame, and the planner finds the foot placement with the new state.[]{data-label="fig:turn"}](A_turning_result){width="0.85\columnwidth"}
Push Recovery while Walking
---------------------------
To validate push recovery, we conduct simulated experiments under large external disturbances and in various directions. Although WBLC is robust to small deviations of the CoM trajectory, for external disturbances we rely on the learned recovery policies described in the theory sections. When the norm of the CoM state error, $$\mathbf{error}= \begin{bmatrix}
\mathbf{x}^{d} - \mathbf{x}\\
0.5 (\dot{\mathbf{x}}^{d} - \dot{\mathbf{x}})
\end{bmatrix},$$ is over a threshold equal to $0.05~\si[per-mode=symbol]{\meter}$ and for longer than 0.02$\si[per-mode=symbol]{\second}$, our planner computes a new trajectory starting from the current CoM state. Instead of setting the new CoM control goal to be the current (disturbed) CoM state, we have experienced that it is better to define a controller goal, $\mathbf{x}^{new}$, equal to $$\mathbf{x}^{new} = \gamma\mathbf{x}^{d} + (1-\gamma)\mathbf{x},$$ where $\gamma$ can be selected heuristically, and we use a value of 0.8. In our tests, we push Valkyrie while she is dynamically walking using various disturbance forces applied for a duration of 0.1 $\si[per-mode=symbol]{\second}$. The maximum disturbance impulse that we apply to Valkyrie is 520 $\si[per-mode=symbol]{\newton}$ for 0.1 $\si[per-mode=symbol]{\second}$. The results are shown in Fig. \[fig:replan\] compared to the undisturbed trajectories. The CoM phase trajectory in the lateral plane shown in Fig. \[fig:replan\] (b) shows that the planner is able to find a new trajectory after an external impulse is applied. The time to compute 15 steps after the disturbance is less than 1$\si[per-mode=symbol]{\milli \second}$ using a dual-core 3.0 GHz Intel i7 processor. At the moment that the replanning process occurs, we also find a new swing foot trajectory that transitions from the original swing trajectory to the new goal.
{width="2.0\columnwidth"}
Fig. \[fig:multi\_test\] shows results of push recoveries while dynamically walking when being subject to various external forces. In all cases, Valkyrie succeeds to sustain the disturbances and continue walking without stopping. The robustness capabilities in this test are competitive to the results of [@Khadiv:2017th] which is not based on statistical learning. In contrast to this state of the art, due to our use of offline learning our planner is able to come up with numerous steps almost instantaneously with respect to the walking time frame.
Conclusion and Discussion
=========================
In this paper, we propose an RL based robust locomotion planner and a new WBC, dubbed WBLC. By utilizing PSP in the RL formulation, we can quickly find locomotion policies for 3D walking. The newly developed WBLC takes into account realistic contact and friction cone constraints. At the same time, WBLC maintains task priorities using projection operators which is missing in previous QP based WBCs. Overall, WBLC simultaneously exploits the benefits of QP based WBC’s and projection based WBC’s, achieving versatility and computational efficiency. Another benefit of our methods is the planning speed. Our locomotion planner almost instantaneously finds a multistep walking trajectory faster than the state of the art. By devising the replanning process during dynamic walking, robots can quickly react to external forces and achieve significant robustness.
One interesting aspect of our planning algorithm is the value function we used in the learning process. In the future we could use this value function as an indicator for walking risk given the disturbed states. Many researchers have suggested indicators for locomotion quality. For example, ZMP [@VUKOBRATOVIC:2004ej] and CP [@Pratt:2006ct] are indicators of balance stability but they don’t take into account other important information such as kinematic constraints or swing time limits. Recently, an allowable CoM acceleration region [@Caron:2015cm] has been proposed for multi-contact stability. However, there is no indication of kinematic or dynamic limitations such as step size or swing time. In contrast our value function takes into account some kinematic and dynamic constraints that could ultimately make it a versatile metric for walking quality evaluation.
In the future, we will experiment with more complex functions to represent learned values and policies (e.g. deep neural network). In this paper, we have focused on finding simple walking patterns. However, complex neural networks, which can represent highly nonlinear and abstract behaviors, can enable more versatile planners. For instance, future planners may be able to traverse rough terrain by exploiting various locomotion modes such as walking, running, or jumping. We also plan to implement the proposed algorithms in real systems and evaluate their performance. In our previous work [@Kim:2016jg], we showed agile bipedal balance with a point-foot biped with series elastic actuators. Since the system is highly unstable by nature, we did not apply external disturbances. We believe that the robustness capabilities we have outlined in this paper may allow us ot accomplish sophisticated behaviors in the real testbeds.
Analytic Solution of the Phase Space Planner {#sec:append_psp}
============================================
When we constraint PIPM dynamics to a piecewise linear height surface, $z = a(x-p_x) + b$, we can find $t_{switch}$ and $p_y$ without numerical integration and bisection search because the system of equations becomes linear, resulting in the following CoM behavior: $$\label{eq:x_state}
\begin{split}
x(t) & = A e^{\omega t} + B e^{-\omega t} + p_x, \\
\dot{x}(t) & = \omega (A e^{\omega t} - B e^{-\omega t} ),
\end{split}$$ where, $$\begin{split}
\omega &= \sqrt{\frac{g}{a p_x + b}}, \\[2mm]
A &= \frac{1}{2}\Big( (x_{0} - p_x) + \frac{1}{\omega}\dot{x}_{0} \Big), \\[2mm]
B &= \frac{1}{2}\Big( (x_{0} - p_x) - \frac{1}{\omega}\dot{x}_{0} \Big).
\end{split}$$ Note that this equation is the same for the $y$ direction. Based on Eq. , we can find an analytical solution for PSP, summarized in Algorithm \[code:Phase\_Space\]. $~\mathbf{x}_{1}$, $\mathbf{y}_{1}$, $\mathbf{x}_{apex,2}$, and $\mathbf{x}_{switch}$ are vector quantities corresponding to the variables $(x_1,\dot{x}_1)$, $(y_1,\dot{y}_1)$, $(x_{apex,2}, \dot{x}_{apex,2})$, and $(x_{switch},\dot{x}_{switch})$.
$ \mathbf{x}_{switch} \gets$ ;\
\
$ t_{switch} \gets $ $ t_{apex} \gets$ $\mathbf{y}_{switch} \gets$ $p_y$ $\gets$
Let us focus on obtaining the step switching time. We can easily manipulate Eq. to analytical solve for the time variable, $$\begin{split}
& x + \frac{1}{\omega}\dot{x} = 2A e^{\omega t} + p_x, \\[2mm]
& x + \frac{1}{\omega}\dot{x} - p_x = 2A e^{\omega t},
\end{split}$$ which renders $$\label{eq:t_eqn}
t = \frac{1}{\omega} \ln \Big( \frac{x + \frac{1}{\omega}\dot{x} - p_x}{2 A} \Big).$$ To find the dynamics, $\dot{x} = f(x)$, which will lead to the switching state solution, let us remove the $t$ term by plugging Eq. into Eq. . $$x = A \frac{x + \frac{\dot{x}}{\omega} - p_x}{2 A} + B \frac{2 A}{x + \frac{\dot{x}}{\omega} - p_x } + p_x$$ $$\begin{aligned}
\frac{1}{2}(x - p_x - \frac{\dot{x}}{\omega}) &= \frac{2 AB}{x + \frac{\dot{x}}{\omega} - p_x} \\
(x - p_x)^2 - \Big(\frac{\dot{x}}{\omega}\Big)^2 &= 4 AB\end{aligned}$$ By performong some algebra we get, $$\begin{split}
\dot{x}^2 &= \omega^2 ( (x - p_x)^2 - 4 AB ), \\[2mm]
\dot{x}^2 &= \omega^2 \Big( (x - p_x)^2 - (x_{0} - p_x)^2 \Big) + \dot{x}_0^2,
\end{split}$$ which yields, $$\label{eq:vel_x}
\dot{x} = \pm \sqrt{\frac{g}{h} \Big( (x - p_x)^2 - (x_{0} - p_x)^2 \Big) + \dot{x}_{0}^2 }.$$ Given two phase trajectories associate with consecutive walking steps, $p_{x,1}$ and $p_{x,2}$ and assuming the robot walks forward, i.e. $\dot{x}_{switch}$ is positive, we calculate the phase space intersection point via continuity of velocities from Eq. : $$\label{eq:x_switch}
\begin{split}
x_{\rm switch}&=\frac{1}{2}\Big( \frac{C}{p_{x,2} - p_{x,1}} + (p_{x,1} + p_{x,2}) \Big)\\
C&=(x_{{0},1}-p_{x,1})^2 - (x_{{0},2} - p_{x,2})^2 + \frac{\dot{x}_{{0},2}^2 - \dot{x}_{{0},1}^2}{\omega^2}
\end{split}$$ We can now find the step switching time by plugging the computed switching position into Eqs and . In addition, we can obtain the timing at the apex velocity from Eq. . The final step is to find the $y$ directional foot placement. We first calculate $\mathbf{y}_{switch}$ by pluggin $t_{switch}$ into the $y$ directional state equation, which has identical form to Eq. . Then, by using the equality that $\dot{y} (t_{\rm apex})=\dot{y}_{apex}$, we can find $p_{y}$, $$\label{eq:yp}
\begin{split}
p_y &= \frac{\dot{y}_{apex}-C}{D},\\[2mm]
C &= \frac{\omega}{2} \big( (y_{switch}+\frac{\dot{y}_{switch}}{\omega})e^{\omega t_{apex}} - \\
&\quad\quad\quad\quad (y_{switch}-\frac{\dot{y}_{switch}}{\omega})e^{-\omega t_{apex}}\big)\\
D &= \frac{\omega}{2}(e^{-\omega t_{apex}} - e^{\omega t_{apex}})
\end{split}$$ After calculating $p_y$, we can easily get $y_{apex}$ and $\dot{y}_{apex}$ by using Eq. .
Equivalent Hierarchy-based Joint Acceleration {#append_b}
=============================================
The joint velocity associated with an operational task ${\mathbf{x}}_{1}$ is $$\begin{split}
\dot{\mathbf{q}} = \bm{J}_{1}^{+} \dot{\mathbf{x}}_{1} + \bm{N}_{1} \dot{\mathbf{q}}_{0}.
\label{eq:qdot_first}
\end{split}$$ The definition of the null-space projection matrix using a pseudo inverse and its time derivative yields the following expression: $$\begin{split}
\bm{N}_{1} = \bm{I} - \bm{J}_{1}^{+} \bm{J}_{1} &\Rightarrow \dot{\bm{N}}_{1} = -\dot{\bm{J}}_{1}^{+} \bm{J}_{1} - \bm{J}_{1}^{+} \dot{\bm{J}}_{1}.
\end{split}$$ The resulting joint acceleration can be obtained by time-derivativating equation (\[eq:qdot\_first\]) as described in [@siciliano1991general] $$\begin{split}
\ddot{\mathbf{q}} &= \bm{J}_{1}^{+} \ddot{\mathbf{x}}_{1} + \dot{\bm{J}}_{1}^{+}\dot{\mathbf{x}}_{1} + \dot{\bm{N}}_{1} \dot{\mathbf{q}}_{0} + \bm{N}_{1}\ddot{\mathbf{q}}_{0} \\
&= \bm{J}_{1}^{+} \ddot{\mathbf{x}}_{1} + \dot{\bm{J}}_{1}^{+} \bm{J}_{1} \dot{\mathbf{q}} + \dot{\bm{N}}_{1} \dot{\mathbf{q}}_{0} + \bm{N}_{1}\ddot{\mathbf{q}}_{0}
\end{split}$$ using the equality $\dot{\bm{J}}_{1}^{+} \bm{J}_{1} = - \dot{\bm{N}}_{1} - \bm{J}_{1}^{+} \dot{\bm{J}}_{1}$ we get $$\begin{split}
\ddot{\mathbf{q}} &= \bm{J}_{1}^{+} \ddot{\mathbf{x}}_{1} -\bm{J}_{1}^{+} \dot{\bm{J}}_{1} \dot{\mathbf{q}} - \dot{\bm{N}}_{1} \dot{\mathbf{q}} + \dot{\bm{N}}_{1} \dot{\mathbf{q}}_{0} + \bm{N}_{1}\ddot{\mathbf{q}}_{0} \\
\end{split}
\label{eq:ddotq_1}$$ This allows us to simplify equation (\[eq:ddotq\_1\]) to $$\begin{split}
\ddot{\mathbf{q}} &= \bm{J}_{1}^{+}\left( \ddot{\mathbf{x}}_{1} - \dot{\bm{J}}_{1} \dot{\mathbf{q}}\right) -\dot{\bm{N}}_{1} \bm{J}_{1}^{+}\dot{\mathbf{x}}_{1} + \bm{N}_{1} \ddot{\mathbf{q}}_0.
\end{split}
\label{eq:qdot_middle}$$ If we consider a secondary task $\mathbf{x}_2$, the term $\dot{\mathbf{q}}_0$ becomes $$\dot{\mathbf{q}}_{0} = \left( \bm{J}_{2} \bm{N}_{1} \right)^{+} \left( \dot{\mathbf{x}}_{2} - \bm{J}_{2} \bm{J}_{1}^{+} \dot{\mathbf{x}}_{1} \right)
\label{eq:qdot_0}$$ because it can be shown that $\bm{N}_{1} \left( \bm{J}_{2} \bm{N}_{1} \right)^{+} = \left( \bm{J}_{2} \bm{N}_{1} \right)^{+}$, we get, $$\begin{split}
\dot{\mathbf{q}}_{0} &= \bm{N}_{1} \dot{\mathbf{q}}_{0} \\
\ddot{\mathbf{q}}_0 &= \bm{N}_{1} \ddot{\mathbf{q}}_0 + \dot{\bm{N}}_{1} \dot{\mathbf{q}}_0 \textrm{.}
\label{eq:qdot_nqdot}
\end{split}$$ From Eq. (\[eq:qdot\_0\]), the term $\ddot{\mathbf{q}}_{0}$ becomes $$\begin{split}
\ddot{\mathbf{q}}_{0} &= \bm{J}_{2|1}^{+} \left( \ddot{\mathbf{x}}_{2} - \dot{\bm{J}}_{2} \bm{J}_{1}^{+} \dot{\mathbf{x}}_{1} - \bm{J}_{2} \dot{\bm{J}}_{1}^{+} \dot{\mathbf{x}}_{1} - \bm{J}_{2} \bm{J}_{1} \ddot{\mathbf{x}}_{1} \right) \\
&+ \dot{\bm{J}}_{2|1}^{+}\left( \dot{\mathbf{x}}_{2} - \bm{J}_{2}\bm{J}_{1}^{+}\dot{\mathbf{x}}_{1} \right)
\end{split}$$ where $ \bm{J}_{2|1}\triangleq \bm{J}_{2} \bm{N}_{1}$ and $$\dot{\bm{J}}_{2|1}^{+} = - \bm{J}_{2|1}^{+} \dot{\bm{J}}_{2|1} \bm{J}_{2|1}^{+}.$$ Then, we can manipulate the equation above defining $\ddot{\mathbf{q}}_{0}$ to yield $$\begin{split}
\ddot{\mathbf{q}}_{0} &= \bm{J}_{2|1}^{+} \left( \ddot{\mathbf{x}}_{2} - \dot{\bm{J}}_{2} \bm{J}_{1}^{+} \dot{\mathbf{x}}_{1} - \bm{J}_{2} \dot{\bm{J}}_{1}^{+} \dot{\mathbf{x}}_{1} - \bm{J}_{2} \bm{J}_{1}^{+} \ddot{\mathbf{x}}_{1} \right)\\
& - \bm{J}_{2|1}^{+} \dot{\bm{J}}_{2|1} \dot{\mathbf{q}}_{0}\\
&= \bm{J}_{2|1}^{+} \left\{ \ddot{\mathbf{x}}_{2} - \dot{\bm{J}}_{2} \dot{\mathbf{q}} - \bm{J}_{2} \bm{J}_{1}^{+} \left(\ddot{\mathbf{x}}_{1} - \dot{\bm{J}}_{1} \dot{\mathbf{q}} \right) \right\} \\
& + \bm{J}_{2|1}^{+} \left( \dot{\bm{J}}_{2} \dot{\mathbf{q}}_{0} - \bm{J}_{2} \bm{J}_{1}^{+} \dot{\bm{J}}_{1} \dot{\mathbf{q}} - \bm{J}_{2} \dot{\bm{J}}_{1}^{+} \dot{\mathbf{x}}_{1} - \dot{\bm{J}}_{2|1} \dot{\mathbf{q}}_{0} \right)
\end{split}$$ For simplicity, we define $ \bm{X}\triangleq\ddot{\mathbf{x}}_{2} - \dot{\bm{J}}_{2} \dot{\mathbf{q}} - \bm{J}_{2} \bm{J}_{1}^{+} \left(\ddot{\mathbf{x}}_{1} - \dot{\bm{J}}_{1} \dot{\mathbf{q}} \right)$. Then the equation on $\ddot{\mathbf{q}}_{0}$ can be further expressed as $$\begin{split}
\ddot{\mathbf{q}}_{0} &= \bm{J}_{2|1}^{+} \bm{X}
+ \bm{J}_{2|1}^{+} \left(-\bm{J}_{2} \bm{J}_{1}^{+} \dot{\bm{J}}_{1} \dot{\mathbf{q}} - \bm{J}_{2} \dot{\bm{J}}_{1}^{+} \dot{\mathbf{x}}_{1} - \bm{J}_{2} \dot{\bm{N}}_{1} \dot{\mathbf{q}}_{0} \right) \\
&= \bm{J}_{2|1}^{+} \bm{X} + \bm{J}_{2|1}^{+} \left\{ -\bm{J}_{2} \bm{J}_{1}^{+} \dot{\bm{J}}_{1} \left( \bm{J}_{1}^{+}\dot{\mathbf{x}}_{1}+ \dot{\mathbf{q}}_{0} \right) \right. \\
& \left. - \bm{J}_{2} \dot{\bm{J}}_{1}^{+} \dot{\mathbf{x}}_{1} + \bm{J}_{2}\left(\dot{\bm{J}}_{1}^{+} \bm{J}_{1} + \bm{J}_{1}^{+} \dot{\bm{J}}_{1} \right) \dot{\mathbf{q}}_{0} \right\} \\
&= \bm{J}_{2|1}^{+} \bm{X} + \bm{J}_{2|1}^{+} \left( -\bm{J}_{2} \bm{J}_{1}^{+} \dot{\bm{J}}_{1} \bm{J}_{1}^{+}\dot{\mathbf{x}}_{1} - \bm{J}_{2} \dot{\bm{J}}_{1}^{+} \dot{\mathbf{x}}_{1} +\bm{J}_{2} \dot{\bm{J}}_{1}^{+} \bm{J}_{1}\dot{\mathbf{q}}_{0} \right)
\end{split}$$ Because $\bm{J}_{1} \dot{\mathbf{q}}_{0} = \bm{0}$, the previous equation becomes $$\begin{split}
\ddot{\mathbf{q}}_{0} &= \bm{J}_{2|1}^{+} \bm{X} + \bm{J}_{2|1}^{+} \left( -\bm{J}_{2} \bm{J}_{1}^{+} \dot{\bm{J}}_{1} \bm{J}_{1}^{+}\dot{\mathbf{x}}_{1} - \bm{J}_{2} \dot{\bm{J}}_{1}^{+} \dot{\mathbf{x}}_{1} \right)\\
&= \bm{J}_{2|1}^{+} \bm{X} + \bm{J}_{2|1}^{+} \left( \bm{J}_{2} \bm{J}_{1}^{+} \bm{J}_{1} \dot{\bm{J}}_{1}^{+}\dot{\mathbf{x}}_{1} - \bm{J}_{2} \dot{\bm{J}}_{1}^{+} \dot{\mathbf{x}}_{1} \right)\\
&=\bm{J}_{2|1}^{+} \bm{X} + \bm{J}_{2|1}^{+} \bm{J}_{2} \left( \bm{J}_{1}^{+} \bm{J}_{1} - \bm{I} \right) \dot{\bm{J}}_{1}^{+} \dot{\mathbf{x}}_{1}\\
&=\bm{J}_{2|1}^{+} \bm{X} - \bm{J}_{2|1}^{+} \bm{J}_{2} \bm{N}_{1} \dot{\bm{J}}_{1}^{+} \dot{\mathbf{x}}_{1}
\textrm{.}
\end{split}$$ Let us develop the term below using the above expression, $$\begin{split}
-\dot{\bm{N}}_{1} \bm{J}_{1}^{+}\dot{\mathbf{x}}_{1} + \bm{N}_{1} \ddot{\mathbf{q}}_0 &= -\dot{\bm{N}}_{1} \bm{J}_{1}^{+}\dot{\mathbf{x}}_{1} + \ddot{\mathbf{q}}_0 \\
&=\bm{J}_{2|1}^{+} \bm{X} - \dot{\bm{N}}_{1} \bm{J}_{1}^{+}\dot{\mathbf{x}}_{1} - \bm{J}_{2|1}^{+} \bm{J}_{2} \bm{N}_{1} \dot{\bm{J}}_{1}^{+} \dot{\mathbf{x}}_{1} \\
&= \bm{J}_{2|1}^{+} \bm{X} + \left(\bm{I}- \bm{J}_{2|1}^{+} \bm{J}_{2} \bm{N}_{1} \right) \bm{N}_{1} \dot{\bm{J}}_{1}^{+} \dot{\mathbf{x}}_{1} \\
&= \bm{J}_{2|1}^{+} \bm{X} + \bm{N}_{2|1} \bm{N}_{1} \dot{\bm{J}}_{1}^{+} \dot{\mathbf{x}}_{1}
\end{split}$$ Thus, equation (\[eq:qdot\_middle\]) becomes $$\begin{split}
\ddot{\mathbf{q}} &= \bm{J}_{1}^{+}\left( \ddot{\mathbf{x}}_{1} - \dot{\bm{J}}_{1} \dot{\mathbf{q}}\right) + \bm{J}_{2|1}^{+} \bm{X} + \bm{N}_{2|1} \bm{N}_{1} \dot{\bm{J}}_{1}^{+} \dot{\mathbf{x}}_{1}\\
&= \bm{J}_{1}^{+}\left( \ddot{\mathbf{x}}_{1} - \dot{\bm{J}}_{1} \dot{\mathbf{q}}\right) + \bm{J}_{2|1}^{+} \bm{X} + \bm{N}_{2|1} \bm{N}_{1} \ddot{\mathbf{q}}_{res}
\end{split}
\label{eq:qdot_final}$$
Related Work {#sec:related_work}
============
Reinforcement Learning based Locomotion Planner
-----------------------------------------------
One of the main challenges for learning robust dynamic locomotion policies is handling the high number of continuous variables describing the motion and force interactions of full humanoid robots. To deal with the dimensionality problem, we review previous work that has greatly inspired us. [@Morimoto:2007eh], solves a periodic locomotion generation problem via RL on a planar biped robot. We advance upon this work by solving the learning problem for 3D robots, avoiding reliance on human walking trajectories, and generating policies for non-periodic gaits. Other important works employ learning as an optimization problem over known locomotion trajectories. In [@Sugimoto:2011ge], a periodic locomotion problems is solved by optimizations of a known stable central patter generated (CPG) walking trajectories using RL. However, no focus is given to dealing with large external disturbances. In addition, our focus is on the generation of trajectories from scratch without prior stable locomotion patterns.
In [@Missura:2014kn], robust walking trajectories to external pushes are achieved based on capture point trajectory optimization via gradient based learning updates. In this work, the capture point method is used as an analytic controller to initiate the learning process with information about foot placement, step-timing, and ZMP controls. Although the authors also show learning of push recovery strategies without previous capture point generated trajectories, our focus is stronger on autonomously learning the locomotion process without reliance on already stable walking gaits. As such, we belief our algorithm is able to learn from scratch recovery strategies in a more generic sense, for instance to recover from pushes in any direction while walking. Like ours, autonomous learning of periodic gaits has been explored before in passive dynamic walkers [@Tedrake:2004ip]. Once more, our focus is on gait generators that can produce non-periodic gaits and tolerate large push disturbances in all directions of motion. The dynamic locomotion community has previously used online optimization methods instead of RL, such as model predictive control (MPC). The main problem of these approaches is the high computational cost. To mitigate this problem, researchers have made significant efforts to develop efficient computational processes. [@Erez:2013cl] used the gradient of a cost function to solve the MPC problem efficiently. [@Khadiv:2017th] linearized the planning problem by optimizing over one step ahead of time. Our approach relying on learned neural networks replaces the need for complex online computations, enabling the generation of hundreds of steps in an instant compared to the stepping time scales. [@Whitman:2009im] proposed a controller for a 12 DoF biped system by using dynamic programming and a lookup table that was obtained offline based on simple models. The multiple policies achieved from each simple model were combined to control the target system. In contrast, our work relies on the generic inverted pendulum locomotion model, and a versatile full-humanoid body controller, i.e., WBLC.
Whole-Body Control
------------------
WBC [@sentis2005synthesis] is a family of multi and prioritized task-space trajectory controllers for humanoid robots that rely on floating-base dynamic and computed torque commands as inputs to the plant. It yields asymptotically stable control policies for multiple tasks with simultaneous control of operational forces when needed. Priorities address resource allocation when two or more task trajectories cannot physically be tracked by the robotic system. It naturally integrates equality constraints such as biarticular transmission constraints [@sentis2013implementation]. Other groups have explored richer versions of WBC with inequality constraints such as joint limit avoidance [@flacco2012motion; @mansard2009unified; @lee2012intermediate], collision avoidance [@kanoun2011kinematic], and singularity avoidance [@moe2015stability]. Several groups used evolved and more practical versions of WBC such as controllers used in the DARPA Robotics Challenge of 2013 and 2014. For instance, [@koolen2013summary; @johnson2015team] incorporate reaction forces as inequality constraints based on solving a quadratic programming optimization problem with desired center of mass trajectories. Treatment of reaction forces as inequality constraints in the WBC communities dates back to the work by [@stephens2010dynamic]. And it showcases one of the weaknesses of our group’s formulation of WBC. In early versions [@sentis2010compliant] we treated reaction forces as equality constraints. Such treatment corresponds to bilateral contact constraints, i.e. assuming that the floor contacts are actually rigid anchors. This is obviously an inaccurate model. One of the main objectives of this paper is to use a realistic unilateral contact model for WBC while maintaining one of its main strengths, efficient prioritized control.
Bipedal and quadrupedal walking capabilities have been devised using WBC. [@hutter2014quadrupedal] demonstrated locomotion of a quadrupedal robot by utilizing hierarchical tasks based on least-square problems. The integration of the versatile capture point (CP) as an operational space of WBC was proposed and controlled either as a constraint or a task for bipedal humanoid robots [@ramos2014whole]. The robot’s Center of Gravity (CoG) has been used as a task controller for a while, such as in [@mistry2007task]. Walking pattern generators have been incorporated into WBC in multiple instances such as in [@carpentier2016versatile]. During the DARPA robotics challenge, several top participants incorporated WBC’s into their strategy for achieving mobile dexterous capabilities. For instance, high-level trajectory optimization and low-level optimization with inverse dynamics were integrated into the framework by [@feng2015optimization].
As stated before, during the DRC several humanoid robots were controlled via WBC including QP solvers for dealing with reaction forces. By introducing QP and task hierarchy (HQP), whole-body motion of humanoid robots could be controlled with the intrinsic reactive advantages of task prioritization [@escande2014hierarchical]. Compared with projection-based WBC algorithms, optimization-based WBC, such as HQP, can incorporate multiple inequality constraints [@saab2013dynamic], which are useful for describing contact conditions such as friction cones [@abe2007multiobjective]. Overall, optimization-based WBC have been a success for practical applications [@koolen2013summary; @feng2015optimization; @kuindersma2014efficiently]. However, their computational cost remains a challenge, specially if being considered as models for motion planning, such as model predictive control. Therefore, efficiency of our newly proposed whole-body controller, dubbed WBLC, is a key consideration of this paper. To achieve the speed boost, we rely on a projection-base formulation. However, it is difficult to incorporate inequality constraints into analytical projection-based methods; thus, our goal is to combine both and also to maintain desired task hierarchy capabilities. The proposed WBLC incorporates an efficient QP, the dimension of which depends only on the number of contact points, and a joint acceleration level controller which only relies on projective operators, thus yielding the speed efficiency that we advocate for.
Acknowledgment {#acknowledgment .unnumbered}
==============
The authors would like to thank the members of the Human Centered Robotics Laboratory at The University of Texas at Austin for their great help and support. This work was supported by the Office of Naval Research, ONR Grant \[grant \#N000141512507\] and NASA Johnson Space Center, NSF/NASA NRI Grant \[grant \#NNX12AM03G\].
[^1]: D. Kim and J. Lee are with the Department of Mechanical Engineering, University of Texas at Austin, Austin, TX, 78712 USA e-mail: .
[^2]: L. sentis is with the Department of Aerospace Engineering and Applied Mechanics, University of Texas at Austin, Austin, TX, 78712
[^3]: Manuscript received April 19, 2005; revised August 26, 2015.
[^4]: Seoul National University Robotics Library. Physics-based simulation. Open-source <http://robotics.snu.ac.kr/srlib/>
[^5]: Open-source <https://rbdl.bitbucket.io>
| {
"pile_set_name": "ArXiv"
} |
The recent discovery of pseudogap in underdoped high $T_{c}$ cuprates has challenged condensed matter physicists for several years. The pseudogap behavior[@Timusk:1999] is observed as strong suppression of low frequency spectral weight below some characteristic temperature $T^{*}$ higher than transition temperature $T_{c}$. This anomalous phenomenon has been observed in angle resolved photoemission spectroscopy (ARPES),[@Ding:1996; @Loeser:1996] specific heat,[@Loram:1993] tunneling,[@Renner:1998] NMR,[@Takigawa:1991] and optical conductivity.[@Homes:1993] One of the most puzzling questions in pseudogap phenomena is why $T^{*}$ has a completely different doping dependence from $T_{c}$, in spite of their possibly close relation. In this paper we demonstrate that induced local spin-singlet amplitude due to short-range spin correlations can cause a normal state pseudogap with $d$-wave symmetry even in the absence of pairing interactions. First of all we argue that there are two energy scales in the problem, because the pseudogap appears as a crossover phenomenon according to experiments. The low energy (or long distance) physics of antiferromagnetic (AF) and superconducting (SC) correlations is well captured by a [*static*]{} mean-field approach, while the relatively high energy (or short distance) physics of the pseudogap is invisible in such a study. Thus we resort to fluctuation theory in order to describe the dynamical nature of the pseudogap, and to determine $T^{*}$ and pseudogap size $\Delta_{pg}$. The mean-field result of the $t-J$ model will be used below solely to find the onset of leading correlations, and to compute mean-field AF and SC order parameters for the calculation of local spin-singlet amplitude. The mean-field $t-J$ Hamiltonian reads $$\begin{aligned}
H_{MF} = \sum_{\vec{k}, \sigma}\varepsilon(\vec{k})
c^{\dag}_{\vec{k},\sigma}
c_{\vec{k},\sigma}
-4Jm\hat{m} -Js(\hat{s}+\hat{s}^{\dag}) \; ,
\label{eq10}\end{aligned}$$ where $\varepsilon(\vec{k}) \simeq -2tx(\cos k_{x}+\cos k_{y})-\mu$ with $x$ the hole density, $\hat{m}=1/(2N)\sum_{\vec{k},\sigma}\sigma c^{\dag}_{\vec{k}+\vec{Q},\sigma}
c _{\vec{k},\sigma}$ and $\hat{s}=1/N\sum_{\vec{k}}\phi_{d}(\vec{k})
c_{\vec{k},\uparrow} c_{-\vec{k},\downarrow}$ with $N$ the total number of lattice sites. $m$, $s$ are mean-field AF, SC order parameters determined from $m=\langle \hat{m} \rangle$ and $s=\langle \hat{s} \rangle$, respectively. $\phi_{d}(\vec{k}) = \cos k_{x}-\cos k_{y}$ and $\vec{Q}$ is the (commensurate) AF wave vector $(\pi,\pi)$ in two dimensions. In this paper we restrict ourselves to a uniform solution which is just enough for our purpose. In a mean-field approximation, mean-field order already sets in when the correlation length reaches roughly one lattice spacing. This forces the above mean-field phase line (Fig. \[fig1\](a)) to be interpreted as the onset of the corresponding short-range correlations. We identify $T_{N}^{MF}$ with another crossover temperature $T^{0}$ at which some magnetic experiments such as Knight shift show their maximum. For the parameter ($t/J=3.0$) used in this paper, short-range spin correlations disappear at $x=x_{c} \simeq 0.19-0.20$ at low temperature.
to 6.5cm [to -5.0cm [ ]{} ]{}
We introduce spin-singlet[@Comment15] correlation function $$\begin{aligned}
\chi_{pp}(i,\tau) = \langle T_{\tau} \Delta_{d}(i,\tau)
\Delta^{\dag}_{d}(0,0) \rangle \; ,
\label{eq20}\end{aligned}$$ where $ \Delta_{d}(i) = \frac{1}{2}\sum_{\delta}g(\delta)
(c_{i+\delta,\uparrow} c_{i,\downarrow}
-c_{i+\delta,\downarrow} c_{i,\uparrow}) $ with $g(\delta) = 1/2$ for $\delta=(\pm 1,0)$, -1/2 for $\delta=(0,\pm 1)$, and 0 otherwise. The spin-singlet correlation function is related to the [*local*]{} spin-singlet amplitude through the sum rule $$\begin{aligned}
\frac{T}{N}\sum_{q}\chi_{pp}(q)e^{-i\nu_{m}0^{-}}
= \langle |\Delta_{d}(0)|^{2} \rangle
\; ,
\label{eq40}\end{aligned}$$ where $q=(\vec{q},i\nu_{m})$ and $\nu_{m}$ is bosonic Matsubara frequencies. $T$ is absolute temperature. In terms of renormalized vertex $V_{pp}$,[@Vilk:1997] we approximate the spin-singlet correlation function $$\begin{aligned}
\chi_{pp}(q) = \frac{ \chi^{0}_{pp}(q)}{1-V_{pp}\chi^{0}_{pp}(q)} \; ,
\label{eq50}\end{aligned}$$ where the irreducible susceptibility is defined as $\chi^{0}_{pp}(q) = \frac{T}{4N}\sum_{k}
(\phi_{d}(\vec{k})+\phi_{d}(\vec{q}-\vec{k}))^{2}
G^{0}(q-k)G^{0}(k)$. $G^{0}(k)$ is the noninteracting Green’s function obtained from Eq. (\[eq10\]) with $J=0$. Now the unknown vertex, $V_{pp}$, is determined by the sum rule Eq. (\[eq40\]).[@Vilk:1997] Hence, an increase in the local spin-singlet amplitude evaluated in the interacting state over that in the noninteracting state leads to a nonvanishing positive value of $V_{pp}$, namely, the enhancement of the correlation function. This (non-perturbative sum rule) approach has been shown to be quite reliable[@Vilk:1997] as long as short range correlations are concerned. In our calculations, the pseudogap appears when the spin-singlet correlation length reaches about 1 lattice constant. The self-energy due to the spin-singlet correlation function is given by[@Vilk:1997] $$\begin{aligned}
\Sigma_{pp}(k) &=& - \frac{1}{4}VV_{pp}
\frac{T}{N}\sum_{q}
\nonumber \\
& & (\phi_{d}(\vec{k})+\phi_{d}(\vec{q}-\vec{k}))^{2}
\chi_{pp}(q)G^{0}(q-k)
\; ,
\label{eq70}\end{aligned}$$ where $V=J$ from Eq. (\[eq10\]). First let us begin by showing the interaction-induced local spin-singlet (solid curve) amplitude (Fig.\[fig1\](b)) evaluated in the mean-field state of the $t-J$ Hamiltonian in a region where $s=0$ (or $T > T^{MF}_{c}$). Since $s=0$, the spin-singlet amplitude is entirely caused by short-range spin correlations in the absence of pairing interactions. Although in general a mean-field state is not accurate for strongly correlated electron systems, certain local and short-range [*static*]{} quantities such as double occupancy and nearest neighbor correlations are reasonably well captured by the mean-field state particularly with AF order (See Ref. [@Kyung:2000-3] for more details). In fact the interaction-induced local spin-singlet amplitude (Eq. (\[eq40\])) is determined most crucially by these quantities. Quite unexpectedly, the local spin-singlet amplitude increases with decreasing doping despite the fact that the mean-field SC order $s$ is absent. The increase of local spin-singlet amplitude traces back to the growing short-range spin correlations with decreasing doping.[@Kyung:2000-3] In Fig. \[fig1\](a) we show the pseudogap temperature $T^{*}$ (filled diamonds) where the single particle spectral function $A(\vec{k},\omega)$ near $\vec{k}=(\pi,0)$ starts to be split into two peaks. $T^{*}$ falls from a high value onto the $T_{c}$ ($\le T_{c}^{MF}$) line instead of sharing a common line with $T_{c}$ in overdoped region. It is not surprising to find that $T^{*}$ closely follows $T_{N}^{MF}=T^{0}$, because in our study the pseudogap is caused by induced local spin-singlet amplitude due to short-range spin correlations, which is reasonably well captured by the mean-field state with AF order. When superconductivity is suppressed by setting $s=0$, $T^{*}$ vanishes near $x_{c}$ where short-range spin correlations disappear. All these features are at least qualitatively consistent with the findings by Tallon and Loram.[@Tallon:2001]
to 6.5cm [to -5.0cm [ ]{} ]{}
Figure \[fig3\](a) shows the pseudogap size $\Delta_{pg}$ (filled diamonds) by setting $s=0$ at $T=0$. In the same figure the pseudogap energy extracted from various experiments by Tallon and Loram[@Tallon:2001] is also shown as empty symbols for comparison . $\Delta_{pg}$ vanishes near $x_{c}$, suggesting the presence of a quantum critical point at a critical doping. The agreement between our results and experiments appears remarkable for such a simple approximation used in this paper. The linear vanishing of $\Delta_{pg}$ near $x_{c}$ is closely related to the corresponding behavior of the induced local spin-singlet amplitude. The total excitation gap (or ARPES leading edge gap or SC gap) $\Delta_{tg}$ at $T=0$ is larger than $\Delta_{pg}$ due to the additional contribution to the local spin-singlet amplitude from $s \ne 0$, which is shown together with $\Delta_{pg}$ in Fig. \[fig3\](b). $\Delta_{pg}$, $\Delta_{tg}$ and their relative ratio $\Delta_{pg} / \Delta_{tg}$ are all monotonically decreasing functions of doping, as shown in the inset of Fig. \[fig3\](b). Since the SC order parameter vanishes at $T_{c}$ (at $T^{MF}_{c}$ in this paper), the SC gap below $T_{c}$ continuously evolves into the normal state pseudogap above $T_{c}$ with the same momentum dependence and magnitude. The pseudogap appears here only as the suppression of low frequency spectral weight in certain physical quantities which are obtained through $A(\vec{k},\omega)$ or its convolution with a relevant vertex. It does not appear as a thermodynamic phase with broken symmetry. The present scenario for the pseudogap predicts that a normal state pseudogap is likely to appear when short-range spin correlations are well established and are not masked by long-range (AF or SC) order. The present results are robust to variations of $t/J=2.5-3.5$[@Comment40] and small to moderate value of $t'$. In summary, we have shown that the induced local spin-singlet amplitude due to short-range spin correlations causes a normal state pseudogap even in the absence of pairing interactions. $T^{*}$ falls from a high value onto the $T_{c}$ line and closely follows $T_{N}^{MF}$. The calculated pseudogap size is in good agreement with experimental results. It would be interesting to see how robust are the features found in this paper, when the no-double-occupancy constraint is strictly imposed on the $t-J$ model and an inhomogeneous solution is used. The author would like to thank Prof. A. M. Tremblay for numerous help and discussions throughout the work. He also thanks Profs. H. Ding, J. W. Loram, J. L. Tallon and T. Timusk for useful discussions. The present work was supported by a grant from the Natural Sciences and Engineering Research Council (NSERC) of Canada and the Fonds pour la formation de Chercheurs et d’Aide à la Recherche (FCAR) of the Québec government.
T. Timusk and B. Statt, Rep. Prog. Phys. [**62**]{}, 61 (1999).
H. Ding, [*et al.*]{}, Nature [**382**]{}, 51 (1996).
A. G. Loeser, [*et al.*]{}, Science [**273**]{}, 325 (1996).
J. W. Loram, [*et al.*]{}, Phys. Rev. Lett. [**71**]{}, 1740 (1993).
Ch. Renner, [*et al.*]{}, Phys. Rev. Lett. [**80**]{}, 149 (1998).
M. Takigawa, [*et al.*]{}, Phys. Rev. B [**43**]{}, 247 (1991).
C. C. Homes, [*et al.*]{}, Phys. Rev. Lett. [**71**]{}, 1645 (1993).
The spin-singlet correlation function (with $d$-wave symmetry) is more general than the $d$-wave pair correlation function in the sense that the spin-singlet correlations can be also induced by short-range spin correlations even in the absence of pairing interactions.
Y. Vilk and A. M. Tremblay, J. Phys. I (France) [**7**]{}, 1309 (1997); B. Kyung, Phys. Rev. B [**63**]{}, 14 502 (2001); B. Kyung, S. Allen, A.-M. S. Tremblay, cond-mat/0010001 and to appear in Phys. Rev. B.
B. Kyung, Phys. Rev. B [**63**]{}, 214 505 (2001).
J. L. Tallon and J. W. Loram, Physica C [**349**]{}, 53 (2001).
For $t=2.5J$ and $3.5J$, $\Delta_{pg}$’s vanish with slightly different slopes and $x_{c}$’s from the experimental results (in opposite ways).
| {
"pile_set_name": "ArXiv"
} |
---
address: |
Department of Physics, University of Southampton, Highfield,\
Southampton SO17 1BJ, UK
author:
- 'T.R. MORRIS'
title: PROPERTIES OF DERIVATIVE EXPANSION APPROXIMATIONS TO THE RENORMALIZATION GROUP
---
This talk is about derivative (or more general momentum) expansions as applied to the renormalization group, in a quantum field theory setting. The motivation is simply this: I want to construct analytic approximation methods with as much reliability and accuracy as possible [*even when there are no obviously small parameters*]{}, $\epsilon=4-D$, $1/N$ with which one could expand perturbatively. (Here, $D$ is space-time dimension, and $N$ is the number of components of the field.) In other words, I want to look for approximations that work in a genuinely setting. Now, as Wilson was instrumental in demonstrating, the continuum limit of a quantum field theory can ly be best understood in terms of the flow of an effective action $S_\Lambda[\phi]$ under lowering an effective U.V. (ultra-violet) cutoff $\Lambda$.[@kogwil] Thus in this framework one works with a flow equation that generically takes the form S\_ = [F]{}\[S\_\], where the cutoff is implemented through some effective U.V. cutoff function $C_{UV}(q,\Lambda)$. (Here $q$ stands for momentum, and the above equation is referred to as the continuous, or momentum space, renormalization group.) Scale invariant continuum limits (thus massless field theories) are then simply given by fixed points: [^1] S\_=0. The massive continuum limits follow from tuning the relevant perturbations around these fixed points. In such a setting one realises that various approximations can be made quite easily that preserve the structure of the continuum limit, while in other frameworks (for example when using truncations of Dyson-Schwinger equations) the continuum limit, equivalently renormalisability, is almost inevitably destroyed.[@erg]
What are the possible approximations? The first thought is to try truncating the space of interactions to just a few operators, however this results in a truncated expansion in powers of the field $\phi$ (about some point). Such an approximation can only be sensible if the field $\phi$ does not fluctuate very much, which is the same as saying that it is close to mean field, in a setting in which weak coupling perturbation theory is anyway valid. Studying the behaviour of truncations in a truly situation, one finds that higher orders cease to converge and thus yield limited accuracy, while there is also no reliability – even qualitatively – since many spurious fixed points are generated.[@trunc]
This situation should be contrasted with truncations to a few operators in the real space renormalization group of spin systems, such as block spin renormalization group of the Ising model. Such truncations were extensively studied in the past,[@rsRG] and could be very accurate.[@burk] A modern variant produces spectacularly accurate results in low dimensions.[@white] The powerful Monte-Carlo renormalization group methods,[@MCrev] are also based on such truncations. In the case of such simple discrete systems however, the expansion effectively results in a short distance expansion of the effective action, for example by keeping only the finite number of interactions linking nearest neighbours, then next-to-nearest neighbours, and so on.
The analogous expansion in our continuum case is, for smooth cutoff functions $C_{UV}$, a derivative expansion of $S_\Lambda[\phi]$ (equivalent to a Taylor expansion in the momenta of its vertices),[@deriv; @twod] while for sharp cutoff functions $C_{UV}(q,\Lambda)=\theta(\Lambda-q)$ it is an expansion in momentum scale $\sim |\partial/\Lambda| \equiv p/\Lambda$, where the coefficients are not analytic in $p^\mu$ but rather, non-trivial functions of the angles between various momenta which must be determined self-consistently through the flow equation.[@truncm] \[The non-analyticity is a purely technical problem that is induced by the non-analyticity of $\theta(\Lambda-q)$.\] At any rate, in both cases such a short distance expansion – [*where no other approximation is made*]{} – seems a particularly natural approximation to try, and in view of the discussion above, sensible results might well be expected providing the Wilson effective action $S_\Lambda$ is ‘sufficiently well behaved’: thus the approximation would fail if the higher derivative terms are not in some sense small, but this would indicate that a description in terms of the given field content is probably itself inappropriate and other degrees of freedom should be introduced. This is an important point, to which I will return later.
Consider the case of $O(N)$ invariant scalar field theory, the so-called $N$-vector model. I shall start by using a sharp cutoff and making the simplest such approximation – keeping only a potential interaction: S\_\~d\^Dx{(\_\^a)\^2+V(,) }. After appropriate scalings to dimensionless combinations, the flow equation is found to be: \[flo\] -V+dV’ -DV=(1+V”)+(N-1)(1+[V’/]{}). Here $\phi\equiv\sqrt{\phi^a\phi^a}$, $\prime\equiv\partial/\partial\phi$, and I have introduced the dimension of the field $\phi$: $d={D/2}-1$. (Since we have thrown away all momentum dependent corrections, $\eta=0$ in this approximation.) The $N=\infty$ case of this equation was already derived by Wegner and Houghton in their paper introducing the sharp-cutoff flow equation,[@wegho] and subsequently the general $N$ case was proposed as a “local potential” approximation by Nicoll, Chang and Stanley.[@nico] It has since been rediscovered by many authors,[@truncm] especially Hasenfratz and Hasenfratz.[@hashas] Nevertheless this equation, and its smooth cutoff sisters, have a number of beautiful properties that have not been pointed out by previous workers.
First note that the fixed point equation for $V(\phi,\Lambda)\equiv V(\phi)$, \[fpe\] dV’ -DV=(1+V”)+(N-1)(1+[V’/]{}), determines [*by itself*]{} at most a countable set [^2] of sensible fixed point potentials, each of which can be identified with approximations to the exact fixed points. This is not obvious because eqn.(\[fpe\]) is a second order ODE (Ordinary Differential Equation) and therefore in some neighbourhood of some starting value $\phi$ we can construct a continuously infinite two parameter set of solutions. Actually, all but a countable number of those solutions are singular! For example, consider the case $D=4$ and $N=4$ (the Higgs field in the Standard Model). Obviously from (\[fpe\]), $V'(0)=0$ is necessary if the potential is not to be singular at $\phi=0$. We can choose a value of $V(0)$ and then (numerically) integrate out to $\phi>0$ using eqn.(\[fpe\]). One discovers almost always that at some point $\phi=\phi_c$ a singularity in $V$ is encountered, after which the potential ceases exist (or at least is complex for $\phi>\phi_c$). The first graph in fig.1 is a plot of $\phi_c$ against $V(0)$. We see that only the (trivial) Gaussian fixed point solution $V(\phi)\equiv0$ exists for all values of the field. If the same is done for the case $D=3$ and $N=1$, we get the second graph in fig.1. In this case there is also one non-trivial non-singular solution, corresponding to the famous Wilson-Fisher fixed point (Ising model universality class). I have checked all cases $D=3,4$ and $N=1,2,3,4$; they reproduce the standard fixed points.[@trunc]
-2cm
=0.9 -2cm
Studying eqn.(\[fpe\]), one can convince oneself that the only way that $V$ can satisfy this equation as $\phi\to\infty$ is if $V(\phi)\sim\phi^{D/d}$. Together with $V'(0)=0$, we now have two boundary conditions and thus we should expect only a countable number of solutions from the second order ODE. For $D=2$, exceptions arise (due to $d=0$), thus for $N=1$ one obtains a semi-infinite continuous line of fixed points with periodic potentials, but these may be identified with critical sine-Gordon models.[@twod] On the other hand, there is no need to impose $V'(0)=0$ when $N=1$. This just allows a constant phase shift on critical sine-Gordon potentials, while in higher than two dimensions the power law constraint on $V$ now holds separately for $\phi\to\infty$ and $\phi\to-\infty$ (with possibly different coefficients). Nevertheless, we have confirmed that in this larger space, there is still only the one non-trivial fixed point in three dimensions.
While all this just reproduces the standard lore, note nevertheless how powerful the method is: the [*infinite dimensional*]{} space of [*all possible potentials*]{} $V(\phi)$ has been searched for continuum limits. Clearly this is much more than is possible with other methods! Also, the continuum is actually accessed [*directly*]{} without the need to go through the construction of introducing an overall cutoff $\Lambda_0$, a bare action $S_{\Lambda_0}$, and then taking the continuum limit $\Lambda_0\to\infty$.
These properties are true also when the approximation is applied to the massive theory. In this case one must determine the form of the perturbations about the fixed point. One can write V(,)=V()+ v\_() \^[-]{}, where eqn.(\[flo\]) is linearised in $\epsilon$ and separation of variables has been used. Now, $v_\lambda(\phi)$ satisfies a linear second order ODE, and once again we appear to have a continuously infinite set of solutions, and for all choices of $\lambda$. In this case the crucial observation is that if, beyond the linearised level, the scale dependence of the perturbation is to be absorbed into an associated coupling $g_\lambda(\Lambda)$, a renormalised coupling corresponding to universal self-similar flow about the fixed point, then $v_\lambda(\phi)$ has to behave as $v_\lambda(\phi)\sim \phi^{(D-\lambda)/d}$ as $\phi\to\infty$. [@hh] For the same reasons as before, this typically allows only a countable number of solutions, but this time we also have linearity, which implies a normalization condition can be set, overconstraining the equations and resulting in quantization of $\lambda$. (Again, $d=0$ provides an exception – resulting in more general perturbations with exponential or periodic behaviour. It is worth remarking that for $d\ne0$, the power laws given above are the unique powers required so that the physical [^3] $v_\lambda$ and $V$ are independent of $\Lambda$, and therefore obtain a non-trivial finite limit as $\Lambda\to0$; this limit gives the Legendre effective potential [@erg] and thus the equation of state.)
Consider now the derivative expansion at $O(\partial^2)$. In this case we need to use a smooth cutoff (as already discussed). The effective action takes the form S\_\~d\^Dx{V()+ (\_\^a)\^2 K()+ (\^a\_\^a)\^2 Z()}(where this last term is required only for $N\ne1$). In this case the fixed point equations are a set of coupled second-order non-linear ODEs, one for each coefficient function ($V$, $K$ and $Z$). This pattern holds to all orders of the derivative expansion. As previously, one can argue for specific power law behaviours for $V$, $K$ and $Z$, and that typically only a countable number of non-singular solutions exist. But now there is another parameter to determine: the critical exponent $\eta$ from the anomalous scaling of $\phi$. The [*exact*]{} renormalization group has a $\phi$ reparametrization invariance [@red; @rie] reflecting the fact that physics is independent of the normalization of the field, and this extra invariance turns the fixed point equations into non-linear eigenvalue equations for $\eta$, because it allows a normalization condition ($K(0)=1$) and thus quantization of $\eta$ (in a similar way to the linear case above for perturbations.) There is a problem however: the derivative expansion generally breaks the reparametrization invariance, with the result that $\eta$, $\nu$, $\omega$ depend on some unphysical parameter such as $K(0)$, the normalization of the kinetic term. [@gol]
Consider the Polchinski form [@pol] of the Wilson flow equation. Schematically, \[Pol\] [S\_]{}=[12]{}[tr]{} [\_[UV]{}]{} { [S\_]{} [S\_]{}-[\^2S\_]{} -2(\^[-1]{}\_[UV]{}) [S\_]{}}, where $\Delta_{UV}(q,\Lambda)=C_{UV}/q^2$, and the total action $S_\Lambda \sim{1\over2}\,\phi.\Delta_{UV}^{-1}.\phi\,
+ S^{int}_\Lambda$. This equation is simply related to the Wilson equation [@kogwil] through $\phi\mapsto\sqrt{C_{UV}}\,\phi$ and ${\cal H}\equiv-S_\Lambda^{int}$, [@deriv; @prep] but in contrast to Wilsons, it has the intuitively nice property that eigen perturbations about the Gaussian fixed point are precisely polynomials in the field and its derivatives. Now in general, the reparametrization symmetry is given by a complicated functional integral transform,[@rie] so it is not surprising that a truncated derivative expansion destroys it – and in fact it is far from clear how to approximate at all in a way which preserves it.
(Let me emphasise that with broken reparametrization invariance, defining $z(\phi)$ through $S^{int}_\Lambda\sim\int\ V(\phi)+\half z(\phi)(\partial_\mu\phi^a)^2 +\cdots$, the results [*depend*]{} on the value $z(0)$. This dependence on $z(0)$ has not been recognized by authors who set $z(0)=0$ with insufficient justification.[@ball])
Fortunately, for two special forms of cutoff function, reparametrization invariance may be linearly realized. These correspond to either sharp cutoff $C_{UV}=\theta(\Lambda-q)$ or a power law smooth cutoff $C_{UV}\sim 1/[1+(q/\Lambda)^{2\kappa+2}]$, where $\kappa$ is some non-negative integer. (The underlying reason is that such cutoffs are left invariant by a subgroup of linearly realised ‘universality symmetries’ that map between different schemes in (\[Pol\]),[@prep] but it would take us too far afield to explain this.) However, a direct derivative/momentum-scale expansion of the Wilson flow equation leads to singular coefficients with both of these cutoffs.[@erg; @twod; @truncm; @ball] The way out of this difficulty is to recognize that, from the form of the right hand side of eqn.(\[Pol\]), the Wilson effective action has a tree structure.[@pol] It is the Taylor expansion of the corresponding propagators that causes the problem.[@erg; @twod; @truncm; @ball] Therefore, to overcome this difficulty we first pull out the one particle irreducible parts.
It can be shown [@erg] that the one particle irreducible parts of $S_\Lambda$ are generated by a Legendre effective action $\Gamma_\Lambda[\phi]$ equipped with infrared cutoff $C_{IR}=1-C_{UV}$, related in the usual way to a partition function except that the bare action is modified to $S_{\Lambda_0}=\half\phi.\Delta_{IR}^{-1}.\phi+ S^{int}_{\Lambda_0}$, where $\Delta_{IR}=C_{IR}/q^2$. Whence, the ‘Legendre flow equation’ follows; [@nici; @erg; @wet] = -[12]{} [tr]{} . In this form, the sharp cutoff limit enjoys the simplest reparametrization invariance:[@truncm] $\phi\mapsto a\,\phi$. For the smooth power law cutoff, certain momentum independent linear transformations on other quantities are also required.[@deriv] Since these transformations are linear and momentum independent, a truncated derivative expansion now [*preserves*]{} the reparametrization invariance, while derivative/momentum-scale expansion of $\Gamma^{int}_\Lambda$ is well-defined with these cutoffs,[@deriv; @truncm] because it results in Taylor expansion of the self-energy, rather than the propagator itself.
Returning now to $O(\partial^2)$, we need the smooth cutoff and choose the integer $\kappa$ as small as possible, to maximise the accuracy of the derivative expansion.[@deriv] Table 1 displays the results obtained in $D=3$ dimensions,[@deriv; @oN] for $\eta$ and $\nu$,[^4] and for the first correction to scaling exponent $\omega$. Only the expected fixed points are found.
\[ht\]
$N$
----- ----------------- --------- ----------------- ----------------- --------- ----------------- ----------------- --------
$O(\partial^2)$ World $O(\partial^0)$ $O(\partial^2)$ World $O(\partial^0)$ $O(\partial^2)$ World
1 .054 .035(3) .66 .618 .631(2) .63 .897 .80(4)
2 .044 .037(4) .73 .65 .671(5) .66 .38 .79(4)
3 .035 .037(4) .78 .745 .707(5) .71 .33 .78(3)
4 .022 .025(4) .824 .816 .75(1) .75 .42 ?
10 .0054 .025 .94 .95 .88 .89 .82 .78
20 .0021 .013 .96 .98 .94 .95 .93 .89
100 .00034 .003 .994 .998 .989 .991 .988 .98
: Critical exponents of the three-dimensional Wilson-Fisher fixed point. The first two orders of the derivative expansion are compared to a combination of the worlds best estimates,[@zinn; @etcN; @N4] with their errors, where available. $\eta$ is identically zero for all $N$ at $O(\partial^0)$.
\[t:2crit\]
For $N=1\cdots4$, the results are fair for the local potential approximation and already quite good at $O(\partial^2)$, with the exception of $\omega$ which gets worse at $O(\partial^2)$ when $N>1$. For the large $N$ cases however, $\omega$ is better estimated, $\nu$ is not improved at $O(\partial^2)$, while $\eta$ is dramatically underestimated – eventually by about a factor of 10. Something is going wrong particularly at large $N$. Actually, the results for $N=\infty$ are guaranteed to be right because in this case the $O(\partial^0)$ approximation is exact,[@wegho; @oN] so the problem only appears in the approach to this limit. I believe that this is an example of a case where all the appropriate fields have [*not*]{} been included in the effective action. Indeed, it is known that at large $N$, a massless bound state field also exists at the critical point.[@zinn; @etcN] We should thus expect that a derivative expansion is ill-behaved, because the vertex functions are hiding within them the effects of this integrated out massless field. To ameliorate this behaviour, we should include the bound state explicitly as an $O(N)$ singlet field, then amongst the new set of fixed points in this enlarged space will be one with the same universal properties as the original $N$ vector model, but with better behaved derivative expansion properties. Similar considerations should apply to fixed points with fermions, particularly since the bound state fields here also correspond to the order parameter (fermion condensate).[@yuri]
The most impressive example so far however, is provided by the case of $D=2$ dimensions. It had been conjectured by Zamalodchikov [@zam] that there should exist an infinite series of multicritical points in the two dimensional $Z_2$ ($\phi\leftrightarrow-\phi$) symmetric theory of a single scalar field, corresponding to the so-called unitary minimal models in CFT (conformal field theory). However a verification of this conjecture is in practice well outside the capabilities of the standard approximation methods: the corresponding $\epsilon$ expansions are so badly behaved as to be useless,[^5] with similar difficulties expected in resummed weak coupling perturbation theory, while lattice methods suffer from difficulties locating and accurately computing the multicritical points in these high dimensional bare coupling constant spaces.[^6] All these methods also get rapidly worse with increasing operator dimension. In constrast, at $O(\partial^2)$, the lowest order at which a fixed point search through $\eta>0$ can be done, we uncover the multicritical points, and [*only*]{} these,[^7] and find an agreement with CFT that [*improves*]{} with increasing multicriticality and dimension.[@twod] The results are displayed in table 2, for the first 10 (multi)critical points and up to the first 10 operators.
\[ht\]
We see that there is a remarkable agreement between these thus lowest order results and CFT, spanning over two orders of magnitude. The worst determined number is $\eta$ for the tricritical point, which is only accurate to 33%, but this gradually improves as $m$ increases. (At $m=11$, $\eta$ is off by 21%.) $\nu$ is worst determined at $m=3$ (13%) after which all are determined to error less than 2% and decreasing with increasing $m$. Indeed, the best determined number is $\nu$ for $m=11$, which is accurate to 0.2%. The worst determined operator dimension is the $3^{\rm rd}$ at $m=5$ (25%), after which errors decrease with increasing $m$ and/or increasing dimension, so that all the rest have errors in the range 9% – 22%. Note that the low dimension operators at high order of multicriticality correspond to renormalization group eigenvalues which agree with CFT to better than 3 significant figures. Fig.2 shows the fixed point solutions for the first three critical points.
We also found some irrelevant operators which cannot be matched to operators in the CFT minimal models. This is the reason for the blank spaces in some corresponding CFT parts of the table. Examples of such operators have also been found (in the correction to scaling at the Ising critical point) by $\epsilon$ expansion and fixed dimension resummed perturbation theory,[@zinn] and argued for in exact treatments.[@schroer] Further work is required to understand their true significance.
=0.9 -2cm
Finally, what are the problems with this approach? The treatment of gauge theory present special problems: the formulation of a flow equation which properly treats the quantum aspects of gauge invariance ly, and secondly, the construction of reliable approximations. The solution to the former must proceed in one of two ways: by allowing the cutoff to break the gauge invariance and then attempting to recover it as the cutoff is removed (via broken Ward identities),[@ell] or by generalizing the flow equations so that gauge invariance is not broken by the cutoff.[@ui] While the first method can be shown to work to all orders in perturbation theory,[@bon] it seems hopeless ly.[@qap] It is surely through an exact preservation of the quantum gauge invariance that real progress will be made; a useable generalization of the pure $U(1)$ case [@ui] may well be possible. Note however, that for the interesting cases (non-Abelian and greater than two dimensions), derivative expansions [*per se*]{} are anyway impractical due to the large number of independent gauge invariant combinations that can be formed even at lowest non-trivial order.[@ui] Instead, much more appropriate forms of approximation deserve study in this case, such as large $N$ methods.
Returning to non-gauge theories, apart from the practical problem that higher orders in the derivative expansion get rapidly more complicated, one problem that this method shares with all other approximations to the renormalization group [@rsRG][^8] is an unphysical dependence on the choice of cutoff function. This dependence is not really a problem however if instead it is used to estimate a rough lower bound on the error of the approximation, and thus test the numerical reliability.[@ball] For example, we can compare the $O(\partial^0)$ $N=1$ results in table 1, to the corresponding results for sharp cutoff, $\nu=.70$ and $\omega=.60$.[@hashas; @deriv] Another problem that has already been mentioned is that generically derivative expansions depend also on one unphysical parameter in the fixed point solutions, due to loss of reparametrization invariance. In a similar way this is not a problem (if one knows about it!), but rather can be used to test numerical reliability in such approximations.[@deriv; @gol] The real question that needs to be answered ultimately, is whether the derivative expansion exists to all orders and converges, because of course if it does, these problems have to lessen and eventually disappear as the expansion is pushed to higher orders. This question is hard to answer in generality, but rather straightforward to analyse perturbatively for a specific theory. In table 3, I show which results converge for the $\beta$ function of the $O(N)$ invariant scalar field theory in $D=4$ dimensions, computed to two loops with different forms of derivative expansion.[@prep] Also summarised are which expansions preserve reparametrization invariance.
\[ht\]
Variant. Repar. Inv.
------------------- ------------- ---------- -----------
one loop two loops
Wilson/Polchinski X X X
Legendre Power $\surd$ $\surd$ X
Legendre Faster X $\surd$ $\surd$
Legendre Sharp $\surd$ $\surd$ $\surd$
: Properties of derivative expansions for different forms of flow equation and cutoff: Wilson or Polchinski flow equation with any smooth cutoff, Legendre flow equation with the Power law cutoff described in the text, Legendre flow equation with some Faster falling cutoffs, and momentum scale expansion of the Legendre flow equation with Sharp cutoff.
A direct derivative expansion of the Wilson effective action does not converge already at one loop: it is again a result of expansion of the propagators inside the effective action (mentioned earlier). On the other hand, the Legendre flow equations give the exact answer to the one loop $\beta$ function already at $O(\partial^0)$.[@erg] Only the last two methods however converge at two loops. In particular, the sharp cutoff case converges very fast at two loops,[@truncm] and since it also preserves a simple reparametrization invariance, further work to overcome the practical difficulties in its implementation [@truncm] certainly seems called for. Although convergence even to all orders in perturbation theory, is no guarantee of convergence ly (truncations in increasing powers of the field trivially must converge at any fixed order of perturbation theory, but as we have seen, fail to do so ly), the natural conjecture is that these last two methods do converge ly. The fact that the Legendre $O(\partial^0)$ approximation is exact at $N=\infty$ [@wegho; @oN] lends further support to this conjecture.[^9]
Of course, negative answers to convergence do not exclude the first two methods from being good model approximations at low orders. Although I have concentrated on the Legendre flow equation with power law cutoff, derivative expansions of the Wilson / Polchinski equation are distinguished by their relative simplicity. Perhaps, by generalising the derivative expansion, one can preserve this simplicity while also preserving more of the structure of the exact renormalization group.
We remind that a large number of references to other work on the lowest order sharp cutoff approximation have been collected.[@truncm] We collect here corresponding smooth cutoff versions not so far mentioned,[@kogwil; @smoo] and similarly attempts to go beyond leading order in the derivative expansion.[@beyo] There are also a number of examples that entertain the idea of derivative expansion but in practice make further truncations.
In conclusion, the derivative (momentum scale) expansion methods – where no other approximation is made – are potentially very powerful, particularly in genuinely settings where all other methods fail. In contrast to more severe truncations, all these variants are robust, in the sense that no spurious solutions have been found, while especially the Legendre flow equation with power law cutoff yields very satisfactory numerical accuracy at low orders. The full potential of these methods is by no means yet realised, and much more theoretical progress on their properties is possible and expected.
Acknowledgments {#acknowledgments .unnumbered}
===============
It is a pleasure to thank the organizers of RG96 for the invitation to speak at this very stimulating conference, several participants – particularly Yuri Kubyshin for helpful conversations, and the SERC/PPARC for financial support through an Advanced Fellowship.
References {#references .unnumbered}
==========
[99]{} K. Wilson and J. Kogut, Phys. Rep. 12C (1974) 75. T.R. Morris, Int. J. Mod. Phys. A9 (1994) 2411. T.R. Morris, Phys. Lett. B334 (1994) 355. “Real-Space Renormalization” (1982), eds. T.W. Burkhardt and J.M.J. van Leeuwen, Springer, Berlin. See particularly T.W. Burkhardt’s review.[@rsRG] S.R. White, Phys. Rev. Lett. 69 (1992) 2863. R.H. Swendsen’s review.[@rsRG] T.R. Morris, Phys. Lett. B329 (1994) 241. T.R. Morris, Phys. Lett. B345 (1995) 139. T.R. Morris, Nucl. Phys. B458\[FS\] (1996) 477. F.J. Wegner and A. Houghton, Phys. Rev. A8 (1973) 401. J.F. Nicoll [*et al*]{}, Phys. Rev. Lett. 33 (1974) 540. A. Hasenfratz and P. Hasenfratz, Nucl. Phys. B270 (1986) 687. T.R. Morris, hep-th/9601128, to be published in Phys. Rev. Lett. F.J. Wegner, J. Phys. C7 (1974) 2098;T.L. Bell and K.G. Wilson, Phys. Rev. B11 (1975) 3431; E.K. Riedel, G.R. Golner and K.E. Newman, Ann.Phys. 161 (1985) 178. G.R. Golner, Phys. Rev. B33 (1986) 7863. J. Polchinski, Nucl. Phys. B231 (1984) 269. T.R. Morris, in preparation. See R.D. Ball [*et al*]{}, Phys. Lett. B347 (1995) 80. J.F. Nicoll and T.S. Chang, Phys. Lett. 62A (1977) 287. C. Wetterich, Phys. Lett. B301 (1993) 90. T.R. Morris and M. Turner, in preparation. J. Zinn-Justin, “Quantum Field Theory and Critical Phenomena” (1993) Clarendon Press, Oxford. S. Ma, Phys. Rev. A10 (1974) 1818; Y. Okabe and M. Oku, Prog. Theor. Phys. 60 (1978) 1277. K. Kanaya and S. Kaya, Phys. Rev. D51 (1995) 2404. Such theories have been considered by J. Comellas, Y. Kubyshin and E. Moreno, hep-th/9512086, hep-th/9601112, and these proceedings. A.B. Zamolodchikov, Yad. Fiz. 44 (1986) 821;J.L. Cardy, in “Phase Transitions and Critical Phenomena”, vol. 11 (1987), eds. C. Domb and J.L. Lebowitz, Academic Press. P.S. Howe and P.C. West, Phys. Lett. B223 (1989) 371;M. Bauer, E. Brézin and C. Itzykson in “Statistical field theory”, vol. 1 (1989), C. Itzykson and J-M. Drouffe, CUP. M. Asorey [*et al*]{}, hep-lat/9410009. B. Schroer, Nucl. Phys. B295\[FS21\] (1988) 586. U. Ellwanger, Phys. Lett. B335 (1994) 364. M. Bonini, M. D’Attanasio and G. Marchesini, Nucl. Phys. B437 (1995) 163; Phys. Lett. B346 (1995) 87; M. Bonini, these proceedings. M. D’Attanasio and T.R. Morris, Phys. Lett. B378 (1996) 213. T.R. Morris, Phys. Lett. B357 (1995) 225. A. Parola, J. Phys. C19 (1986) 5071; V.I. Tokar, Phys. Lett. A104 (1984) 135; G. Felder, Comm. Math. Phys. 111 (1987) 101; G. Zumbach, Nucl. Phys. B413 (1994) 754 and 771; A.E. Filippov and S.A. Breus, Phys. Lett. A158 (1991) 300, Physica A192 (1993) 486; A.E. Filippov, J. Stat. Phys. 75 (1994) 241. G.R. Golner, Phys. Rev. B8 (1973) 339; A.E. Filippov and A.V. Radievsky, Phys. Lett. A169 (1992) 195.
[^1]: once all quantities have been rewritten in terms of dimensionless quantities, using $\Lambda$
[^2]: $D=2$ is an exception: see later
[^3]: in the original dimensionful variables.
[^4]: from which all other exponents (excepting correction to scaling exponents) follow, because all hyper-scaling relations are here exactly preserved
[^5]: already with the tricritical point, $O(\epsilon^2)$ underestimates $\eta$ by a factor $\sim1/100$. The situation gets factorially worse as multicriticality is increased.[@badE]
[^6]: Indeed to date, only the lattice computation of the two lowest operator dimensions around the tricritical point, has been attempted.[@asor]
[^7]: The search was restricted to real $Z_2$–symmetric $V$ and $K$, with $K(0)>0$.
[^8]: and analogously perturbation theory to a certain extent, through scheme dependence.
[^9]: the same is not true of truncations in powers of the field, to any order.
| {
"pile_set_name": "ArXiv"
} |
[**Solvable ${\cal PT}$ symmetric Hamiltonians** ]{}
Miloslav Znojil
Ústav jaderné fyziky AV ČR, 250 68 Řež, Czech Republic\
e-mail: [email protected],
Abstract {#abstract .unnumbered}
========
Within the so called ${\cal PT}$ symmetric version of quantum mechanics a brief review of the exactly solvable models is given. Distinction is made between the curved and straight coordinate lines, between their unbounded (aperiodic) and bounded (periodic) choices, and between the completely and partially solvable cases.
\[KEYWORDS\] {#keywords .unnumbered}
============
\[ ${\cal PT}$ symmetry, Schrödinger equation in complex domain, exact solutions\]
0.5cm
\[AMS 1991 Mathematics Subject Classification\] {#ams-1991-mathematics-subject-classification .unnumbered}
===============================================
81P10 81Q05 81R40 47B50 34A20 15A57
Introduction
============
${\cal PT}$ symmetric quantum mechanics [@BBMM] has independently been proposed and used as a methodical laboratory in quantum physics by several groups of authors: By Caliceti et al [@CGM] in perturbation theory, by Bessis et al [@DB] in field theory and by Andrianov et al [@ACDI] in supersymmetric context.
This short review will pay attention to the (partially as well as completely) exactly solvable models within this framework, with emphasis on the results obtained by the present author.
Complete solvability on curved paths
====================================
One of the first exactly and completely solvable examples of a ${\cal PT}$ symmetric system has been found by Cannata et al [@ACDI] and re-discovered by Bender et al [@BBJS] more than one year later. Its modified Schrödinger bound state problem is defined on certain curved, left-right symmetric “generalized coordinate" lines in the complex plane. Mathematically, it is defined via exponential potential and proves exactly solvable in terms of Bessel functions. Due to its relationship to a power-law forces in the large exponent limit, it can be most simply interpreted as a certain smooth and non-Hermitian ${\cal PT}$ symmetric analogue of the current square well.
Recently, a double-well counterpart of the latter set of models has been shown exactly solvable, in terms of Laguerre polynomials, in ref. [@8]. In a way similar to the above “single well" example its paths of integration are the same, curved complex lines again. Their spectra exhibit a puzzling and highly unexpected feature of certain coupling-dependent re-arrangements mediated by “unavoided" crossings at critical points. The phenomenon reflects the non-Hermiticity of the Hamiltonian [@Herbst].
Via a suitable Liouvillean change of the variables in the above double-well-like differential Schrödinger equation one can immediately obtain another Laguerre-related solvable system with the potential of the Coulombic single-pole form. In this case the deformation of the integration path plays the beneficial role of a natural regularization prescription. At the same time it also leads to the necessity of working with the complex charges in a way described in ref. [@3]. Also the related spectrum of energies exhibits certain unexpected features: Positivity, a coexistence of the growth and decrease with the increasing coupling strength, etc.
A clear mathematical explanation of behaviour as well as a consistent and/or possible physical interpretation of the counterintuitive models of this type remain unclarified up to now. Much better situation emerges in the case of the integration curves defined as the left-right-symmetric straight lines.
Straight paths
==============
Using the language of the so called Kustaanheimo - Steifel transformation (cf. [@Pogosyan] for detailed references) both the above curvilinear examples can be shown equivalent to the ${\cal PT}$ symmetric harmonic oscillator of ref. [@10] with a centrifugal term regularized by the mere downward complex shift of the (full) real axis, (- + x\^2 -2icx + ) (x) = (E+c\^2) (x), (x) L\_2(-,). \[SE\] This oscillator with the Laguerre-polynomial normalizable solutions \_[(n)]{}(x) = [N]{} (x-ic)\^[+1/2]{}e\^[-(x-ic)\^2/2]{} L\^[()]{}\_n , n = 0, 1, …\[waves\] possesses the non-equidistant spectrum of energies $ E=E_{(\pm
n)}=4n+2 \pm 2 \alpha$ and represents a certain unperturbed limit of the quartic oscillator models of Buslaev and Grecchi [@BG]. These authors, unfortunately, did not notice the existence of the “quasi-even", $_{(-n)}-$signed half of the spectrum. This omission can be easily corrected. One just introduces a “two-to-one" isospectrality correspondence between the respective Hermitian and non-Hermitian anharmonic oscillator models of ref. [@BG].
Interpretation of the models “living" on the straight lines becomes significantly facilitated by the easier identification and interpretation of their complex components [@12]. Immediate purely analytic constructions recover, e.g., the existence of models which are in a one-to-one correspondence to the so called shape-invariant real forces in one dimension (cf. their presentation in ref. [@7]) and on the half line (their ${\cal
PT}$ symmetric counterparts were described and listed in ref. [@4]).
The situation is reviewed in ref. [@2a]. A fully general form of this type of analytic constructions dates back to the introduction of the so called Natanzon potentials and, in the present context, is thoroughly analyzed and described in ref. [@1].
Models with periodic boundary conditions
========================================
A new and promising development of the ${\cal PT}$ symmetric considerations has been recently inspired by the study of the two- and three-particle models [@Milos]. The ${\cal PT}$ symmetrization of the Hamiltonians has been again conjectured to be sufficient for keeping their spectrum real. The related “weakening of the Hermiticity" finds a natural generalization in the new context.
Particular attention has been paid to the possible non-Hermitian generalizations of the well known Calogero model [@Calogero]. In this setting the separability of the underlying partial differential Schrödinger equation in the hyperspherical coordinates helps us to reduce the problem to the “hyperangular" ordinary differential equation defined on a finite interval. In this way, in the simplest cases one has to solve the complexified ordinary differential equations of the generalized Pöschl-Teller type, ( - + + ) () = E() \[angular\] on an interval $\phi \in (-M\,\pi/2, M\,\pi/2)$ with a suitable integer $M$. These equations can be solved exactly in terms of the hypergeometric functions [@Khare].
The strongly repulsive singularities at $\phi_j=j\,\pi/2$ are currently not penetrable [@Fluegge]. Here, they become regularized in a ${\cal PT}$ symmetric manner which parallels a few older constructions on the unbounded intervals [@4]. The quasi-symmetric and quasi-antisymmetric solutions arise from certain [*ad hoc*]{} boundary conditions [@Milos; @comment].
For the most elementary illustration let us now choose $\lambda=0$ and $M=2$. Then, the differential equation (\[angular\]) possesses the two independent hypergeometric solutions, \^[()]{}()=()\^[1/2]{} \_2F\_1(u\^[()]{},v\^[()]{};1;\^2), =+1/2>0 where $2u^{(\pm)}=1/2- \beta \pm \alpha$ and $2v^{(\pm)}=1/2+
\beta \pm \alpha$. On the boundary of convergence $\sin^2 \phi =
1$ the matching of the logarithmic derivatives is equivalent to the termination of this series to the Gegenbauer polynomials, ()=\_[(k)]{}()=()\^[1/2]{} C\_k\^[1/2]{}(), k = 0, 1, … . The construction also quantizes the energies and gives them in the closed form, E=E\_[(k)]{}= (k + 1/2)\^2, k = 0, 1, … . This set of eigenvalues is composed of the two subsets in a way which resembles the above-mentioned non-equidistant spectrum of the ${\cal PT}$ symmetrized singular harmonic oscillator (\[SE\]) [@10].
Conclusions and outlook
=======================
Historically, one of the first persuasive manifestations of the merits and power of the ${\cal PT}$ symmetry has been offered by Bender and Boettcher [@BB] who discovered the quasi-exact (i.e., incomplete) solvability of the most common and popular quartic polynomial oscillators.
The plausible reasons of the unexpected delay of such an “obvious" observation are closely related to the above-mentioned “forgotten" energies. One has to keep in mind the “spontaneous" regularity of the singular potentials within the new formalism. This observation was made explicit in our paper [@6] where the presence of the two additional singular terms has still been shown compatible with the quasi-exact solvability of quartic potentials.
In the quasi-exact context the changes of variables can play the same role as in the completely solvable models. This has been illustrated by the particular constructions of the decadic model [@2] and of the harmonic + Coulomb superposition [@11]. Further work in this direction is in progress [@sextic].
Acknowledgements {#acknowledgements .unnumbered}
================
Research assisted by the Grant Agency of the Academy of Sciences of the Czech Republic, contract Nr. A 1048004.
[00]{}
C. M. Bender, S. Boettcher, K. A. Milton and P. N. Meisinger, J. Math. Phys. [40]{} (1999) 2201.
E. Calicetti, S. Graffi and M. Maioli, Commun. Math. Phys. [75]{} (1980) 51;\
G. Alvarez, J. Phys. A: Math. Gen. 27 (1995) 4589;\
M. F. Fernández, R. Guardiola, J. Ros and M. Znojil, J. Phys. A 31 (1998) 10105.
D. Bessis and J. Zinn-Justin, private communication (1992);\
C. M. Bender and A. Turbiner, Phys. Lett. [A 173]{} (1993) 442;\
C. M. Bender and K. A. Milton, Phys. Rev. D 55 (1997) R3255;\
C. M. Bender and S. Boettcher, Phys. Rev. Lett. [ 24]{} (1998) 5243;\
E. Delabaere and F. Pham, Phys. Lett. A [**250**]{} (1998) 25 and 29.
A. A. Andrianov, F. Cannata, J.-P. Dedonder and M. V. Ioffe, Int. J. Mod. Phys. A [14]{} (1999) 2675;\
F. Cannata, G. Junker and J. Trost, Phys. Lett A [246]{} (1998) 219;\
B. Bagchi and R. Roychoudhury, J. Phys. A [33]{} (2000) L1;\
M. Znojil, F. Cannata, B. Bagchi and R. Roychoudhury, Phys. Lett. B 483 (2000) 284.
C. M. Bender, S. Boettcher, H. F. Jones and Van M. Savage, J. Phys. A [32]{} (1999) 6771.
M. Znojil, Phys. Lett. A 264 (1999) 108.
I. Herbst, private communication.
M. Znojil and G. Levai, Phys. Lett. A 271 (2000) 327.
P. Kustaanheimo and E. Steifel, J. Reine Angew. Math. 218 (1965) 204;
E. G. Kalnins, W. Miller Jr. and G. S. Pogosyan, arXiv: quant-ph/9906055.
M. Znojil, Phys. Lett. A 259 (1999) 220.
V. Buslaev and V. Grecchi, J. Phys. A [36]{} (1993) 5541.
M. Znojil, J. Phys. A: Math. Gen. 32 (1999) 7419;\
M. F. Fernández, R. Guardiola, J. Ros and M. Znojil, J. Phys. A 32 (1999) 3105.
M. Znojil, J. Phys. A: Math. Gen. 33 (2000) L61.
M. Znojil, J. Phys. A: Math. Gen. 33 (2000) 4561.
M. Znojil, Rendiconti del Circ. Mat. di Palermo, to appear.
G. Levai and M. Znojil, J. Phys. A: Math. Gen. 33 (2000) 7165.
M. Znojil and M. Tater, Complex Calogero model with real energies, arXiv: quant-ph/0010087, submitted to J. Phys. A.
F. Calogero, J. Math. Phys. 10 (1969) 2191;
A. Turbiner, Mod. Phys. Lett. A 13 (1998) 1473 with further references.
F. Cooper, A. Khare and U. Sukhatme, Phys. Rep. 251 (1995) 267.
S. Flügge, Practical Quantum Mechanics I (Springer, Berlin, 1971), p 89.
M. Znojil, Phys. Rev. A 61 (2000) 066101.
C. M. Bender and S. Boettcher, J. Phys. A [31]{} (1998) L273.
M. Znojil, J. Phys. A 33 (2000) 4203.
M. Znojil, J. Phys. A: Math. Gen. 33 (2000) 6825.
M. Znojil, J. Phys. A: Math. Gen. 32 (1999) 4563.
F. Gemperle, M. Znojil and F. Cannata, in preparation.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
In this paper, we propose a Boltzmann-type kinetic description of mass-varying interacting multi-agent systems. Our agents are characterised by a microscopic state, which changes due to their mutual interactions, and by a label, which identifies a group to which they belong. Besides interacting within and across the groups, the agents may change label according to a state-dependent Markov-type jump process. Hence the mass of each group is not conserved. We derive general kinetic equations for the joint interaction/label switching processes in each group. Moreover, for prototypical birth/death dynamics we characterise the transient and equilibrium kinetic distributions of the groups via a Fokker-Planck asymptotic analysis. Finally, we introduce and discuss, both analytically and numerically, a new model for the contagion of infectious diseases with quarantine based on this non-conservative kinetic framework.
[**Keywords:**]{} Boltzmann-type equations, Markov-type jump processes, transition probabilities, Fokker-Planck asymptotics, contagion of infectious diseases, quarantine
[**Mathematics Subject Classification:**]{} 35Q20, 35Q70, 35Q84
author:
- 'Nadia Loy[^1]'
- 'Andrea Tosin[^2]'
bibliography:
- 'LnTa-nonconservative.bib'
date: |
Department of Mathematical Sciences “G. L. Lagrange”\
Politecnico di Torino, Italy
title: 'Non-conservative Boltzmann-type kinetic equations for multi-agent systems with label switching'
---
Introduction
============
Boltzmann-type kinetic equations are a valuable tool to model multi-agent systems, their success being confirmed by a great variety of modern applications which take advantage of the formalism of the collisional kinetic theory. The literature in the field is constantly growing as witnessed by very recent contributions to econophysics [@dimarco2020PREPRINT-a], human ecology [@dimarco2019JSP-a; @dimarco2020PREPRINT-b; @toscani2019PRE], vehicular traffic with autonomous vehicles [@piccoli2019ZAMP_preprint; @tosin2019MMS; @tosin2020MCRF], opinion formation [@fraia2020UMI_preprint; @pareschi2019JNS], biology [@loy2019JMB; @loy2020KRM; @preziosi2020PREPRINT].
Kinetic models of multi-agent systems are based on a revisitation of the methods of the classical kinetic theory, however with remarkable differences due to the different nature of the physical systems at hand. Classical kinetic theory deals mostly with the dynamics of gas molecules and their elastic collisions, which conserve microscopically both the momentum and the kinetic energy of the pairs of colliding molecules. Instead, interactions in multi-agent systems often do not conserve either the first or the second statistical moment of the distribution function, which does not only have consequences on the physical interpretation of the dynamics but also on the techniques required to investigate it mathematically. On the other hand, the virtually ubiquitous characteristic of kinetic models of multi-agent systems is the fact that the total number of agents does not change in time. This is equivalent to the conservation of mass in the collisions among gas molecules and allows one to regard the kinetic distribution function as a probability density function. Consequently, the kinetic equations may be derived from stochastic microscopic interaction dynamics by appealing to probabilistic arguments.
In this paper, we are instead interested in multi-agent dynamics which do not conserve necessarily the number of agents. Indeed, many applications in population dynamics involve non-conservative dynamics, which in the abstract may be assimilated to “birth” and “death” processes, i.e. to the appearance and disappearance of interacting agents on the basis of the microscopic state that they are currently expressing. In particular, we focus on the so-called *label switching* process, that we may summarise as follows: while interacting, agents may change a label which denotes their membership of a particular group/category within the whole population. As a consequence, the number of agents in each group varies in time. This process clearly affects the interactions in each group, both because the number of interacting agents changes and because the distribution of the microscopic states in each group is altered by the introduction or the removal of agents.
Recently, other contributions dealing with label switching in particle systems have been proposed, see e.g., [@albi2019M3AS; @morandotti2020SIMA]. While sharing some conceptual analogies with our microscopic agent dynamics, these works focus on mean-field descriptions, which in principle may be regarded as particular cases of Boltzmann-type kinetic descriptions. Indeed, in some regimes Boltzmann-type “collisional” models might be approximated by mean-field models but in general they contain a much richer variety of trends depending on the ranges of the parameters. It is worth pointing out that one of the main goals of the kinetic theory is precisely to unravel the links between the microscopic parameters and the emerging aggregate trends of a system. For the reasons above, they are in general accounted for more thoroughly by a Boltzmann-type approach, which is therefore interesting to investigate systematically in presence of label switching. This is the aim and the main novelty of this work compared to other studies about similar topics.
In more detail, the paper is organised as follows: in Section \[sect:preliminaries\] we review separately some basic facts about the Boltzmann-type description of binary interactions and Markov-type jump processes, which in this context provide a proper mathematical framework to formalise the label switching process. In Section \[sect:int+labswitch\] we derive the Boltzmann-type equations with binary interactions and label switching starting from a stochastic microscopic description of the interaction and relabelling processes. Within this formalism, in Section \[sect:death.birth\] we investigate the transient and asymptotic trends of birth and death processes of interacting agents, which can be regarded as the prototypes of a wide range of non-conservative particle dynamics. In Section \[sect:contagion\] we propose a concrete application of the mathematical structures previously developed to a problem of contagion of infectious diseases with quarantine. Finally, in Section \[sect:numerical\] we show some numerical simulations of the kinetic model of Section \[sect:contagion\] obtained by means of a Monte Carlo particle algorithm that we derive straightforwardly from the stochastic microscopic description introduced in Section \[sect:int+labswitch\]. The algorithm is reported in Appendix \[app:nanbu\].
Preliminaries on labelled interacting agents {#sect:preliminaries}
============================================
Let us consider a large system of agents described by a microscopic state $v\in{\mathbb{R}}_+$ representing a non-negative physical quantity. Extensions to negative and possibly also bounded microscopic states are mostly a matter of technicalities using the very same ideas presented in this paper. The agents may belong to different groups or categories identified by a discrete label $x\in{\mathcal{I}}=\{1,\,\dots,\,n\}$, that the agents may change as a result of a Markov-type jump process. Such a stochastic *label switching* process is defined by a transition probability $$T=T(t,\,v;\,x\vert y)\in [0,\,1] \qquad \forall\,v\in{\mathbb{R}}_+,\ x,\,y\in{\mathcal{I}},\ t>0,
\label{trans_prob}$$ namely the probability that an agent with state $v$ at time $t$ switches from label $y$ to label $x$. In order for $T(t,\,v;\,\cdot\vert y)$ to be a conditional probability density, it has to satisfy the following further property: $$\int_{\mathcal{I}}T(t,\,v;\,x\vert y)\,dx=1 \qquad \forall\,v\in{\mathbb{R}}_+,\ y\in{\mathcal{I}},\ t>0.$$ A label switch corresponds therefore to a migration of an agent to a different group, however in such a way that the total mass of the agents in the system is conserved.
\[rem:discrete\] Since the variable $x$ is discrete, the mapping $x\mapsto T(t,\,v;\,x\vert y)$ is a discrete probability measure. Consequently, we actually have $$\int_{\mathcal{I}}T(t,\,v;\,x\vert y)\,dx=\sum_{i=1}^{n}T(t,\,v;\,i\vert y).$$
Agents within the same group, i.e. with the same label, are assumed to be indistinguishable. Their microscopic state $v$ evolves in consequence of *binary interactions* with either other agents of the same group or agents belonging to a different group. We will take into account the possibility that the interactions among agents with the same label differ from those among agents with different labels. In general, if $v,\,v_\ast\in{\mathbb{R}}_+$ denote the pre-interaction states of any two interacting agents, their post-interaction states $v',\,v_\ast'$ will be given by general *linear* microscopic rules of the form $$v'=p_1v+q_1v_\ast, \qquad v_\ast'=p_2v_\ast+q_2v,
\label{eq:binary_gen}$$ where $p_i,\,q_i\in{\mathbb{R}}_+$ for $i=1,\,2$ are either deterministic or stochastic coefficients. As a particular relevant sub-case, we will consider *symmetric* interactions, namely those with $p_1=p_2$ and $q_1=q_2$. In this case, it will be generally sufficient to refer only to the representative rule $$v'=pv+qv_\ast
\label{eq:binary_gen_sym}$$ with $p,\,q\in{\mathbb{R}}_+$ independent random variables.
Boltzmann-type description of the interaction dynamics {#sect:int.dyn}
------------------------------------------------------
It is known that an aggregate description of the (sole) interaction dynamics inspired by the principles of statistical mechanics can be obtained by introducing a distribution function $f=f(t,\,v)\geq 0$ such that $f(t,\,v)dv$ gives the proportion of agents having at time $t$ a microscopic state comprised between $v$ and $v+dv$. Such a distribution function satisfies a Boltzmann-type kinetic equation, which in weak form reads $$\frac{d}{dt}\int_{{\mathbb{R}}_+}\varphi(v)f(t,\,v)\,dv=\mu\int_{{\mathbb{R}}_+}\int_{{\mathbb{R}}_+}\ave{\varphi(v')-\varphi(v)}f(t,\,v)f(t,\,v_\ast)\,dv\,dv_\ast,
\label{eq:boltz}$$ where $\varphi:{\mathbb{R}}_+\to{\mathbb{R}}$ is an observable quantity (test function) and $\mu>0$ is the interaction frequency, here assumed to be constant. This equation expresses the fact that the time variation of the expectation of $\varphi$ (left-hand side) is due to the mean variation of $\varphi$ in a binary interaction (right-hand side). For a detailed derivation of we refer the interested reader to [@pareschi2013BOOK].
Choosing $\varphi(v)=1$ we obtain $$\frac{d}{dt}\int_{{\mathbb{R}}_+} f(t,\,v)\,dv=0,$$ which means that the total mass of the agents is conserved in time by the interactions . Therefore, we say that the interaction dynamics are *conservative*.
Choosing instead $\varphi(v)=v^n$, $n=1,\,2,\,\dots$, we obtain the evolution of the statistical moments of $f$. For instance, with $\varphi(v)=v$ we find that the mean $m(t):=\int_{{\mathbb{R}}_+}vf(t,\,v)\,dv$ evolves according to $$\frac{dm}{dt}=\ave{p+q-1}m.$$ If $\ave{p+q}=1$ then $m$ is conserved in time while if $\ave{p+q}>1$ or $\ave{p+q}<1$ then $m$ either blows to infinity or decreases to zero exponentially fast in time.
The trend of the total energy $E(t):=\int_{{\mathbb{R}}_+}v^2f(t,\,v)\,dv$ is obtained from $\varphi(v)=v^2$ and turns out to be ruled by the equation $$\frac{dE}{dt}=\ave{p^2+q^2-1}E+2\ave{pq}m^2.$$
From the trend of $m$ and $E$ we can also infer that of the internal energy $e(t):=E(t)-m^2(t)$, namely the variance of the distribution $f$: $$\frac{de}{dt}=\ave{p^2+q^2-1}e+\ave{(p+q-1)^2}m^2.$$
Notice that if $m$ is conserved in time and $m\neq 0$ then both $E$ and $e$ tend asymptotically to a finite non-vanishing value if $\ave{p^2+q^2}<1$.
Kinetic description of the label switching {#sect:label.switch}
------------------------------------------
If we consider only the label switch process then the evolution of the distribution function $f=f(t,\,x)\geq 0$ of the agents with label $x$ at time $t$ can be modelled by a standard kinetic equation describing a Markov-type jump process [@loy2020CMS]: $$\partial_tf(t,\,x)=\lambda\left(\int_{\mathcal{I}}T(t;\,x\vert y)f(t,\,y)\,dy-f(t,\,x)\right),
\label{eq:mark_proc_1}$$ where $\lambda>0$ is the (constant) switch rate. In weak form reads $$\frac{d}{dt}\int_{\mathcal{I}}\psi(x)f(t,\,x)\,dx=\lambda\int_{\mathcal{I}}\int_{\mathcal{I}}(\psi(x)-\psi(y))T(t;\,x\vert y)f(t,\,y)\,dx\,dy,
\label{eq:jump_proc_weak}$$ where $\psi:{\mathcal{I}}\to{\mathbb{R}}$ is another observable quantity (test function).
Since $x\in{\mathcal{I}}$ is discrete, we may conveniently represent the distribution function $f$ as $$f(t,\,x)=\sum_{i=1}^{n}f_i(t)\delta(x-i),
\label{eq:f.delta.label_switch}$$ where $\delta(x-i)$ is the Dirac distribution centred in $x=i$ and $f_i=f_i(t)\geq 0$ is the probability that an agent is labelled by $x=i$ at time $t$. In this way, we reconcile the weak form with the convention introduced in Remark \[rem:discrete\], for actually becomes $$\sum_{i=1}^{n}\psi(i)f_i'(t)=\lambda\sum_{i=1}^{n}\sum_{j=1}^{n}(\psi(i)-\psi(j))T(t;\,i\vert j)f_j(t).$$ Choosing $\psi$ such that $\psi(i)=1$ for a certain $i\in{\mathcal{I}}$ and $\psi(x)=0$ for all $x\in{\mathcal{I}}\setminus\{i\}$ we get in particular $$f_i'=\lambda\left(\sum_{j=1}^{n}T(t;\,i\vert j)f_j-f_i\right), \qquad i=1,\,\dots,\,n.
\label{eq:fi.jump}$$
Kinetic description of interactions with label switching {#sect:int+labswitch}
========================================================
We now want to derive a kinetic equation for the joint distribution function $f=f(t,\,x,\,v)\geq 0$, such that $f(t,\,x,\,v)dv$ gives the proportion of agents labelled by $x\in{\mathcal{I}}$ and having microscopic state comprised between $v$ and $v+dv$ at time $t$. The discreteness of $x$ allows us to represent $f$ as $$f(t,\,x,\,v)=\sum_{i=1}^{n}f_i(t,\,v)\delta(x-i),$$ where $f_i=f_i(t,\,v)\geq 0$ is the distribution function of the microscopic state $v$ of the agents with label $i$ and, in particular, $f_i(t,\,v)dv$ is the proportion of agents with label $i$ whose microscopic state is comprised between $v$ and $v+dv$ at time $t$.
Since both the interactions and the label switching conserve the total mass of the system, we may assume that $f(t,\,x,\,v)$ is a probability distribution, namely: $$\int_{{\mathbb{R}}_+}\int_{\mathcal{I}}f(t,\,x,\,v)\,dx\,dv=\sum_{i=1}^{n}\int_{{\mathbb{R}}_+} f_i(t,\,v)\,dv=1 \qquad \forall\,t>0.
\label{eq:f.prob}$$ Notice, however, that the $f_i$’s are in general not probability density functions because their $v$-integral varies in time due to the label switching. We denote by $$\rho_i(t):=\int_{{\mathbb{R}}_+} f_i(t,\,v)\,dv
\label{eq:rhoi}$$ the mass of the group of agents with label $i$, thus $0\leq\rho_i(t)\leq 1$ and $$\sum_{i=1}^{n}\rho_i(t)=1 \qquad \forall\,t>0.$$
Let $(X_t,\,V_t)\in{\mathcal{I}}\times{\mathbb{R}}_+$ be a pair of random variables denoting the label and the microscopic state of a representative agent of the system at time $t$. The joint probability distribution of such a pair is $f(t,\,x,\,v)$. During a sufficiently small time interval $\Delta{t}>0$ the agent may or may not change the pair $(X_t,\,V_t)$ depending on whether a label switch and/or a binary interaction with another agent takes place. We express this discrete-in-time random process as $$\begin{aligned}
\begin{aligned}[c]
X_{t+\Delta{t}} &= (1-\Theta)X_t+\Theta J_t, \\
V_{t+\Delta{t}} &= (1-\Xi)V_t+\Xi V'_t,
\end{aligned}
\label{eq:micro.rules.gen}\end{aligned}$$ where $J_t$, $V'_t$ are random variables describing the new label after a label switch and the new microscopic state after a binary interaction, respectively, while $\Theta,\,\Xi\in\{0,\,1\}$ are Bernoulli random variables, which we assume independent of all the other variables appearing in , discriminating whether a label switch and a binary interaction take place ($\Theta,\,\Xi=1$) or not ($\Theta,\,\Xi=0$) during the time interval $\Delta{t}$. In particular, we set $${\operatorname{Prob}}(\Theta=1)=\lambda\Delta{t}, \qquad {\operatorname{Prob}}(\Xi=1)=\mu\Delta{t}
\label{eq:bernoulli}$$ where $\lambda$, $\mu$ are the frequencies introduced in Sections \[sect:int.dyn\], \[sect:label.switch\] and $\Delta{t}\leq\min\{\frac{1}{\lambda},\,\frac{1}{\mu}\}$ for consistency. The underlying assumption is that the longer the time interval $\Delta{t}$ the higher the probability that a label switch and/or a binary interaction takes place. Notice that $\frac{1}{\lambda}$, $\frac{1}{\mu}$ can be understood as the mean waiting times between two successive label switches/binary interactions, respectively.
The random variable $J_t\in{\mathcal{I}}$ models the Markov-type jump process leading to a label switch. If $P(t,\,j,\,v)$ denotes the joint probability distribution of the pair $(J_t,\,V_t)$ then $$P(t,\,j,\,v)=\int_{\mathcal{I}}T(t,\,v;\,j\vert y)f(t,\,y,\,v)\,dy,$$ where $T(t,\,v;\,j\vert y)$ is the transition probability .
The random variable $V'_t\in{\mathbb{R}}_+$ gives instead the new microscopic state after a binary interaction with another agent described by the pair $(Y_t,\,V^\ast_t)\in{\mathcal{I}}\times{\mathbb{R}}_+$. In order to account for possibly different interaction rules depending on the labels of the interacting agents, we define $$V'_t:=\delta_{X_t,Y_t}\bar{V}'_t+(1-\delta_{X_t,Y_t})\tilde{V}'_t
\label{eq:micro.rule_exch}$$ where $$\delta_{X,Y}=
\begin{cases}
1 & \text{if } X=Y \\
0 & \text{if } X\neq Y
\end{cases}$$ is the Kronecker delta. In particular, $\bar{V}'_t,\,\tilde{V}'_t\in{\mathbb{R}}_+$ represent the outcomes of binary interactions between agents with the same and different labels, respectively. They will be both of the form , namely $$\bar{V}'_t=\bar{p}V_t+\bar{q}V^\ast_t, \qquad \tilde{V}'_t=\tilde{p}V_t+\tilde{q}V^\ast_t$$ with $\bar{p},\,\bar{q},\,\tilde{p},\,\tilde{q}\in{\mathbb{R}}$ either deterministic or independent random coefficients.
Let now $\phi=\phi(x,\,v)$ be an observable quantity defined on ${\mathcal{I}}\times{\mathbb{R}}_+$. From , , together with the assumed independence of $M,\,N$, we see that the mean variation rate of $\phi$ in the time interval $\Delta{t}$ satisfies $$\begin{aligned}
&\frac{\ave{\phi(X_{t+\Delta{t}},\,V_{t+\Delta{t}})}-\ave{\Phi(X_t,\,V_t)}}{\Delta{t}}= \\
&\qquad\qquad\phantom{+} \frac{(1-\lambda\Delta{t})(1-\mu\Delta{t})\ave{\phi(X_t,\,V_t)}+\mu\Delta{t}(1-\lambda\Delta{t})\ave{\phi(X_t,\,V'_t)}}{\Delta{t}} \\
&\qquad\qquad +\dfrac{\lambda\Delta{t}(1-\mu\Delta{t})\ave{\phi(J_t,\,V_t)}+\lambda\mu{\Delta{t}}^2\ave{\phi(J_t,\,V'_t)}-\ave{\phi(X_t,\,V_t)}}{\Delta{t}},\end{aligned}$$ whence we deduce the instantaneous time variation of the average of $\phi$ in the limit $\Delta{t}\to 0^+$ as $$\frac{d}{dt}\ave{\phi(X_t,\,V_t)}=\lambda\ave{\phi(J_t,\,V_t)}+\mu\ave{\phi(X_t,\,V'_t)}-(\lambda+\mu)\ave{\phi(X_t,\,V_t)}.$$ Notice that the simultaneous change of label and microscopic state, i.e. the term $\ave{\phi(J_t,\,V'_t)}$, turns out to be a higher order effect in time disregarded in this limit equation. Owing to , we further obtain $$\begin{aligned}
\begin{aligned}[b]
\frac{d}{dt}\ave{\phi(X_t,\,V_t)} &= \lambda\ave{\phi(J_t,\,V_t)} \\
&\phantom{=} +\mu\ave{\delta_{X_t,Y_t}\phi(X_t,\,\bar{V}'_t)}+\mu\ave{(1-\delta_{X_t,Y_t})\phi(X_t,\,\tilde{V}'_t)} \\
&\phantom{=} -(\lambda+\mu)\ave{\phi(X_t,\,V_t)}.
\end{aligned}
\label{eq:mean_obs}\end{aligned}$$
We consider now that $$\ave{\phi(J_t,\,V_t)}=\int_{{\mathbb{R}}_+}\int_{{\mathcal{I}}}\phi(i,\,v)P(t,\,i,\,v)\,di\,dv=\sum_{i=1}^{n}\sum_{j=1}^{n}\int_{{\mathbb{R}}_+}\phi(i,\,v)T(t,\,v;\,i|j)f_j(t,\,v)\,dv$$ and that $$\begin{aligned}
\ave{\delta_{X_t,Y_t}\phi(X_t,\,\bar{V}'_t)} &= \sum_{i=1}^{n}\int_{{\mathbb{R}}_+}\int_{{\mathbb{R}}_+}\ave{\phi(i,\,\bar{v}')}f_i(t,\,v)f_i(t,\,v_\ast)\,dv\,dv_\ast \\
\ave{(1-\delta_{X_t,Y_t})\phi(X_t,\,\tilde{V}'_t)} &= \sum_{i=1}^{n}\sum_{\substack{j=1 \\ j\neq i}}^{n}\int_{{\mathbb{R}}_+}\int_{{\mathbb{R}}_+}\ave{\phi(i,\,\tilde{v}')}f_i(t,\,v)f_j(t,\,v_\ast)\,dv\,dv_\ast,\end{aligned}$$ where $\ave{\cdot}$ denotes the average with respect to the possibly random coefficients $\bar{p}$, $\bar{q}$, $\tilde{p}$, $\tilde{q}$ contained in $\bar{v}'$ and $\tilde{v}'$. As typically done in kinetic theory [@cercignani1988BOOK], in writing these interaction terms we assume the *propagation of chaos*, which allows us to perform the factorisation $f(t,\,x,\,v,\,y,\,v_\ast)=f(t,\,x,\,v)f(t,\,y,\,v_\ast)$ of the two-particle distribution function. Hence from we deduce the following equation: $$\begin{aligned}
\frac{d}{dt}\sum_{i=1}^{n}\int_{{\mathbb{R}}_+}\phi(i,\,v)f_i(t,\,v)\,dv &= \lambda\sum_{i=1}^{n}\sum_{j=1}^{n}\int_{{\mathbb{R}}_+}\phi(i,\,v)T(t,\,v;\,i|j)f_j(t,\,v)\,dv \\
&\phantom{=} +\mu\sum_{i=1}^{n}\int_{{\mathbb{R}}_+}\int_{{\mathbb{R}}_+}\ave{\phi(i,\,\bar{v}')}f_i(t,\,v)f_i(t,\,v_\ast)\,dv\,dv_\ast \\
&\phantom{=} +\mu\sum_{i=1}^{n}\sum_{\substack{j=1 \\ j\neq i}}^{n}\int_{{\mathbb{R}}_+}\int_{{\mathbb{R}}_+}\ave{\phi(i,\,\tilde{v}')}f_i(t,\,v)f_j(t,\,v_\ast)\,dv\,dv_\ast \\
&\phantom{=} -(\lambda+\mu)\sum_{i=1}^{n}\int_{{\mathbb{R}}_+}\phi(i,\,v)f_i(t,\,v)\,dv,\end{aligned}$$ which has to hold for every $\phi:{\mathcal{I}}\times{\mathbb{R}}_+\to{\mathbb{R}}$. Choosing $\phi(x,\,v)=\psi(x)\varphi(v)$ with $\psi$ such that $\psi(i)=1$ for a certain $i\in{\mathcal{I}}$ and $\psi(x)=0$ for all $x\in{\mathcal{I}}\setminus\{i\}$ and exploiting to merge the loss term (last term on the right-hand side) with the other terms on the right-hand side, we finally obtain the following system of equations for the $f_i$’s: $$\begin{aligned}
\begin{aligned}[b]
\frac{d}{dt}\int_{{\mathbb{R}}_+}\varphi(v)f_i(t,\,v)\,dv &= \lambda\int_{{\mathbb{R}}_+}\varphi(v)\left(\sum_{j=1}^{n}T(t,\,v;\,i\vert j)f_j(t,\,v)-f_i(t,\,v)\right)dv \\
&\phantom{=} +\mu\int_{{\mathbb{R}}_+}\int_{{\mathbb{R}}_+}\ave{\varphi(\bar{v}')-\varphi(v)}f_i(t,\,v)f_i(t,\,v_\ast)\,dv\,dv_\ast \\
&\phantom{=} +\mu\sum_{\substack{j=1 \\ j\neq i}}^{n}\int_{{\mathbb{R}}_+}\int_{{\mathbb{R}}_+}\ave{\varphi(\tilde{v}')-\varphi(v)}f_i(t,\,v)f_j(t,\,v_\ast)\,dv\,dv_\ast, \quad i=1,\,\dots,\,n.
\end{aligned}
\label{eq:boltz.fi}\end{aligned}$$
For each $i\in{\mathcal{I}}$, this is a non-conservative kinetic equation. Indeed, letting $\varphi(v)=1$ we discover that the mass $\rho_i$ of the agents with label $i$, cf. , evolves according to $$\frac{d\rho_i}{dt}+\lambda\rho_i=\lambda\sum_{j=1}^{n}\int_{{\mathbb{R}}_+}T(t,\,v;\,i\vert j)f_j(t,\,v)\,dv,$$ which depends explicitly on the label switch process.
Particularisations and generalisations of
------------------------------------------
#### Same interaction rules {#same-interaction-rules .unnumbered}
If there is a common interaction rule among individuals with the same and with different labels, i.e. if $\bar{v}'=\tilde{v}'=:v'$, then simplifies as $$\begin{aligned}
\frac{d}{dt}\int_{{\mathbb{R}}_+}\varphi(v)f_i(t,\,v)\,dv &= \lambda\int_{{\mathbb{R}}_+}\varphi(v)\left(\sum_{j=1}^{n}T(t,\,v;\,i\vert j)f_j(t,\,v)-f_i(t,\,v)\right)dv \\
&\phantom{=} +\mu\sum_{j=1}^{n}\int_{{\mathbb{R}}_+}\int_{{\mathbb{R}}_+}\ave{\varphi(v')-\varphi(v)}f_i(t,\,v)f_j(t,\,v_\ast)\,dv\,dv_\ast, \quad i=1,\,\dots,\,n,\end{aligned}$$ where $v'$ is given e.g., by .
#### Interactions among agents with the same label only {#interactions-among-agents-with-the-same-label-only .unnumbered}
We may reproduce the situation in which only agents with the same label interact by letting $\tilde{v}'=v$ in . This corresponds to saying that interactions among agents with different labels do not produce a change of microscopic state, hence they are actually “non-interactions”. Consequently, simplifies as $$\begin{aligned}
\begin{aligned}[b]
\frac{d}{dt}\int_{{\mathbb{R}}_+}\varphi(v)f_i(t,\,v)\,dv &= \lambda\int_{{\mathbb{R}}_+}\varphi(v)\left(\sum_{j=1}^{n}T(t,\,v;\,i\vert j)f_j(t,\,v)-f_i(t,\,v)\right)dv \\
&\phantom{=} +\mu\int_{{\mathbb{R}}_+}\int_{{\mathbb{R}}_+}\ave{\varphi(v')-\varphi(v)}f_i(t,\,v)f_i(t,\,v_\ast)\,dv\,dv_\ast, \quad i=1,\,\dots,\,n
\end{aligned}
\label{eq:boltz.fi-same_label}\end{aligned}$$ with $v'$ given e.g., by .
#### Interactions with a background {#interactions-with-a-background .unnumbered}
The agents may change their microscopic state because of interactions with a fixed background rather than with other agents. In this case, the state of the background is usually described by a random variable $z\in{\mathbb{R}}_+$ distributed according to a given probability density $\zeta_i=\zeta_i(z):{\mathbb{R}}_+\to{\mathbb{R}}_+$, see e.g., [@furioli2017M3AS], which may depend on the agent label. The interaction rule takes the form $$v'=pv+qz,$$ which implies that becomes a *linear* kinetic equation: $$\begin{aligned}
\frac{d}{dt}\int_{{\mathbb{R}}_+}\varphi(v)f_i(t,\,v)\,dv &= \lambda\int_{{\mathbb{R}}_+}\varphi(v)\left(\sum_{j=1}^{n}T(t,\,v;\,i\vert j)f_j(t,\,v)-f_i(t,\,v)\right)dv \\
&\phantom{=} +\mu\int_{{\mathbb{R}}_+}\int_{{\mathbb{R}}_+}\ave{\varphi(v')-\varphi(v)}f_i(t,\,v)\zeta_i(z)\,dv\,dz, \quad i=1,\,\dots,\,n.\end{aligned}$$
#### Non-symmetric interaction rules {#non-symmetric-interaction-rules .unnumbered}
If the interaction rules are not symmetric, i.e. if they are inspired by the prototype with $p_1\neq p_2$ and $q_1\neq q_2$ both in the case of agents with the same label and in the case of agents with different labels, then, following [@pareschi2013BOOK], equation generalises as $$\begin{aligned}
\frac{d}{dt}\int_{{\mathbb{R}}_+}\varphi(v)f_i(t,\,v)\,dv &= \lambda\int_{{\mathbb{R}}_+}\varphi(v)\left(\sum_{j=1}^{n}T(t,\,v;\,i\vert j)f_j(t,\,v)-f_i(t,\,v)\right)dv \\
&\phantom{=} +\frac{\mu}{2}\int_{{\mathbb{R}}_+}\int_{{\mathbb{R}}_+}\ave{\varphi(\bar{v}')+\varphi(\bar{v}_\ast')-\varphi(v)-\varphi(v_\ast)}f_i(t,\,v)f_i(t,\,v_\ast)\,dv\,dv_\ast \\
&\phantom{=} +\frac{\mu}{2}\sum_{\substack{j=1 \\ j\neq i}}^{n}\int_{{\mathbb{R}}_+}\int_{{\mathbb{R}}_+}\ave{\varphi(\tilde{v}')+\varphi(\tilde{v}_\ast')-\varphi(v)-\varphi(v_\ast)}f_i(t,\,v)f_j(t,\,v_\ast)\,dv\,dv_\ast\end{aligned}$$ for $i=1,\,\dots,\,n$.
Death and birth processes {#sect:death.birth}
=========================
We now use the kinetic equations derived in the previous section to study *death and birth processes*. These are the prototypes and the building blocks of general non-conservative dynamics in a wide range of applications, one of which will be illustrated in the next Section \[sect:contagion\]. Therefore, it is particularly meaningful to investigate their properties in the kinetic framework of interacting multi-agent systems.
To be definite, we consider $n=2$ labels: $i=1$ denotes interacting or “living” agents whilst $i=2$ denotes inert or “dead” agents. The total mass of the agents is conserved but the mass of the agents with either label may change in time in consequence of changes of label, i.e. “deaths” (transitions from $i=1$ to $i=2$) or “births” (transitions from $i=2$ to $i=1$).
Since we consider the agents labelled with $i=2$ as inert, we implicitly mean that they do not interact either with one another or with the agents labelled with $i=1$. Therefore, the reference equation for this application is with $v'=v$ for $i=2$.
Death
-----
We begin by considering the death process only, in which only transitions from $i=1$ to $i=2$ are possible. Therefore, the transition probabilities describing the Markov-type jump process may be chosen as $$\begin{array}{ll}
T(t,\,v;\,1\vert 2)=0, & T(t,\,v;\,2\vert 2)=1 \\
T(t,\,v;\,2\vert 1)=\beta(t,\,v), & T(t,\,v;\,1\vert 1)=1-\beta(t,\,v)
\end{array}
\label{eq:trans_prob.death}$$ with $0\leq\beta(t,\,v)\leq 1$ for all $v\in{\mathbb{R}}_+$ and $t>0$. From , the evolution equations for the distribution functions $f_1$, $f_2$ take then the form $$\begin{aligned}
\begin{aligned}[b]
\frac{d}{dt}\int_{{\mathbb{R}}_+}\varphi(v)f_1(t,\,v)\,dv &= -\lambda\int_{{\mathbb{R}}_+}\varphi(v)\beta(t,\,v)f_1(t,\,v)\,dv \\
&\phantom{=} +\mu\int_{{\mathbb{R}}_+}\int_{{\mathbb{R}}_+}\ave{\varphi(v')-\varphi(v)}f_1(t,\,v)f_1(t,\,v_\ast)\,dv\,dv_\ast
\end{aligned}
\label{eq:death_f1_weak}\end{aligned}$$ and $$\frac{d}{dt}\int_{{\mathbb{R}}_+}\varphi(v)f_2(t,\,v)\,dv=\lambda\int_{{\mathbb{R}}_+}\varphi(v)\beta(t,\,v)f_1(t,\,v)\,dv.
\label{eq:death_f2_weak}$$
### Mass balance {#sect:death.mass}
Letting $\varphi(v)=1$ in , yields the time evolution of the masses $\rho_1$, $\rho_2$ of the two groups of agents: $$\begin{aligned}
& \frac{d\rho_1}{dt}=-\lambda\int_{{\mathbb{R}}_+}\beta(t,\,v)f_1(t,\,v)\,dv \label{eq:death.mass_balance} \\
& \frac{d\rho_2}{dt}=\lambda\int_{{\mathbb{R}}_+}\beta(t,\,v)f_1(t,\,v)\,dv. \nonumber\end{aligned}$$ Actually, since $\rho_1(t)+\rho_2(t)$ is constant, if we assume a unitary total mass we may replace the second equation simply by $\rho_2(t)=1-\rho_1(t)$.
If the transition probability $\beta$ does not depend on $v$, i.e. $\beta=\beta(t)$, then we get in particular $$\rho_1(t)=\rho_{1,0}\exp{\left(-\lambda\int_0^t\beta(s)\,ds\right)},$$ $\rho_{1,0}\in [0,\,1]$ being the prescribed mass $\rho_1$ at the initial time $t=0$. From here we see that $\rho_1$ tends to vanish asymptotically in time if e.g., $\beta$ does not depend on $t$ or if it approaches a constant non-zero value for large times. Conversely, if $\beta$ vanishes definitively from a certain time $t=t_0$ on then a residual mass of agents with label $i=1$ remains for large times.
If $\beta$ features a full dependence on $v$ and $t$, we cannot deduce from an explicit expression for $\rho_1(t)$. Nevertheless, we observe that if there exists $\beta_0>0$ such that $\beta(t,\,v)\geq\beta_0$ for all $v\in{\mathbb{R}}_+$ and all $t>0$ then $$\rho_1(t)\leq\rho_{1,0}e^{-\lambda\beta_0 t},$$ which implies that $\rho_1$ still vanishes for $t\to +\infty$. Consequently, we deduce $f_1(t,\,\cdot)\to 0$ in $L^1({\mathbb{R}}_+)$ for $t\to +\infty$.
### Quasi-invariant limit and asymptotic distributions {#sect:death.quasi-invariant_limit}
One of the most interesting issues in the study of kinetic models is the characterisation of the stationary distributions arising asymptotically for $t\to +\infty$, which depict the emergent behaviour of the system. For conservative kinetic equations this is typically carried out by means of asymptotic procedures, which, in suitable regimes of the parameters of the microscopic interactions, transform a Boltzmann-type integro-differential equation into a partial differential equation usually more amenable to analytical investigations. An effective asymptotic procedure is the so-called *quasi-invariant limit*, which leads to *Fokker-Planck-type* equations.
The idea behind the quasi-invariant limit is that one studies a regime in which the post-interaction state $v'$ is close enough to the pre-interaction state $v$, so that interactions produce a small transfer of microscopic state between the interacting agents. This concept was first introduced in the kinetic literature on multi-agent systems in [@cordier2005JSP; @toscani2006CMS] for binary collisions and in [@furioli2017M3AS] for the interactions with a fixed background and has its roots in the concept of *grazing collisions* studied in the classical kinetic theory [@villani1998ARMA].
In the present context, as we are considering both microscopic interactions and label switching, we extend this procedure to quasi-invariant microscopic dynamics which encompass both quasi-invariant interactions and quasi-invariant transition probabilities.
In , after introducing a small parameter $0<\epsilon\ll 1$, we scale the coefficients as $p\to p^\epsilon$, $q\to q^\epsilon$, where $p^\epsilon$, $q^\epsilon$ are random variables such that $$\begin{array}{ll}
\ave{p^\epsilon}=1-\epsilon, & {\operatorname{Var}}(p^\epsilon)=\kappa\epsilon \\[2mm]
\ave{q^\epsilon}=\epsilon, & {\operatorname{Var}}(q^\epsilon)=\kappa\epsilon^{1+\theta}
\end{array}
\label{eq:quasi-invariant_scaling}$$ and $\kappa,\,\theta>0$ are constant parameters. These choices are motivated by the following considerations: for $\epsilon\to 0^+$, on one hand $p^\epsilon$, $q^\epsilon$ converge in law to the constants $p=1$, $q=0$, respectively, thus in the regime of small $\epsilon$ the interaction is quasi-invariant. On the other hand, for finite $\epsilon>0$ it results $\ave{p^\epsilon+q^\epsilon}=1$ and, if $\epsilon$ is sufficiently small, $\ave{(p^\epsilon)^2+(q^\epsilon)^2}=1+(\kappa-2)\epsilon+o(\epsilon)$, hence $\ave{(p^\epsilon)^2+(q^\epsilon)^2}<1$ if $\kappa<2$. Therefore, owing to the discussion set forth in Section \[sect:int.dyn\], in the regime of small $\epsilon$ the scaling allows one to observe physical dynamics with conserved mean and internal energy evolving towards a finite non-zero value. Moreover, considering that the variation of the microscopic state due to the interaction is $v'-v=(p^\epsilon-1)v+q^\epsilon v_\ast$, we further observe that the parameter $$\kappa={\left\lvert\frac{{\operatorname{Var}}(p^\epsilon-1)}{\ave{p^\epsilon-1}}\right\rvert}$$ is the ratio between the stochastic and the deterministic average contributions of the $v$-coefficient $p^\epsilon$ to the post-interaction variation of the microscopic state $v$ itself. Conversely, since ${\operatorname{Var}}(q^\epsilon)=o({\operatorname{Var}}(p^\epsilon))$ for $\epsilon\to 0^+$, the scaling implies that the stochastic contribution of the $v_\ast$-coefficient $q^\epsilon$ to the variation of the microscopic state $v$ is negligible in the limit with respect to that of $p^\epsilon$.
As far as the transition probability is concerned, we scale $\beta$ as $$\beta^\epsilon(t,\,v)=\epsilon\beta(t,\,v),$$ so that from we deduce $p(t,\,v;\,2\vert 1)\to 0$ and $p(t,\,v;\,1\vert 1)\to 1$ when $\epsilon\to 0^+$, meaning that also the label switching tends to be quasi-invariant (in probability) for $\epsilon$ small enough.
To compensate for the smallness of each interaction and each label switching, we simultaneously scale the corresponding rates as $$\lambda=\mu=\frac{1}{\epsilon},
\label{eq:lambda_mu.scaling}$$ which imply a high number of interactions and instances of label switch per unit time when $\epsilon\approx 0$.
Let us denote by $f_1^\epsilon(t,\,v)$ the distribution function of the group $i=1$ parametrised by the scaling parameter $\epsilon$. From we deduce that it satisfies $$\begin{aligned}
\begin{aligned}[b]
\frac{d}{dt}\int_{{\mathbb{R}}_+}\varphi(v)f_1^\epsilon(t,\,v)\,dv &= -\int_{{\mathbb{R}}_+}\varphi(v)\beta(t,\,v)f_1^\epsilon(t,\,v)\,dv \\
&\phantom{=} +\frac{1}{\epsilon}\int_{{\mathbb{R}}_+}\int_{{\mathbb{R}}_+}\ave{\varphi(v')-\varphi(v)}f_1^\epsilon(t,\,v)f_1^\epsilon(t,\,v_\ast)\,dv\,dv_\ast.
\end{aligned}
\label{eq:q.i_inter}\end{aligned}$$ Now, let $\varphi$ be a smooth and compactly supported function. Expanding the difference $\varphi(v')-\varphi(v)$ in Taylor series about $v$ and using with $p=p^\epsilon$, $q=q^\epsilon$ like in we get $$\begin{aligned}
\begin{aligned}[b]
\frac{d}{dt}\int_{{\mathbb{R}}_+}\varphi(v)f_1^\epsilon(t,\,v)\,dv &= -\int_{{\mathbb{R}}_+}\varphi(v)\beta(t,\,v)f_1^\epsilon(t,\,v)\,dv \\
&\phantom{=} +\int_{{\mathbb{R}}_+}\int_{{\mathbb{R}}_+}\varphi'(v)(v_\ast-v)f_1^\epsilon(t,\,v)f_1^\epsilon(t,\,v_\ast)\,dv\,dv_\ast \\
&\phantom{=} +\frac{\kappa\rho_1^\epsilon(t)}{2}\int_{{\mathbb{R}}_+}\varphi''(v)v^2f_1^\epsilon(t,\,v)\,dv+R_\varphi(f_1^\epsilon,\,f_1^\epsilon)(t,\,v),
\end{aligned}
\label{eq:FP+R}\end{aligned}$$ where the remainder $R_\varphi(f_1^\epsilon,\,f_1^\epsilon)$ satisfies[^3] (cf. [@cordier2005JSP] for similar calculations) $$\begin{aligned}
\begin{aligned}[b]
{\left\lvertR_\varphi(f_1^\epsilon,\,f_1^\epsilon)(t,\,v)\right\rvert} &\lesssim \Vert\varphi''\Vert_\infty\left(\epsilon+\epsilon^\theta\right)\int_{{\mathbb{R}}_+}v^2f_1^\epsilon(t,\,v)\,dv \\
&\phantom{\lesssim} +\Vert\varphi'''\Vert_\infty\left(\sqrt{\epsilon}+\epsilon^{\frac{1+3\theta}{2}}+\epsilon^2\right)\int_{{\mathbb{R}}_+}v^3f_1^\epsilon(t,\,v)\,dv.
\end{aligned}
\label{eq:remainder}\end{aligned}$$
In view of the scaling , we can standardise $p^\epsilon$, $q^\epsilon$ as $$p^\epsilon=1-\epsilon+\sqrt{\kappa\epsilon}Z, \qquad q^\epsilon=\epsilon+\sqrt{\kappa\epsilon^{1+\theta}}Z_\ast,$$ where $Z,\,Z_\ast$ are two independent random variables with zero mean and unitary variance, which we assume to be such that $\ave{{\left\lvertZ\right\rvert}^3},\,\ave{{\left\lvertZ_\ast\right\rvert}^3}<+\infty$. Thanks to this representation, setting $\varphi(v)=v^2,\,v^3$ in we further discover, after some algebraic calculations using in particular the inequalities $ab\leq\frac{1}{2}(a^2+b^2)$ and $ab^2\leq\frac{2}{3}(a^3+b^3)$ for $a,\,b\geq 0$, that $$\begin{aligned}
& \frac{d}{dt}\int_{{\mathbb{R}}_+}v^2f_1^\epsilon(t,\,v)\,dv\lesssim\left(1+\epsilon+\epsilon^\theta\right)\int_{{\mathbb{R}}_+}v^2f_1^\epsilon(t,\,v)\,dv \\
& \frac{d}{dt}\int_{{\mathbb{R}}_+}v^3f_1^\epsilon(t,\,v)\,dv\lesssim\left(1+\sqrt{\epsilon}+\epsilon+\epsilon^2+\epsilon^\theta+\epsilon^\frac{1+3\theta}{2}+\epsilon^{1+\theta}\right)
\int_{{\mathbb{R}}_+}v^3f_1^\epsilon(t,\,v)\,dv,\end{aligned}$$ which imply that, for all fixed $t>0$, the terms $\int_{{\mathbb{R}}_+}v^2f_1^\epsilon(t,\,v)\,dv$, $\int_{{\mathbb{R}}_+}v^3f_1^\epsilon(t,\,v)\,dv$ remain bounded when $\epsilon\to 0^+$. Therefore, from we infer $$R_\varphi(f_1^\epsilon,\,f_1^\epsilon)\xrightarrow{\epsilon\to 0^+}0.$$
Let us assume now that $(f_1^\epsilon)$ converges in $C({\mathbb{R}}_+;\,L^1({\mathbb{R}}_+)\cap L^1({\mathbb{R}}_+;\,v\,dv))$, possibly up to subsequences, to a distribution function $f_1$ when $\epsilon\to 0^+$. Hence we have in particular $$\begin{aligned}
& \rho_1^\epsilon(t)=\int_{{\mathbb{R}}_+}f_1^\epsilon(t,\,v)\,dv\xrightarrow{\epsilon\to 0^+}\rho_1(t)=\int_{{\mathbb{R}}_+}f_1(t,\,v)\,dv \\
& \rho_1^\epsilon(t)m_1^\epsilon(t)=\int_{{\mathbb{R}}_+}vf_1^\epsilon(t,\,v)\,dv\xrightarrow{\epsilon\to 0^+}\rho_1(t)m_1(t)=\int_{{\mathbb{R}}_+}vf_1(t,\,v)\,dv,\end{aligned}$$ where $m_1^{(\epsilon)}(t):=\frac{1}{\rho_1^{(\epsilon)}(t)}\int_{{\mathbb{R}}_+}vf_1^{(\epsilon)}(t,\,v)\,dv$ denotes the first moment of the distribution function $f_1^{(\epsilon)}$. Then, passing to the limit $\epsilon\to 0^+$ in we obtain the limit equation $$\begin{aligned}
\frac{d}{dt}\int_{{\mathbb{R}}_+}\varphi(v)f_1(t,\,v)\,dv &= -\int_{{\mathbb{R}}_+}\varphi(v)\beta(t,\,v)f_1(t,\,v)\,dv \\
&\phantom{=} +\int_{{\mathbb{R}}_+}\int_{{\mathbb{R}}_+}\varphi'(v)(v_\ast-v)f_1(t,\,v)f_1(t,\,v_\ast)\,dv\,dv_\ast \\
&\phantom{=} +\frac{\kappa\rho_1(t)}{2}\int_{{\mathbb{R}}_+}\varphi''(v)v^2f_1(t,\,v)\,dv,\end{aligned}$$ which, by integration by parts and recalling the compactness of the support of $\varphi$, can be recognised as a weak form of the following Fokker-Planck equation with non-constant coefficients and reaction term: $$\partial_tf_1=\frac{\kappa\rho_1(t)}{2}\partial_v^2(v^2f_1)+\rho_1(t)\partial_v\bigl((v-m_1(t))f_1\bigr)-\beta(t,\,v)f_1.
\label{eq:FP.death.f1}$$
The same quasi-invariant limit applied to leads to $$\partial_tf_2=\beta(t,v)f_1.
\label{eq:FP.death.f2}$$
If we look for the stationary distributions, say $f_1^\infty$, $f_2^\infty$, in the quasi-invariant regime we find that they satisfy the system of equations $$\begin{cases}
\dfrac{\kappa\rho_1^\infty}{2}\partial_v^2(v^2f_1^\infty)+\rho_1^\infty\partial_v\bigl((v-m_1^\infty)f_1^\infty\bigr)-\beta^\infty(v)f_1^\infty=0 \\[2mm]
\beta^\infty(v)f_1^\infty=0,
\end{cases}$$ where the symbols $\rho_1^\infty$, $m_1^\infty$ have an obvious meaning while $\beta^\infty(v):=\lim_{t\to +\infty}\beta(t,\,v)$.
If $\beta^\infty$ is not identically zero then the second equation implies $f_1^\infty(v)=0$, which is clearly also a solution of the first equation. This is consistent with the idea that, in the long run, all living agents labelled with $i=1$ die by switching to the label $i=2$. Conversely, if $\beta^\infty\equiv 0$ then we distinguish two cases:
1. if $\beta(t,\,v)\to 0^+$ for $t\to +\infty$ but there exists $\beta_0=\beta_0(t)>0$ such that $\beta(t,\,v)\geq\beta_0(t)$ for all $t>0$ and all $v\in{\mathbb{R}}_+$ and moreover $\int_{{\mathbb{R}}_+}\beta_0(t)\,dt=+\infty$ (i.e., roughly speaking, $\beta(t,\,v)$ tends to zero slowly enough in time) then from with the quasi-invariant scaling $\lambda=\frac{1}{\epsilon}$, $\beta_0^\epsilon(t)=\epsilon\beta_0(t)$ we deduce $$\rho_1(t)\leq\rho_{1,0}\exp{\left(-\int_0^t\beta_0(s)\,ds\right)}\xrightarrow{t\to +\infty}0,$$ whence we obtain again the stationary distribution $f_1^\infty(v)=0$;
2. if there exists $t_0>0$ such that $\beta(t,\,v)\equiv 0$ for $t\geq t_0$ then from $t_0$ onwards the masses $\rho_1$, $\rho_2$ are conserved. Moreover, owing to the quasi-invariant scaling , also the first moment $m_1$ is conserved. Indeed from with $\varphi(v)=v$ we get, for $t>t_0$, $\frac{d}{dt}(\rho_1^\epsilon m_1^\epsilon)=0$ and the result follows passing to the limit $\epsilon\to 0^+$. Hence $\rho_1^\infty=\rho_1(t_0)>0$, $m_1^\infty=m_1(t_0)$ and the stationary distribution $f_1^\infty$ satisfies $$\frac{\kappa}{2}\partial_v(v^2f_1^\infty)+(v-m_1(t_0))f_1^\infty=0,$$ whose unique solution with mass $\rho_1(t_0)$ and first moment $m_1(t_0)$ is $$f_1^\infty(v)=\rho_1(t_0)\frac{(2\kappa m_1(t_0))^{2\kappa+1}}{\Gamma(2\kappa+1)}\cdot\frac{e^{-\frac{2\kappa m_1(t_0)}{v}}}{v^{2(\kappa+1)}},$$ namely an inverse-gamma-type distribution with shape parameter $2\kappa+1$ and scale parameter $2\kappa m_1(t_0)$. Notice however that the exact determination of $\rho_1(t_0)$, $m_1(t_0)$ requires to solve the transient dynamics described by up to $t=t_0$. The same is also necessary for the determination of $f_2^\infty(v)=f_2(t_0,\,v)$.
### An explicitly solvable case
Further insights into the solutions to , can be obtained in the particular case in which the transition probability $\beta(t,\,v)$ is constant, say $\beta(t,\,v)\equiv\beta_0>0$. Then from with $\varphi(v)=1,\,v$, together with the quasi-invariant scaling $\lambda=\frac{1}{\epsilon}$, $\beta_0^\epsilon=\epsilon\beta_0$ plus , we obtain respectively, for $\epsilon\to 0^+$, $$\frac{d\rho_1}{dt}=-\beta_0\rho_1, \qquad \frac{d}{dt}(\rho_1m_1)=-\beta_0\rho_1m_1,$$ which imply $\rho_1(t)=\rho_{1,0}e^{-\beta_0t}$ and $m_1\equiv\text{constant}$, i.e. the first moment of $f_1$ is conserved in time. In this situation, it is reasonable to look for a self-similar solution of the form $$f_1(t,\,v)=\frac{\rho_1(t)}{m_1}g\!\left(\frac{v}{m_1}\right),$$ where $g:{\mathbb{R}}_+\to{\mathbb{R}}_+$ is such that $$\int_{{\mathbb{R}}_+}g(v)\,dv=1, \qquad \int_{{\mathbb{R}}_+}vg(v)\,dv=1.
\label{eq:g.self-similar}$$ Plugging into , we discover that $g$ satisfies the following stationary Fokker-Planck equation: $$\frac{\kappa}{2}\partial_v^2(v^2g)+\partial_v\bigl((v-1)g\bigr)=0,
\label{eq:FP.g.death}$$ whose unique solution with unitary mass is $$g(v)=\frac{(2\kappa)^{2\kappa+1}}{\Gamma(2\kappa+1)}\cdot\frac{e^{-\frac{2\kappa}{v}}}{v^{2(\kappa+1)}},
\label{eq:g}$$ namely an inverse-gamma distribution with shape parameter $2\kappa+1$ and scale parameter $2\kappa$. Consequently, we determine $$f_1(t,\,v)=\rho_{1,0}e^{-\beta_0t}\frac{(2\kappa m_1)^{2\kappa+1}}{\Gamma(2\kappa+1)}\cdot\frac{e^{-\frac{2\kappa m_1}{v}}}{v^{2(\kappa+1)}}$$ and from $$f_2(t,\,v)=f_{2,0}(v)+\rho_{1,0}\left(1-e^{-\beta_0t}\right)\frac{(2\kappa m_1)^{2\kappa+1}}{\Gamma(2\kappa+1)}\cdot\frac{e^{-\frac{2\kappa m_1}{v}}}{v^{2(\kappa+1)}},$$ where $f_{2,0}(v)\geq 0$ with $\int_{{\mathbb{R}}_+}f_{2,0}(v)\,dv=1-\rho_{1,0}$ is the initial distribution function of the agents with label $i=2$. These solutions provide the exact evolution of the system under the joint label switching and interaction processes.
Birth
-----
We consider now the birth process, in which the group $i=1$ composed of interacting agents accepts new incomes from the inert group $i=2$. The transition probabilities may therefore be chosen as $$\begin{array}{ll}
T(t,\,v;\,1\vert 2)=\beta(t,\,v), & T(t,\,v;\,2\vert 2)=1-\beta(t,\,v) \\
T(t,\,v;\,2\vert 1)=0, & T(t,\,v;\,1\vert 1)=1
\end{array}$$ with $0\leq\beta(t,\,v)\leq 1$ for all $t>0$ and all $v\in{\mathbb{R}}_+$. The kinetic equations describing the evolution of $f_1$ and $f_2$ can be deduced from , considering that the agents of the population $i=2$ do not interact. Therefore, we have $$\begin{aligned}
\begin{aligned}[b]
\frac{d}{dt}\int_{{\mathbb{R}}_+}\varphi(v)f_1(t,\,v)\,dv &= \lambda\int_{{\mathbb{R}}_+}\varphi(v)\beta(t,\,v)f_2(t,\,v)\,dv \\
&\phantom{=} +\mu\int_{{\mathbb{R}}_+}\int_{{\mathbb{R}}_+}\ave{\varphi(v')-\varphi(v)}f_1(t,\,v)f_1(t,\,v_\ast)\,dv\,dv_\ast
\end{aligned}
\label{eq:birth_f1_weak}\end{aligned}$$ and $$\frac{d}{dt}\int_{{\mathbb{R}}_+}\varphi(v)f_2(t,\,v)\,dv=-\lambda\int_{{\mathbb{R}}_+}\varphi(v)\beta(t,\,v)f_2(t,\,v)\,dv.
\label{eq:birth_f2_weak}$$ Notice that in this case from we obtain explicitly $$f_2(t,\,v)=f_{2,0}(v)\exp\left(-\lambda\int_0^t \beta(s,\,v)\,ds\right),$$ which can be possibly plugged into to obtain a self-consistent equation for the sole distribution function $f_1$.
### Mass balance {#sect:birth.mass_balance}
The evolution of the mass of the two populations is obtained with $\varphi(v)=1$ in , and reads $$\begin{aligned}
& \frac{d\rho_1}{dt}=\lambda\int_{{\mathbb{R}}_+}\beta(t,\,v)f_2(t,\,v)\,dv \\
& \frac{d\rho_2}{dt}=-\lambda\int_{{\mathbb{R}}_+}\beta(t,\,v)f_2(t,\,v)\,dv.\end{aligned}$$
Like in Section \[sect:death.mass\], if $\beta$ does not depend on $v$, i.e. $\beta=\beta(t)$, we determine explicitly $$\rho_2(t)=\rho_{2,0}\exp{\left(-\lambda\int_0^t\beta(s)\,ds\right)}$$ and consequently, recalling that $\rho_{1,0}+\rho_{2,0}=1$, $$\rho_1(t)=1-\rho_{2,0}\exp{\left(-\lambda\int_0^t\beta(s)\,ds\right)}.$$ In this case, if $\int_{{\mathbb{R}}_+}\beta(t)\,dt=+\infty$ then $\rho_2\to 0$ and $\rho_1\to 1$ for $t\to +\infty$, i.e. the whole population $i=2$ is born in the long run. If instead $\beta$ vanishes definitively from a certain time $t=t_0$ onwards then a residual mass of agents in $i=2$ remains and $\rho_1<1$ for $t\to +\infty$.
If $\beta$ features a full dependence on $t$ and $v$ then it is not possible to determine explicitly the evolution of $\rho_1$, $\rho_2$. Nevertheless, if there exists $\beta_0=\beta_0(t)\geq 0$ such that $\beta(t,\,v)\geq\beta_0(t)$ for all $t\geq 0$ and all $v\in{\mathbb{R}}_+$ then we may estimate $$\rho_1(t)\geq 1-\rho_{2,0}\exp{\left(-\lambda\int_0^t\beta_0(s)\,ds\right)}, \qquad \rho_2(t)\leq\rho_{2,0}\exp{\left(-\lambda\int_0^t\beta_0(s)\,ds\right)},$$ which still imply $\rho_1\to 1$ and $\rho_2\to 0$ as $t\to +\infty$ if $\int_0^{+\infty}\beta_0(t)\,dt=+\infty$.
### Quasi-invariant limit and explicit solutions
The same quasi-invariant scaling , of Section \[sect:death.quasi-invariant\_limit\] applied to , produces in this case $$\partial_tf_1=\frac{\kappa\rho_1(t)}{2}\partial_v^2(v^2f_1)+\rho_1(t)\partial_v\bigl((v-m_1(t))f_1\bigr)+\beta(t,\,v)f_2
\label{eq:FP.birth.f1}$$ and $$\partial_tf_2=-\beta(t,\,v)f_2.
\label{eq:FP.birth.f2}$$
From , we obtain that at the steady state it results $\beta^\infty(v)f_2^\infty=0$, hence from that $f_1^\infty$ solves $$\frac{\kappa}{2}\partial_v^2(v^2f_1^\infty)+\partial_v\bigl((v-m_1^\infty)f_1^\infty\bigr)=0.$$ Hence we deduce that $f_1^\infty$ is always an inverse-gamma-type distribution with mean $m_1^\infty$ and mass $\rho_1^\infty=\lim_{t\to+\infty}\rho_1(t)$. Using the arguments of Section \[sect:birth.mass\_balance\], we observe that $\rho_1^\infty=1$ whenever $\beta(t,\,v)$ tends to zero slowly enough for $t\to +\infty$. Otherwise, if $\beta(t,\,v)\equiv 0$ for $t\geq t_0$ then $\rho_1^\infty=\rho_1(t_0)\leq 1$.
Let us now consider , in the case of constant $\beta$, say $\beta(t,\,v)=\beta_0>0$ for all $t\geq 0$ and all $v\in{\mathbb{R}}_+$. Then yields $f_2(t,\,v)=f_{2,0}(v)e^{-\beta_0t}$ and the first moment $m_2$ is conserved, say $m_2(t)=m$ for all $t\geq 0$. Next, from with $\varphi(v)=v$, under the scaling , and in the quasi-invariant limit $\epsilon\to 0^+$, we deduce $$\frac{dm_1}{dt}=\beta_0\frac{\rho_2}{\rho_1}(m-m_1),$$ which shows that $m_1(t)\to m$ for $t\to +\infty$ and, if $m_{1,0}=m$, that $m_1$ is in turn conserved and equals $m$ at all times. Therefore, it makes sense to look for a self-similar solution of of the form $$f_1(t,\,v)=\frac{\rho_1(t)}{m}g\!\left(\frac{v}{m}\right),$$ where $g:{\mathbb{R}}_+\to{\mathbb{R}}_+$ satisfies . Plugging into we get $$\frac{\kappa}{2}\partial_v^2(v^2g)+\partial_v((v-1)g)=\frac{\beta_0m}{\rho_1^2}\left(\frac{\rho_2}{m}g-f_2\right)
\label{eq:FP.g.birth}$$ whence, choosing the initial shape of $f_2$ as $f_{2,0}(v)=\frac{\rho_{2,0}}{m}g\!\left(\frac{v}{m}\right)$ so that $f_2(t,\,v)=\frac{\rho_2(t)}{m}g\!\left(\frac{v}{m}\right)$, we recover for $g$ the Fokker-Planck equation . This allows us to conclude that the time evolution of $f_1$, $f_2$ is given explicitly by $$f_1(t,\,v)=\left(1-\rho_{2,0}e^{-\beta_0 t}\right)\frac{(2\kappa m)^{2\kappa+1}}{\Gamma(2\kappa+1)}\cdot\frac{e^{-\frac{2\kappa m}{v}}}{v^{2(\kappa+1)}}, \qquad
f_2(t,\,v)=\rho_{2,0}e^{-\beta_0 t}\frac{(2\kappa m)^{2\kappa+1}}{\Gamma(2\kappa+1)}\cdot\frac{e^{-\frac{2\kappa m}{v}}}{v^{2(\kappa+1)}}.$$
If at the initial time $f_2$ is not the $g$-shaped distribution above then cannot be solved explicitly due to the time-varying reaction term on the right-hand side. Nevertheless, since $f_2(t,\,v)\to 0$, $\rho_1(t)\to 1$ and $\rho_2(t)\to 0$ when $t\to +\infty$, we formally infer that for large times $g$ still solves . Consequently, we at least characterise the stationary distributions as $$f_1^\infty(v)=\frac{(2\kappa m)^{2\kappa+1}}{\Gamma(2\kappa+1)}\cdot\frac{e^{-\frac{2\kappa m}{v}}}{v^{2(\kappa+1)}}, \qquad f_2^\infty(v)=0.$$
A kinetic model of the contagion of infectious diseases with quarantine {#sect:contagion}
=======================================================================
Models for the spreading of infectious diseases are a prominent example of non-conservative dynamics in which agents switch from one group to another depending on their infection condition. Classical population dynamics models take a super-macroscopic point of view and describe the time evolution of the total number of susceptible ($S$), infected ($I$) and recovered ($R$) individuals assuming that the probability of a contagious encounter between a susceptible and an infected individual is proportional to the size of either group.
As an example, we show that the popular SIR model may be derived from the kinetic description of the pure Markov-type microscopic jump process presented in Section \[sect:label.switch\]. Let $x=1$ label the susceptible individuals, $x=2$ the infected individuals and $x=3$ the recovered individuals. The kinetic distribution function $f$ may be represented by setting $n=3$ in , $f_1$, $f_2$, $f_3$ being the probabilities (normalised masses) of the three groups of individuals at time $t$. They satisfy the system of equations in which, up to a time scaling, we may conveniently assume $\lambda=1$. If we further specify the transition probabilities as $$\begin{array}{lll}
T(t;\,1\vert 1)=1-\beta f_2(t), & T(t;\,2\vert 1)=\beta f_2(t), & T(t;\,3\vert 1)=0 \\
T(t;\,1\vert 2)=0, & T(t;\,2\vert 2)=1-\gamma, & T(t;\,3\vert 2)=\gamma \\
T(t;\,1\vert 3)=0, & T(t;\,2\vert 3)=0, & T(t;\,3\vert 3)=1 \\
\end{array}$$ with $0\leq\beta,\,\gamma\leq 1$, we obtain from $$\begin{cases}
f_1'=-\beta f_1f_2 \\
f_2'=\beta f_1f_2-\gamma f_2 \\
f_3'=\gamma f_2,
\end{cases}$$ which has indeed the form of the SIR model. Notice that all the transition probabilities are constant but those associated with the label switch $1\to 2$ from susceptible to infected, which are proportional to the density of infected individuals.
As this example demonstrates, this type of models does not describe the progression of the contagion as the result of actual microscopic contacts among the individuals but simply as a consequence of their jumps from one label to another. The kinetic framework introduced in the previous sections allows us to conceive a new model, in which individual contacts are explicitly taken into account and the progression of the contagion is described statistically without the need for a sharp compartmentalisation of the population. The labelling of the agents can then profitably be used for different purposes, such as e.g., to distinguish individuals put in isolation when recognised as contagious.
Let us assume that each individual is characterised by a microscopic state $v\in{\mathbb{R}}_+$ representing their *infectiveness* and moreover that there are two groups of people: those labelled by $x=1$, who have not been diagnosed with the infection yet, and those labelled by $x=2$, who have been detected as infected and quarantined. Undiagnosed people ($x=1$) interact with one another and switch to label $x=2$ once diagnosed. Conversely, quarantined people ($x=2$) do not interact due to isolation and possibly switch back to label $x=1$ when their infectiveness has decreased enough. To describe these dynamics, the reference framework is provided by the kinetic equation , which takes into account interactions within the same labels only. In particular, we model the transition probabilities as
$$\begin{aligned}
{2}
& T(v;\,1\vert 1)=1-\alpha(v), && \qquad T(v;\,2\vert 1)=\alpha(v) \label{eq:T_quarantine.1->2} \\
& T(v;\,1\vert 2)=\beta(v), && \qquad T(v;\,2\vert 2)=1-\beta(v), \label{eq:T_quarantine.2->1}
\end{aligned}$$
\[eq:T\_quarantine\]
where $0\leq\alpha(v),\,\beta(v)\leq 1$ are the probabilities that an individual with infectiveness $v$ is diagnosed and quarantined or gets back from quarantine, respectively. We assume that these probabilities are time-independent and moreover that $\alpha$ is non-decreasing and $\beta$ is non-increasing in $v$. Consequently, the equations for the distribution functions $f_1$, $f_2$ of the undiagnosed and quarantined people read: $$\begin{aligned}
\begin{aligned}[b]
\frac{d}{dt}\int_{{\mathbb{R}}_+}\varphi(v)f_1(t,\,v)\,dv &= \lambda\int_{{\mathbb{R}}_+}\varphi(v)\Bigl(\beta(v)f_2(t,\,v)-\alpha(v)f_1(t,\,v)\Bigr)dv \\
&\phantom{=} +\mu\int_{{\mathbb{R}}_+}\int_{{\mathbb{R}}_+}\ave{\varphi(v')-\varphi(v)}f_1(t,\,v)f_1(t,\,v_\ast)\,dv\,dv_\ast
\end{aligned}
\label{eq:undiagnosed}\end{aligned}$$ and $$\begin{aligned}
\begin{aligned}[b]
\frac{d}{dt}\int_{{\mathbb{R}}_+}\varphi(v)f_2(t,\,v)\,dv &= \lambda\int_{{\mathbb{R}}_+}\varphi(v)\Bigl(\alpha(v)f_1(t,\,v)-\beta(v)f_2(t,\,v)\Bigr)dv \\
&\phantom{=} +\mu\int_{{\mathbb{R}}_+}\int_{{\mathbb{R}}_+}\ave{\varphi(v')-\varphi(v)}f_2(t,\,v)f_2(t,\,v_\ast)\,dv\,dv_\ast.
\end{aligned}
\label{eq:quarantined}\end{aligned}$$
As far as the interaction rules are concerned, we assume that undiagnosed people ($x=1$) may infect other undiagnosed people depending on their current infectiveness. Specifically, we set $$v'=(1-\nu_1+\eta)v+\nu_2v_\ast
\label{eq:int.undiagnosed}$$ where $\nu_1,\,\nu_2\in (0,\,1)$ are exchange rates among the individuals modelling the contagion dynamics and $\eta\in (\nu_1-1,\,+\infty)$ is a centred (i.e. $\ave{\eta}=0$) random variable accounting for random fluctuations in the individual infectiveness. Notice that is of the form with stochastic $p=1-\nu_1+\eta$ and deterministic $q=\nu_2$.
Conversely, we assume that quarantined people ($x=2$) may only recover from the infection due to the lack of contacts with other individuals: $$v'=(1-\gamma+\xi)v,
\label{eq:int.quarantined}$$ where $\gamma\in (0,\,1)$ is the rate of recovery and $\xi\in (\gamma-1,\,+\infty)$ is another centred random variable independent of $\eta$. Also is of the form with stochastic $p=1-\gamma+\xi$ and deterministic $q=0$.
Recently, another kinetic model dealing with the spreading of infectious diseases and which may be seen as a particular case of the general non-conservative Boltzmann-type framework presented in this paper has been proposed [@dimarco2020PREPRINT-a]. However, contagion dynamics are not the main focus of the work, indeed they are described by a kinetic rephrasing of the SIR model analogous to that discussed ad the beginning of this section. The focus is instead on the impact of infectious diseases on the socio-economic status of the individuals.
Constant transition probabilities {#sect:alpha_beta.const}
---------------------------------
We investigate first the simple case of constant $\alpha,\,\beta$, which allows for a detailed analysis of the trend of model -.
Assume $$\lambda=\mu=1,$$ meaning that the label switch and interaction dynamics take place on the same time scale. Setting $\varphi(v)=1,\,v$ in , and using the interaction rules , , we obtain the following equations for the evolution of the zeroth and first moments of $f_1,\,f_2$:
[align]{} & =-(+)\_1 \[eq:rho1\]\
& \_2=1-\_1 \[eq:rho2\]\
& (\_1m\_1)=-(+\_1-\_2)\_1m\_1+\_2m\_2 \[eq:rho1m1\]\
& (\_2m\_2)=\_1m\_1-(+)\_2m\_2. \[eq:rho2m2\]
From , , together with the natural initial conditions $\rho_{1,0}=1$, $\rho_{2,0}=0$, we obtain $$\rho_1(t)=\frac{\beta}{\alpha+\beta}\left(1+\frac{\alpha}{\beta}e^{-(\alpha+\beta)t}\right), \qquad
\rho_2(t)=\frac{\alpha}{\alpha+\beta}\left(1-e^{-(\alpha+\beta)t}\right),
\label{eq:Tconst.rho}$$ whence $\rho_1^\infty=\frac{\beta}{\alpha+\beta}$ and $\rho_2^\infty=\frac{\alpha}{\alpha+\beta}$. Therefore, quarantine apparently settles as a persistent condition involving a fixed fraction $\rho_2^\infty$ of the population. This may indicate that the infection cannot be eradicated and becomes endemic.
Nevertheless, we obtain a clearer picture by considering the further piece of information provided by , , which unveil the evolution of the mean infectivenesses $m_1,\,m_2$ of undiagnosed and quarantined individuals. The only equilibrium is clearly $\rho_1^\infty m_1^\infty=\rho_2^\infty m_2^\infty=0$, which, owing to the previous result, implies $m_1^\infty=m_2^\infty=0$. With classical arguments of linear stability, we see that this equilibrium is globally asymptotically stable if $$\alpha>\alpha_\dagger:=\max\left\{\frac{\beta+\gamma}{\gamma}(\nu_2-\nu_1),\,\nu_2-\nu_1-(\beta+\gamma)\right\},
\label{eq:alpha}$$ i.e. if the quarantine rate is large enough. In this case, the system evolves towards the configuration $$f_1^\infty(v)=\frac{\beta}{\alpha+\beta}\delta(v), \qquad f_2^\infty(v)=\frac{\alpha}{\alpha+\beta}\delta(v),$$ which represents actually a complete eradication of the infection from the population because all individuals in both $x=1$ and $x=2$ reach the state $v=0$. Conversely, if is violated then eradication may not occur because the equilibrium $m_1^\infty=m_2^\infty=0$ may become unstable. In particular, either $m_1$ or $m_2$ may diverge asymptotically to $+\infty$.
As a first example, let us consider $\nu_1=\nu_2$ in , which reduces to $\alpha>0$. Thus any arbitrarily small quarantine rate leads, in the long run, to the eradication of the infection. However, the smallness of $\alpha$ affects considerably the speed of convergence to such an equilibrium. Indeed, for $\alpha\to 0^+$ the eigenvalues $\omega_1,\,\omega_2$ of the linear system - can be approximated as $$\omega_1=-\frac{\alpha\gamma}{\beta+\gamma}+o(\alpha), \qquad \omega_2=-(\beta+\gamma)+o(1),$$ which shows that the convergence of $m_1$ to $0$ may be particularly slow.
As a second example, let us consider $\nu_1=0$, $\nu_2>0$ in . Hence an individual can only get more infected by the contact with another infected individual. Notice that without label switching, i.e. for $\alpha=\beta=0$, such a microscopic interaction would lead to a blow up of the mean infectiveness $m_1$ in time, indeed $\ave{p+q}=1+\nu_2>1$ (cf. Section \[sect:int.dyn\]). Conversely, thanks to the label switching, $m_1$ converges to zero if, according to , $$\alpha>\max\left\{\frac{\beta+\gamma}{\gamma}\nu_2,\,\nu_2-(\beta+\gamma)\right\}.$$ Clearly, the higher the contagion rate $\nu_2$ the more promptly infected individuals have to be diagnosed and quarantined.
Variable transition probabilities: Two-scale dynamics and hydrodynamic limit {#sect:alpha_beta.var}
----------------------------------------------------------------------------
For variable $\alpha$, $\beta$, a regime which allows us to gain some insights into the trends of model - is when the label switching and the interaction processes take place on two well separated time scales. Assume $$\lambda=1, \qquad \mu=\frac{1}{\delta},$$ where $0<\delta\ll 1$ is a small parameter. This implies that interactions among the agents with the same label are much more frequent than changes of label. In view of this argument, we can split e.g., as $$\begin{aligned}
&\frac{d}{dt}\int_{{\mathbb{R}}_+}\varphi(v)f_1(t,\,v)\,dv=\frac{1}{\delta}\int_{{\mathbb{R}}_+}\int_{{\mathbb{R}}_+}\ave{\varphi(v')-\varphi(v)}f_1(t,\,v)f_1(t,\,v_\ast)\,dv\,dv_\ast \\
&\frac{d}{dt}\int_{{\mathbb{R}}_+}\varphi(v)f_1(t,\,v)\,dv=\int_{{\mathbb{R}}_+}\varphi(v)\Bigl(\beta(v)f_2(t,\,v)-\alpha(v)f_1(t,\,v)\Bigr)dv.\end{aligned}$$ By introducing the new time scale $$\tau:=\frac{t}{\delta}$$ and defining $\tilde{f}_1(\tau,\,v):=f_1(t,\,v)$, we rewrite the system above as
[align]{} &\_[\_+]{}(v)\_1(,v)dv=\_[\_+]{}\_[\_+]{}\_1(,v)\_1(,v\_)dvdv\_ \[eq:split.interaction.f1\]\
&\_[\_+]{}(v)f\_1(t,v)dv=\_[\_+]{}(v)((v)f\_2(t,v)-(v)f\_1(t,v))dv, \[eq:split.switching.f1\]
that we interpret as follows: while the conservative interaction dynamics reach rapidly an equilibrium on the quick time scale $\tau$, the label switching dynamics are basically frozen on the slow time scale $t$.
The equilibrium on the $\tau$-scale is especially interesting when $\nu_1=\nu_2$ in , for then the contagion dynamics do not only conserve the mass $\rho_1$ of the undiagnosed individuals but also their mean infectiveness $m_1$. Indeed, in this case we have $\ave{p+q}=1$, cf. Section \[sect:int.dyn\]. As a consequence, the $\tau$-asymptotic distribution produced by is parametrised by both these macroscopic quantities on the $t$-scale and can be expressed in self-similar form as $$\frac{\rho_1(t)}{m_1(t)}g_1\!\left(\frac{v}{m_1(t)}\right),$$ where $g_1:{\mathbb{R}}_+\to{\mathbb{R}}_+$ satisfies the normalisation conditions $$\int_{{\mathbb{R}}_+}g_1(v)\,dv=1, \qquad \int_{{\mathbb{R}}_+}vg_1(v)\,dv=1.$$
The distribution $g_1$ may be explicitly determined from in a suitable quasi-invariant regime. For instance, if we scale the coefficients of the contagion dynamics as $\nu_1=\nu_2=\epsilon$ and ${\operatorname{Var}}(\eta^2)=\kappa\epsilon$ then for $\epsilon>0$ small we are in the conditions . By further scaling the interaction frequency in to $\frac{1}{\epsilon}$, which does not change the equilibrium distributions but compensate for the little effect of each interaction, in the limit $\epsilon\to 0^+$ we obtain that $g_1$ satisfies and is therefore given by .
Notice that the distribution is *heavy-tailed*, indeed for $v\to +\infty$ it decays polynomially (rather than exponentially) to zero with trend given by $v^{-2(\kappa+1)}$. This implies that the cumulative distribution function $G_1=G_1(v)$ of the infectiveness satisfies, for $v\gg 1$, $$1-G_1(v)=\int_v^{+\infty}g_1(v)\,dv\simeq\frac{1}{v^{2\kappa+1}}$$ and in turn this indicates that the probability that an individual is highly infective is in general not negligible. Borrowing the terminology from econophysics, the exponent $2\kappa+1>1$ may be assimilated to a *Pareto index* [@gualandi2018ECONOMICS].
Interestingly, it has been proposed that at the onset of the COVID-19 pandemic most contagions were caused by a very small number of highly infective individuals, a phenomenon known as *super-spreading* [@adam2020; @bi2020LID; @he2020NM]. In view of the previous considerations, it is clear that the curve takes into account the presence of such individuals, which in turn suggests that the simple contagion model is a good one to explain the emergence of consistent infectiveness distributions.
An analogous splitting argument applied to produces
[align]{} &\_[\_+]{}(v)\_2(,v)dv=\_[\_+]{}\_[\_+]{}\_2(,v)\_2(,v\_)dvdv\_ \[eq:split.interaction.f2\]\
&\_[\_+]{}(v)f\_2(t,v)dv=\_[\_+]{}(v)((v)f\_1(t,v)-(v)f\_2(t,v))dv \[eq:split.switching.f2\]
with $\tilde{f}_2(\tau,\,v):=f_2(t,\,v)$. We now let $\rho_2(t)g_2(v)$ denote the $\tau$-asymptotic distribution generated by , where $g_2:{\mathbb{R}}_+\to{\mathbb{R}}_+$ satisfies only the normalisation condition $$\int_{{\mathbb{R}}_+}g_2(v)\,dv=1$$ because the recovery dynamics conserve only the mass of the quarantined individuals.
On the whole, on the $t$-scale we express $$f_1(t,\,v)=\frac{\rho_1(t)}{m_1(t)}g_1\!\left(\frac{v}{m_1(t)}\right), \qquad f_2(t,\,v)=\rho_2(t)g_2(v)
\label{eq:split.f1.f2}$$ so that, plugging these distributions into , and taking the conservation relationship $\rho_1(t)+\rho_2(t)=1$ into account, we obtain the evolution of the macroscopic parameters $\rho_1$, $\rho_2$, $m_1$ on the slow time scale $t$ analogously to what happens in classical kinetic theory with the *hydrodynamic limit*: $$\begin{cases}
\dfrac{d\rho_1}{dt}=\left(\displaystyle{\int_{{\mathbb{R}}_+}}\beta(v)g_2(v)\,dv\right)\rho_2-\left(\displaystyle{\int_{{\mathbb{R}}_+}}\alpha(m_1v)g_1(v)\,dv\right)\rho_1 \\
\rho_2=1-\rho_1 \\
\dfrac{d}{dt}(\rho_1m_1)=\left(\displaystyle{\int_{{\mathbb{R}}_+}}v\beta(v)g_2(v)\,dv\right)\rho_2-\left(\displaystyle{\int_{{\mathbb{R}}_+}}v\alpha(m_1v)g_1(v)\,dv\right)\rho_1m_1.
\end{cases}
\label{eq:split.hydro}$$
It is not difficult to check that , together with the microscopic dynamics , produces $g_2(v)=\delta(v)$. Indeed, is such that $\ave{p+q}=1-\gamma<1$, which according to Section \[sect:int.dyn\] implies $m_2\to 0^+$ asymptotically in time. Therefore, specialises as $$\begin{cases}
\dfrac{d\rho_1}{dt}=\beta(0)(1-\rho_1)-\left(\displaystyle{\int_{{\mathbb{R}}_+}}\alpha(m_1v)g_1(v)\,dv\right)\rho_1 \\
\dfrac{d}{dt}(\rho_1m_1)=-\left(\displaystyle{\int_{{\mathbb{R}}_+}}v\alpha(m_1v)g_1(v)\,dv\right)\rho_1m_1,
\end{cases}$$ whose steady states are given by $$\beta(0)-\left(\beta(0)+\displaystyle{\int_{{\mathbb{R}}_+}}\alpha(m_1^\infty v)g_1(v)\,dv\right)\rho_1^\infty=0, \quad
\left(\displaystyle{\int_{{\mathbb{R}}_+}}v\alpha(m_1^\infty v)g_1(v)\,dv\right)\rho_1^\infty m_1^\infty=0.$$
Assume $\beta(0)>0$, then from the first equation we deduce $\rho_1^\infty>0$. Assume also that the mapping $v\mapsto\alpha(v)$ is strictly increasing with $\alpha(0)=0$. Then from the second equation we get $m_1^\infty=0$, considering that the integral term does not vanish if $m_1^\infty>0$. Indeed, from the assumed monotonicity of $\alpha$ we have $$\int_{{\mathbb{R}}_+}v\alpha(m_1^\infty v)g_1(v)\,dv\geq\alpha\!\left(\frac{m_1^\infty}{2}\right)\int_{\frac{1}{2}}^{+\infty}vg_1(v)\,dv$$ and moreover, from the normalisation properties of $g_1$, $$\int_{\frac{1}{2}}^{+\infty}vg_1(v)\,dv=\int_{{\mathbb{R}}_+}vg_1(v)\,dv-\int_0^{\frac{1}{2}}vg_1(v)\,dv\geq 1-\frac{1}{2}\int_0^{\frac{1}{2}}g_1(v)\,dv\geq\frac{1}{2}.$$
In conclusion, we obtain $\rho_1^\infty=1$ and $m_1^\infty=0$, which from produce $$f_1^\infty(v)=\delta(v), \qquad f_2^\infty(v)=0.$$ Hence the quarantine leads, in the long run, to a full recovery of the population $x=1$.
Numerical tests {#sect:numerical}
===============
Parameter $N$ $\lambda$ $\Delta{t}$ $\nu_2$ $\gamma$ $\alpha_\dagger$
----------- -------- ----------- ------------- --------- ---------- ------------------
Value $10^6$ $1$ $0.1$ $0.2$ $0.3$ $0.46$
: Parameters kept constant in all numerical tests of Section \[sect:numerical\][]{data-label="tab:constant"}
Parameter Figure \[fig:Tconst\_stable\] Figure \[fig:Tconst\_unstable\] Figure \[fig:Tvar\_nu1=nu2\_mu=10\] Figure \[fig:Tvar\_nu1=nu2\_mu=1\] Figure \[fig:Tvar\_nu1=0\_mu=10\] Figure \[fig:Tvar\_nu1=0\_mu=1\]
----------- ------------------------------- --------------------------------- ------------------------------------- ------------------------------------ ----------------------------------- ----------------------------------
$\mu$ $1$ $1$ $10$ $1$ $10$ $1$
$\alpha$ $0.8$ $0.2$ $0.8\left(1-e^{-v}\right)$ $0.8\left(1-e^{-v}\right)$ $0.8\left(1-e^{-v}\right)$ $0.8\left(1-e^{-v}\right)$
$\beta$ $0.4$ $0.4$ $0.4 e^{-v}$ $0.4 e^{-v}$ $0.4 e^{-v}$ $0.4 e^{-v}$
$\nu_1$ $0$ $0$ $0.2$ $0.2$ $0$ $0$
: Parameters varying from test to test of Section \[sect:numerical\][]{data-label="tab:variable"}
We show now some numerical solutions to model - with interaction rules , , which confirm the findings of Sections \[sect:alpha\_beta.const\], \[sect:alpha\_beta.var\] and allow us to explore also cases not explicitly covered by the previous theoretical study.
To solve the kinetic equations - numerically we use a modified version of the Nanbu-Babovski Monte Carlo algorithm [@bobylev2000PRE; @pareschi2001ESAIMP; @pareschi2013BOOK], see Algorithm \[alg:nanbu\] in Appendix \[app:nanbu\], which includes the transfer of agents from one label to another. Algorithm \[alg:nanbu\] is based on a direct implementation of the time discrete stochastic microscopic processes with $\Theta,\,\Xi$ distributed according to , $J_t$ conditionally distributed according to and $V_t'$ defined like either or depending on the population label. The algorithm consists of four main blocks:
lines \[line:int1.start\]–\[line:int1.end\] implement the microscopic interactions in $x=1$;
lines \[line:int2.start\]–\[line:int2.end\] implement the microscopic rule in $x=2$;
lines \[line:jump1.start\]–\[line:jump1.end\] implement the jumps from $x=1$ to $x=2$ with transition probabilities ;
lines \[line:jump2.start\]–\[line:jump2.end\] implement the jumps from $x=2$ to $x=1$ with transition probabilities .
In Table \[tab:constant\] we list the parameters of the algorithm and of the model that we keep fixed in all numerical tests. In Table \[tab:variable\] we list instead those that we vary from test to test. In all numerical tests we prescribe as initial conditions: $$f_{1,0}(v)=\mathbb{1}_{[0,\,1]}(v), \qquad f_{2,0}(v)=0.$$ Hence, $f_1^0$ is the uniform distribution in $[0,\,1]$ (from which we sample initially the particles with label $x=1$ in Algorithm \[alg:nanbu\]) while no agents are quarantined at $t=0$.
Figures \[fig:Tconst\_stable\], \[fig:Tconst\_unstable\] refer to the case of constant transition probabilities discussed in Section \[sect:alpha\_beta.const\]. In particular, in Figure \[fig:Tconst\_stable\] where $\alpha>\alpha_\dagger$ we recover both the density trends predicted by and the decay to zero of the mean infectivenesses. Conversely, in Figure \[fig:Tconst\_unstable\] where $\alpha<\alpha_\dagger$ we see that the density trends are still reproduced at the particle level but this time the mean infectivenesses quickly blow to infinity as predicted by the qualitative analysis.
Figures from \[fig:Tvar\_nu1=nu2\_mu=10\] to \[fig:Tvar\_nu1=0\_mu=1\] refer instead to the case of variable transition probabilities discussed in Section \[sect:alpha\_beta.var\]. Specifically, we set $$\alpha(v)\propto 1-e^{-v}, \qquad \beta(v)\propto e^{-v},$$ which are respectively a monotonically increasing and a monotonically decreasing function with $\alpha(0)=0$ and $\beta(0)>0$. This way we reproduce exactly the conditions of the qualitative analysis of Section \[sect:alpha\_beta.var\].
In Figures \[fig:Tvar\_nu1=nu2\_mu=10\], \[fig:Tvar\_nu1=nu2\_mu=1\] we set $\nu_1=\nu_2$, which makes the mean infectiveness of the undiagnosed agents ($x=1$) conserved in time. Furthermore, in Figure \[fig:Tvar\_nu1=nu2\_mu=10\] we consider the regime $\mu\gg\lambda$, which produces a label switching-driven hydrodynamic evolution of the densities and mean infectivenesses based on a local-in-time equilibrium of the interactions. The Monte Carlo numerical solution confirms the theoretical predictions obtained in Section \[sect:alpha\_beta.var\] by means of the hydrodynamic splitting - and -: in the long run, $\rho_1\to 1$ and $\rho_2\to 0$ with $m_1,\,m_2\to 0$. Conversely, in Figure \[fig:Tvar\_nu1=nu2\_mu=1\] we consider the regime $\mu=\lambda$, which does not allow for a hydrodynamic splitting of the kinetic equations because the interactions and the label switching take place on the same time scale. Although in Section \[sect:alpha\_beta.var\] we have not explored this case, from the numerical results we observe that the qualitative trends of both the densities and the mean infectivenesses are very similar to those obtained for $\mu\gg\lambda$. In particular, up to a slower rate of convergence in time, the asymptotic states are the same.
\
In Figures \[fig:Tvar\_nu1=0\_mu=10\], \[fig:Tvar\_nu1=0\_mu=1\] we finally examine the case $\nu_1=0$ in the frame of variable transition probabilities, that we have not investigated in Section \[sect:alpha\_beta.var\]. This corresponds to infection dynamics such that an individual may only get more infected after coming into contact with another infected individual. In Figure \[fig:Tvar\_nu1=0\_mu=10\] we illustrate the case $\mu\gg\lambda$: since interactions are much quicker than label switches, the mean infectiveness of the undiagnosed individuals ($x=1$) tends to blow up locally in time without any possibility for the quarantine to affect this rapid growth. As a result, in the long run the whole population tends to be quarantined (Figure \[fig:Tvar\_nu1=0\_mu=10\]a), however with no effect on the control of the infection. Indeed, also the mean infectiveness of the quarantined individual rapidly blows up (Figure \[fig:Tvar\_nu1=0\_mu=10\]b). Finally, in Figure \[fig:Tvar\_nu1=0\_mu=1\] we illustrate the regime $\mu=\lambda$: now, similarly to the case of Figure \[fig:Tconst\_stable\], the quarantine can successfully control the spreading of the infection because the contagion among the individuals take place on the same time scale of the label switches. Nevertheless, due to the infection-dependent transition probabilities, the infection is not completely eradicated in time but becomes endemic. In particular, Figure \[fig:Tvar\_nu1=0\_mu=1\]a shows that in the long run a fixed percentage of individuals is systematically quarantined and Figure \[fig:Tvar\_nu1=0\_mu=1\]b confirms that such a quarantine is not fictitious like in the numerical test illustrated in Figure \[fig:Tconst\_stable\]. Indeed, the quarantined population is not fully healthy like in Figure \[fig:Tconst\_stable\]b because its asymptotic mean infectiveness is strictly positive. Figure \[fig:Tvar\_nu1=0\_mu=1\]c show the normalised histograms which approximate numerically the asymptotic distributions $f_1^\infty,\,f_2^\infty$ and which illustrate these conclusions with an even greater statistical detail.
Conclusions
===========
In this paper, we have shown how Boltzmann-type kinetic models with label switching can be derived from stochastic microscopic dynamics accounting for the superposition of conservative interactions and non-conservative state-dependent relabelling of the agents. Remarkably, such a derivation has yielded straightforwardly a simple and efficient Monte Carlo particle scheme for the numerical approximation of the resulting non-conservative kinetic equations.
For prototypical non-conservative dynamics, such as death and birth processes, we have been able to characterise explicitly both the transient and the equilibrium (“Maxwellian”) kinetic distributions in the special regime of sufficiently small parameters (quasi-invariant regime) by means of Fokker-Planck asymptotics and self-similar solutions.
Moreover, we have applied our non-conservative kinetic framework to the construction of a new model of the contagion of infectious diseases with quarantine, which describes from a statistical point of view the interplay among:
the microscopic dynamics of contact and contagion among the individuals of a community;
the isolation of individuals diagnosed as infected;
the reintroduction in the community of quarantined individuals diagnosed as recovered.
In particular, the isolation and the reintroduction are regarded as label switches modelled on an infectiveness-dependent probabilistic basis. Thanks to its kinetic structure, this model depends on a relatively small number of parameters. Yet, it shows a quite rich variety of trends, which suggest clearly the impact of the microscopic features of the system on either the success or the failure of the quarantine as a control strategy of the global spreading of the infection. More importantly, the kinetic structure of the model has allowed us to address analytically several significant regimes by taking advantage of powerful methods of the kinetic theory, such as the hydrodynamic limit. This way, we have obtained a precise characterisation of the role of the microscopic parameters in the emergence of either global trend.
Acknowledgements {#acknowledgements .unnumbered}
================
This research was partially supported by the Italian Ministry for Education, University and Research (MIUR) through the “Dipartimenti di Eccellenza” Programme (2018-2022), Department of Mathematical Sciences “G. L. Lagrange”, Politecnico di Torino (CUP: E11G18000350001) and through the PRIN 2017 project (No. 2017KKJP4X) “Innovative numerical methods for evolutionary partial differential equations and applications”.
NL acknowledges support from “Compagnia di San Paolo” (Torino, Italy)
Both authors are members of GNFM (Gruppo Nazionale per la Fisica Matematica) of INdAM (Istituto Nazionale di Alta Matematica), Italy.
Numerical algorithm {#app:nanbu}
===================
Fix $\Delta{t}\leq\min\{\frac{1}{\lambda},\,\frac{1}{\mu}\}$
[^1]: `[email protected]`
[^2]: `[email protected]`
[^3]: Here and henceforth we use the notation $a\lesssim b$ to mean that there exists a constant $C>0$, independent of $\epsilon$ and whose specific value is unimportant, such that $a\leq Cb$.
| {
"pile_set_name": "ArXiv"
} |
---
address:
- 'Institute of Physics, Technische Universität Dortmund, D-44221 Dortmund, Germany'
- 'Institute of Statistics, Ruhr-Universität Bochum, D-44780 Bochum, Germany'
author:
- Sabrina Einecke
- Katharina Proksch
- Nicolai Bissantz
- Fabian Clevermann
- Wolfgang Rhode
bibliography:
- 'astrobib.bib'
title: 'Uncertainty limits on solutions of inverse problems over multiple orders of magnitude using bootstrap methods: An astroparticle physics example'
---
Uncertainty limits ,Bootstrap ,Unfolding ,Inverse problem
| {
"pile_set_name": "ArXiv"
} |
---
author:
- |
Laurent Bruneau\
\
Department of Mathematical Methods in Physics\
Warsaw University\
Hoża 74, 00-682, Warszawa, Poland
title: The ground state problem for a quantum Hamiltonian model describing friction
---
**Abstract:** In this paper, we consider the quantum version of the hamiltonian model describing friction introduced in [@BDB]. This model consists of a particle which interacts with a bosonic reservoir representing a homogeneous medium through which the particle moves. We show that if the particle is confined, then the Hamiltonian admits a ground state if and only if a suitable infrared condition is satisfied. The latter is violated in the case of linear friction, but satisfied when the friction force is proportional to a higher power of the particle speed.
Introduction {#sec:intro}
============
In [@BDB] we introduced a classical Hamiltonian model of a particle moving through a homogeneous dissipative medium at zero temperature in such a way that the particle experiences an effective *linear* friction force proportional to its velocity. The medium consists at each point in the space of a vibration field with which the particle exchanges energy and momentum. More precisely the Hamiltonian is given by $$\begin{aligned}
\label{ham:class}
H(q,p,\phi,\pi) & = & \frac{p^2}{2}+V(q)+ \frac{1}{2}\int_{\R^d}dx \int_{\R^n} dy\, c^2|\nabla_y \phi(x,y)|^2+|\pi(x,y)|^2 \nonumber\\
& & \qquad \qquad +\int_{\R^d}dx \int_{\R^n}dy\, \rho_1(x-q)\rho_2(y)\phi(x,y),\end{aligned}$$ where $V$ is an external potential, $c$ represents the speed of the wave propagation in the “membranes” and the functions $\rho_1$ and $\rho_2$ determine the coupling between the particle and the field and are smooth radial functions with compact support.
We studied the asymptotic behaviour of the particle motion for two categories of potentials: linear ones (which means constant external force) and confining ones. We proved that under suitable assumptions (on the initial conditions), for $c$ sufficiently large and, most importantly, $n=3,$ the particle behaves asymptotically as if its motion was governed by the effective equation $$\ddot{q}(t)+\gamma \dot{q}(t)= -\nabla V(q(t)),$$ where the friction coefficient $\gamma$ is non negative and is explicit in terms of the parameters of the model: $$\label{eq:frictioncoef}
\gamma:= \frac{\pi}{c^3} |\hat{\rho}_2(0)|^2\int_{\R^n}d\xi \int_{\R^{d-1}}d\eta\, |\hat{\rho}_1(|\xi|,\eta)|^2.$$ If $V=-F\cdot q,$ which means that we apply a constant external force $F$ to the particle, then this particle reaches exponentially fast (with rate $\gamma$) an asymptotic velocity $v(F)=\frac{F}{\gamma}$ which is proportional to the applied force (at least for small forces). This is, in particular, at the origin of Ohm’s law. On the other hand, if $V$ is confining, the particle stops at one of the critical points of the potential, the convergence rate being still exponential (but with rate $\frac{\gamma}{2}$ as expected from the effective equation).
In [@BDB] we mostly concentrated on linear friction. This is why the $n=3$ assumption was required. However, for other values of $n \, (>3)$, our model still describes friction. Indeed, the reaction force of the environment on a particle moving with velocity $v$ takes the form $-\gamma |v|^{n-3}v$ (at least for small $v$ and where $\gamma$ is defined in (\[eq:frictioncoef\])). One can therefore see that we have linear friction when $n=3,$ and otherwise a friction force which is proportional to some other power of the velocity of the particle.
Such models, where a small system interacts with a *large* environment, are called open systems. The reason for studying those models is usually to have a Hamiltonian description of dissipative phenomena. There exist several mechanisms leading to dissipation. Among them, two important, and very different, mechanisms are radiation damping and friction (which can be linear or not). As far as radiation damping is concerned, there exist many models, which are more or less related to electromagnetism. One example is the “classical Nelson model” $$\begin{aligned}
H_{\textrm{nels}}(q,p,\phi,\pi) & = & \frac{p^2}{2}+V(q)+ \frac{1}{2}\int_{\R^d}dx \left( |\nabla
\phi(x)|^2+|\pi(x)|^2\right)\\
& & \qquad \qquad \qquad \qquad \qquad \qquad +\int_{\R^d}dx \rho(x-q)\phi(x),\end{aligned}$$ which has been studied in [@KKS] (except for the kinetic energy of the particle which was $\sqrt{p^2+1}$ instead of $\frac{p^2}{2}$). This model describes a particle interacting with a scalar radiation field, and exhibits radiation damping. Concerning friction, although there exist various Hamiltonian models in the literature, ours is the only one we are aware of that describes the friction produced by the motion of the particle through a homogeneous medium. In particular, the coupling between the medium and the particle is translationally invariant and hence non-linear in the particle variable. This means that no dipole approximation is assumed and is essential for a correct treatment of a constant external force field. Despite the formal similarity between our model and the classical Nelson model, we want to stress once again that they describe physically totally different phenomena. This is reflected in mathematical differences that will become apparent below.
Our goal in this paper is to begin the study of the quantum version of the model (\[ham:class\]). Since the speed of the wave propagation will not play any role in our paper, we take it equal to 1. The quantum Hamiltonian can then be written as follows $$\begin{aligned}
H & = & \left( -\Delta +V\right) \otimes \one + \one \otimes \int dx\, dk\, \omega(x,k)a^*(x,k)a(x,k)\\
& & \qquad \qquad \qquad \qquad +\int dx\, dk\, \frac{\rho_1(x-Q)\hat{\rho}_2(k)}{\sqrt{2\omega(x,k)}}\otimes a^*(x,k)+ h.c.,\end{aligned}$$ where $a$ and $a^*$ are the usual annihilation and creation operators on the bosonic Fock space $\mathcal{F}(L^2(\R^{d+n},dx\, dk)),$ and $\omega(x,k)=|k|$ is the bosons dispersion relation. In this paper, we start with the study of confining potentials, which are less difficult. More precisely, we deal with the question of existence of a ground state, which is essential before studying questions such as scattering theory or return to equilibrium for example. If a Hamiltonian is bounded from below, we say that it admits a ground state if the infimum of its spectrum is an eigenvalue. We call ground state energy this infimum and ground state any corresponding eigenvector if it exists. We will prove that such a ground state exists provided the following *infrared condition* is satisfied: $$\int_{\R^n} dk\, \frac{|\hat{\rho}_2(k)|^2}{|k|^3} < +\infty.$$ This condition will be used to control the number of bosons which have low energy (soft bosons). Let us suppose that $\hat{\rho}_2(0)\neq 0.$ Indeed, this is the only interesting case since the friction coefficient $\gamma$ vanishes together with $\hat{\rho}_2(0)$ (see (\[eq:frictioncoef\])). Then, there exists a ground state if the infrared condition is satisfied (Theorem \[thm:gdstate\]). One can see that this condition is fulfilled when the friction is non-linear. On the other hand, for linear friction, there is generically no ground state (Proposition \[prop:noground\]). Thus, we have a class of models, depending on a parameter $n,$ describing friction phenomena, linear or proportional to a power of the velocity of the particle, for which we are able to say wether they admit a ground state or not.
As in the classical case, our model looks very similar to the Nelson model, and more generally to the Pauli-Fierz models (following the terminology of [@DG]), in which a (small) quantum system interacts with a scalar bosonic field, although they lead to very different dissipative phenomena. We will recall some basic facts about Fock spaces and describe the quantum version of the model in Sect. \[sec:model\], while, in Sect. \[sec:results\], we state our main results.
To prove the existence of a ground sate, we follw the standard strategy: we first prove the result for coupling to a massive field and then we let the mass tend to zero. We study the massive case in Sect. \[sec:massifs\] along the lines of [@BFS1]-[@BFS2]-[@GJ]: we first constrain the model to a finite box (|x|<L) and then control the error terms as $L$ goes to infinity. This has to be done with care since in the interaction term, the norm of $\rho_1(x-Q)$ as an operator on $L^2(\R^d)$ does not decrease with $x.$ In order to control this problem, we will need to use the exponential decay of the spectral projectors in the $Q$ variable. Furthermore, a second difficulty that arises is the following. The cutoff in space is equivalent to discretizing the model in the momentum variable. It is therefore equivalent to study the “cutoff” Hamiltonians and Hamiltonians “discretized” in momentum. In the case of the models for radiation damping, those discrete Hamiltonians are quite easy to study. Indeed, the free discrete Hamiltonian has then purely discrete spectrum (the energy of the bosons can only take a discrete number of values), and, because the interaction is relatively bounded with respect to it, so has the full discrete Hamiltonian. Now, this will not be the case in our model because the energy of the bosons only depends on $k$ and not on the discretized momentum which comes from the fact that the dispersion relation $\omega$ only depends on $k.$ This is the main mathematical difference with the models for radiation damping. One then has to control the momentum of the bosons in the “$x$-direction”. A careful study of the discrete Hamiltonians will therefore be needed (Sect. \[ssec:disc\]). In Sect. \[sec:nonmassif\] we first prove Theorem \[thm:gdstate\]. To do this, we adapt the proof of [@G] to our model. In particular, we will need to control the momentum in “$x$” of the bosons. Moreover, we will also have to take into account that the norm of $\rho_1(x-Q)$ as an operator on $L^2(\R^d)$ does not depend on $x$ and is therefore not square integrable with respect to this variable. Once again, this will require to use the exponential decay of the spectral projectors in the $Q$ variable. Finally, we also prove Proposition \[prop:noground\].
We said that one of the main assumptions for the existence of a ground state in the massless case was the so-called *infrared condition* (see Sect. \[sec:results\]). In Sect. \[sec:infrarouge\], we will present some classical interpretation of this infrared condition.
Description of the model {#sec:model}
========================
Fock spaces {#ssec:fock}
-----------
In this section, we give a rather general (and brief) presentation of the different objects we will use in this paper. It will in particular allow us to fix notations. The reader will find a more detailed description in *e.g.* [@DG]-[@RS2].
Let $\mathfrak{h}$ be a complex Hilbert space, which is often called the $1-$particle space. Given $f,g$ in $\mathfrak{h},$ we denote by $\langle f;g\rangle$ their scalar product. It is chosen to be antilinear in the first variable and linear in the second variable. For $m \in \N,$ we define the $m-$particle sector as the $m$-fold symmetric tensor product of $\mathfrak{h}:\ \mathfrak{h}_m= \otimes^m_s \mathfrak{h},$ with $\mathfrak{h}_0= \C.$ We then define the Fock space over $\mathfrak{h}$ to be the direct sum $$\mathcal{F}(\mathfrak{h}):= \oplus_{m=0}^{\infty} \mathfrak{h}_m.$$ We will denote by $\Omega=(1,0,\dots)$ the vacuum vector and by $a^*$ and $a$ the usual creation/annihilation operators on $\mathcal{F}(\mathfrak{h})$ ([@RS2], Chapter X.7).
In the case where $\mathfrak{h}= L^2(\R^{\nu}),$ we can rewrite those operators in the following way: $$a^*(h)=\int_{\R^{\nu}} dk\, h(k)a^*(k), \quad a(h)=\int_{\R^{\nu}} dk\, \bar{h}(k)a(k),$$ where $a^*(k)$ and $a(k)$ are the distributional creation and annihilation operators. They satisfy the usual canonical commutation relations: $$\label{ccr2}
[a(k),a^*(k')] = \delta(k-k'), \quad \left[a(k),a(k') \right] = [a^*(k),a^*(k')]=0.$$ Finally, given an operator $b$ on $\mathfrak{h},$ we define: $$\begin{aligned}
\label{2quantif}
\mbox{d}\Gamma(b)&:& \mathcal{F}(\mathfrak{h}) \to \mathcal{F}(\mathfrak{h}) \nonumber \\
\mbox{d}\Gamma(b)|_{\mathfrak{h}_m}&:=& \sum_{j=1}^m \underbrace{\one\otimes \dots \otimes \one}_{j-1} \otimes b \otimes
\underbrace{\one \otimes \dots \otimes \one}_{m-j},\end{aligned}$$ and $$\label{2quantif2}
\Gamma(b):\ \mathcal{F}(\mathfrak{h}) \to \mathcal{F}(\mathfrak{h}), \quad
\Gamma(b)|_{\mathfrak{h}_m}:= \ \underbrace{b\otimes \dots \otimes b}_m.$$ The operator $\mbox{d}\Gamma(b)$ is called the second quantization of the operator $b.$ Note that when $b$ is selfadjoint, we have the following relation $$e^{i\textrm{d}\Gamma(b)}=\Gamma(e^{ib}).$$ An operator which plays an important role is the number operator $N:=\mbox{d}\Gamma(\one).$
Description of the model {#ssec:modq}
------------------------
We can now introduce the quantum version of the model introduced in Sect. \[sec:intro\]. The dynamics of the particle is given by the Schrödinger operator $H_p=-\Delta +V$ on $L^2(\R^d).$ Troughout this paper we will only consider confining potentials, so that $H_p$ has a compact resolvent and purely discrete spectrum.
The Hilbert space for the environment will be the bosonic Fock space over $L^2(\R^{d+n},dx\, dk).$ In what follows, we will just write $$\label{def:fieldspace}
\mathcal{F}:= \mathcal{F} (L^2(\R^{d+n}, dx\, dk)).$$ The Hamiltonian of the field is given by $$\label{hamq:champ}
H_f:= \mbox{d}\Gamma(\omega),$$ where $\omega$ is the multiplication operator on $L^2(\R^{d+n},dx\, dk)$ by the function $$\label{reldispers}
\omega: (x,k)\in \R^d\times \R^n\rightarrow \omega(x,k)=|k| \in [0,+\infty[.$$ The function $\omega$ depends only on $k,$ so we will write $\omega(k)$ for $\omega(x,k).$ It is well known that one can rewrite $H_f$ using the creation and annihilation operators as follows: $$\label{hamq:champ2}
H_f= \int_{\R^{d+n}} dx \, dk \, \omega(k) a^*(x,k)a(x,k).$$
We can now describe the full system. The Hilbert space is the tensor product of the particle space and of the environment one, namely: $$\label{def:space}
\mathcal{H}:= L^2(\R^d)\otimes \mathcal{F},$$ and the free Hamiltonian ([*i*.e.]{} without interaction) is given by: $$\label{hamq:libre}
H_0:=H_p \otimes \one + \one \otimes H_f.$$
The interaction term is given by $$\begin{aligned}
\label{hamq:interaction}
H_I & := & \int dx\, dk\,\, \rho_1(x-Q)\frac{\hat{\rho}_2(k)}{\sqrt{2\omega(k)}}\otimes a^*(x,k) \nonumber \\
& & \qquad \qquad \qquad + \rho_1(x-Q)\frac{\bar{\hat{\rho}}_2(k)}{\sqrt{2\omega(k)}}\otimes a(x,k),\end{aligned}$$ where $\rho_1$ and $\rho_2$ are two smooth functions with compact support and spherical symetry, and $\rho_1(x-Q)$ is the multiplication operator on $L^2(\R^d)$ by the function $\rho_1(x-\cdot).$
Finally, the full Hamiltonian of the interacting system is therefore $$\label{hamq:total}
H:= H_0+H_I.$$
Main results {#sec:results}
============
Selfadjointness {#ssec:aa}
---------------
From now, we will suppose that $n\geq 3.$ We first give the precise condition we impose on the potential $V:$
> \(C) $V \in L^2_{loc}(\R^d), \lim_{|q|\to \infty} V(q)=+\infty.$
This hypothesis ensures that $H_p$ is well defined and is selfadjoint on $\mathcal{D}(H_p)=
\{\psi \in L^2(\R^d)| H_p\psi \in L^2(\R^d)\}$ ([@RS2], Theorem X.28). We also know that $H_f$ is selfadjoint on its domain $\mathcal{D}(H_f)$ ([@RS1], Chapter VIII.10). One then easily proves that $H_0$ is essentially selfadjoint on $\mathcal{D}(H_p)\otimes \mathcal{D}(H_f)$ ([@RS1], Chapter VIII.10). We now have the following result
\[prop:aa\] Suppose that $n\geq 3,$ and $V$ satisfies condition (C). Then $H$ is selfadjoint on $\mathcal{D}(H)=\mathcal{D}(H_0)$. Moreover, $H$ is essentially selfadjoint on any core for $H_0,$ and it is bouded from below.
This is in the standard way a consequence of the Kato-Rellich theorem ([@RS2], Theorem X.12). The only ingredient needed is that $H_I$ is infinitesimally $H_0$-bounded, which follows from the following lemma.
\[lem:a-estimates\] Under the hypothesis of Proposition \[prop:aa\], for all $\Psi \in \mathcal{D}(H_0)$, we have: $$(i) \qquad \| \int dx\, dk \frac{\hat{\rho}_2(k)}{\sqrt{\omega (k)}}\rho_1(x-Q)\otimes a(x,k) \Psi
\|^2_{\mathcal{H}} \qquad \qquad \qquad \qquad \qquad$$ $$\qquad \qquad \qquad \qquad \leq \Big[ \int dx\, dk |\rho_1(x)|^2 \frac{|\hat{\rho}_2(k)|^2}{\omega(k)^2}
\Big] \|(\one \otimes H_f^{\frac{1}{2}})\Psi \|^2_{\mathcal{H}}.$$
$$(ii) \qquad \| \int dx\, dk \frac{\hat{\rho}_2(k)}{\sqrt{\omega (k)}}\rho_1(x-Q)\otimes a^*(x,k) \Psi
\|^2_{\mathcal{H}} \qquad \qquad \qquad \qquad \qquad$$ $$\qquad \qquad \qquad \qquad \leq \Big[ \int dx dk |\rho_1(x)|^2 \frac{|\hat{\rho}_2(k)|^2}{\omega(k)^2} \Big]
\|(\one \otimes H_f^{\frac{1}{2}})\Psi \|^2_{\mathcal{H}}$$ $$\qquad \qquad \qquad \qquad \qquad \qquad \qquad +\Big[ \int dx\, dk |\rho_1(x)|^2 \frac{|\hat{\rho}_2(k)|^2}{\omega(k)} \Big] \| \Psi\|^2_{\mathcal{H}}.$$
\(i) Such kind of estimates are well known [@A1]-[@BFS1]-[@DJ] and are sometimes called $N_{\tau}-$ estimates.
\(ii) The $n \geq 3$ hypothesis ensures that the integrals on the right-hand side of both inequalities converge.
[**Proof of Lemma \[lem:a-estimates\]:**]{} We use the fact that $\mathcal{H}$ is isomorphic to $L^2(\R^d, dq, \mathcal{F}).$ We then have: $$\begin{aligned}
& & \| \int dx\, dk \frac{\hat{\rho}_2(k)}{\sqrt{\omega (k)}}\rho_1(x-Q)\otimes a(x,k)
\Psi \|^2_{\mathcal{H}}\\
& = & \int_{\R^d} dq \| \int dx\, dk \frac{\hat{\rho}_2(k)}{\sqrt
{\omega (k)}}\rho_1(x-q) a(x,k)\Psi(q)\|^2_{\mathcal{F}}\\
& = & \int_{\R^d} dq \| a(g_q)\Psi(q)\|^2_{\mathcal{F}},\end{aligned}$$ where $g_q$ is the function $
g_q(x,k)= \frac{\hat{\rho}_2(k)}{\sqrt{\omega (k)}}\rho_1(x-q).
$ For all $q\in \R^d$, we have (see [@BFS1]): $$\begin{aligned}
\|a(g_q)\Psi (q)\|^2_{\mathcal{F}} & = & \| \int dx\, dk\, g_q(x,k) a(x,k)\Psi (q)\|^2_{\mathcal{F}} \\
& \leq & \left( \int dx\, dk\, \frac{|g_q(x,k)|}{\sqrt{\omega(k)}} \sqrt{\omega(k)} \|a(x,k) \Psi (q)\|
\right)^2 \\
& \leq & \Big[ \int dx\, dk \frac{|g_q(x,k)|^2}{\omega(k)} \Big] \int dx\, dk\, \omega(k) \|a(x,k) \Psi
(q)\|^2. \end{aligned}$$ But $$\int dx\, dk\, \omega(k) \|a(x,k) \Psi (q)\|^2= \langle \Psi (q); H_f \Psi (q)\rangle,$$ so, $$\begin{aligned}
\|a(g_q)\Psi (q)\|^2_{\mathcal{F}} & \leq & \Big[ \int dx\, dk \frac{|g_q(x,k)|^2}{\omega(k)} \Big]
\|H^{1/2}_f \Psi (q)\|^2_{\mathcal{F}}\\
& \leq & \Big[ \int dx\, dk |\rho_1(x)|^2 \frac{|\hat{\rho}_2(k)|^2}
{\omega(k)^2} \Big] \|H^{1/2}_f \Psi (q)\|^2_{\mathcal{F}}.\end{aligned}$$ Finally, $$\begin{aligned}
& & \| \int dx\, dk \frac{\hat{\rho}_2(k)}{\sqrt{\omega (k)}}\rho_1(x-Q)\otimes a(x,k)
\Psi \|^2_{\mathcal{H}}\\
& \leq & \Big[ \int dx\, dk |\rho_1(x)|^2 \frac{|\hat{\rho}_2(k)|^2}
{\omega(k)^2} \Big] \int_{\R^d} dy \|H^{1/2}_f \Psi (y)\|^2_{\mathcal{F}}\\
& \leq & \Big[ \int dx\, dk |\rho_1(x)|^2 \frac{|\hat{\rho}_2(k)|^2}
{\omega(k)^2} \Big] \| (\one \otimes H^{1/2}_f)\Psi \|^2_{\mathcal{H}},\end{aligned}$$ which proves $(i).$ One proves $(ii)$ in a similar way.
Existence of a ground state {#ssec:gdstate}
---------------------------
Let $E_0$ denote the ground state energy of $H.$ It is well known that one of the main obstacles to the existence of a ground state, in those models where a particle interacts with a field, comes from the so-called *infrared catastroph*, which is due to the behaviour of $\omega(k)$ for small $k$ and in particular to the fact that $\omega(0)=0.$ We will then need the following condition on the coupling:
> (IR) $\int_{\R^n} dk\, \frac{|\hat{\rho}_2(k)|^2}{\omega(k)^3}<+\infty.$
This is what we call the infrared condition. We prove the following theorem, which is the main result of our paper:
\[thm:gdstate\] Suppose $n\geq 3,$ $V$ satisfies hypothesis (C), and $\hat{\rho}_2$ satisfies (IR). Then $H$ has a ground state.
As we said in the introduction, this (IR) condition is satisfied when the friction is non-linear but not if it is linear. On the other way, in the case of the Nelson model, the same kind of condition is necessary and sufficient to have a ground state [@G]-[@LMS]. It is then reasonable to think this is also true for our model. Indeed, we will prove that if the infrared condition is violated, then there is no ground state but provided the following additional condition is satisfied $$\hat{\rho}_1(0)\neq 0,$$ which means that the total charge of the particle does not vanish. More precisely, we prove the following result:
\[prop:noground\] Suppose $n\geq 3,$ $V$ satisfies hypothesis (C), $\hat{\rho}_2$ does not satisfy (IR) and $\hat{\rho}_1(0)\neq 0,$ then $H$ has no ground state.
To prove Theorem \[thm:gdstate\], we will need to study some “intermediate” models, and in particular to consider *massive* bosons and to “discretize” space. The term *massive* means that, instead of $\omega(k),$ we will consider a function $\omega_m(k)$ satisfying $$(H_{\omega})\left\{ \begin{array}{l} \nabla \omega_m \in L^{\infty}(\R^n), \\ \lim_{|k|\to \infty} \omega_m(k)=+\infty,
\\ \inf \omega_m(k)=m>0. \end{array} \right.$$ Our proof will use different methods developed in the literature [@BFS1]-[@BFS2]-[@DG]-[@G]-[@GJ].
Finally, we would like to emphasize that all the Hamiltonians we will deal with have the same structure as (\[hamq:total\]) and so, a similar result to the one of Proposition \[prop:aa\] is available for each of them.
Ground state for massive bosons {#sec:massifs}
===============================
Our goal in this section is to prove a first result similar to Theorem \[thm:gdstate\] but in the case of massive bosons (Theorem \[thm:massif\], Sect. \[ssec:massif\]). We use the same approach as in [@GJ] and [@BFS1]. The idea is first to consider a finite box ($|x|<L$) and then to control the remaining part as $L$ goes to infinity. We will see, in Sect. \[ssec:massif\], that the “cutoff” model so obtained can be written in the form (\[hdisc\]). We therefore first study models of this latter type (Theorem \[thm:disc\]).
Discrete models {#ssec:disc}
---------------
### Description {#sssec:description}
We consider Hamiltonians of te form $$\begin{aligned}
\label{hdisc}
H^{\d} & := & H_p \otimes \one + \one \otimes \sum_{l \in \Z^d} \int_{\R^n} dk\, \omega_m(k) a^*_l(k)a_l(k) \nonumber \\
& & \qquad \qquad \qquad + \sum_{l \in \Z^d} \int_{\R^n} dk\, (\beta_l(k) \otimes a^*_l(k) + \bar{\beta}_l(k) \otimes a_l(k))\nonumber\\
& = & H_0^{\d} + W^{\d}, \end{aligned}$$ on the space $$\label{esp-disc1}
\mathcal{H}^{\d}:= L^2(\R^d)\otimes \mathcal{F}\left(l^2(\Z^d)\otimes L^2(\R^n)\right),$$ and where the $\beta_l(k)$ satisfy\
> $(C_{\beta}) \quad \beta_l(k)= \zeta_l \frac{\hat{\rho}_2(k)}{\sqrt{2\omega_m(k)}}$ where $\zeta_l$ is a multiplication operator on $L^2(\R^d)$ such that $\sup_l \| |l|^s\zeta_l \| < +\infty$ for all $s>0$,\
$a_l(k)$ and $a^*_l(k)$ are the annihilation and creation operators on the space $\mathcal{F}\left(l^2(\Z^d)\otimes L^2(\R^n)\right),$ and for $l=(l_1,\dots, l_d)\in \Z^d,$ $|l|:=\sup_i |l_i|.$
We would like to stress that one can consider the Hamiltonians of the form (\[hdisc\]) as models similar to ours, but with only a discrete set of “membranes” (situated at each $l \in \Z^d$) rather than a continuous one.
Now, let $E^{\d}_0$ denote the ground state energy for $H^{\d}.$ We will prove the following:
\[thm:disc\] $\sigma_{ess}(H^{\d}) \subset \left[ E^{\d}_0+m, +\infty \right[.$ In particular, $H^{\d}$ has a ground state $\phi^{\d}_0$.
### Cutoff models {#sssec:disc-cutoff}
In the following, $M$ will be a non negative number. On $\mathcal{H}^{\d},$ we define $$\begin{aligned}
\label{hdiscl1}
H^{\d}(M) & := & H_0^{\d} + \sum_{|l|\leq M} \int_{\R^n} dk\, (\beta_l(k) \otimes a^*_l(k) + \bar{\beta}_l(k) \otimes
a_l(k))\\
& = & H_0^{\d} + W^{\d}(M). \nonumber\end{aligned}$$ We also define $$\begin{aligned}
\label{hdiscl2}
\tilde{H}^{\d}(M) & := & H_p \otimes \one + \one \otimes \sum_{|l|\leq M} \int_{\R^n} dk\, \omega_m(k) a^*_l(k)a_l(k)+W^{\d}(M)\quad \\
& = & \tilde{H}^{\d}_0(M) + W^{\d}(M), \nonumber\end{aligned}$$ as an operator on the space $$\label{esp-disc2}
\mathcal{H}^{\d}_M:= L^2(\R^d)\otimes \mathcal{F}\left(l^2(\Lambda_M)\otimes L^2(\R^n)\right),$$ where $\Lambda_M=\{l \in \Z^d, |l|\leq M\},$ so that $l^2(\Lambda_M)$ is a finite dimensional space. Let $E^{\d}_0(M)$ (resp. $\tilde{E}^{\d}_0(M)$) be the ground state energy for $H^{\d}(M)$ (resp. $\tilde{H}^{\d}(M)$). Our goal is to get informations on $H^{\d}$ from the ones we will have on $H^{\d}(M)$ (taking the limit $M \to +\infty$). Thus, we first prove a result similar to Theorem \[thm:disc\], but for $H^{\d}(M).$
\[prop:sp-disc\] $\sigma_{ess}(H^{\d}(M)) \subset \left[ E^{\d}_0(M)+m, +\infty \right[.$ In particular, $H^{\d}(M)$ has a ground state $\phi^{\d}_0(M).$ Moreover, $E^{\d}_0(M)=\tilde{E}^{\d}_0(M).$
\[lem:sp-disc\] $\sigma_{ess}(\tilde{H}^{\d}(M)) \subset \left[ \tilde{E}^{\d}_0(M)+m, +\infty \right[.$ In particular, $\tilde{H}^{\d}(M)$ has a ground state $\tilde{\phi}^{\d}_0(M).$
[**Proof of Lemma \[lem:sp-disc\]:**]{} The set $\Lambda_M$ is finite. If its cardinal was one, we would have exactly the model studied in [@DG], and the lemma would correspond to their Theorem 4.1. Having finitely many elements does not change anything and the result can be proven the same way.
[**Proof of Proposition \[prop:sp-disc\]:**]{} The proposition follows immediately from the preceding lemma using an identification between $\mathcal{H}^{\d}_M$ and some subspace of $\mathcal{H}^{\d},$ [@GJ]. Indeed, one can write $$l^2(\Z^d)\simeq l^2(\Lambda_M) \oplus l^2(\Lambda_M^c),$$ $\Lambda_M^c$ denoting the complement of $\Lambda_M$ in $\Z^d,$ so one has $$\mathcal{F}\left(l^2(\Z^d) \otimes L^2(\R^n)\right) \simeq \mathcal{F}\left(l^2(\Lambda_M) \otimes L^2(\R^n)\right) \otimes
\mathcal{F}\left(l^2(\Lambda_M^c) \otimes L^2(\R^n)\right).$$ And finally $$\mathcal{H}^{\d} \simeq \mathcal{H}^{\d}_M \otimes \mathcal{F}\left(l^2(\Lambda_M^c) \otimes L^2(\R^n)\right).$$ One can then identify $\mathcal{H}^{\d}_M$ with $\mathcal{H}^{\d}_M \otimes \Omega^c_M$ where $\Omega^c_M$ is the vacuum of $\mathcal{F}\left(l^2(\Lambda_M^c) \otimes L^2(\R^n)\right).$ We can rewrite $\mathcal{H}^{\d}$ as $$\mathcal{H}^{\d} = \bigoplus_{j=0}^{+\infty} \left(\mathcal{H}^{\d}_M \otimes^j_s
\left(l^2(\Lambda_L^c) \otimes L^2(\R^n)\right) \right) = \bigoplus_{j=0}^{+\infty} \mathcal{H}^{(j)}.$$ Actually, we have $$\mathcal{H}^{\d}_M=\mathcal{H}^{(0)} \quad \textrm{and} \quad (\mathcal{H}^{\d}_M)^{\perp}= \bigoplus_{j=1}^{+\infty} \mathcal{H}^{(j)}.$$ One sees that the $\mathcal{H}^{(j)}$ are invariants for $H^{\d}(M).$ But, on $\mathcal{H}^{(j)},$ one has $$\begin{aligned}
H^{\d}(M) & = & \tilde{H}^{\d}(M) \otimes \one + \one \otimes \sum_{|l|>L} \int_{\R^n} dk\, \omega_m(k)
a^*_l(k)a_l(k) \\
& \geq & \tilde{H}^{\d}(M) \otimes \one + mj,\end{aligned}$$ and on $\mathcal{H}^{(0)},$ $$H^{\d}(M) = \tilde{H}^{\d}(M) \otimes \one.$$ Then, we have $$\sigma \left( H^{\d}(M)|_{\mathcal{H}^{\d}_M}\right) =\sigma \left( \tilde{H}^{\d}(M)\right) \quad \textrm{and}
\quad \sigma_{ess} \left( H^{\d}(M)|_{\mathcal{H}^{\d}_M}\right) =\sigma_{ess} \left( \tilde{H}^{\d}(M)\right),$$ and also $$\sigma_{ess} \left( H^{\d}(M)|_{(\mathcal{H}^{\d}_M)^{\perp}}\right) \subset \sigma \left(
H^{\d}(M)|_{(\mathcal{H}^{\d}_M)^{\perp}}\right) \subset \left[ \tilde{E}^{\d}_0(M)+m, +\infty \right[,$$ which ends the proof. Moreover, one can remark that $\phi^{\d}_0(M)=\tilde{\phi}^{\d}_0(M)\otimes \Omega^c_M.$
### Removing the cutoff {#sssec:disc-remove}
We first prove some convergence results as $M$ goes to infinity.
\[prop:st-res-disc\] $H^{\d}(M)$ converges to $H^{\d}$ in the strong resolvent sens.
We have $$H^{\d}-H^{\d}(M) = W^{\d}-W^{\d}(M)
= \sum_{|l|>M} \int_{\R^n} dk\, \beta_l(k)\otimes a^*_l(k)+ \bar{\beta}_l(k)\otimes a_l(k).$$ Let $\psi \in D(H^{\d}_0).$ Using condition $(C_{\beta}),$ one has $$\begin{aligned}
\| \sum_{|l|>M} \int_{\R^n} dk\, \bar{\beta}_l(k)\otimes a_l(k) \psi \| & \leq & \frac{C(s)}{1+M^s}
\| \sum_{|l|>M} \int_{\R^n} dk\, \one \otimes a_l(k) \psi \| \\
& \leq & \frac{C(s)}{1+M^s} \| (\one \otimes N^{\d})^{\frac{1}{2}} \psi \| \ \leq \ \frac{C(s)}{1+M^s} \| (H^{\d}_0)^{\frac{1}{2}} \psi \|.\end{aligned}$$ Then, using the commutation relations (\[ccr2\]), we have $$\begin{aligned}
\| \sum_{|l|>M} \int_{\R^n} dk\, \beta_l(k)\otimes a^*_l(k) \psi \|^2 & = & \| \sum_{|l|>M} \int_{\R^n} dk\,
\bar{\beta}_l(k)\otimes a_l(k) \psi \|^2 \\
& & \qquad \qquad + \left( \sum_{|l|>M} \int_{\R^n} dk\, |\beta_l(k)|^2 \right) \| \psi \|^2.\end{aligned}$$ Finally, one gets $$\|H^{\d} \psi -H^{\d}(M) \psi \| \leq \frac{2C(s)}{1+M^s} \| (H^{\d}_0)^{\frac{1}{2}} \psi \| + \left( \sum_{|l|>M}
\int_{\R^n} dk\, |\beta_l(k)|^2 \right)^{\frac{1}{2}} \| \psi \|.$$ Using condition $(C_{\beta}),$ one shows that the right hand side tends to zero as $M$ goes to infinity. So, $H^{\d}(M)$ converges strongly to $H^{\d}$ and then also in the strong resolvent sens ([@RS1], Theorem VIII.25).
\[prop:vp-disc\] $E^{\d}_0(M)$ is a decreasing function of $M$ which tends to $E^{\d}_0.$
We know that, if $\phi^{\d}_0(M)$ is a ground state for $H^{\d}(M),$ then $\phi^{\d}_0(M)=\tilde{\phi}^{\d}_0(M)\otimes \Omega^c_M,$ and so $$\forall l\in \Lambda_M^c, \forall k\in \R^n, a_l(k) \phi^{\d}_0(M)=0.$$ Let $M'>M,$ $$\begin{aligned}
E^{\d}_0(M') & \leq & \langle \phi^{\d}_0(M); H^{\d}(M') \phi^{\d}_0(M) \rangle \\
& \leq & \underbrace{\langle \phi^{\d}_0(M); H^{\d}(M) \phi^{\d}_0(M) \rangle}_{= E^{\d}_0(M)}+ \underbrace{\langle
\phi^{\d}_0(M); (W^{\d}(M')-W^{\d}(M)) \phi^{\d}_0(M) \rangle}_{=0}.\end{aligned}$$ So the function $E^{\d}_0(M)$ decreases. With the same argument, one proves that $E^{\d}_0(M)\geq E^{\d}_0.$ Then $E^{\d}_0(M)$ converges to some $E_{\infty}\geq E^{\d}_0.$ But $E^{d}_0\in \sigma (H^{\d})$ and $H^{\d}(M)$ converges to $H^{\d}$ in the strong resolvent sens, so ([@RS1], Theorem VIII.24), $$\forall M>0, \exists E(M)\in \sigma (H^{\d}(M))/ E(M)\to E^{d}_0.$$ Using the fact that $E^{\d}_0(M)$ is the ground state energy of $H^{\d}(M),$ we finally get $E_{\infty}= E^{\d}_0.$
\[prop:norm-cvg-disc\] Let $\Delta$ be an interval bounded from above. For all $s>0,$ there exists $K(s,\Delta)>0$ such that $$\| \chi_{\Delta}(H^{\d})(W^{\d}-W^{\d}(M))\chi_{\Delta}(H^{\d})\| \leq \frac{K(s,\Delta)}{1+M^s}.$$
Let $\phi, \psi \in \mathcal{H}^{\d}.$ We have $$\begin{aligned}
& & \left|\langle \phi; \chi_{\Delta}(H^{\d})(W^{\d}-W^{\d}(M))\chi_{\Delta}(H^{\d}) \psi \rangle \right| \\
& = & |\langle \phi; \chi_{\Delta}(H^{\d}) (\sum_{|l|>M} \int_{\R^n} dk \, \beta_l(k)\otimes a^*_l(k)
+\bar{\beta}_l(k) \otimes a_l(k)) \chi_{\Delta}(H^{\d}) \psi \rangle | \\
& \leq & |\langle \chi_{\Delta}(H^{\d}) \phi; (\sum_{|l|>M} \int_{\R^n} dk \, \bar{\beta}_l(k) \otimes
a_l(k)) \chi_{\Delta}(H^{\d}) \psi \rangle | \\
& & \qquad \qquad \qquad + |\langle (\sum_{|l|>M} \int_{\R^n} dk \, \bar{\beta}_l(k) \otimes a_l(k))\chi_{\Delta}(H^{\d})
\phi; \chi_{\Delta}(H^{\d}) \psi \rangle |\end{aligned}$$ $$\begin{aligned}
& \leq & \| \phi \| \times \| (\sum_{|l|>M} \int_{\R^n}dk\, \bar{\beta}_l(k)\otimes a_l(k))\chi_{\Delta}(H^{\d})\psi \| \\
& & \qquad \qquad \qquad \qquad + \| \psi \| \times \| (\sum_{|l|>M} \int_{\R^n}dk\, \bar{\beta}_l(k)\otimes a_l(k))\chi_{\Delta}(H^{\d})\phi \| \\
& \leq & \frac{C(s)}{1+M^s} \left( \| \phi \| \times \| (\one \otimes N^{\d})^{\frac{1}{2}} \chi_{\Delta}(H^{\d}) \psi \| + \|
\psi \| \times \| (\one \otimes N^{\d})^{\frac{1}{2}} \chi_{\Delta}(H^{\d}) \phi \| \right).\end{aligned}$$ But $\Delta$ is bounded from above, $\one \otimes N^{\d} \leq \frac{1}{m} H_0^{\d}$ and $W^{\d}$ is relatively $H_0^{\d}$ bounded, so $(\one \otimes N^{\d})^{\frac{1}{2}} \chi_{\Delta}(H^{\d})$ is a bounded operator. Finally, one has $$\left|\langle \phi; \chi_{\Delta}(H^{\d})(W^{\d}-W^{\d}(M))\chi_{\Delta}(H^{\d}) \psi \rangle \right| \leq \frac{2C(s)
\|(N^{\d})^{\frac{1}{2}} \chi_{\Delta}(H^{\d})\|}{1+M^s} \| \phi \| \times \| \psi \|,$$ which ends the proof.
[**Proof of Theorem \[thm:disc\]:**]{}
We use the method of [@BFS2]. Given an operator $A,$ $[A]_-$ will denote its negative part and $\textrm{Tr}(A)$ its trace. To prove the theorem, it suffices to show that, for all $\epsilon>0,$ we have $$\textrm{Tr} \left\{ [ H^{\d} -E^{\d}_0-m+\epsilon]_- \right\} >-\infty.$$ Let $\epsilon >0,$ and $\Delta= ]-\infty, E^{\d}_0+m-\epsilon[.$ Then $$[ H^{\d} -E^{\d}_0-m+\epsilon]_- = \chi_{\Delta}(H^{\d}) ( H^{\d} -E^{\d}_0-m+\epsilon) \chi_{\Delta}(H^{\d}),$$ and so $$\begin{aligned}
\textrm{Tr} \left\{ [ H^{\d} -E^{\d}_0-m+\epsilon]_- \right\} & = & \textrm{Tr} \left\{ \chi_{\Delta}(H^{\d}) (H^{\d}
-E^{\d}_0-m+\epsilon) \chi_{\Delta}(H^{\d}) \right\} \\
& = & \textrm{Tr} \{ \chi_{\Delta}(H^{\d}) (H^{\d}(M)-E^{\d}_0(M)-m+\epsilon \\
& & \qquad +W^{\d}-W^{\d}(M) -E^{\d}_0+E^{\d}_0(M)) \chi_{\Delta}(H^{\d}) \}.\end{aligned}$$ But $$E^{\d}_0(M) \to E^{\d}_0 \quad \textrm{and} \quad \| \chi_{\Delta}(H^{\d})(W^{\d}-W^{\d}(M))\chi_{\Delta}(H^{\d})\| \to
0,$$ using Propositions \[prop:vp-disc\] and \[prop:norm-cvg-disc\], so, for $M$ large enough, we have $$\begin{array}{l}
\textrm{Tr} \left\{ [ H^{\d} -E^{\d}_0-m+\epsilon]_- \right\}\\
\qquad \qquad \qquad \geq \textrm{Tr} \left\{ \chi_{\Delta}(H^{\d})
(H^{\d}(M) -E^{\d}_0(M)-m+\frac{\epsilon}{2}) \chi_{\Delta}(H^{\d}) \right\} \\
\qquad \qquad \qquad \geq \textrm{Tr} \left\{ \chi_{\Delta}(H^{\d}) [H^{\d}(M) -E^{\d}_0(M)-m+\frac{\epsilon}{2}]_- \chi_{\Delta}(H^{\d}) \right\} \\
\qquad \qquad \qquad \geq \textrm{Tr} \left\{ [H^{\d}(M) -E^{\d}_0(M)-m+\frac{\epsilon}{2}]_- \right\}>-\infty\\
\end{array}$$ where in the last step we used Proposition \[prop:sp-disc\].
Continuous models {#ssec:massif}
-----------------
In this section, we are interested in the model introduced in Sect. \[ssec:modq\] but for massive bosons, *i.e.* the function $\omega(k)$ is replaced by $\omega_m(k)$ satisfying $(H_{\omega}).$ We then consider, on $\mathcal{H},$ the following Hamiltonian: $$\begin{aligned}
\label{h-mass}
H_m & := & H_p \otimes \one + \one \otimes \int_{\R^d} dx\, \int_{\R^n} dk\, \omega_m(k) a^*(x,k)a(x,k) \nonumber \\
& & \qquad +\int_{\R^d} dx\, \int_{\R^n} dk\, \rho_1(x-Q)\frac{\hat{\rho}_2(k)}{\sqrt{2\omega_m(k)}} \otimes a^*(x,k) \nonumber \\
& & \qquad \qquad \qquad \qquad +\rho_1(x-Q)\frac{\bar{\hat{\rho}}_2(k)}{\sqrt{2\omega_m(k)}} \otimes a(x,k) \\
& = & H^0_m+ W_m. \nonumber\end{aligned}$$
We denote by $E_m$ the ground state energy of $H_m.$ The main result of this section is the
\[thm:massif\] $\sigma_{ess}(H_m) \subset [E_m+m, +\infty[.$ In particular, $H_m$ has a ground state $\phi_m.$
The strategy of the proof is very similar to the one of the previous section. However, one has to be more careful with the estimates when removing the cutoff because the norm of $\rho_1(x-Q)$ as an operator on $L^2(\R^d)$ does not decrease with $x,$ even worse, it does not depend on it. To control this problem, we will use the exponential decay of the spectral projectors in the $Q$ variable, which will be obtained via the Agmon method (see Sect. \[sssec:expbound\])
### Cutoff models {#sssec:cont-cutoff}
Let $j$ be a smooth function with compact support on $\R^d$ such that $$0 \leq j(x) \leq 1, \quad j(x)=1 \ \ \rm{for} \ |x|\leq 1/2, \quad \rm{and} \quad j(x)=0 \ \
\rm{for} \ |x|\geq 3/4.$$ For all $L>0,$ we define $j_L(x)=j(\frac{x}{L})$ and $\bar{j}_L(x)=1-j_L(x).$ We then define $$\begin{aligned}
\label{h-massl1}
H_m(L) & := & H_m^0+ \int_{\R^d} dx\, \int_{\R^n} dk\, \rho_1(x-Q)j_L(x)\frac{\hat{\rho}_2(k)}{\sqrt{2\omega_m(k)}}
\otimes a^*(x,k) \nonumber \\
& & \qquad \qquad \qquad +\rho_1(x-Q)j_L(x)\frac{\bar{\hat{\rho}}_2(k)}{\sqrt{2\omega_m(k)}} \otimes a(x,k)\\
& = & H_m^0 + W_m(L) \nonumber\end{aligned}$$ on $\mathcal{H}.$ Using the definition of $j_L,$ one can, in $W_m(L),$ replace $\int_{\R^d} dx$ by $\int_{[-L,L]^d} dx.$ Finally, we define $$\label{h-massl2}
\tilde{H}_m(L):= H_p\otimes \one+ \one \otimes \int_{[-L,L]^d} dx\, \int_{\R^n} dk\, \omega_m(k) a^*(x,k)a(x,k)+ W_m(L)$$ on $L^2(\R^d)\otimes \mathcal{F}\left(L^2([-L,L]^d)\otimes L^2(\R^n)\right).$ We denote by $E_m(L)$ and $\tilde{E}_m(L)$ the ground state energies of those two operators respectively.
We have “cut” the Hamiltonian $H_m$ in the $x$ variable. We are now in a finite volume box. If we consider the variable $p,$ conjugate to $x,$ this is equivalent to “discretizing” the problem. One has to note that here the variable $p$ is discrete: $p\in \Z^d.$ If $$a^*_p(k)= \frac{1}{(2L)^{\frac{d}{2}}}\int_{[-L,L]^d} dx \, e^{ipx} a^*(x,k),\quad
a_p(k)= \frac{1}{(2L)^{\frac{d}{2}}}\int_{[-L,L]^d} dx \, e^{-ipx} a(x,k)$$ and $$\beta_p= \frac{1}{(2L)^{\frac{d}{2}}}\int_{[-L,L]^d} dx \, \rho_1(x-Q)j_L(x)$$ denote the Fourier coefficients of $a^*(x,k), a(x,k)$ and $\rho_1(x-Q)j_L(x)$ respectively, the problem can now be written as follows $$\begin{aligned}
\tilde{H}_m(L)& = & H_p\otimes \one+ \one \otimes \sum_{p \in \Z^d} \int_{\R^n} dk\, \omega_m(k) a^*_p(k)a_p(k) \\
& & \qquad + \sum_{p \in \Z^d} \int_{\R^n} dk\, (\beta_p \frac{\hat{\rho}_2(k)}{\sqrt{2\omega_m(k)}}\otimes a^*_p(k)
+ \bar{\beta}_p \frac{\bar{\hat{\rho}}_2(k)}{\sqrt{2\omega_m(k)}}\otimes a_p(k)),\end{aligned}$$ which has the form (\[hdisc\]). If the $\beta_p$ satisfy $(C_{\beta}),$ we will then have the following result:
\[prop:sp-mass1\] $\forall L>0, \sigma_{ess}(\tilde{H}_m(L)) \subset [\tilde{E}_m(L)+m, +\infty[.$
Finally, a splitting of $L^2(\R^d)$ into $L^2([-L,L]^d)\oplus L^2(\R^d\backslash [-L,L]^d)$ together with the argument of the previous section will lead to the
\[prop:sp-mass2\] $\sigma_{ess}(H_m(L)) \subset [E_m(L)+m, +\infty[.$ In particular, $H_m(L)$ has a ground state $\phi_m(L).$
So, it remains to check that the $\beta_p$ satisfy the condition $(C_{\beta}).$ The function $j_L$ is zero for $|x|>L$ and $\rho_1$ has compact support (in a ball of radius $R_1$), so $$\forall |q|>L+R_1, \forall x\in \R^d, \, \rho_1(x-q)j_L(x)=0.$$ Then, for all $p$ in $\Z^d,$ $\beta_p$ is a multiplication operator by a compactly supported function. Moreover, the function $\rho_1(x-q)j_L(x)$ is $C^{\infty},$ so its Fourier coefficients decay faster than any power of $p.$ Those two facts ensure us that $\sup_p \sup_q \big|\beta_p(q)|p|^n\big|<C_n(L)<+\infty$ and so condition $(C_{\beta})$ is satisfied. To prove Theorem \[thm:massif\], it remains to control the limit $L\to +\infty.$
### Exponential bounds {#sssec:expbound}
\[prop:exp-decr\] Let $\Delta$ be a bounded from above interval. For any $\alpha>0,$ there exists $M(\alpha,\Delta)>0$ such that
- $\| (e^{\alpha|Q|}\otimes \one) \chi_{\Delta}(H_m(L))\| \leq M(\alpha, \Delta).$
- $\| (e^{\alpha|Q|}\otimes \one) \chi_{\Delta}(H_m)\| \leq M(\alpha, \Delta).$
- $\| (e^{\alpha|Q|}\otimes \one) \chi_{\Delta}(H)\| \leq M(\alpha, \Delta).$
This bound is uniform in $L$ and $m.$ The proof is exactly the same as the one of Theorem II.1 of [@BFS1]. The only difference is that $\sigma_{ess}(H_p)=\emptyset,$ which makes things easier and in particular one does not need any condition on $\alpha$ or on the supremum of the interval $\Delta.$
For any $R>0,$ we now define $$\label{nb-photproches}
N(|x|\leq R):= \int_{|x|\leq R}dx\, \int_{\R^n}dk\, a^*(x,k)a(x,k),$$ and $$\label{nb-photloin}
N(|x|>R):= \int_{|x|>R}dx\, \int_{\R^n}dk\, a^*(x,k)a(x,k).$$ $N(|x|\leq R)$ is the number of bosons inside the ball centered at the origin and of radius $R$ (in the $x$ variable), and $N(|x|>R)$ is the number of bosons outside this ball. We will prove that the number of these “far away” bosons decays exponentially fast with $R.$ More precisely, we have the following estimate:
\[prop:photdecr\] For any $\alpha>0,$ there exists $C(\alpha)>0$ such that $$\label{eq:photdecr}
\langle \phi_m(L); \one \otimes N(|x|>R) \phi_m(L)\rangle \leq C(\alpha) e^{-\alpha R}$$ uniformly in $L.$
The idea is to adapt the proof of [@BFS1]. What is new in our model is that we need an explicit control on the number of “far away” bosons in the $x$ direction, even for massive bosons. Fot that purpose, we use the following lemma which comes from the well known pullthrough formula (see *e.g.* [@G]):
\[lem:pullthrough\] $\|\one \otimes a(x,k) \phi_m(L)\| \leq \frac{1}{\omega_m(k)}\|
\rho_1(x-Q)j_L(x)\frac{\hat{\rho}_2(k)}{\sqrt{2\omega_m(k)}}\otimes \one \phi_m(L)\|.$
[**Proof of Proposition \[prop:photdecr\] : **]{} Let $\alpha >0,$ $$\begin{aligned}
& & \langle \phi_m(L); \one \otimes N(|x|>R) \phi_m(L)\rangle \\
& = & \int_{|x|>R}dx\, \int_{\R^n} dk\, \| \one \otimes a(x,k) \phi_m(L)\|^2 \\
& \leq & \int_{|x|>R}dx\, \int_{\R^n} dk\, \frac{1}{\omega_m^2(k)} \| \rho_1(x-Q)j_L(x)
\frac{\hat{\rho}_2(k)}{\sqrt{2\omega_m(k)}}\otimes \one \phi_m(L)\|^2 \\
& \leq & \int_{|x|>R}dx\, \int_{\R^n} dk\, \frac{|\hat{\rho}_2(k)|^2}{2\omega_m^3(k)} \| \rho_1(x-Q) j_L(x)
e^{-\alpha |Q|}\|^2_{\mathcal{B}(L^2)} \times \|e^{\alpha |Q|}\otimes \one \phi_m(L)\|^2.\end{aligned}$$
The function $\hat{\rho}_2$ is a Schwartz function and $\omega_m$ is bounded from below by $m>0,$ so the integral with respect to the $k$ variable converges. Now we recall that the function $\rho_1$ has compact support in the ball of radius $R_1,$ so, for any given $x\in \R^d,$ we have $$\begin{aligned}
\|\rho_1(x-Q)e^{-\alpha |Q|}\|_{\mathcal{B}(L^2)} & = & \sup_{q\in \R^d} |\rho_1(x-q)e^{-\alpha |q|}| \\
& = & \sup_{|q-x|\leq R_1} |\rho_1(x-q)e^{-\alpha |q|}| \;
\leq \| \rho_1 \|_{\infty} e^{\alpha R_1} e^{-\alpha |x|}.\end{aligned}$$ Thus $$\int_{|x|>R} dx\, \|\rho_1(x-Q)e^{-\alpha |Q|}\|^2_{\mathcal{B}(L^2)} \leq \| \rho_1 \|^2_{\infty} e^{2\alpha
R_1} \int_{|x|>R} dx\, e^{-2\alpha |x|}
\leq K(\alpha) e^{-\alpha R}.$$ And so, finally, $$\langle \phi_m(L); \one \otimes N(|x|>R) \phi_m(L)\rangle \leq K'(\alpha)e^{-\alpha R} \|e^{\alpha |Q|}\otimes \one \phi_m(L)\|^2.$$
But, for any $L,$ we have $E_m(L)\leq E^0_{p}$ where $E^0_{p}$ is the ground state energy of $H_p.$ Indeed, if $\psi^0_p$ is the ground state of $H_p,$ we have $$E_m(L) \leq \langle \psi^0_p \otimes \Omega ; H_m(L)\; \psi^0_p \otimes \Omega\rangle = E_p^0.$$ Take now $\Delta= ]-\infty, E_p^0],$ one can then write $\phi_m(L)= \chi_{\Delta}(H_m(L)) \phi_m(L),$ and so $$\begin{aligned}
\|e^{\alpha |Q|}\otimes \one \phi_m(L)\|^2 & \leq & \|e^{\alpha |Q|}\otimes \one \; \chi_{\Delta}(H_m(L))\|^2 \|\phi_m(L)\| \\
& \leq & M(\alpha, \Delta)^2,\end{aligned}$$ which ends the proof.
We finally give an estimate similar to the one of Proposition \[prop:norm-cvg-disc\].
\[prop:norm-cvg-mass\] Let $\Delta$ and $\alpha$ be as in Proposition \[prop:exp-decr\], then there exists $K(\alpha, \Delta)$ such that $$\|\chi_{\Delta}(H_m) (W_m-W_m(L))\chi_{\Delta}(H_m)\| \leq K(\alpha, \Delta) e^{-\alpha L}.$$
We follow the scheme of the proof of Proposition \[prop:exp-decr\] using estimates similar to the ones of the previous proposition. Let $\phi, \psi \in \mathcal{H},$ $$\begin{aligned}
& & |\langle \phi; \chi_{\Delta}(H_m) (W_m-W_m(L))\chi_{\Delta}(H_m)\psi \rangle | \\
& \leq & |\langle (e^{2\alpha |Q|}\otimes \one) \chi_{\Delta}(H_m)\phi; \Big(\int_{|x|>\frac{L}{2}}dx\, \int_{\R^n}dk\,
e^{-2\alpha |Q|}\rho_1(x-Q)\bar{j}_L(x) \\
& & \qquad \qquad \qquad \qquad \qquad \qquad \qquad \times \frac{\bar{\hat{\rho}}_2(k)}{\sqrt{2\omega_m(k)}} \otimes
a(x,k)\Big) \chi_{\Delta}(H_m)\psi \rangle |\\
& & \qquad + |\langle (e^{2\alpha |Q|}\otimes \one) \chi_{\Delta}(H_m)\psi; \Big(\int_{|x|>\frac{L}{2}}dx\, \int_{\R^n}dk\,
e^{-2\alpha |Q|}\rho_1(x-Q)\bar{j}_L(x) \\
& & \qquad \qquad \qquad \qquad \qquad \qquad \quad \qquad \times \frac{\bar{\hat{\rho}}_2(k)}{\sqrt{2\omega_m(k)}} \otimes
a(x,k)\Big) \chi_{\Delta}(H_m)\phi \rangle |. \end{aligned}$$ We consider only the first term, the other one being similar. $$\begin{aligned}
& & |\langle (e^{2\alpha |Q|}\otimes \one) \chi_{\Delta}(H_m)\phi; \Big( \int_{|x|>\frac{L}{2}}dx\, \int_{\R^n}dk\,
e^{-2\alpha |Q|}\rho_1(x-Q)\bar{j}_L(x) \\
& & \qquad \qquad \qquad \qquad \qquad \qquad \qquad \times \frac{\bar{\hat{\rho}}_2(k)}{\sqrt{2\omega_m(k)}} \otimes
a(x,k)\Big) \chi_{\Delta}(H_m)\psi \rangle |\\
& \leq & \|(e^{2\alpha |Q|}\otimes \one) \chi_{\Delta}(H_m)\phi\| \times \| (\int_{|x|>\frac{L}{2}}dx\, \int_{\R^n}dk\,
e^{-2\alpha |Q|}\rho_1(x-Q)\bar{j}_L(x) \\
& & \qquad \qquad \qquad \qquad \qquad \qquad \qquad \quad \times \frac{\bar{\hat{\rho}}_2(k)}{\sqrt{2\omega_m(k)}} \otimes
a(x,k))\chi_{\Delta}(H_m)\psi\| \\
& \leq & M(2\alpha, \Delta) \|\phi \| \left[ \int_{|x|>\frac{L}{2}}dx\, \int_{\R^n}dk\, \| e^{-2\alpha |Q|}\rho_1(x-Q)
\bar{j}_L(x) \frac{\bar{\hat{\rho}}_2(k)}{\sqrt{2\omega_m(k)}}\|^2_{\mathcal{B}(L^2)} \right]^{\frac{1}{2}}\\
& & \qquad \qquad \qquad \qquad \qquad \times \left[ \int_{|x|>\frac{L}{2}}dx\, \int_{\R^n}dk\,
\| \one \otimes a(x,k) \chi_{\Delta}(H_m)\psi\|^2 \right]^{\frac{1}{2}}\\
& \leq & C(\alpha, \Delta) e^{-\alpha L} \| \phi \| \times \| (\one \otimes N)^{\frac{1}{2}}\chi_{\Delta}(H_m)\psi \|. \end{aligned}$$ The result then follows as in the discrete case.
### Removing the cutoff {#sssec:cont-remove}
\[prop:st-res-mass\] $H_m(L)$ converges to $H_m$ in the strong resolvent sens.
As for Proposition \[prop:st-res-disc\], it suffices to show that $H_m(L)$ converges strongly to $H_m$. Let $\psi \in D(H_m^0),$ $$\begin{aligned}
& & \| H_m \psi -H_m(L)\psi \|^2_{\mathcal{H}}\\
& = & \| W_m \psi -W_m(L)\psi \|^2_{\mathcal{H}} \\
& = & \int_{|q|>\frac{L}{2}-R_1} dq\, \Big\| \Big( \int_{|x|>\frac{L}{2}}dx\, \int_{\R^n}dk\, \rho_1(x-q)\bar{j}_L(x)
\frac{\hat{\rho}_2(k)}{\sqrt{2\omega_m(k)}} a^*(x,k)\\
& & + \int_{\R^d} dq\, \int_{|x|>\frac{L}{2}}dx\, \int_{\R^n}dk\,
\rho_1(x-q)\bar{j}_L(x) \frac{\bar{\hat{\rho}}_2(k)}{\sqrt{2\omega_m(k)}} a(x,k)\Big) \psi(q)\Big\|^2_{\mathcal{F}}.\end{aligned}$$ With similar computations as the ones of Proposition \[prop:st-res-disc\], we get $$\| H_m \psi -H_m(L)\psi \|^2_{\mathcal{H}} \leq C \int_{|q|>\frac{L}{2}-R_1} dq\, \|N^{\frac{1}{2}}\psi
(q)\|^2_{\mathcal{F}} +\| \psi(q)\|^2_{\mathcal{F}}.$$ But $N^{\frac{1}{2}}\psi(q)$ and $\psi(q)$ belong to $L^2(\R^d,\mathcal{F}),$ so the right-hand side tends to zero as $L$ goes to infinity.
\[prop:vp-mass\] $E_m(L)$ converges to $E_m$ as $L$ goes to infinity.
Remember that $\phi_m(L)$ is a ground state of $H_m(L).$ We have $$\begin{aligned}
E_m & \leq & \langle \phi_m(L); H_m \phi_m(L) \rangle \\
& \leq & E_m(L) +\langle \phi_m(L); (W_m-W_m(L)) \phi_m(L) \rangle \\
& \leq & E_m(L) + 2 \mathcal{R}e \big( \langle \phi_m(L); \int_{|x|>\frac{L}{2}}dx\, \int_{\R^n}dk\, \rho_1(x-Q)
\bar{j}_L(x) \\
& & \qquad \qquad \qquad \qquad \qquad \qquad \qquad \times \frac{\hat{\rho}_2(k)}{\sqrt{2\omega_m(k)}}\otimes a(x,k)
\phi_m(L) \rangle \big)\\
& \leq & E_m(L) + 2 \mathcal{R}e \big( \langle e^{\alpha |Q|}\otimes \one \phi_m(L); \int_{|x|>\frac{L}{2}}dx\,
\int_{\R^n}dk\, e^{-\alpha |Q|}\rho_1(x-Q) \\
& & \qquad \qquad \qquad \qquad \qquad \qquad \qquad \times \bar{j}_L(x)\frac{\hat{\rho}_2(k)}{\sqrt{2\omega_m(k)}}\otimes a(x,k)
\phi_m(L) \rangle \big). \end{aligned}$$ Then, the same computation as in Proposition \[prop:norm-cvg-mass\] leads to $$\begin{aligned}
E_m & \leq & E_m(L) + K(\alpha) e^{-\frac{\alpha L}{2}} \langle \phi_m(L); \one \otimes N(|x|>\frac{L}{2}) \phi_m(L)\rangle\\
& \leq & E_m(L) + C(\alpha) e^{-\alpha L}.\end{aligned}$$ So, the function $E_m(L)$ is bounded from below (and from above by $E_p^0$). Then there exists a sequence $L_n$ and $E_{\infty}$ such that $$\lim_{n\to +\infty} E_m(L_n)= E_{\infty}\geq E_m.$$ Now, $H_m(L_n)$ converges to $H_m$ in the strong resolvent sens and $E_m \in \sigma(H_m)$, so, for all $n,$ there exists $E(L_n)\in \sigma(H_m(L_n))$ such that $$\lim_{n\to +\infty} E(L_n)= E_m.$$ But $E(L_n)$ is bigger than $E_m(L_n)$ for all $n,$ so finally $E_m=E_{\infty}.$ The function $E_m(L)$ is then bounded with only one accumulating point $E_m$, which proves that the function converges to this point.
[**Proof of Theorem \[thm:massif\] : **]{} The proof is identical to the one of Theorem \[thm:disc\].
Another way to prove our results concerning the massive case would be to use the ideas of [@DG]-[@GLL]. The idea is to prove that $E_m$ is not in the essential spectrum using the Weyl criterion. For that purpose, one proves that, given a normed sequence $\psi_j$ tending weakly to zero, $$\label{eq:weyl}
\liminf_{j\to \infty} \langle \psi_j;(H_m-E_m)\psi_j \rangle>0.$$ The philosophy is that, if $\psi_j$ tends weakly to zero, it must “escape to infinity” in some way. In our model, if it escapes in the particle part, with the number of bosons or with their momentum in the $y$ direction (that is when $k$ tends to infinity), then the energy grows necessarily and (\[eq:weyl\]) is certainly satisfied. Now, if it escapes with far away bosons, either in “space” (that is in the $x$ or $y$ direction) or in “momentum in the $x$ direction”, the idea is that those bosons do not interact with the particle and so each of them has an energy at least $m.$ A Weyl sequence can then exist only for $E\geq E_m+m.$ A precise writing of such a proof would imply a control on the momentum of the bosons in the $x$ direction, which is the new element of our model. In our proof, such a control already exists but appears in a hidden way in Proposition \[prop:norm-cvg-disc\]. Finally, we would like to emphasize that writing a proof using this other method would not be much shorter.
Proof of the main results {#sec:nonmassif}
=========================
The goal of this section is to prove the results of Sect. \[sec:results\]. We start with Theorem \[thm:gdstate\]. We adapt the method of [@G]. We will insist on the differences with this paper. The idea is to approach (in a way which has to be made precise) $H$ with Hamiltonians for which we know that they have a ground state and then to obtain the same result for $H.$ More precisely, we will use the following lemma:
\[lem:ah\] ([@AH], Lemma 4.9) Let $H, H_n (n\in \N)$ be selfadjoint operators on a Hilbert space $\mathcal{H}.$ We suppose that
- $\forall n \in \N, H_n$ has a ground state $\psi_n$ with ground state energy $E_n,$
- $H_n$ tends to $H$ in the strong resolvent sens,
- $\lim_{n \to +\infty} E_n=E,$
- *w*$-\lim_{n\to +\infty} \psi_n=\psi \neq 0.$
Then $\psi$ is a ground state of $H$ with ground state energy $E.$
Infrared cutoff {#ssec:ircutoff}
---------------
We denote by $\chi_{\sigma\leq\omega(k)}$ the caracteristic function of the set $\{k\in \R^n |\sigma\leq\omega(k)\}.$ For any $\sigma>0,$ we then define $$\begin{aligned}
\label{hamq:cutoff}
H^{\sigma} & := & H_0+ \int_{\R^d}dx \int_{\R^n}dk\, \rho_1(x-Q)\frac{\hat{\rho}_2(k)}{\sqrt{2\omega(k)}}\chi_{\sigma\leq\omega(k)}(k)\otimes a^*(x,k)\nonumber \\
& & \qquad \qquad \qquad \qquad \qquad +\rho_1(x-Q)^*\frac{\bar{\hat{\rho}}_2(k)}{\sqrt{2\omega(k)}}\chi_{\sigma\leq\omega(k)}(k)\otimes a(x,k) \nonumber\\
& = & H_0+ H_{I,\sigma}, \end{aligned}$$ where $H_0$ is the free Hamiltonian defined in (\[hamq:libre\]). We want to use Lemma \[lem:ah\] with $H$ and $H^{\sigma_n}$ where $\sigma_n$ is some sequence going to zero.
We consider a function $\tilde{\omega}_{\sigma}(k)$ satisfying $$\left\{ \begin{array}{l} \nabla \tilde{\omega}_{\sigma} \in L^{\infty}(\R^n), \\
\tilde{\omega}_{\sigma}(k)= \omega(k) \quad \textrm{si} \quad \omega(k)\geq \sigma ,\\ \inf
\tilde{\omega}_{\sigma}(k)\geq \frac{\sigma}{2}>0, \end{array} \right.$$ and we define $$\tilde{H}^{\sigma}= H_p\otimes \one + \one \otimes \mbox{d}\Gamma(\tilde{\omega}_{\sigma}) + H_{I,\sigma}.$$ Then we have the following result:
\[prop:cutoff\] For any $\sigma>0, H^{\sigma}$ has a ground state $\psi_{\sigma}.$ We denote by $E_{\sigma}$ its ground state energy.
To prove this result we use the following lemma:
([@G], Lemma 3.2) $H^{\sigma}$ has a ground state if and only if $\tilde{H}^{\sigma}$ has one.
[**Proof of Proposition \[prop:cutoff\] : **]{} According to the previous lemma, it suffices to show that $\tilde{H}^{\sigma}$ has a ground state. But $\tilde{H}^{\sigma}$ is a Hamiltonian of the form studied in Sect. \[ssec:massif\], so, according to Theorem \[thm:massif\], it has a ground state.
\[prop:st-res\] $H^{\sigma}$ tends to $H$ in the norm resolvent sens.
We use Lemma A.2 of [@G] which says that it suffices to show that $Q^{\sigma}$ converges to $Q$ in the topology of $\mathcal{D}(Q),$ where $Q^{\sigma}$ and $Q$ are the quadratic forms associated to $H^{\sigma}$ and $H.$ But, with a similar computation to the one of Lemma \[lem:a-estimates\], one has $$\begin{aligned}
|Q(u,v)-Q^{\sigma}(u,v)| & \leq & \left( \int_{\R^d} dx\, \int_{\omega(k)\leq \sigma} dk\,
\frac{\rho_1(x-q)^2|\hat{\rho}_2(k)|^2}{2\omega^2(k)}\right)^{\frac{1}{2}} \\
& & \qquad \qquad \qquad \times (Q(u,u)\|v\|+Q(v,v)\|u\|).\end{aligned}$$
\[cor:vp-cvg\] $\lim_{\sigma \to 0} E_{\sigma}=E_0.$
As in the massive case, one has $E_{\sigma}\leq E_p^0$ for all $\sigma>0.$
Using Propositions \[prop:cutoff\] and \[prop:st-res\] together with Corollary \[cor:vp-cvg\], one can see that the operators $H^{\sigma}$ and $H$ satisfy assumptions $(i)-(ii)-(iii)$ of Lemma \[lem:ah\]. So, it remains to check condition $(iv)$ and Theorem \[thm:gdstate\] will be proven.
Uniform estimates {#ssec:unif}
-----------------
\[lem:nrj-bound\] There exists $C_1>0$ such that for all $\sigma>0,$ $$\langle \psi_{\sigma}; H_0 \psi_{\sigma}\rangle \leq C_1.$$
This inequality comes from the fact that $H_{I,\sigma}$ is relatively $H_0$ bounded with infinitesimal bound, uniformly with respect to $\sigma>0.$ Of course, we need an estimate on the number of soft bosons, estimate which uses the infrared condition (IR).
\[lem:nb-bound\] There exists $C_2>0$ such that for all $\sigma>0,$ $$\langle \psi_{\sigma}; \one \otimes N \psi_{\sigma}\rangle \leq C_2.$$
As in Lemma \[lem:pullthrough\], one can show that $$\label{eq:pullthr2}
\|\one \otimes a(x,k) \psi_{\sigma}\| \leq \frac{1}{\omega(k)}\| \rho_1(x-Q)\frac{\hat{\rho}_2(k)}{\sqrt{2\omega_m(k)}}
\chi_{\omega(k)\geq \sigma}(k)\otimes \one \psi_{\sigma} \|.$$ Thus, $$\begin{aligned}
\langle \psi_{\sigma}; \one \otimes N \psi_{\sigma}\rangle & = & \int_{\R^d}dx\, \int_{\R^n} dk\, \| \one \otimes a(x,k)
\psi_{\sigma}\|^2_{\mathcal{H}} \\
& \leq & \int_{\R^d}dx\, \int_{\omega(k)\geq \sigma} dk\, \frac{1}{\omega^2(k)} \| \rho_1(x-Q)
\frac{\hat{\rho}_2(k)}{\sqrt{2\omega(k)}}\otimes \one \psi_{\sigma}\|^2_{\mathcal{H}} \\
& \leq & \int_{\R^d} dq \int_{\R^d}dx\, \int_{\omega(k)\geq \sigma} dk\, \frac{|\hat{\rho}_2(k)|^2}{2\omega^3(k)}
|\rho_1(x-q)|^2 \|\psi_{\sigma}(q)\|^2_{\mathcal{F}}\\
& \leq & \| \rho_1\|^2_2 \left( \int_{\R^n} dk\, \frac{|\hat{\rho}_2(k)|^2}{2\omega^3(k)} \right)
\int_{\R^d} dq\, \|\psi_{\sigma}(q)\|^2_{\mathcal{F}}
\leq C_2.\end{aligned}$$
We have obtained a control on the total number of bosons. However, we will also need some control (uniform with respect to $\sigma$) on the number of “far away bosons”, that is on the following quantities: $\langle \psi_{\sigma}; N(|x|>R) \psi_{\sigma} \rangle$, $\langle
\psi_{\sigma}; N(|y|>S) \psi_{\sigma} \rangle$ and $\langle \psi_{\sigma}; N(|p|>P) \psi_{\sigma} \rangle$ where $$N(|x|>R)=\int_{|x|>R}dx\, \int_{\R^n}dk\, a^*(x,k)a(x,k),$$ $$N(|y|>S)=\int_{|y|>S}dx\, \int_{\R^n}dy\, \tilde{a}^*(x,y)\tilde{a}(x,y),$$ $$N(|p|>P)=\int_{|p|>P}dp\, \int_{\R^n}dk\, \hat{a}^*(p,k)\hat{a}(p,k).$$ The operators $\tilde{a}$ and $\tilde{a}^*$ come from $a$ and $a^*$ via a partial Fourier transform in the $k$ variable, and the operators $\hat{a}$ and $\hat{a}^*$ via a partial Fourier transform in the $x$ variable. We then prove a result similar to Proposition \[prop:photdecr\]:
\[lem:nb-decay1\] For any $\alpha>0,$ there exists $C(\alpha)>0$ such that $$\langle \psi_{\sigma}; \one \otimes N(|x|>R) \psi_{\sigma}\rangle \leq C(\alpha) e^{-\alpha R}.$$
The proof of this lemma is exactly the same to the one of Proposition \[prop:photdecr\]. This lemma gives us a control on the number of “far away” bosons in the $x$ direction. Similarly one can control the number of bosons whose momentum in the $x$ direction is large:
\[lem:nb-decay2\] For any $s>0,$ there exists $C(s)>0$ such that $$\langle \psi_{\sigma}; \one \otimes N(|p|>P) \psi_{\sigma}\rangle \leq \frac{C(s)}{1+P^s}.$$
Using (\[eq:pullthr2\]) and a computation similar to the one of Proposition \[prop:photdecr\], one gets $$\langle \psi_{\sigma}; \one \otimes N(|p|>P) \psi_{\sigma}\rangle \leq \left( \int dk\,
\frac{|\hat{\rho}_2(k)|^2}{2\omega(k)^3} \right)\times \left( \int_{|p|>P} dp |\hat{\rho}_1(p)|^2\right),$$ and the result follows .
Finally, to control $N(|y|>S),$ we use the following result noting that $\mbox{d}\Gamma(1-F_S(y))\leq N(|y|>\frac{S}{2}).$
\[lem:y-decay\] Let $F\in C^{\infty}_0(\R^n)$ such that $$0 \leq F(y) \leq 1, \quad F(y)=1 \ \ \rm{for} \ |y|\leq 1/2, \quad \rm{and} \quad F(y)=0 \ \ \rm{for} \
|y|\geq 1.$$ Let $F_S(y)=F(\frac{|y|}{S}).$ Then $$\lim_{\sigma\to 0, S\to +\infty} \langle \psi_{\sigma}; \textrm{d}\Gamma(1-F_S(y))\psi_{\sigma} \rangle=0.$$
There is a similar result in [@G] (Lemma 4.5), and we essentially follow its proof. The main difference is that the norm of $\rho_1(x-Q)$ as an operator on $L^2(\R^d)$ does not depend on $x$ and is therefore not square integrable with respect to this variable. As in Sect. \[sssec:expbound\], to control this problem, we will use the exponential decay of the spectral projectors in the $Q$ variable (Proposition \[prop:exp-decr\]). First, one easily sees that $$\label{dgamma-repr}
\mbox{d}\Gamma(1-F_S(y))=\int dx\, dk\, a^*(x,k)(1-F(\frac{|D_k|}{S}))a(x,k).$$ We recall that for any $\sigma$ one has $$a(x,k)\psi_{\sigma}= (E_{\sigma}-H^{\sigma}-\omega(k))^{-1} \frac{\rho_1(x-Q)\hat{\rho}_2(k)}{\sqrt{2\omega(k)}}
\chi_{\sigma\leq\omega(k)}(k)\psi_{\sigma}.$$ Then one can prove ([@G], Prop 4.4) that $$\lim_{\sigma \to 0} a(x,k)\psi_{\sigma}- (E_0-H-\omega(k))^{-1}
\frac{\rho_1(x-Q)\hat{\rho}_2(k)}{\sqrt{2\omega(k)}}\psi_{\sigma}=0$$ in $L^2(\R^{d+n},dx\, dk;\mathcal{H}).$ Using this together with (\[dgamma-repr\]), we then have $$\begin{array}{l}
\langle \psi_{\sigma}; \mbox{d}\Gamma(1-F_S(y))\psi_{\sigma} \rangle \\
\qquad = \int dx\, dk\, \langle (E_0-H-\omega(k))^{-1}\frac{\rho_1(x-Q)\hat{\rho}_2(k)}{\sqrt{2\omega(k)}}\psi_{\sigma}; \\
\qquad \qquad \qquad \qquad (1-F(\frac{|D_k|}{S}))(E_0-H-\omega(k))^{-1}\frac{\rho_1(x-Q)\hat{\rho}_2(k)}{\sqrt{2\omega(k)}}
\psi_{\sigma}\rangle + o(\sigma^0)\\
\qquad \leq \|(E_0-H-\omega(k))^{-1}\frac{\rho_1(x-Q)\hat{\rho}_2(k)}{\sqrt{2\omega(k)}}\psi_{\sigma}\|_{L^2(\R^{d+n};\mathcal{H})} \\
\qquad \qquad \times \|(1-F(\frac{|D_k|}{S}))(E_0-H-\omega(k))^{-1}\frac{\rho_1(x-Q)\hat{\rho}_2(k)}{\sqrt{2\omega(k)}}
\psi_{\sigma}\|_{L^2(\R^{d+n};\mathcal{H})}+ o(\sigma^0)\\
\qquad \leq \|(E_0-H-\omega(k))^{-1}\frac{\rho_1(x-Q)e^{-\alpha|Q|}\hat{\rho}_2(k)}{\sqrt{2\omega(k)}}\|_{L^2(\R^{d+n};\mathcal{B}(\mathcal{H}))}
\times \|e^{\alpha|Q|} \psi_{\sigma}\|_{\mathcal{H}}\\
\qquad \qquad \times \|(1-F(\frac{|D_k|}{S}))(E_0-H-\omega(k))^{-1} \frac{\rho_1(x-Q)e^{-\alpha|Q|}\hat{\rho}_2(k)}{\sqrt{2\omega(k)}}\|_{L^2(\R^{d+n};\mathcal{B}(\mathcal{H}))}\\
\qquad \qquad \qquad \times \|e^{\alpha|Q|} \psi_{\sigma}\|_{\mathcal{H}}+ o(\sigma^0).\\
\end{array}$$ We check that $(E_0-H-\omega(k))^{-1}\frac{\rho_1(x-Q)e^{-\alpha|Q|}\hat{\rho}_2(k)}{\sqrt{2\omega(k)}}$ belongs to $L^2(\R^{d+n};\mathcal{B}(\mathcal{H})),$ using the fact that $\|(E_0-H-\omega(k))^{-1}\|\leq \omega(k)^{-1}$ and condition (IR). Thus $$\lim_{S\to +\infty} \|(1-F(\frac{|D_k|}{S}))(E_0-H-\omega(k))^{-1}
\frac{\rho_1(x-Q)e^{-\alpha|Q|}\hat{\rho}_2(k)}{\sqrt{2\omega(k)}}\|_{L^2(\R^{d+n};\mathcal{B}(\mathcal{H}))}=0.$$ Moreover $\|e^{\alpha|Q|} \psi_{\sigma}\|_{\mathcal{H}}$ is uniformly bounded (w.r.t $\sigma$), which can be proven as for $\|e^{\alpha|Q|} \psi_m(L)\|_{\mathcal{H}}$ (see Sect. \[ssec:massif\]), and the result follows.
Proof of Theorem \[thm:gdstate\] {#ssec:proof}
--------------------------------
We have seen that the only thing we had to check was condition $(iv)$ of Lemma \[lem:ah\]. The unit ball of $\mathcal{H}$ is weakly compact, so there exists a sequence $\sigma_n \to 0$ and $\psi \in \mathcal{H}$ such that $\psi_{\sigma_n}$ converges weakly to $\psi.$ It then suffices to prove that $\psi\neq 0.$ The idea is to find a compact operator $K$ such that for any $n$ large enough one has such an estimate: $$\label{eq:unif-est}
\|K \psi_{\sigma_n}\| \geq \delta>0.$$ This will ensure that $\psi$ is non zero. Indeed, $K$ is compact, so $K \psi_{\sigma_n}$ tends strongly to $K\psi.$ If $\psi$ was zero then $\|K \psi_{\sigma_n}\|$ would go to zero, which enters in contradiction with (\[eq:unif-est\]).
Let us then take $F\in C^{\infty}_0(\R^n)$ and $G\in C^{\infty}_0(\R^d)$ satisfying the conditions of Lemma \[lem:y-decay\]. Remembering that $p$ is the variable conjugate to $x,$ *i.e.* $p=-i\nabla_x$ on $L^2(\R^{d+n}, dx\, dk),$ one has the following inequalities: $$\label{eq:gammaf}
(1-\Gamma(F_S(y)))^2 \leq (1-\Gamma(F_S(y))) \leq \mbox{d}\Gamma(1-F_S(y)),$$ $$\label{eq:gammag1}
(1-\Gamma(G_R(x)))^2 \leq (1-\Gamma(G_R(x))) \leq \mbox{d}\Gamma(1-G_R(x)) \leq N(|x|>\frac{R}{2}),$$ $$\label{eq:gammag2}
(1-\Gamma(G_P(p)))^2 \leq (1-\Gamma(G_P(p))) \leq \mbox{d}\Gamma(1-G_P(p)) \leq N(|p|>\frac{P}{2}).$$ Finally, let $\chi(s\leq s_0)$ be a function with support in $\{ |s|\leq s_0\}$ and equal to $1$ in $\{ |s|\leq \frac{s_0}{2}\}.$ For any non negative $\theta, P, R$ and $S,$ we define $$\label{def:compact}
K(\theta,P,R,S):= \chi(N\leq \theta)\chi(H_0\leq \theta)\Gamma(F_S(y))\Gamma(G_R(x))\Gamma(G_P(p)).$$ The assumptions on $F, G, \chi$ and $\omega$ ensure that $K(\theta,P,R,S)$ is compact for any $\theta, P, R$ and $S.$
Using Lemmas \[lem:nrj-bound\] and \[lem:nb-bound\], there exists $\theta_0>0$ such that, for all $n,$ one has: $$\|(1-\chi(N\leq \theta))\psi_{\sigma_n}\|\leq \frac{1}{10}, \|(1-\chi(H_0\leq \theta))\psi_{\sigma_n}\|\leq \frac{1}{10}.$$ Likewise, using Lemmas \[lem:nb-decay1\] and \[lem:nb-decay2\] together with inequalities (\[eq:gammag1\]) and (\[eq:gammag2\]), there exist $R_0,P_0>0$ such that, for all $n,$ one has: $$\|(1-\Gamma(G_R(x)))\psi_{\sigma_n}\| \leq \frac{1}{10}, \|(1-\Gamma(G_P(p)))\psi_{\sigma_n}\| \leq \frac{1}{10}.$$ Finally, using Lemma \[lem:y-decay\] and (\[eq:gammaf\]), there exist $S_0>0$ and $n_0$ such that, for all $n\geq n_0,$ one has: $$\|(1-\Gamma(F_S(y)))\psi_{\sigma_n}\| \leq \frac{1}{10}.$$ Then, for any $n\geq n_0:$ $$\begin{aligned}
\|\psi_{\sigma_n}\| & \leq & \|(1-\chi(N\leq \theta_0))\psi_{\sigma_n}\|+\|\chi(N\leq \theta_0)(1-\chi(H_0\leq
\theta_0))\psi_{\sigma_n}\| \\
& & + \|\chi(N\leq \theta_0)\chi(H_0\leq \theta_0)(1-\Gamma(G_{R_0}(x)))\psi_{\sigma_n}\| \\
& & + \|\chi(N\leq \theta_0)\chi(H_0\leq \theta_0)\Gamma(G_{R_0}(x))(1-\Gamma(G_{P_0}(p)))\psi_{\sigma_n}\| \\
& & + \|\chi(N\leq \theta_0)\chi(H_0\leq \theta_0)\Gamma(G_{R_0}(x))\Gamma(G_{P_0}(p))(1-\Gamma(F_{S_0}(y)))\psi_{\sigma_n}\| \\
& & + \|K(\theta_0,P_0,R_0,S_0)\psi_{\sigma_n}\| \\
& \leq & \frac{1}{2} +\|K(\theta_0,P_0,R_0,S_0)\psi_{\sigma_n}\|. \end{aligned}$$ But $\|\psi_{\sigma_n}\|=1$ for all $n,$ thus $$\|K(\theta_0,P_0,R_0,S_0)\psi_{\sigma_n}\| \geq \frac{1}{2},$$ for any $n\geq n_0,$ which is an estimate of the form (\[eq:unif-est\]).
Proof of Proposition \[prop:noground\] {#ssec:noground}
--------------------------------------
The idea of the proof is adapted from [@DG2]. Once again, one of the main tools is the pullthrough formula, which comes from the commutator between $H$ and annihilation operators $$\label{eq:pull}
[H,\one \otimes a(x,k)]=-\omega(k) \one \otimes a(x,k)-\rho_1(x-Q)\frac{\hat{\rho}_2(k)}{\sqrt{2\omega(k)}}
\otimes \one.$$ In order to get our result we will need to use this formula taking into account the membranes alltogether, which, on a formal level, means that we will integrate the previous formula over the “$x$-space”. It is therefore more convenient to look at the Hamiltonian not in the $(x,k)$ variables but in the $(p,k)$ variables, where $p$ is the variable conjugate to $x$ *via* Fourier transform, and then consider the value $p=0$. In such variables, the pullthrough formula just becomes $$\label{eq:pull2}
[H,\one \otimes \hat{a}(p,k)]=-\omega(k) \one \otimes \hat{a}(p,k)-\hat{\rho}_1(p)e^{-ipQ}
\frac{\hat{\rho}_2(k)}{\sqrt{2\omega(k)}}\otimes \one.$$
Suppose now that $\Psi\in \mathcal{H}$ satisfies $H\Psi=E_0\Psi,$ where $E_0$ is the ground state energy of $H.$ We will show that $\Psi=0.$ We apply equation (\[eq:pull2\]) on such a vector. One then gets the following equality $$\one \otimes \hat{a}(p,k)\ \Psi=-(H+\omega(k)-E_0)^{-1}\left( \frac{\hat{\rho}_1(p)e^{-ipQ}
\hat{\rho}_2(k)}{\sqrt{2\omega(k)}}\otimes \one\right) \Psi.$$ We denote with an exponent $(m)$ the component of a vector in the $m$-particle sector. We have, for any $m$, $$(\one \otimes \hat{a}(p,k)\ \Psi)^{(m)}(p_1,k_1,\dots,p_m,k_m)=\Psi^{(m+1)}(p,k,p_1,k_1,\dots,p_m,k_m)$$ and the righthand side is square integrable with respect to all its arguments because $\Psi \in \mathcal{H}.$ Therefore, for all $m,$ $$\Phi^{(m)}(p,k):= \left( -(H+\omega(k)-E_0)^{-1} \frac{\hat{\rho}_1(p)e^{-ipQ}
\hat{\rho}_2(k)}{\sqrt{2\omega(k)}}\otimes \one \ \Psi \right)^{(m)}$$ is square integrable with respect to $(p,k).$ On the other hand, it is a continuous function on $\R^d\times (\R^n-\{0\}).$ Then, for any $p_0\in \R^d, \, \Phi^{(m)}(p_0,k)$ is a well defined function of $k$ and it is square integrable. As we have said previously, we consider the value $p_0=0.$ But $$\Phi^{(m)}(0,k)= \frac{\hat{\rho}_1(0)\hat{\rho}_2(k)}{\sqrt{2}\omega(k)^{\frac{3}{2}}} \Psi^{(m)},$$ which is not square integrable if the infrared condition is violated, unless $\hat{\rho}_1(0)\Psi^{(m)}=0.$ By assumption, $\hat{\rho}_1(0)\neq 0,$ so $\Psi^{(m)}=0$ for all $m$ which means that $\Psi=0.$
A classical interpretation of the infrared problem {#sec:infrarouge}
==================================================
In this section, we would like to say a few words about the infrared problem. We know that this condition is necessary and sufficient for the existence of a ground state in the case of the Nelson model, and sufficient and “almost” necessary in our model.
On the other side, in [@A2], the author shows that if we consider another *ad hoc* representation of the canonical commutation relations the Nelson model without infrared condition has a ground state. In some sens, this representation regularizes the infrared singularity and is of course not unitarily equivalent to the Fock one. One could think that the same approach should work in our case. However, it turns out that this is not true. To explain why, we will briefly explain the physical origin of this representation. It will allow us to see that this procedure can apply to our model but does not have the same regularising effect.
We would like to explain the idea which is behind this change of representation coming back to classical mechanics [@DB]. It will then allow us to show the difference between the Nelson model and ours. We thus consider a classical Hamiltonian of the form $$H=\frac{1}{2}\int_A d\mu(\alpha) (\omega(\alpha)^2 \phi(\alpha)^2 +\pi(\alpha)^2),$$ where $\omega(\alpha)$ is some almost everywhere non negative function. We will simply write $X$ instead of $(\phi,\pi).$ The Hamiltonian flow can be written as $$X_t=\Phi_t X_0= \cos(\omega t)X_0- \sin(\omega t) JX_0, \quad \rm{where} \ \
J= \left( \begin{array}{rr} 0 & -\omega^{-1} \\ \omega & 0 \\ \end{array} \right).$$ One can see that $J^2=-\one.$
Now, for any $s\in \R,$ we define $$\mathcal{D}(\omega^s):= \{ \phi \in L^2 | \omega^s \phi \in L^2\},$$ and $[\mathcal{D}(\omega^s)]$ its closure for the norm $\| \phi \|_s= \| \omega^s \phi \|_{L^2}.$ Given $s,r \in \R,$ we define $$\mathcal{H}_{s,r}:=[\mathcal{D}(\omega^s)]\times [\mathcal{D}(\omega^r)].$$ The operator $J$ is well defined on $\mathcal{H}_{s,r}$ (as a bounded operator) if and only if $r=s-1.$ On the other side, the symplectic form $$\sigma(X_1,X_2)=\int_A d\mu(\alpha)(\phi_1(\alpha) \pi_2(\alpha)-\phi_2(\alpha) \pi_1(\alpha))$$ is meaningfull only on spaces of the form $\mathcal{H}_{s,-s}.$ So $J$ and $\sigma$ are both well defined only on $$\mathcal{H}=[\mathcal{D}(\omega^{\frac{1}{2}})]\times [\mathcal{D}(\omega^{-\frac{1}{2}})].$$ On $\mathcal{H}$, we consider the following complex structure $\langle \,\, ; \,\,\rangle$ defined as $$\langle X_1;X_2\rangle= \sigma(X_1,JX_2)+ i \sigma(X_1,X_2).$$ One can then identify $\mathcal{H}$ and $L^2(A, d\mu,\C)$ *via* the following isometry: $$(\phi,\pi)\in \mathcal{H}\to \sqrt{\omega(\alpha)}\phi(\alpha)+ \frac{i}{\sqrt{\omega(\alpha)}}\pi(\alpha) \in
L^2(A, d\mu,\C).$$ and if we define $$a(\alpha):=\frac{1}{\sqrt2} \left( \sqrt{\omega(\alpha)}\phi(\alpha)+ \frac{i}{\sqrt{\omega(\alpha)}}\pi(\alpha)\right) \in
L^2(A, d\mu,\C),$$ one can rewrite $H$ as $$H=\int_A \omega(\alpha) a^*(\alpha)a(\alpha) d\mu(\alpha).$$ The Poisson bracket associated to $\sigma$ is $
\{ \phi(\alpha),\pi(\alpha)\}= \delta_{\mu}(\alpha-\alpha'),
$ where $\delta_{\mu}$ is defined by $$\int_A f(\alpha') \delta_{\mu}(\alpha-\alpha')d\mu(\alpha')=f(\alpha).$$ It is easy to see that $
\{a(\alpha),a^*(\alpha')\}=-i \delta_{\mu}(\alpha-\alpha').
$ Those relations are the classical equivalent of the relations (\[ccr2\]). To write the quantum version of this model, one then consider the Fock space over $L^2(A, d\mu,\C).$
Consider now the Nelson model, *i.e.* $A=\R^d$ and $d\mu(\alpha)= dk.$ If we consider a particle which interacts with this field and which is on the other hand submitted to a confining potential $V$ such that $\min V=V(0),$ the equilibrium point of the system which correpsonds to the minimum of the energy can be written as $(q_*,\phi_*,p_*,\pi_*)=(0,-\frac{\hat{\rho}}{\omega^2},0,0)$ and $(\phi_*,\pi_*)$ is in $\mathcal{H}$ if and only if $\frac{\hat{\rho}}{\omega^{3/2}}\in L^2(\R^d),$ which is exactly the condition (IR). One should remind that, for this model, this condition is necessary and sufficient to have a ground state. In other words, the minimum $(q_*,\phi_*,p_*,\pi_*)$ of the classical Hamiltonian belongs to $\mathcal{H}$ if and only if (IR) is satisfied. In this way, one can say that the condition to have a ground state is the same on both classical and quantum level.
The representation considered in [@A2] corresponds, on the classical level, to the affine space “$\tilde{\mathcal{H}}=\mathcal{H}+(\frac{\hat{\rho}}{\omega^2},0)$”, more precisely, one considers the following symplectic transformation: $$\tilde{q}=q,\tilde{\phi}=\phi+\frac{\hat{\rho}}{\omega^2},\tilde{p}=p,\tilde{\pi}=\pi.$$ Here, $\tilde{\phi}$ represents the difference between the field and its equilibrium position. We already note that $0\notin \tilde{\mathcal{H}}$ if and only if (IR) is not satisfied. One then sees that the two equilibrium points, before ($\phi_*=0$) and after ($\phi_*=-\frac{\hat{\rho}}{\omega^2}$) having turned on the interaction with the particle, are not in the same space. This is this phenomenon which, on the quantum level, expresses that the ground state exists but in another representation, non-equivalent to the Fock one. One sometimes reads that the “ground state is not in the Fock space” (within the context of “Van Hove Hamiltonians” for example [@D]-[@Fr]-[@VH]).
In those new variables, the Hamiltonian of the whole system then writes $$\begin{aligned}
\label{tildeham}
\tilde{H}(\tilde{q},\tilde{\phi},\tilde{p},\tilde{\pi}) & = & \frac{1}{2} \int (\omega^2 \tilde{\phi}^2+\tilde{\pi}^2)
+\int \hat{\rho}(k)(e^{-ikq}-1)\tilde{\phi}\nonumber \\
& & +\frac{\tilde{p}^2}{2}+ V(\tilde{q})\underbrace{-\int
\frac{|\hat{\rho}(k)|^2e^{-ikq}}{\omega^2}}_{=W(\tilde{q})\,\, \textrm{bounded}}+ \underbrace{\int
\frac{|\hat{\rho}(k)|^2}{2\omega^2}}_{=\textrm{constant}}.\end{aligned}$$ If we want to study the system near the new equilibriun, one then has to chose the phase space such that $\tilde{\phi}=0$ belongs to it. It is then natural to study $\tilde{H}$ not on $\tilde{\mathcal{H}}$ but on $\mathcal{H}.$ In other words, to make sure that the equilibrium point for the interacting system is in the phase space, one has to consider another space. If one does so, condition (IR) is then satisfied even for $d=3$: $$\frac{\hat{\rho}(k)(e^{-ikq}-1)}{\omega^{\frac{3}{2}}}\in L^2(\R^d).$$ Then, if one quantizes $\tilde{H}$, one obtains a model in which a ground state exists. This is precisely what Arai does in [@A2], but without explaining it this way. However, the same transformation in our model does not make things “better.” Indeed, condition (IR) becomes, after the same transform: $$\frac{[\rho_1(x-q)-\rho_1(x)]\hat{\rho}_2(k)}{\omega(k)^{\frac{3}{2}}} \in L^2(\R^d\times\R^n),$$ which is still not satisfied if $n=3$ unless $\hat{\rho}_2(0)=0.$
**Acknowledgments:** Part of this work was supported by the Postdoctoral Training Program HPRN-CT-2002-0277. The author wishes to thank S. De Bièvre for many enjoyable discussions and useful comments.
[9999999]{}
Araï A., On a model of a harmonic oscillator coupled to a quantized, massless, scalar field, I, Journal of Math. Phys. [**22**]{}, 2539-2548 (1981). Araï A., Ground state of the massless Nelson model without infrared cutoff in a Non-Fock representation, Rev. Math. Phys [**13**]{}, 1075-1094 (2001). Araï A., Hirokawa M., On the existence and uniqueness of ground states of a generalized spin-boson model, J. Func. Anal. [**151**]{}, 455-503 (1997). Bruneau L., De Bièvre S., A Hamiltonian model for linear friction in a homogeneous medium, Comm. Math. Phys. **229**, 511-542 (2002). Bach V., Fröhlich J., Sigal I, Quantum electrodynamics of confined non-relativistic particles, Adv.Math. [**137**]{}, 299-395 (1998). Bach V., Fröhlich J., Sigal I, Spectral analysis for systems of atoms and molecules coupled to the quantized radiation field, Comm. Math. Phys. **207**, 249-290 (1999). Dereziński J., Van Hove Hamiltonians - exactly solvable models of the infrared and ultraviolet problem, preprint. De Bièvre S., *private communication* and Classical and quantum linear oscillator fields, in preparation. Dereziński J., Gérard C., Asymptotic completeness in quantum field theory. Massive Pauli-Fierz Hamiltonians, Rev. Math. Phys. [**11**]{}, 383-450 (1999). Dereziński J., Gérard C., Scattering theory of infrared divergent Pauli-Fierz Hamiltonians, in preparation. Dereziński J., Jakšić V., Spectral theory of Pauli-Fierz operators, J. Func. Anal. [**180**]{}, 243-327 (2001). Friedrichs K.O., Mathematical aspects of quantum theory of fields, New-York (1953). Gérard C., On the existence of ground states for massless Pauli-Fierz Hamiltonians, Ann. Henri. Poincaré. **1**, 443-459 (2000). Glimm J., Jaffe A., The $\lambda(\phi^4)_2$ quantum field theory without cutoffs II. The field operators and the approximate vacuum, Ann. Math. **91**, 362-401 (1970). Griesemer M., Lieb E.H., Loss M., Ground states in non-relativistic quantum electrodynamics, Invent. Math. **145**, 557-595 (2001). Komech A., Kunze M., Spohn H., Long-time asymptotics for a classical particle interacting with a scalar wave field, Comm. Partial Differential Equation **22**, 307-335 (1997). Lörinczi J., Minlos R.A., Spohn H., The infrared behaviour in Nelson’s model of a quantum particle coupled to a massless scalar field, Ann. Henri. Poincaré. **3**, 269-295 (2002). Reed M., Simon B., Methods of modern mathematical physics (vol 1), Academic Press, London (1976). Reed M., Simon B., Methods of modern mathematical physics (vol 2), Academic Press, London (1976). Van Hove L., Les difficultés de divergences pour un modèle particulier de champ quantifié, Physica **18**, 145-152 (1952).
| {
"pile_set_name": "ArXiv"
} |
---
author:
- Feng Yuan Ramesh Narayan
title: Hot Accretion Flows Around Black Holes
---
psfig.sty
accretion disks, active galactic nuclei, active galactic feedback, black holes, black hole X-ray binaries, jet, outflow
Prospects and remaining open questions {#prospect}
======================================
The discovery of the self-similar ADAF solution 20 years ago (Narayan & Yi 1994), and the subsequent development of the ADAF model of hot accretion flows (Abramowicz et al. 1995, Narayan & Yi 1995b, Chen et al. 1995), triggered a flurry of activity which has contributed greatly to our understanding of the dynamics and thermodynamics of hot accretion flows, as well as the recognition that these flows are relevant for numerous astrophysical objects: Sgr A\*, low luminosity AGNs, BHBs.
Despite the impressive progress of the last two decades, there are presently more questions than answers in this field. Below, in the authors’ view, are some of the more important questions:
- How are electrons and ions heated in a hot accretion flow? What particle energy distributions do these processes generate? What role do non-thermal particles play in the dynamical and radiative properties of the system?
- Are there processes in addition to Coulomb collisions which transfer energy from ions to electrons, and how do they influence the temperatures of the two species?
- How strong are mass outflows from hot accretion flows, and how does the mass accretion rate at the black hole ($\dot{M}_{\rm BH}$) depend on boundary conditions at large radius? What if gas is supplied from an external two-phase (or even multi-phase) medium?
- Why do hot accretion flows produce jets, whereas cool thin disks apparently do not? What role does the black hole, especially its spin, play in determining the properties of the jet? What fraction of the observed radiation comes from the jet versus the hot accretion flow?
- How efficiently do hot accretion flows advect large-scale ordered magnetic field towards the center, and how often do accreting black holes approach the “magnetically arrested disk” limit?
- Why and how do state transitions in black hole binaries occur? What are the physical processes responsible for converting cold optically thick gas into hot optically thin gas in a “truncated thin disk and hot inner accretion flow” configuration, and how do they relate to the hysteresis phenomenon? How do the same processes behave in the case of supermassive black holes?
- What is the thermal state of a hot accretion flow when $\dot{M}_{\rm
BH}$ is close to the upper limit for a hot solution? Does the accreting gas become a two-phase medium, and what observational signatures do the hot and cold phases produce?
- What determines whether a hot accretion flow produces a steady jet or an episodic jet, and why are the latter often associated with the hard to soft state transition in black hole binaries? How does this map to supermassive black holes, AGN jets and the radio loud/quiet dichotomy?
- What is the angular distribution of mass, momentum and energy outflow from a hot accretion flow around a supermassive black hole, and how do they determine the efficiency of feedback processes?
- How does the relative importance of mechanical versus radiative feedback depend on $\dot{M}_{\rm BH}$ and other parameters of the accretion flow?
- What are the properties of hot accretion flows around compact objects with a surface?[^1]
0.2cm [The authors are grateful to M. C. Begelman, C. Done, J.-P. Lasota, J. E. McClintock, J. P. Ostriker, J. Poutanen, E. Quataert, J. Stone, A. Tchekhovskoy and A. Zdziarski for helpful comments on the manuscript. This work was supported by grants 11133005 and 11121062 from the NSFC (FY) and AST1312651 from the NSF (RN).]{}
0.5cm
[^1]: An object with a surface introduces two important modifications compared to the black hole flows considered in this review. First, since gas comes to rest at the stellar surface, the inner boundary condition on the dynamical equations is very different and will result in vastly different density, velocity, pressure, etc. at small radii. Second, radiation from the surface will Compton-cool the hot accreting gas and modify its temperature. There have been only a few applications of the ADAF model to accreting neutron stars and white dwarfs. For lack of space, we have not reviewed this work here. Much more could be done in this area.
| {
"pile_set_name": "ArXiv"
} |
Ł ł
=-15mm=-2mm amssym.def
**A Note on Quantum Entanglement and PPT**
Shao-Ming Fei$^{1,2}$ and Xianqing Li-Jost$^3$
$^1$Department of Mathematics, Capital Normal University, Beijing 100037, P.R. China\
$^2$Institut f[ü]{}r Angewandte Mathematik, Universit[ä]{}t Bonn, 53115 Bonn, Germany\
E-mail: [email protected]\
$^3$Max-Planck-Institute for Mathematics in the Sciences, 04103 Leipzig, Germany\
E-mail: [email protected]
1 true cm
Abstract
We study quantum states for which the PPT criterion is both sufficient and necessary for separability. We present a class of $3\times 3$ bipartite mixed states and show that these states are separable if and only if they are PPT.
Keywords: PPT, Entanglement, Separability
PACS numbers: 03.65.Bz, 89.70.+c
Introduction
============
Quantum entanglement has been recently the subject of much study as a potential resource for communication and information processing [@nielsen]. Thus characterization and quantification of entanglement become an important issue. Entanglement of formation (EOF) [@eof] and concurrence [@concurrence] are two well defined quantitative measures of quantum entanglement. For two-quibt systems it has been proved that EOF is a monotonically increasing function of the concurrence and an elegant formula for the concurrence was derived analytically by Wootters [@wotters]. However with the increasing dimensions of the subsystems the computation of EOF and concurrence becomes formidably difficult. A few explicit analytic formulae for EOF and concurrence have been found only for some special symmetric states [@Terhal-Voll2000].
In fact if one only wants to know wether a state is separable or not, it is not necessary to compute the exact values of the measures for quantum entanglement. The estimation of lower bounds of entanglement measures can be just used to judge the separability of a quantum state [@chenk]. There are also many separability criteria, e.g., PPT (positive partial transposition) criterion [@peres], realignment [@ChenQIC03] and generalized realignment criteria [@chenkai], as well as some necessary and sufficient operational criteria for low rank density matrices [@hlvc00]. Further more, separability criteria based on local uncertainty relation [@hofmann] and the correlation matrix [@julio] of the Bloch representation for a quantum state have been derived, which are strictly stronger than or independent of the PPT and realignment criteria.
The PPT criterion is generally a necessary condition for separability. It becomes sufficient only for the cases $2\times2$ and $2\times3$ bipartite systems [@2231]. Other states of such property are the Schmidt-correlated (SC) states [@SC], which are the mixtures of pure states sharing the same Schmidt bases and naturally appear in a bipartite system dynamics with additive integrals of motion [@SC1]. In this paper we consider another special class of $3\times 3$ quantum mixed states. We show that for the states in this class, PPT is both necessary and sufficient for separability.
A class of quantum states and PPT
=================================
We consider $3\times 3$ quantum mixed states given by \[rho\] = |X><X| + ’ |X’><X’| + ” |X”><X”|, where $\lambda + \lambda' + \lambda''=1$, $0<\lambda , \lambda',
\lambda''<1$, $|X>, |X'>,|X''>$ are orthonormal vectors, \[XXX\] |X> = (,0,0,0,,0,0,0,)\^t,\
|X’>= (0,’,0,0,0,’,’,0,0)\^t,\
|X”>= (0,0,”,”,0,0,0,”,0)\^t, where $\alpha,\beta,\gamma,\alpha',\beta',\gamma',\alpha'',\beta'',\gamma''$ are non-zero complex numbers, $t$ stands for transposition. If we take the basis $|1>=(1,0,0)^t$, $|2>=(0,1,0)^t$, $|3>=(0,0,1)^t$, then $|X(\alpha,\beta,\gamma)>\equiv |X>
=\alpha|11>+\beta|22>+\gamma|33>$, $|X'> =(I\otimes
P)|X(\alpha',\beta',\gamma')>$, $|X''> =(I\otimes
P^2)|X(\alpha'',\beta'',\gamma'')>$, where $P=\left(\ba{ccc}0&0&1\\1&0&0\\0&1&0\ea\right)$ is the permutation operator.
[Theorem:]{} State $\rho$ is separable if and only if it is PPT.
To prove the theorem we first note that after partial transposition $\rho$ has the form $$\ba{l}
\rho^{pt}=\\[2mm]
\left(
\begin{array}{ccccccccc}
\lambda\alpha\alpha^*&0&0&0&0&\lambda''\alpha''\beta''^*&0&\lambda'\alpha'\gamma'^*&0\\[3mm]
0&\lambda'\alpha'\alpha'^*&0&\lambda\alpha\beta^*&0&0&0&0&\lambda''\alpha''\gamma''^*\\[3mm]
0&0&\lambda''\alpha''\alpha''^*&0&\lambda'\alpha'\beta'^*&0&\lambda\alpha\gamma^*&0&0\\[3mm]
0&\lambda\alpha^*\beta&0&\lambda''\beta''\beta''^*&0&0&0&0&\lambda'\beta'\gamma'^*\\[3mm]
0&0&\lambda'\alpha'^*\beta'&0&\lambda\beta\beta^*&0&\lambda''\beta''\gamma''^*&0&0\\[3mm]
\lambda''\alpha''^*\beta''&0&0&0&0&\lambda'\beta'\beta'^*&0&\lambda\beta\gamma^*&0\\[3mm]
0&0&\lambda\alpha^*\gamma&0&\lambda''\beta''^*\gamma''&0&\lambda'\gamma'\gamma'^*&0&0\\[3mm]
\lambda'\alpha'^*\gamma'&0&0&0&0&\lambda\beta^*\gamma&0&\lambda''\gamma''\gamma''^*&0\\[3mm]
0&\lambda''\alpha''^*\gamma''&0&\lambda'\beta'^*\gamma'&0&0&0&0&\lambda\gamma\gamma^*
\end{array}
\right).
\ea$$ $\rho^{pt}$ is hermitian. The non-negativity of $\rho^{pt}$, $\rho^{pt}\geq 0$, implies that $<\psi|\rho^{pt}|\psi> \geq 0$ for all vector $|\psi>\in H\otimes H$, which is equivalent to the non-negativity of the following three $3\times 3$ matrices: A\_1=(
[ccc]{} \^\*&”””\^\*&’’’\^\*\
””\^\*”&’’’\^\*&\^\*\
’’\^\*’&\^\*&”””\^\*
), A\_2=(
[ccc]{} ’’’\^\*&\^\*&”””\^\*\
\^\*&”””\^\*&’’’\^\*\
””\^\*”&’’\^\*’&\^\*
), and A\_3=(
[ccc]{} ”””\^\*&’’’\^\*&\^\*\
’’\^\*’&\^\*&”””\^\*\
\^\*&””\^\*”&’’’\^\*
).
The non-negativity of $A_1$ is equivalent to the following inequalities: $$\begin{aligned}
\label{rcl}\label{11}
\lambda\lambda'|\alpha|^2|\beta'|^2&\geq&\lambda''^2|\alpha''|^2|\beta''|^2,\\\label{12}
\lambda\lambda''|\alpha|^2|\gamma'''|^2&\geq&\lambda'^2|\alpha'|^2|\gamma'|^2,\\\label{13}
\lambda'\lambda''|\beta'|^2|\gamma''|^2&\geq&\lambda^2|\gamma|^2|\beta|^2\end{aligned}$$ and \[14\] ’”||\^2|’|\^2|”|\^2+2’ ”Re’”\^\*”\^\*’\^\* -\^3||\^2||\^2||\^2\
-’\^3|’|\^2|’|\^2| ’|\^2-”\^3|”|\^2|”|\^2|”|\^20. Similarly the non-negativity of $A_2$ and $A_3$ give rise to $$\begin{aligned}
\label{21}
\lambda'\lambda''|\alpha'|^2|\beta''|^2\geq\lambda^2|\alpha|^2|\beta|^2,\\ \label{22}
\lambda\lambda'|\alpha'|^2|\gamma|^2\geq\lambda''^2|\alpha''|^2|\gamma''|^2,\\\label{23}
\lambda\lambda''|\gamma|^2|\beta''|^2\geq\lambda'^2|\gamma'|^2|\beta'|^2,\end{aligned}$$ \[24\] ’”|’|\^2|”|\^2||\^2+2’ ”Re”\^\*’\^\*”’\^\* -\^3||\^2||\^2||\^2\
-’\^3|’|\^2|’|\^2| ’|\^2-”\^3|”|\^2|”|\^2|”|\^2 0, and $$\begin{aligned}
\label{31}
\lambda\lambda''|\alpha''|^2|\beta|^2\geq\lambda'^2|\alpha'|^2||\beta'|^2,\\\label{32}
\lambda'\lambda''|\alpha''|^2|\gamma'|^2\geq\lambda^2|\alpha|^2|\gamma|^2,\\\label{33}
\lambda\lambda'|\beta|^2|\gamma'|^2\geq\lambda''^2|\beta''|^2|\gamma''|^2,\end{aligned}$$ \[34\] ’”|”|\^2||\^2|’|\^2+2’ ”Re’\^\*’\^\*””\^\*-\^3||\^2||\^2||\^2\
-’\^3|’|\^2|’|\^2| ’|\^2-”\^3|”|\^2|”|\^2|”|\^20.
We can show that the inequalities (\[11\])-(\[13\]), (\[21\])-(\[23\]), (\[31\])-(\[33\]) are equalities. In fact if (\[11\]) is an inequality, $\lambda\lambda'|\alpha|^2|\beta'|^2 > \lambda''^2|\alpha''|^2|\beta''|^2$, then from (\[21\]) and (\[31\]), one would have, by multiplying $\lambda'|\alpha'|^2$ on both sides, $\lambda'|\alpha'|^2\lambda''^2|\alpha''|^2|\beta''|^2
<\lambda\lambda'^2|\alpha|^2|\alpha'|^2|\beta'|^2
\le\lambda\lambda''|\alpha''|^2|\beta|^2\lambda|\alpha|^2
\le \lambda'|\alpha'|^2\lambda''^2|\alpha''|^2|\beta''|^2$, which contradicts. Therefore (\[11\])-(\[13\]), (\[21\])-(\[23\]), (\[31\])-(\[33\]) become $$\begin{aligned}
\label{41}
\sqrt{\lambda\lambda'}\alpha\beta'=\lambda''\alpha''\beta'' e^{i\theta''_1},~
\sqrt{\lambda\lambda''}\alpha\gamma''=\lambda'\alpha'\gamma' e^{i\theta'_2},~
\sqrt{\lambda''\lambda'}\gamma''\beta'=\lambda\gamma\beta e^{i\theta_3}\\\label{42}
\sqrt{\lambda''\lambda'}\alpha'\beta''=\lambda\alpha\beta e^{i\theta_1},~
\sqrt{\lambda\lambda'}\alpha'\gamma=\lambda''\alpha''\gamma'' e^{i\theta''_2},~
\sqrt{\lambda\lambda''}\gamma\beta''=\lambda'\gamma'\beta' e^{i\theta'_3}\\\label{43}
\sqrt{\lambda\lambda''}\alpha''\beta=\lambda'\alpha'\beta' e^{i\theta'_1},~
\sqrt{\lambda''\lambda'}\alpha''\gamma'=\lambda\alpha\gamma e^{i\theta_2},~
\sqrt{\lambda\lambda'}\gamma'\beta=\lambda''\gamma''\beta'' e^{i\theta''_3},\end{aligned}$$ where all $\theta\in[-\pi,\pi]$. From (\[41\])-(\[43\]), (\[14\]), (\[24\]) and (\[34\]) become $$\begin{aligned}
Re(\alpha'\alpha''^*\beta''\beta^*\gamma\gamma'^*)\geq|\alpha|^2|\beta'|^2|\gamma''|^2,\\
Re(\alpha\alpha''^*\beta'\beta^*\gamma''\gamma'^*)\geq|\alpha'|^2|\beta''|^2|\gamma|^2,\\
Re(\alpha'\alpha^*\beta'^*\beta''\gamma''^*\gamma)\ge|\alpha''|^2|\beta|^2|\gamma'|^2.\end{aligned}$$
Applying these 9 equalities one obtains also
[cc]{} |’”\^\*”\^\*’\^\*|&|’||”||”||||||’|\
=|||||||||||| &=||\^2|’|\^2|”|\^2.
Similarly one has |”\^\*’\^\*”’\^\*||’|\^2|”|\^2||\^2, |’\^\*’\^\*””\^\*||”|\^2||\^2|’|\^2. By using (\[41\])-(\[43\]), we get $$\begin{aligned}
\label{51}
Re(\alpha'\alpha''^*\beta''\beta^*\gamma\gamma'^*) = \alpha'\alpha''^*\beta''\beta^*\gamma\gamma'^*=
|\alpha|^2|\beta'|^2|\gamma''|^2,\\\label{52}
Re(\alpha\alpha''^*\beta'\beta^*\gamma''\gamma'^*)=\alpha\alpha''^*\beta'\beta^*\gamma''\gamma'^* =
|\alpha'|^2|\beta''|^2|\gamma|^2,\\\label{53}
Re(\alpha'\alpha^*\beta'^*\beta''\gamma''^*\gamma) =\alpha'\alpha^*\beta'^*\beta''\gamma''^*\gamma =
|\alpha''|^2|\beta|^2|\gamma'|^2.\end{aligned}$$ From (\[51\]) and (\[52\]), we have $$\frac{\alpha'\beta''\gamma}{\alpha\beta'\gamma''}=\frac{|\alpha|^2|
\beta'|^2|\gamma''|^2}{|\alpha'|^2|\beta''|^2|\gamma|^2}
=\frac{\lambda^2|\alpha|^2\cdot\frac{1}{\lambda'\lambda''}|\beta|^2|\gamma|^2}{|\alpha'|^2|
\beta''|^2|\gamma|^2}=1.$$ While from (\[51\]) and (\[53\]), we have $$(\frac{\alpha''\beta\gamma'}{\alpha\beta'\gamma''})^*=1.$$ Therefore $$\alpha'\beta''\gamma=\alpha\beta'\gamma''=\alpha''\beta\gamma'$$ and, from (\[41\])-(\[43\]), $$\theta_1=\theta_2=\theta_3\equiv\theta ~,\theta'_1=\theta'_2=\theta'_3\equiv\theta'~,
\theta''_1=\theta''_2=\theta''_3\equiv\theta''.$$
Now by using these PPT conditions of $\rho$ we prove that $\rho$ has a pure state decomposition $\rho =\sum_{l=1}^{l=3}|\psi_l><\psi_l|$ such that all states $|\psi_l>$, $l=1,2,3$, are separable. $|\psi_l>$ can be generally expressed as $|\psi_l>=\sum_m^3 U_{ml}|X_m>=\sum_{ij}^{3} a_{ij}^{l}|ij>$ for some $a_{ij}^{l}\in\Cb$ under some basis $|ij>$, where $U_{ml}$ are the entries of a $3\times 3$ unitary matrix $U$, $|X_1>=\sqrt{\lambda}|X>$, $|X_2>=\sqrt{\lambda'}|X'>$, $|X_3>=\sqrt{\lambda''}|X''>$. We denote $B_l$ the $3\times 3$ matrix with entries $a_{ij}^{l}$. Suppose the matrix $U$ has the following form U= (
[ccc]{} u\_1&u\_2&u\_3\
u’\_1 &u’\_2&u’\_3\
u”\_1 &u”\_2&u”\_3\
) = (
[ccc]{} c\_1 e\^[i]{} &c\_2 e\^[i]{}&c\_3 e\^[i]{}\
c’\_1 e\^[i’]{}&c’\_2 e\^[i’]{}&c’\_3 e\^[i’]{}\
c”\_1 e\^[i”]{}&c”\_2 e\^[i”]{} &c”\_3 e\^[i”]{}
), where according to the unitary condition of $U$, \[ud\] \_[l=1]{}\^[l=3]{} c\_lc\_l\^\*=\_[l=1]{}\^[l=3]{}c’\_lc’\^\*\_l = \_[l=1]{}\^[l=3]{}c”\_lc”\^\*\_l = 1, \_[l=1]{}\^[l=3]{}c\_lc’\^\*\_l= 0, \_[l=1]{}\^[l=3]{}c\_lc”\^\*\_l = 0, \_[l=1]{}\^[l=3]{}c’\_lc”\^\*\_l= 0. Then $B_l$, $l=1,2,3$, has the following form B\_l=(
[ccc]{} u\_l&’u’\_l&”u”\_l\
”u”\_l&u\_l&’u’\_l\
’u’\_l&”u”\_l&u\_l
). It is straightforward to verify that the matrix $B_l$ has rank one if \[rankone\] c\_l\^2 e\^[i2]{}=c’\_lc”\_l.
As $0 < rank(B_l B_l^+) \leq rank (B_l)rank (B_l^+)$, if the rank of $B_l$ is one, matrix $B_l B_l^+$ has also rank one and $|\psi_l>$ is separable. Therefore if we can find $c_l$, $c'_l$, $c''_l$, $l = 1,2,3$, satisfying the unitary condition (\[ud\]) and the rank one condition (\[rankone\]), then $\rho = \lambda |X><X| + \lambda' |X'><X'| + \lambda'' |X''><X''|$ has separable pure state decomposition, $\rho =\sum_{l=1}^{l=3}|\psi_l><\psi_l|$ and $\rho$ is then separable if it is PPT. We show now that there exist such coefficients $c_i$, $c'_i$, $c''_i$, $i = 1,2,3$, satisfying both the unitary condition and the rank-one condition. Set $c_l = \frac{1}{\sqrt{3}}e^{i\varphi_l}$, $c'_l =
\frac{1}{\sqrt{3}}e^{i\varphi'_l}$, $c''_l =
\frac{1}{\sqrt{3}}e^{i\varphi''_l}$, $l = 1,2,3$, with $\varphi_l$, $\varphi'_l$, $\varphi''_l$, $l = 1,2,3$, satisfying $$\varphi_1-\varphi'_1=\xi',~ \varphi_2-\varphi'_2
=\xi'+ \frac{2\pi}{3},~ \varphi_3-\varphi'_3=\xi'-\frac{2\pi}{3},$$ $$\varphi_1-\varphi''_1=\xi'' + \frac{2\pi}{3},~ \varphi_2-\varphi''_2=\xi'',~
\varphi_3-\varphi''_3=\xi''-\frac{2\pi}{3},$$ for some real numbers $\xi'$ and $\xi''$. Then the unitary conditions (\[ud\]) are satisfied.
The rank-one conditions require that $-2\varphi_l + \varphi'_l+\varphi''_l= 2\theta,~
-2\varphi'_l + \varphi_l+\varphi''_l= 2\theta',~
-2\varphi''_l + \varphi_l+\varphi'_l= 2\theta''$, which can be realized by simply choosing $\xi'=\frac{2}{3}\theta'-\frac{2}{3}\theta$, $\xi''=-\frac{2}{3}\theta'-\frac{4}{3}\theta- \frac{2\pi}{3}$. Therefore if state $\rho$ is PPT, then it is separable. In fact due to that there is still freedom in choosing the parameters $\varphi_l$, $\varphi'_l$, $\varphi''_l$, $l=1,2,3$, there exist many separable pure state decompositions.
Conclusions and remarks
=======================
We have studied a special kind of bipartite quantum mixed states. For this class of states, the PPT criterion is both sufficient and necessary for separability. Here the states we concerned are rank three ones on $3\times 3$ bipartite systems. It has been shown that any bipartite entangled states of rank three are distillable [@rankthree], that is, there is no rank three bipartite bound entangled states. Therefore if the state $\rho$ is not PPT, i.e. conditions $$\frac{\lambda}{\sqrt{\lambda'\lambda''}}|\alpha\beta\gamma|
=\frac{\lambda'}{\sqrt{\lambda\lambda''}}|\alpha'\beta'\gamma'|
=\frac{\lambda''}{\sqrt{\lambda\lambda'}}|\alpha''\beta''\gamma''|$$ are not satisfied, $\rho$ must be not only entangled, but also distillable. This gives an example that a separable state could directly become a distillable state when some parameters varies continuously. There could be no bound entangled states between separable states and distillable states. Above all, with a similar construction of states (\[XXX\]), rank $2k+1$ states on $(2k+1)\times (2k+1)$, $k\in \Nb$, bipartite system can be obtained. Analogous investigations could be applied to get similar results.
[**Acknowledgments**]{} This work is supported by the NSFC 10675086, NSFC 10875081, KZ200810028013 and NKBRPC (2004CB318000).
[18]{}
Nielsen M A, Chuang I L. Quantum Computation and Quantum Information. Cambridge: Cambridge University Press, (2000).
C. H. Bennett, D. P. DiVincenzo and J. A. Smolin, et al. Phys. Rev. A 54, 3824(1996).\
M. B. Plenio and S. Virmani, Quant. Inf. Comp. 7, 1(2007).
A. Uhlmann Phys. Rev. A 62 032307(2000);\
P. Rungta, V. Bu$\breve{z}$ek, and C. M. Caves, et al. Phys. Rev. A 64, 042315(2001);\
S. Albeverio and S. M. Fei, J. Opt. B: Quantum Semiclass Opt, 3, 223-227(2001).
W. K. Wootters, Phys. Rev. Lett. 80, 2245 (1998).
Terhal B M, Vollbrecht K G H, Phys. Rev. Lett., 85, 2625-2628(2000);\
S.M. Fei, J. Jost, X.Q. Li-Jost and G.F. Wang, Phys. Lett. A 310, 333-338(2003);\
S.M. Fei and X.Q. Li-Jost, Rep. Math. Phys. 53, 195-210(2004);\
S.M. Fei, Z.X. Wang and H. Zhao, Phys. Lett. A 329, 414-419(2004);\
P. Rungta and C.M. Caves, Phys. Rev. A 67, 012307(2003).
K. Chen, S. Albeverio and S.M. Fei, Phys. Rev. Lett. 95, 040504(2005);\
K. Chen, S. Albeverio and S.M. Fei, Phys. Rev. Lett. 95, 210501(2005);\
X. H. Gao, S. M. Fei and K. Wu, Phys. Rev. A 74, 050303(R)(2007);\
J. I. de Vicente, Phys. Rev. A 75, 052320 (2007);\
C. J. Zhang, Y. X. Gong, Y. S. Zhang, and G. C. Guo, Phys. Rev. A 78, 042308(2008);\
H.P. Breuer, J. Phys. A 39, 11847(2006);\
E. Gerjuoy, Phys. Rev. A 67, 052308(2003);\
Y. C. Ou, H. Fan and S. M. Fei, Phys. Rev. A 78, 012311(2008).
A. Peres, Phys. Rev. Lett. 77, 1413 (1996).
K. Chen and L. A. Wu, Quant. Inf. Comput. 3, 193 (2003);\
O. Rudolph, Phys. Rev. A 67, 032312 (2003);\
K. Chen and L. A. Wu, Phys. Lett. A 306, 14 (2002).
S. Albeverio, K. Chen and S. M. Fei, Phys. Rev. A 68, 062313 (2003).
P. Horodecki, M. Lewenstein, G. Vidal and I. Cirac, Phys. Rev. A 62, 032310 (2000);\
S. Albeverio, S. M. Fei and D. Goswami, Phys. Lett. A 286, 91 (2001);\
S. M. Fei, X. H. Gao, X. H. Wang, Z. X. Wang and K. Wu, Phys. Lett. A [300]{}, 555 (2002).
H. F. Hofmann and S. Takeuchi. Phys, Rev. A 68, 032103 (2003);\
O. Gühne, M. Mechler, G. Toth and P. Adam, Phys. Rev. A 74, 010301(R) (2006);\
O. Gühne, Phys. Rev. Lett. 92, 117903 (2004);\
O. Gühne, P. Hyllus, O. Gittsovich, and J. Eisert, Phys. Rev. Lett. 99, 130504 (2007).
J. I. de Vicente, Quantum Inf. Comput. 7, 624 (2007);\
A. S. M. Hassan and P. S. Joag, Quantum Inf. Comput. 8, 0773 (2008).
M. Horodecki, P. Horodecki and R. Horodecki, Phys. Lett. A [223]{}, 1 (1996).
E.M. Rains, Phys. Rev. A 60, 179(1999).\
M.J. Zhao, S.M. Fei and Z.X. Wang, Phys. Lett. A 372, 2552(2008).
M. Khasin, R. Koslo®, Phys. Rev. A 76, 012304(2007).
L. Chen and Y.X. Chen, Phys. Rev. A78, 022318(2008).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Code-switching refers to the usage of two languages within a sentence or discourse. It is a global phenomenon among multilingual communities and has emerged as an independent area of research. With the increasing demand for the code-switching automatic speech recognition (ASR) systems, the development of a code-switching speech corpus has become highly desirable. However, for training such systems, very limited code-switched resources are available as yet. In this work, we present our first efforts in building a code-switching ASR system in the Indian context. For that purpose, we have created a Hindi-English code-switching speech database. The database not only contains the speech utterances with code-switching properties but also covers the session and the speaker variations like pronunciation, accent, age, gender, etc. This database can be applied in several speech signal processing applications, such as code-switching ASR, language identification, language modeling, speech synthesis etc. This paper mainly presents an analysis of the statistics of the collected code-switching speech corpus. Later, the performance results for the ASR task have been reported for the created database.'
address: |
Department of Electronics and Electrical Engineering,\
Indian Institute of Technology Guwahati, Guwahati-781039, India.
author:
- 'Ganji Sreeram, Kunal Dhawan and Rohit Sinha'
bibliography:
- 'refs.bib'
title: 'Hindi-English Code-Switching Speech Corpus'
---
code-switching, speech corpus, automatic speech recognition
Introduction
============
In multilingual communities, the speakers often switch or mix between two or more languages or language varieties during communication in their day to day lives. In linguistics, this phenomenon is referred to as code-switching [@Gumperz_1982_Discourse; @nilep2006code]. This phenomenon poses some interesting research challenges to speech recognition [@Lyu_2006_Speech; @Bhuvanagirir_2012_Mixed; @ahmed2012automatic], language identification [@Lyu_2008_Language] and language modelling [@Cao_2010_Semantics; @Yeh_2010_Integrated; @hamed2017building] domains. Over the years, due to urbanization and geographical distribution, people have moved from one place to another for a better livelihood. Hence, communicating in two or more languages helps to interact better with people from different places and cultures. There are many reasons for the occurrence of code-switching. The people belonging to the bilingual communities say that the main reason for code-switching between languages is due to the lack of words in the vocabulary of that particular native language [@Grosjean_1982_Life]. According to [@myers1993social; @malik1994socio; @milroy1995one; @dey2014hindi], some possible reasons for code-switching are: (i) to qualify the message by emphasizing specific words, (ii) to convey a personalized message, (iii) to maintain confidentiality during verbal communication, (iv) to show expertise, authority, status, etc. Another reason for code-switching is to enrich communication between speakers without any change in the situation. Hence, a native language speaker actively embeds meanings into the conversation by mixing non-native language words [@su2001code]. Based on the locations of the non-native words, code-switching can be broadly classified into two modes. When the switching happens within the sentence it is referred to as the *intra-sentential* code-switching and the one predominantly happening at the sentence boundary is referred to as the *inter-sentential* code-switching [@myers1989codeswitching]. Intra-sentential mode of switching is a common phenomenon and has become an identifying characteristic in bilingual communities.
Over the years, the English language has become the most widely spoken language in the world. After gaining independence from the British rule, though the Indian constitution declared Hindi as the primary official language, the usage of English was continued as a secondary language for its dominance in administration, education, and law [@malhotra1980hindi]. Thereby, the urban population has started a trend to communicate in English for economic and social purposes. Over the years, substantial code-switching to English while speaking Hindi, as well as many other Indian languages, has become a common feature [@kumar1986certain; @Smita2009]. Note that, $41.1\%$ of the Indian population are native Hindi speakers and hence the switching between Hindi and English is very common. Also, in the recent past, the researchers have reported that the native language of the speaker influences the foreign (non-native) language acquisition [@flege1995second]. In India, English is taught in schools from elementary level across the country, but very few schools are able to impart correct English pronunciations devoid of native language influences to their pupils. The recent works [@bali2014; @Das_2015_Code] have highlighted that the code-switching phenomenon is also observed in chats, comments, and messages posted on the social media sites like Facebook, Twitter, WhatsApp, YouTube, etc. Table \[cs\_ex\] shows a few example sentences of different modes of code-switching while highlighting the differences in the contextual information carried by the non-native words. In Type-1 intra-sentential code-switching, the non-native language words either occur in sequence or form a phrase, thus carry some contextual information. Whereas, in Type-2 case, the non-native language words are embedded into the native language sentences in such a manner that virtually no contextual information could be derived from those words. Also, during code-switching, we observe that the majority of the sentences belong to Type-2 intra-sentential mode. However, due to lack of availability of the domain-specific resources, the research activity is somewhat limited.
The monolingual automatic speech recognition (ASR) systems may be capable of recognizing a few words from a foreign language but are unable to handle a significant amount of code-switching in the data. On account of the existence of different variants of English pronunciations and code-switching effects, the development of an ASR system for Hindi-English (Hinglish) code-switching speech data is a challenging task. To the best of our knowledge, there is no large-sized Hinglish corpus available for carrying out the research. Towards addressing that constraint, we recently created a Hinglish corpus covering all typical sources of variations such as accent, session, channel, age, gender, etc. In this work, we describe the details of that corpus and also present basic experimental evaluation is done on the same.
{width="8.8cm"}
\[cs\_ex\]
The remainder of this paper is organized as follows: In Section \[literature\], we review the code-switching corpora currently reported in the literature. In Section \[sec:database\], the details about Hinglish speech and text corpus along with that of the necessary lexical resources for developing the Hinglish ASR system, are presented in detail. The experimental evaluations using the created Hinglish corpus has been presented in Section \[sec:experiments\]. The paper is concluded in Section \[sec:conclusion\].
Literature Review on Code-switching Corpora {#literature}
===========================================
In literature, a few code-switching speech corpora are already reported and they happen to cover different native and non-native language combinations. In the following, we briefly review those code-switching corpora while summarizing their salient attributes.
- The CUMIX Cantonese-English code-switching speech corpus developed by Joyce Y. C. Chan, et al., at the Chinese University of Hong Kong [@chan2005development]. It contains code-switched speech utterances read by the speakers. The database contains $17$ hours of data read by $40$ speakers.
- A small Mandarin-Taiwanese code-switching speech corpus was developed for testing purpose in [@lyu2006speech] by Dau-cheng Lyu and Ren-yuan Lyu. The corpus contains $4000$ Mandarin-Taiwanese code-switching utterances recorded from $16$ speakers.
- The English-Spanish code-switching speech corpus was compiled by Franco J. C. and Solorio at the University of Texas [@franco2007baby]. The corpus contains $40$ minutes of transcribed spontaneous conversations of $3$ speakers.
- The SEAME is a Mandarin-English code-switching conversational speech corpus developed by Dau-Cheng Lyu and Tien Ping Tan from Nanyang Technological University, Singapore, and Universiti Sains Malaysia [@lyu2010seame; @vu2012first]. The database contains $63$ hours of spontaneous Mandarin-English code-switching interview and conversational speech uttered by $157$ Singaporean and Malaysian speakers.
- Han-Ping Shen, et al., developed the CECOS, a Chinese-English code-switching speech corpus at the National Cheng Kung University in Taiwan [@shen2011cecos]. It contains $12.1$ hours of speech data collected from $77$ speakers uttering prompted code-switching sentences.
- A small Hindi-English code-switching speech corpus was collected by Anik Dey and Pascale Fung at Hong Kong University of Science and Technology. This corpus is primarily made up of student interview speech [@dey2014hindi]. It is about $30$ minutes of data collected from $9$ speakers.
- A corpus of Sepedi-English code-switching speech corpus was created by the South African CSIR [@modipa2013implications]. The database consists of $10$ hours of prompted speech, sourced from radio broadcasts and read by $20$ Sepedi speakers.
- Emre Yılmaz, et al., developed FAME!, a Frisian-Dutch code-switching speech corpus of radio broadcast speech at Radboud University, Nijmegen [@yilmaz2016longitudinal]. The recordings are collected from the archives of Omrop Fryslan, the regional public broadcaster of the province Fryslan. The database covers almost a $50$ years time span.
- The Malay-English corpus developed by Basem H. A. Ahmed, et al., consists of $100$ hours of Malaysian Malay-English code-switching speech data from $120$ Chinese, $72$ Malay and $16$ Indian speakers. [@ahmed2012automatic].
- MediaParl is a Swiss accented bilingual database developed by David Imseng, et al. contains recordings in both French and German as they are spoken in Switzerland. The data was recorded at the Valais Parliament. Valais is a bi-lingual Swiss canton with many local accents and dialects [@imseng2012mediaparl].
- The FACST, a French-Arabic speech corpus consists of records of code-switching read and conversational utterances by $20$ bilingual adult speakers who tend to code-switch in their daily lives [@amazouz2018french]. It is about $7.30$ hours of data.
- A South African speech corpus containing English-isiZulu, English-isiXhosa, English-Setswana, and English-Sesotho code-switching speech utterances is created from South African soap operas by Ewald van der Westhuizen and Thomas Niesler. The soap opera speech is typically fast, spontaneous and may express emotion, with a speech rate higher than prompted speech in the same languages [@van2018first].
- The Arabic-English is recently developed by Injy Hamed, et al., by conducting the interviews with $12$ participants [@hamed2018collection].
From the literature review, it can be noted that very small sized code-switching acoustic and linguistic resources have been available so far covering the Indian context. This motivated us to create moderate sized Hinglish resources so that current technological advances in acoustic and language modeling can be explored for Hinglish ASR task.
Creation of Hinglish Corpus {#sec:database}
===========================
This section describes the details of the creation of Hinglish (code-switching) corpus. Firstly, we describe the context and means employed for the creation of Hinglish sentences. Secondly, the details of the procedure followed by the speakers while recording the speech data corresponding to the created Hinglish sentences, are described. Finally, the creation of the lexical resources is discussed.
Hinglish text transcripts {#subsec:text_data}
-------------------------
For the experimental purpose, the Hinglish code-switching text data has been collected by crawling a few blogging websites [^1]$^{,}$[^2]$^{,}$[^3]$^{,}$[^4] having different contexts. The crawled data is normalized into meaningful sentences and further processed to remove extra spaces, special characters, emoticons, etc. Data thus obtained is used for training the language models, creating the lexicon and also as the text transcription for recording the acoustic data. The salient details of the Hinglish code-switching text corpus created is summarized in Table \[tab:database\]
Hinglish speech corpus
----------------------
Hinglish code-switching acoustic data is recorded over the phone from speakers belonging to different states in India. A consultant was hired for enrolling the speakers to call a toll free number from their mobile phones. The speakers called from various acoustic environments such as home, office, etc. Each speaker was given $100$ unique sentences taken from the above-processed text data. These $100$ sentences are partitioned into $5$ groups which contain $20$ sentences each. Each speaker is requested to record those $5$ groups in $5$ different sessions in order to capture the session variations such as emotions, environment, etc. It is worth highlighting that the duration of the sentences given to each speaker varies from $2$$-$$30$ seconds. Each speaker took about $10$ minutes to complete recording the $20$ sentences in each session. On an average, to complete recording the $100$ sentences, each speaker took about $50$ minutes. The volunteering speakers were compensated with $250$ for their time and effort.
The speech data is recorded at $8$ kHz sampling frequency and a bit rate of $128$. This set of speech files was manually inspected and pruned. At the end of the data collection phase, the Hinglish code-switching database contained $7,005$ utterances in total spoken by $71$ speakers.
Development of the lexical resources
------------------------------------
For the creation of a lexicon for development of an ASR system for Hinglish data, a unified phone list has been created for Hindi and English words. Also, a unique word list is extracted from the $13,071$ sentences obtained from Sub-section \[subsec:text\_data\]. The phone level transcription for those words has been done manually. Thus created lexicon covers all the pronunciation variations.
Statistical Analysis of the Database
====================================
This section provides the statistical analysis of the Hinglish code-switching speech corpus. The following subsection provides information about the speakers in the database. Later, a description of the size and linguistic features of the database is provided.
Speaker information
-------------------
In order to collect the Hinglish code-switching speech data, the field data consultant recruited speakers from Indian Institute of Technology Guwahati (IITG) who are natively from different states of India. A total of $71$ speakers are involved in the development of this database. To model a robust ASR system for Hinglish code-switching data, we need to have a database that covers variations due to different geographical distribution, gender, age, etc. Aiming at this, we have collected the database which covers all such variations. The details of the database are discussed below.
### Geographical distribution
Since the speakers residing in IITG are from different states of India and from different geographical locations, diversity in the acoustic data is guaranteed. The geographical distribution of the speakers is shown as a pie-chart in Figure \[fig:loc\_dist\]. The area-wise distribution of the speakers involved in this study is provided in Table \[lab:table1\].
### Age information
The Hinglish code-switching speech data has been recorded from the speakers between $20$ to $64$ years of age. The age distribution is shown as a bar-diagram in Figure \[fig:age\_dist\].
coordinates [ ($<$, 4.22) (20, 2.81) (21, 8.45) (22, 2.81) (23, 5.63) (24, 5.63) (25, 9.85) (26, 4.22) (27, 18.30) (28, 7.04) (29, 15.49) (30, 5.63) (31, 4.22) (32, 1.40) ($>$, 2.81) ]{};
### Gender information
The Hinglish code-switching speech data is recorded from $27$ female speakers and $44$ male speakers resulting in a total of $71$ speakers from different states of India. The gender distribution is shown as a pie-chart in Figure \[fig:gen\_dist\].
Experimental Evaluation and Discussion {#sec:experiments}
======================================
The Hinglish code-switching database has been validated by developing an ASR system. For this purpose, the recorded $7005$ utterances are partitioned into training and testing sets containing $5,500$ and $1,505$, respectively. Later, the GMM/DNN based acoustic models are trained using the training set by employing the Kaldi toolkit [@povey2011kaldi]. A $3$-gram language model (LM) is trained over the entire text data obtained from Sub-section \[subsec:text\_data\] after excluding those sentences that are used in testing. Therefore, $13071-1505=11,566$ number of sentences are used for training the LM. For developing the $3$-gram LM, we have employed the IRSTLM toolkit [@federico2008irstlm]. The evaluation results in terms of percentage word error rate (%WER) are given in Table \[eval\]. The DNN-based acoustic model with $3$-gram LM resulted in the best $\%WER$ score when compared to other models.
Parameter tuning
----------------
The context-dependent GMM acoustic models are trained by tuning the number of senones. After tuning, the number of senones is set to be $2500$. The Gaussian mixtures per senone are set to be $8$ in all the cases. The DNN based acoustic models are trained with $5$ hidden layers and $1024$ nodes with *tanh* as non-linearity function in each of the hidden layers. These models are trained with $20$ epochs and mini-batch size of $128$.
**Model** **Features** **%WER**
----------- ------------------ ----------
Mono MFCC 53.51
Tri1 MFCC 33.52
Tri2 MFCC + LDA 32.73
Tri3 MFCC + LDA + SAT 27.20
DNN MFCC + LDA + SAT 25.40
: Evaluation of Hinglish code-switching speech corpus in context od ASR task. The performance results in terms of percentage word error rate (%WER) are reported.[]{data-label="eval"}
Conclusion {#sec:conclusion}
==========
In this work, the procedure followed to develop a Hinglish code-switching speech database has been presented. It contains $7,005$ utterances spoken by $71$ speakers from different parts of India. The database has been validated by developing an ASR system. The collection of the database is still in progress.
Acknowledgment
==============
The authors wish to acknowledge with gratitude for the financial assistance received towards data collection from an ongoing project grant no. 11(18)/2012-HCC(TDIL) from the Ministry of Electronics and Information Technology, Govt. of India.
[^1]: https://shoutmehindi.com
[^2]: https://notesinhinglish.blogspot.in
[^3]: https://www.techyukti.com
[^4]: http://www.learncpp.com
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present new spectroscopic observations of the early type, double-lined eclipsing binary V1441Aql. The radial velocities and the available photometric data obtained by $ASAS$ is analysed for deriving the parameters of the components. The components of V1441Aql are shown to be a B3IV primary with a mass M$_p$=8.02$\pm$0.51 M$_{\odot}$ and radius R$_p$=7.33$\pm$0.19 R$_{\odot}$ and a B9 III secondary with a mass M$_s$=1.92$\pm$0.14 M$_{\odot}$ and radius R$_s$=4.22$\pm$0.11 R$_{\odot}$. Our analyses show that V1441Aql is a double-contact system with rapidly rotating components. Based on the position of the components plotted on the theoretical Hertzsprung-Russell diagram, we estimate that the ages of V1441Aql is about 30Myr, neglecting the effects of mass exchange between the components. Using the UBVJHK magnitudes and interstellar absorption we estimated the mean distance to the system V1441Aql as 550$\pm$25pc.'
address:
- 'Ege University, Science Faculty, Astronomy and Space Sciences Dept., 35100 Bornova, İzmir, Turkey '
- 'Osservatorio Astrofisico di Catania, Via S. Sofia 78, 95123 Catania, Italy'
author:
- 'Ö. Çak[i]{}rl[i]{}'
- 'C. Ibanoglu'
- 'E. Sipahi'
- 'A. Frasca'
- 'G. Catanzaro'
title: Analysis of the Massive Eclipsing Binary V1441Aql
---
stars: binaries: eclipsing – stars: fundamental parameters – stars: binaries: spectroscopic – stars:V1441Aql
Introduction {#sec:intro}
============
High-mass stars are much less frequent than intermediate- or low-mass stars due to both the star formation process, which gives rise to an initial mass function declining with the mass [e.g. @Salpeter1955; @Kroupa2001], and to their shorter evolutionary times. However, high-mass stars are very important because they can affect their surroundings with their winds, their strong radiation fields, and their catastrophic death as supernovae, chemically enriching their environment and triggering star formation. They usually form within the dense cores of stellar clusters and/or associations where dynamical interactions play an important role. It is widely believed that massive stars may be the product of collisions between two or more intermediate-mass stars. This idea is supported by the fact that a large fraction of massive stars harbour close companions, as failed mergers. It has been recently estimated that at least 50% of massive stars are member of binary or multiple star system [@San12]. This lucky occurrence allows to directly measure the masses by means of their radial velocity (RV) curves. In many cases, spectral lines of both components are visible (SB2 systems), allowing to derive the orbital parameters like the period, $P_{\rm orb}$, the projected semi-major axes, $a_{1,2}\sin i$, and the masses, $M_{1,2}\sin^3i$, apart from the factor $\sin^3i$. If our line-of-sight is close to the orbital plane and fractional radii of the components are not too small the stars display mutually eclipses. The orbital inclination and fractional radii of the component stars can be determined by the analysis of photometric light curves. Therefore, eclipsing binaries are unique targets for determining the masses and radii from their combined light curves and radial velocities analyses. Nevertheless, absolute radii were measured only for a rather small number of early-type B-stars which are members of eclipsing binary systems [@Hil04; @Tor10; @Iba03; @Iban13]. Thus, we started a systematic observing program devoted to the spectroscopic study of close eclipsing binary systems with at least one hot component.
V1441Aquilae
------------
The eclipsing character of V1441Aquilae (HD 177624; BD+09$^{o}$3979; HIP 93732; V=6.90, B-V=0.19 mag) was discovered by the $Hipparcos$ satellite mission [@Per97], the primary eclipse having an amplitude of 0.09mag. The depth of the secondary minimum is nearly half of the primary one. V1441Aql was first recognized as a double-lined spectroscopic binary by @Hil80. They classified it as a B3V star with semi-amplitudes of radial velocities of about 82 and 196 [kms$^{-1}$]{}. The first eclipse light curve (LC) was roughly revealed by the $Hipparcos$ observations. @Kaz99 designated it as V1441Aql in the [74th Name-list of Variable Stars]{}, and classified it as an EB. In recent years, large-scale photometric surveys such as All Sky Automated Survey (ASAS, @Poj02) have been conducted with the main aim of looking for transiting exoplanets. The valuable by-product of these searches has been the very large number of well sampled eclipsing binary LCs. One of the eclipsing binaries observed at the Las Campanas Observatory as part of the ASAS was V1441Aql. Although the phase coverage is very good the photometric accuracy of the observations is only around 0.05 mag.
The orbital period of V1441Aql was determined from the spectroscopic observations as 2.374148 days by @Hil80. The first photometric observations of the system were made by the $Hipparcos$ mission and an orbital period of about 2.374 days was estimated [@Lef09]. The $ASAS$ photometry permits a redetermination of the orbital period of the system. A periodogram analysis performed with PERIOD04 [@Len05], which was applied to all the data obtained by $ASAS$. We derive the following ephemeris
$$Min I(HJD)=2\,455\,100.4153(32)+2^d.3741896(16) \times E$$
where the standard deviations in the last significant digits are given in parentheses.
Observation
===========
Optical spectroscopic observations of V1441Aql was obtained at the TUBITAK National Observatory using the Turkish Faint Object Spectrograph Camera (TFOSC)[^1] attached to the 1.5m telescope. The observations were made from July 22, 2012 to August 3, 2013, under good seeing conditions. Further details on the telescope and the spectrograph can be found at http://www.tug.tubitak.gov.tr. The wavelength coverage of each spectrum was 4000-9000 Å in 11 orders, with a resolving power of $\lambda$/$\Delta \lambda$ $\sim$7000 at 6500 Å. The average signal-to-noise ratio (S/N) was $\sim$120. We also obtained high S/N spectra of early type standard stars 1Cas (B0.5IV), HR153 (B2IV), $\tau$Her (B5IV), 21Peg (B9.5V) and $\alpha$Lyr (A0V) which were used as templates in derivation of the radial velocities.
The echelle spectra were extracted from the raw images following standard reduction steps involving electronic bias subtraction, flat field division, cosmic rays removal, optimal extraction of the echelle orders, and wavelength calibration thanks to the emission lines of a Th-Ar lamp. The reduction was performed using tasks of the IRAF package[^2]
Radial velocities and atmospheric parameters
============================================
The time-resolved spectroscopic dataset consists of 17 observations for V1441Aql. We have measured radial velocities (RVs) from the spectra, focusing on spectral segments containing the He[i]{} $\lambda$5876 (order4) and $\lambda$6678 (order3) lines which are the most prominent un-blended features in our spectra, apart from the Balmer lines. We have employed the standard cross-correlation method for measuring the velocities of the component stars of the systems. The numerical cross-correlation technique [@Sim74; @Ton79] is a standard approach for measuring RVs from the spectra of close binary systems. Cross-correlation analyses were made using the spectra of $\tau$Her and 21Peg as templates. The principle spectral features showing splitting due to binarity were the He[i]{} lines at $\lambda\lambda$5876 and 6678. We used also order 9, containing the He[i]{} $\lambda$4471 line, for a few measurements of the radial velocities. The spectra taken close to the conjunctions, which display no double-lined feature, were disregarded. The Balmer lines were not used in the measurements of radial velocities due to their extremely broad profiles.
We obtained 17 radial velocities for each component of V1441Aql. The average radial velocities and their associated standard errors derived from the spectral segments containing He[i]{} $\lambda\lambda$4471, 5876, and 6678 lines are presented in Table1, along with the observation date and orbital phase. The mean error of radial velocities is 3.6kms$^{-1}$ for the primary, and 8.5 kms$^{-1}$ for the secondary star of V1441Aql. The RVs are plotted against the orbital phase in Fig.1, where the filled squares represent the primaries and the empty squares the secondaries, respectively. Examination of the $ASAS$ light curve show no evidence for any eccentricity in the orbits of both systems. Therefore, we have assumed circular orbits and analysed the RVs using the [RVSIM]{} software programme [@Kan07]. Final orbital parameters are presented in Table2.
Intermediate-resolution optical spectroscopy permits us to derive most of the fundamental stellar parameters, such as the projected rotational velocity ($v\,sin\,i$), spectral type (S$_p$), luminosity class, effective temperature ($T_{\rm eff}$), surface gravity ($log~g$), and metallicity (\[Fe/H\]).
The width of the cross-correlation function (CCF) is a good tool for the measurement of projected rotational velocity ($v\sin i$) of a star. We use a method developed by @Pen96 to estimate the $v\sin i$ of each star composing the investigated SB2 systems from its CCF peak by a proper calibration based on a spectrum of a narrow-lined star with a similar spectral type. Per each system, the rotational velocities of the components were obtained by measuring the FWHM of the CCF peak related to each component in five high-S/N spectra acquired near the quadratures, where the spectral lines have the largest Doppler-shift. The CCFs were used for the determination of $v\,sin\,i$ through a calibration of the full-width at half maximum (FWHM) of the CCF peak as a function of the $v\,sin\,i$ of artificially broadened spectra of slowly rotating standard star (21Peg, $v\sin i \simeq$14 kms$^{-1}$, e.g., @Roy02) acquired with the same setup and in the same observing night as the targets systems. The limb darkening coefficient was fixed at the theoretically predicted values, 0.42 for both systems [@Van93]. We calibrated the relationship between the CCF Gaussian width and $v\,sin\,i$ using the @Con77 data sample. This analysis yielded projected rotational velocities for the components of V1441Aql as $v_{\rm P}\sin i$=196 kms$^{-1}$, and $v_{\rm S}\sin i=101$ kms$^{-1}$. The mean deviations were 4 and 7 kms$^{-1}$, for the primary and secondary, respectively, between the measured velocities for different lines.
{width="9.5cm"}
-------------- -------- ------------- ---------- ------------- ---------- --
HJD 2400000+ Phase
$V_{\rm P}$ $\sigma$ $V_{\rm S}$ $\sigma$
56131.3242 0.2137 $-54$ 3 $ 167$ 8
56132.2925 0.6216 $ 38$ 3 $-157$ 8
56133.4035 0.0895 $-33$ 3 $ 89$ 7
56134.4926 0.5483 $ 11$ 2 $ -54$ 8
56135.4443 0.9491 $ 11$ 2 $ -66$ 5
56136.2708 0.2972 $-51$ 5 $ 166$ 7
56137.2913 0.7270 $ 44$ 5 $-189$ 9
56147.3444 0.9614 $ 9$ 3 $ -71$ 7
56167.4232 0.4185 $-31$ 2 $ 77$ 7
56506.4817 0.2287 $-61$ 3 $ 167$ 4
56506.4993 0.2361 $-55$ 3 $ 178$ 5
56507.2803 0.5651 $ 12$ 4 $ -60$ 9
56507.2957 0.5716 $ 17$ 4 $ -67$ 11
56507.3107 0.5779 $ 21$ 4 $ -91$ 8
56507.4161 0.6222 $ 27$ 4 $-141$ 9
56507.4338 0.6297 $ 37$ 6 $-137$ 8
56507.4494 0.6363 $ 33$ 5 $-133$ 8
-------------- -------- ------------- ---------- ------------- ---------- --
: Heliocentric radial velocities of V1441Aql. The columns give the heliocentric Julian date, orbital phase and the radial velocities of the two components with the corresponding standard deviations.
-------------------------- ----------------- ----------------- -- -- -- -- -- -- --
Parameter
Primary Secondary
$k$ (kms$^{-1}$) 44$\pm$2 184$\pm$4
$V_\gamma$ (kms$^{-1}$)
Average O-C (kms$^{-1}$) 3.6 7.5
$a\sin i$ ($R_{\odot}$) 2.064$\pm$0.002 8.631$\pm$0.006
$M\sin^3i$ ($M_{\odot}$) 2.353$\pm$0.140 0.563$\pm$0.041
-------------------------- ----------------- ----------------- -- -- -- -- -- -- --
: Results of the radial velocity analysis for V1441Aql.
We also performed a spectral classification for the components of the systems using COMPO2, an IDL code for the analysis of high-resolution spectra of SB2 systems written by one of us [see, e.g., @Frasca2006] and adapted to the TFOSC spectra. This code searches for the best combination of two reference spectra able to reproduce the observed spectrum of the system. We give, as input parameters, the radial velocities and projected rotational velocities $v\sin i$ of the two components of each system, which were already derived. The code then finds, for the selected spectral region, the spectral types and fractional flux contributions that better reproduce the observed spectrum, i.e. which minimize the residuals in the collection of difference (observed$-$composite) spectra. For this task we used reference spectra taken from the @Val04 $Indo-U.S.\ Library\ of\ Coude\ Feed\ Stellar\ Spectra$ (with a a resolution of $\approx$1Å) that are representative of stars with spectral types from late-O type to early-A, and luminosity classes V, IV, and III. The atmospheric parameters of these reference stars were recently revised by @Wu2011.
We selected 198 reference spectra spanning the ranges of expected atmospheric parameters, which means that we have searched for the best combination of spectra among 39204 possibilities per each spectrum. The observed spectra of V1441Aql in the $\lambda\lambda$6525–6720 spectral region were best represented by the combination of HD179761 (B8 II-III) and HD182568 (B3 IV). However, we have adopted, for each component, the spectral type and luminosity class with the highest score in the collection of the best combinations of templates, where the score takes into account the goodness of the fit expressed by the minimum of the residuals. We have thus derived a spectral type of B3 subgiant for the primary and B9 giant for the secondary star of V1441Aql, with an uncertainty of about 1 spectral subclass. The effective temperature and surface gravity of the two components of each system are obtained as the weighted average of the values of the best spectra at phases near to the quadratures combinations of templates adopting a weight $w_i=1/\sigma_i^2$, where $\sigma_i$ is the average of residuals for the $i$-th combination. The standard error of the weighted mean was adopted for the atmospheric parameters. Both stars appear to have a solar metallicity, within the errors. The atmospheric parameters obtained by the code and their standard errors are reported in Table3. The observed spectra of V1441Aql at phases near to the quadratures are shown in Fig.2 together with the combination of two reference spectra which gives the best match.
------------------------ ---------------- ----------------- -- -- -- -- -- -- --
Parameter
Primary Secondary
Spectral type B(3$\pm$0.5)IV B(9$\pm$0.5)III
$T_{\rm eff}$ (K) 18760$\pm$950 11670$\pm$650
$\log~g$ ($cgs$) 3.77$\pm$0.05 3.95$\pm$0.17
$v\sin i$ (kms$^{-1}$) 196$\pm$4 101$\pm$7
------------------------ ---------------- ----------------- -- -- -- -- -- -- --
: Spectral types, effective temperatures, surface gravities, and rotational velocities of the components estimated from the spectra of V1441Aql.
![Two spectra of V1441Aql near opposing quadrature phases, at phases 0.294 and 0.728. The wavelength limits are 6525-6715 Å, which include the H$\alpha$ and He[i]{} $\lambda$6678 lines. The deeper lines in each spectra refer to the primary star. Vertical axis is the normalized flux.](fig2.ps){width="12.5cm"}
Analyses of the light curve
===========================
We extracted the V-band light curve for V1441Aql (ASAS190509+0938.5) from the All Sky Automated Survey database [@Poj02]. The light curve composed of 375 photometric points, has a very good phase coverage and displays two equally spaced minima whose depths are about 0.09 and 0.04 mag. The observed light curve is reminiscent of an ellipsoidal variable with tidally distorted stars. V1441Aql was also observed by the $Hipparcos$ satellite [@Per97]. 69 Hp magnitudes were obtained for V1441Aql. The light curve is similar to that obtained by $ASAS$ but its shape is poorly defined due to small number of measurements. @Mal06 estimate an Hp magnitude of 6.91 at the light maximum and a primary minimum with a depth of 0.09 mag in the $Hipparcos$ light curve. The light curve obtained by the $ASAS$ is shown in Fig.3, where the vertical axis is the normalized flux.
The effective temperature of the primary star and the mass-ratio of the system are key parameters needed for the analysis of the ligth curve. The effective temperatures of both components have already been determined from the spectra as 18760$\pm$950 K and 11670$\pm$650K as well as a mass-ratio of 0.239$\pm$0.009 was derived from the radial velocity curve solution. Another approach to estimating the temperature of the primary star is to use the observed colours of the system. The apparent visual magnitude and colour indices were given as mean of nine measures by @Mor71 as $V$=6$^m$.90, ($U-B$)=$-$0$^m$.36, ($B-V$)=0$^m$.19 and ($V-R$)=0$^m$.23 with an uncertainty of $\pm$0.01mag. The quantity $Q$=($U-B$) $-$ 0.72$(B-V)$ of the Johnson’s UBV photometric system is independent of interstellar extinction. We compute the reddening-free index as $Q$=$-$0.497 for the system, which corresponds to a B4 star [@Hov04], consistent with the spectral classification from the spectra. The effective temperature for the primary star, derived from the spectra, corresponds to an intrinsic colour of ($B-V$)=$-$0$^m$.22. A preliminary analysis of the light curve yields a light ratio of $l_{s}$/$l_{p}$=0.15 for the V-passband. Using the intrisic colour of the primary star, the light ratio and observed colour of 0$^m$.19 we compute an intrinsic composite colour for the system as ($B-V$)=$-$0$^m$.21. Then we estimate the reddening for the system as $E_{(B-V)}$=0$^m$.40.
Parameters V1441Aql
------------------------------- ------------------- -- -- --
$i^{o}$ 41.65$\pm$0.58
$T_{\rm eff_1}$ (K) 18760\[Fix\]
$T_{\rm eff_2}$ (K) 11650$\pm$300
$\Omega_1$ 2.327$\pm$0.030
$\Omega_2$ 2.327$\pm$0.030
$r_1$ 0.4557$\pm$0.0065
$r_2$ 0.2620$\pm$0.0033
$\frac{L_{1}}{(L_{1}+L_{2})}$ 0.8721$\pm$0.0035
$\sum(O-C)^{2}$ 0.0370
$N$ 375
$\sigma$ 0.0100
: Final solution parameters for the double-contact model of V1441Aql.
Since the $Hipparcos$ data contain only 69 H$_{p}$ points with uncertainties larger than the $ASAS$ ones, we did not attempt to analyze them for determination of the orbital parameters. We started to analyze the $ASAS$ light curve using the Wilson-Devinney code [hereafter WD; e.g., @Wil71; @Wil79; @Wil06] as implemented in the software [phoebe]{} [@Prs05]. The WD code is widely used for determination of the orbital parameters of the eclipsing binaries. To run the code we need some initial elements. The logarithmic limb-darkening coefficients were interpolated in effective temperature and surface gravity from the coefficients tabulated by @Van93. The initial limb-darkening coefficients were taken as $x_1$=0.45, $y_1$=0.24, $x_2$=0.59, $y_2$=0.29 for the hotter and cooler star, respectively. They are updated at every iteration. The gravity-brightening coefficients $g_1$=$g_2$=1.0 and albedos $A_1$=$A_2$=1.0 were fixed for both components, as appropriate for stars with radiative atmospheres. We started with the $Mode-2$ option of the WD code, designed for detached binaries, for the analysis of light curves as described by @Wil06. The adjustable parameters in the light curve fitting were the orbital inclination $i$, the effective temperature of the secondary star $T_{\rm eff_2}$, the monochromatic luminosity of the primary L$_1$, and the zero-epoch offset. We could not obtain a good fit of the light curve with a detached configuration. This led us to try with a semi-detached configuration. We thus applied $Mode-5$ option of the WD code, in which the secondary star fills its lobe. The convergence was obtained with the primary star overfilling its lobe. Then we tried the $Mode-3$ (for overcontact systems, the stars are in geometrical contact without being in thermal contact). This solution resulted in that both stars are well inside their lobes. Then we applied $Mode-4$ (primary star fills its lobe) but we arrived at a result that the secondary star is overfilling its lobe.
Finally, the solution with the $Mode-6$ option (for double contact systems, DCSs) resulted in acceptable parameters. @Wil79 defined DCSs as a binaries in which both stars fill their Roche lobes and at least one rotates faster than synchronously, so that the components do not touch, even at one point. The fits of the computed light curves obtained with $Mode-6$ to the observations are satisfactory with the smallest sum of residuals squared. The out-of-eclipse part of the observed light curve is now better reproduced. We have already measured the projected rotational velocities from the spectral lines for the components of V1441Aql. Preliminary analysis using the $Mode-6$ of WD code yielded orbital parameters for the component stars. We have computed projected rotational velocities of the components. A comparison of the observed rotational velocities with those computed velocities shows that the stars rotate faster than the synchronous values. Therefore we repeated our analysis taking $F_1$=1.8 and $F_2$=1.6.
The parameters of our final solutions are listed in Table4. The orbital inclination of the system, effective temperatures, fractional mean radii (equivalent volume) of the components and fractional luminosity of the primary star in the V band were given in this table. The uncertainties assigned to the adjusted parameters are the internal errors provided directly by the code. The squared sum of residuals, $\sum(O-C)^{2}$, the number of data points, $N$, and the standard deviation, $\sigma$, of the observed light curve are quoted in the last three lines of Table4, respectively. The computed light curve is overplotted to the observations in Fig.3. This solution indicates for V1441Aql a grazing eclipse lasting about half an hour.
{width="12.5cm"}
Results and discussion
======================
Based on the results of radial velocities and light curves analyses we have calculated the physical properties of the V1441Aql. For this purpose, we used the $JKTABSDIM$ code developed by @Sou05. This code is now widely used for derivation of the absolute parameters of the eclipsing binary stars’ components. It calculates complete error budgets using a perturbation algorithm. The fundamental stellar parameters for the components such as masses, radii, luminosities and their standard deviations have been derived using this code. The astrophysical parameters of the components, and other properties for the components of V1441Aql are presented in Table5.
The separation between the components was found to be 16.09$\pm$0.34 R$_{\odot}$ for V1441Aql. The masses were measured to precision of about 6–7% for the compoents. On the other hand the radii of the components have been derived with a precision of better than 5%. The accuracy of any parameter of an eclipsing binary system depends mainly on the coverage of the both spectroscopic and photometric observations and their precision. In addition, the light curve solutions are more accurate for totally eclipsing systems. The light curve of V1441Aql shows instead a very shallow grazing eclipse. Despite these drawbacks, the physical parameters of the components of system could be determined with sufficient precision. We note that the effective temperatures of the secondary stars derived from the spectra are in good agreement with those obtained from the light curve analyses.
The luminosities and absolute bolometric magnitudes are calculated directly from the radii and and effective temperatures of the components. The effective temperature of 5777 K and the absolute bolometric magnitude of 4.74 mag were adopted for the Sun [e.g., @Dri00]. The bolometric corrections were interpolated from the tables of @Flo96. The V-band magnitudes of the systems at out-of-eclipse phases are taken as 6.90 and 7.32 for V1441Aql. We have calculated the absolute visual magnitudes for the components using the fractional luminosities and bolometric corrections given in Table4 and 5. Combining these values with the interstellar absorption of 1.27mag for V1441Aql and we have estimated the distances to the systems as 550$\pm$25.
In the log $T_{\rm eff}$–$\log L/L_{\odot}$ (left panel) and log $T_{\rm eff}$-logg planes (right panel) we have plotted the positions of the components (Fig.4), with 1-$\sigma$ error bars. The filled and empty squares refer to the components of V1441Aql. The evolutionary tracks and isochrones for the non-rotating single stars with solar composition are taken from @Eks12.
The most striking result from analyses of the radial velocities and light curve is the double-contact configuration of V1441Aql. Double-contact systems (DCSs) were described in detail by @Wil79. He showed that in the double-contact systems the massive stars do not rotate synchronously. @Wil85 estimated for the first time the rotation rate of a star in an eclipsing binary using photometric observations. They showed that the primary star of the double-contact system RZSct rotates about at 6.7 times the synchronous value. Very recently, @Ter14 discussed the double-contact nature of the eclipsing binary system TTHer. They have already shown that the observed radial velocities and light curves could only be represented with a non-synchronous rotation of the primary star. Their analyses resulted in that the more massive star rotates at 1.25 times the synchronous value. Our measured projected rotational velocities for V1441Aql show that both components rotate faster than the synchronous values, amounting to about 1.7$v_{syn}\sin i$.
The comparison with the evolution of single stars, shown in Fig.4, indicates that the age of V1441Aql is longer than 30 Myr. Both stars have evolved away from the main sequence, i.e. exhausted the hydrogen in their cores. The primary star is probably in the shell hydrogen burning phase. While the more massive star appears to be crossing the Hertzsprung gap in the HR diagram, the less massive star is close to the base of giant branch. In addition, the secondary component is seen as an over-luminous and hotter star with respect to its mass. This may indicate that the secondary star lost the outer envelope during the evolution in past. At present, it has a larger radius and therefore has a higher luminosity, similar to the donors in the semi-detached Algol-type binaries.
We calculated the synchronization time for V1441Aql as about 2 Myr using the hydrodynamic damping formalism of @Tas97. The Tassouls’ synchronization time is about 15 times shorter than the estimated age of the binary. Both observational [@Gla08] and theoretical [@Der10] studies showed that the mass-gainers of the semi-detached close binaries rotate faster than synchronous. There is a consensus that the most plausible reason for this asynchronous rotation of the primaries is the high values of the mass transfer rate. @Der10 recently showed a relation between mass transfer rate and asynchronous rotation velocity of the gainers in the Algol-type systems. The double-contact nature of V1441Aql with asynchronous components indicates that it has passed through rapid phase of mass transfer.
------------------------- ------------- ----------------- ----------------- -- -- -- -- -- --
Parameter Units
Primary Secondary
Mass M$_{\odot}$ 8.02$\pm$0.51 1.92$\pm$0.14
Radius R$_{\odot}$ 7.33$\pm$0.19 4.22$\pm$0.11
$T_{\rm eff}$ K 18760$\pm$950 11650$\pm$300
$\log(L/L_{\odot})$ 3.779$\pm$0.091 2.471$\pm$0.047
${\ensuremath{\log g}}$ $cgs$ 3.611$\pm$0.016 3.471$\pm$0.023
$Sp.Type$ B3IV B9III
$M_{bol}$ mag $-4.70\pm$0.23 $-1.43\pm$0.12
$BC$ mag $-1.75$ $-0.61$
$M_{V}$ mag $-2.95\pm$0.11 $-0.82\pm$0.07
$v\sin i_{\rm cal}$ kms$^{-1}$ 104$\pm$3 60$\pm$2
$v\sin i_{\rm obs}$ kms$^{-1}$ 196$\pm$4 101$\pm$7
$d$ pc 560$\pm$32 548$\pm$22
------------------------- ------------- ----------------- ----------------- -- -- -- -- -- --
{width="14cm"}
Summary
=======
V1441Aql is close eclipsing binaries containing high-mass stars. We carried out spectroscopic observations of system. The atmospheric parameters of the components in the eclipsing pairs have been determined from their spectra. The spectra were analyzed using cross-correlation for measuring the radial velocities of both components and with an ad-hoc code for deriving their atmospheric parameters. Moreover, the $ASAS$ and $HIPPARCOS$ light curves were modeled using the WD code. The physical parameters for the system is measured to accuracies of 6-7% in mass, and 5% in radius. V1441Aql seems to a double-contact system with asynchronously-rotating components. The distances to the systems V1441Aql is estimated as 550$\pm$25pc. A comparison of physical parameters of the components with the theoretical models of single stellar evolution models has been made. An age of about 30 Myr is estimated for V1441Aql.
Acknowledgments {#acknowledgments .unnumbered}
===============
We thank to TÜB[İ]{}TAK National Observatory (TUG) for a partial support in using RTT150 telescope with project number [11BRTT150-198]{}. We thank to EB[İ]{}LTEM Ege University Research Center for a partial support with project number [2013/BIL/018]{}. We also thank to the staff of the Bak[i]{}rl[i]{}tepe observing station for their warm hospitality. This study is supported by Turkish Scientific and Technology Council under project number [112T263]{}. This research was also partly supported by the Scientific Research Projects Coordination Unit of Istanbul University. Project number 3685. We thank Canakkale Onsekiz Mart University Astrophysics Research Center and Ulupinar Observatory together with Istanbul University Observatory Research and Application Center for their support and allowing use of IST60 telescope. This work was partially supported by the Italian [*Ministero dell’Istruzione, Università e Ricerca*]{} (MIUR). The following internet-based resources were used in research for this paper: the NASA Astrophysics Data System; the SIMBAD database operated at CDS, Strasbourg, France; and the ar$\chi$iv scientific paper preprint service operated by Cornell University.
Conti, P. S., & Ebbets, D., 1977, ApJ, 213, 438 Dervi[ş]{}o[ǧ]{}lu, A., Tout, C. A., & Ibano[ǧ]{}lu, C. 2010, MNRAS, 406, 1071 Drilling J. S., Landolt A. U., 2000, Allen’s astrophysical quantities, 4th ed. Edited by Arthur N. Cox. ISBN: 0-387-98746-0. Publisher: New York: AIP Press; Springer, 2000, p.381 Ekstr[ö]{}m S., Georgy C., Eggenberger P., et al., 2012, A&A, 537, A146 Flower, P. J., 1996, ApJ, 469, 355 Frasca, A., Guillout, P., Marilli, E. 2006, A&A, 454, 301 Glazunova, L. V.,Yushchenko, A. V., Tsymbal, V. V., et al. 2008, AJ, 136, 1736 Hilditch, R. W., 2004, ASPC, 318, 198 Hill, G., & Fisher, W. A., 1980, PDAO Victoria, 15, 409 Hovhannessian R. K., 2004, Ap, 47, 499 Ibanoglu C., Cak[i]{}rl[i]{}, [Ö]{}., & Sipahi, E., 2013a, MNRAS, 436, 750 Ibanoglu, C., [Ç]{}ak[i]{}rl[i]{}, [Ö]{}., & Sipahi, E., 2013b, NewA, 25, 68 Kane, S. R., Schneider, D. P., & Ge, J., 2007, MNRAS, 377, 1610 Kazarovets E. V., Samus N. N., Durlevich O. V., et al., 1999, IBVS, 4659, 1 Kroupa, P. 2001, MNRAS, 322, 231 Lef[è]{}vre, L., Marchenko, S. V., Moffat, A. F. J., & Acker, A., 2009, A&A, 507, 1141 Lenz, P., & Breger, M., 2005, Communications in Asteroseismology, 146, 53 Malkov, O. Y., Oblak, E., Snegireva, E. A., & Torra, J., 2006, A&A, 446, 785 Moreno, H., 1971, A&A, 12, 442 Penny, L. R., 1996, ApJ, 463, 737 Perryman M. A. C., Lindegren L., Kovalevsky J. et al., 1997, A&A, 323, L49 Pojmanski, G., 2002, AcA,52, 397 Prša A., Zwitter T., 2005, ApJ, 628, 426P Royer F., Gerbaldi M., Faraggiana R., & Gomez A. E. 2002, A&A, 381, 105 Sana, H., Dunstall, P. R.,H[é]{}nault-Brunet, V., et al. 2012, Proceedings of a Scientific Meeting in Honor of Anthony F. J. Moffat, ASPC, 465, 284 Salpeter, E. E. 1955, ApJ, 121, 161 Simkin, S. M., 1974, A&A, 31, 129 Southworth, J., Maxted, P. F. L., & Smalley, B., 2005, A&A, 429, 645 Tassoul, M., & Tassoul, J.-L. 1997, ApJ, 481, 363 Terrell, D., & Nelson, R. H., 2014, ApJ, 783, 35 Tonry, J., & Davis, M. 1979, AJ, 84, 1511 Torres G., Andersen J., & Gim[é]{}nez A., 2010, A&ARv, 18, 67 Valdes, F., Gupta, R., Rose, J. A., Singh, H. P., & Bell, D. J., 2004, ApJS, 152, 251 van Hamme, W., 1993, AJ, 106, 2096 Wilson, R. E., 1979, ApJ, 234, 1054 Wilson, R. E., 2006, Astrophysics of Variable Stars, ASPC, 349, 71 Wilson, R. E., & Devinney, E. J., 1971, ApJ, 166, 605 Wilson, R. E., van Hamme, W.,& Pettera, L. E., 1985, ApJ, 289, 748 Wu, Y., Singh, H. P., Prugniel, P., Gupta, R., & Koleva, M. 2011, A&A, 525, 71
[^1]: http://http://tug.tubitak.gov.tr/tr/teleskoplar/tfosc
[^2]: IRAF is distributed by the National Optical Astronomy Observatory, which is operated by the Association of Universities for Research in Astronomy, Inc. (AURA), under cooperative agreement with the National Science Foundation.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Short-hard $\gamma$-ray bursts (SHBs) may arise from gravitational wave (GW) driven mergers of double neutron star (DNS) systems. DNSs may be “primordial” or can form dynamically by binary exchange interactions in globular clusters during core-collapse. For primordial binaries, the time delay between formation and merger is expected to be short, $\tau\sim0.1$ Gyr, implying that the redshift distribution of merger events should follow that of star-formation. We point out here that for dynamically formed DNSs, the time delay between star-formation and merger is dominated by the cluster core-collapse time, rather than by the GW inspiral time, yielding delays comparable to the Hubble time. We derive the redshift distribution of merger events of dynamically formed DNSs, and find it to differ significantly from that typically expected for primordial binaries. The observed redshift distribution of SHBs favors dynamical formation, although a primordial origin cannot be ruled out due to possible detection biases. Future red-shift observations of SHBs may allow to determine whether they are dominated by primordial or dynamically formed DNSs.'
author:
- 'Clovis Hopman, Dafne Guetta, Eli Waxman and Simon Portegies Zwart'
title: |
The redshift distribution of short $\gamma$-ray bursts from dynamically formed\
neutron star binaries
---
Introduction
============
Observations of $\gamma$-ray bursts (GRBs) indicate that they divide into two classes (Kouveliotou et al. [@Kea93]). GRBs of one class are of relatively long ($\gtrsim2\!-\!200$ sec) duration and have softer spectra. Long-soft GRBs occur in star forming galaxies with high redshift $z$ (van Paradijs et al. [@Par97]), and their association in several cases with type Ibc SNe ([@G98; @Stanek03; @Hjorth03; @Malesani04; @Campana06]) suggests that they are the result of core collapse SN explosions of massive stars ([@Woosley93; @Pac98; @MacFadyen99]). The second class of GRBs have short duration ($<2$ sec) and harder spectra. Afterglows of short-hard GRBs (SHBs) have only recently been observed (e.g., Gehrels et al.; [@Gea05]; Bloom et al. [@Bea05]; Berger et al. [@BPea05]; Fox et al. [@Gea05]), and this has led to the first identifications of SHB host galaxies. In contrast to long GRBs, SHBs were found to occur in at least some cases in elliptical galaxies with very low star formation rates (SFRs), of order $\lesssim0.1{M_{\odot}}{\,{\rm yr^{-1}}}$. It is therefore unlikely that the progenitors of SHBs are also massive stars, since these have very short ($\sim\,{\rm few}\,{{\,\rm Myr}}$) life times.
The gravitational wave (GW) driven merger of a double neutron star (DNS) may lead to a SHB (Goodman 1986; Paczyński 1986; Eichler et al. [@ELP89]; Narayan, Paczynski & Piran [@NPP92]). If the time-lag $\tau$ between DNS formation and the merger is large (comparable to a Hubble time), it forms a natural explanation why SHBs occur in galaxies where the SFR is very low.
We consider two formation mechanisms for DNSs. If two massive stars are born as a binary system, a DNS may form after the super nova (SN) explosion of both components (“primordial” DNSs). Population synthesis models of massive binaries yield typically short merger times ($\tau\!\approx\!0.1{{\,\rm Gyr}}$), implying that the SHB $z$-distribution function (DF) closely follows the SFR history (e.g., Bloom, Sigurdsson & Pols [@BSP99]; Belczynski et al. [@Bel01], [@Bel06]). The kick-velocity of the DNS is low ($\lesssim50{\rm km\,s^{-1}}$; Dewi, Podsiadlowski & Pols [@DPP05]), and the spatial distribution of SHB mergers should follow the light distribution of their host galaxy (Portegies Zwart & Yungelson [@PY98]; Bloom et al. [@BSP99]).
Another possibility is that at the moment of star formation the neutron stars (NSs) are [*not*]{} in the same binary system, but one of the NSs is in a binary with a low mass main sequence (MS) star. In globular clusters (GCs), such binaries are likely to have an exchange interaction with a single neutron star (Sigurdsson & Phinney [@Sig95]; Efremov [@Efr00]; Grindlay, Portegies Zwart & McMillan [@GPZM06], hereafter GPZM06), and thus form DNSs. A significant fraction ($\sim30\%$) of all NS mergers in the Universe may stem from such dynamically formed systems (GPZM06).
For exchange interactions to occur, the stellar density must be very high. The delay time for dynamically formed DNSs is therefore mainly determined by the time until core-collapse (CC) of GCs, which is typically comparable to the Hubble time (§\[s:tau\]). We show that the predicted $z$-DF of SHBs is different for primordial and dynamically formed DNSs, so that future $z$ observations may determine which formation channel (if any) is dominant. To date, the few SHBs with $z$-detection favor dynamical formation, although possible biases and small number statistics make a final conclusion at this point impossible (§\[s:Nz\]). We discuss our results in §\[s:discussion\].
Distribution of merger times {#s:tau}
============================
The delay time $\tau$ is a sum of the time $t_{cc}$ until the dynamical formation of a DNS during core-collapse, and the time $t_{\rm GW}$ until the DNS merges. Here we determine the resulting delay function $(dp/d\tau)_{\rm dyn}$ of dynamically formed DNSs.
GPZM06 performed scattering experiments with [scatter3]{} and [sigma3]{} in the [Starlab]{} environment (Portegies Zwart et al. [@Por01]) to determine the cross-section for the formation of NS binaries when a NS interacts with a (NS, MS) binary. We use the orbital parameters of the resulting binaries to determine the resulting DF of GW merger times $dp_{\rm GW}/dt_{\rm GW}$, which was found to be well fitted by $dp_{\rm GW}/dt_{\rm GW}\propto t_{\rm
GW}^{-1.1}$. Typical merger times $t_{\rm GW}$ are very short compared to the Hubble time.
The DNS formation rate per GC can be estimated as $\Gamma\!\approx\!
4\,{\rm Gyr^{-1}}\,n_6 v_1(N_{\rm pr}/20)$, where $n\!=\!10^6{\,\mathrm{pc}}^{-3}n_{6}$ is the number density of NS stars, $v\!=\!10{{\,\rm km}}{\,{\rm s^{-1}}}v_1$ is the velocity dispersion, and $N_{\rm pr}$ is the number of progenitor binaries containing one NS (GPZM06). Since $\Gamma\!\propto\! n$, DNSs form when the GC is very dense, i.e., during the CC phase. The delay between the formation of the GC and the SHB is the sum of the delay time $t_{\rm GW}$ and the time $t_{cc}$ between formation of the GC and CC.
For the DF of CC times we make the following assumptions: the CC times of the GCs in our Galaxy are representative for the whole Universe; the relation between $t_{cc}$ and the half mass relaxation time is only a function of concentration, and it is well approximated by the relation given by Quinlan ([@Q96]); the formation rate of GCs is proportional to the total SFR. We (conservatively) neglect repeated phases of CC (“gravothermal oscillations”; Sugimoto & Bettwieser [@SB83]; Makino [@M96]), which would lead to even lower redshifts.
We use the half-mass relaxation times given by Harris ([@H96]) to find the cumulative DF $P_{cc}(<t_{cc})$ for GCs with CC times smaller than $t_{cc}$ (Fig. \[f:tcc\]). The time $t_{cc}$ for a GC between formation and CC is somewhat uncertain. Results in the literature for single mass systems without binaries agree approximately (e.g., Quinlan [@Q96]; Joshi, Nave & Rasio [@Joh01]; Baumgardt [@Bau01]) . A spectrum of masses can significantly decrease $t_{cc}$ (Gürkan, Freitag, & Rasio [@GFR04]), while primordial binaries increase $t_{cc}$ (Fregeau et al. [@FGR03]). The estimate by Quinlan ([@Q96]) leads to most GCs having $t_{cc}> t_H$ (Fig. \[f:tcc\]), consistent with the observation that most $(\sim80\%)$ GCs in the Galaxy have not yet experienced CC.
The probability function of the total time $\tau=t_{\rm GW}+t_{cc}$ to be in the interval $(\tau, \tau+d\tau)$ is given by $$\label{eq:Ftau}
\left({dp\over d\tau}\right)_{\rm dyn}= {d\over d\tau}\int_0^{\tau}dt_{cc}{dp_{cc}\over dt_{cc}}\int_0^{\tau-t{cc}}dt_{\rm GW} {dp_{\rm GW}\over dt_{\rm GW}}.$$ The resulting delay function is shown in figure (\[f:delay\]); we weighed the CC times by the GC luminosity. Since the GW inspiral time DF diverges towards small times, $(dp/d\tau)_{\rm dyn}$ is mostly determined by the DF of of CC times. It is also shown in fig. (\[f:delay\]) that the delay time DF $(dp/d\tau)_{\rm dyn}$ is not modified significantly by choosing $t_{cc}$ which is smaller than that of fig. \[f:tcc\] by a factor of 10. This demonstrates the robustness of the conclusion that for dynamical DNSs, the delay times are comparable to the Hubble time.
In contrast to the delay function for primordial DNSs, we find that [*the delay function of dynamically formed DNSs grows with the delay time*]{}. For delay times shorter than 10 Gyr, the average delay time is $\bar{\tau}\approx6{{\,\rm Gyr}}$.
{width="1.0\columnwidth"}
{width="1.0\columnwidth"}
Redshift distribution {#s:Nz}
=====================
We derive the predicted $z$-DF of SHBs, which depends on the event rate of SHBs as a function of $z$, the luminosity DF of SHBs, the delay function, and the detection threshold.
The intrinsic $z$-distribution
------------------------------
For DNS mergers, the [*intrinsic*]{} (as opposed to observed) SHB rate is given by the convolution of the star formation rate ${\rm SFR}(z)$ with the distribution $dp(\tau)/d\tau$ of time delays,
$$\label{rSHB}
N_{\rm intr}(z)\propto \int_z^{\infty}dz'{dt\over dz'}{\rm
SFR}(z'){dp\over dt}\left[t(z)-t(z')\right]$$
(e.g. Guetta & Piran ([@GP05] \[hereafter GP05\]; Nakar, Gal-Yam & Fox [@NGF06]); $t(z)$ is the age of the Universe as function of $z$. We employ the SF2 model of Porciani & Madau (2001) $$\begin{aligned}
\label{SFR}
{\rm SFR}_{\rm PM}(z) \propto { 23 e^{3.4z}[\Omega_M(1+z)^3+\Omega_k(1+z)^2+\Omega_{\Lambda}]^{1/2} \over({e^{3.4z}+22}) (1+z)^{3/2}},\nonumber\\
&&\end{aligned}$$ and the Rowan-Robinson ([@Row99]) SFR,
$$\label{RR}
{\rm SFR}_{\rm RR}(z) \propto
\left\{ \begin{array}{ll}
10^{0.75 z} & z<1 \nonumber \\
10^{0.75} & z>1.
\end{array}
\right.$$
In figure (\[tdelay\]) we show the intrinsic SHB $z$-DF $N(z)$ for $(dp/d\tau)_{\rm
prim}\propto1/\tau$ (as possibly appropriate for primordial DNSs, Bloom et al. [@BSP99]; Belczynski et al. [@Bel06]) and $dp/d\tau=(dp/d\tau)_{\rm dyn}$ (Eq. \[\[eq:Ftau\]\]).
{width="1.0\columnwidth"}
The SHB luminosity function
---------------------------
For the [*observed*]{} $z$-DF, the DF of peak luminosities and the minimal observable flux $P_{\rm lim}$ are required. Here we follow the procedure outlined in Guetta, Piran & Waxman ([@GPW04]). We consider a broken power law peak luminosity function (LF) with lower and upper limits, $1/\Delta_1$ and $\Delta_2$, respectively: $$\label{Lfun} \Phi_o(L) d\log L =C_0 d\log L \left\{
\begin{array}{ll}
(L/L^*)^{-\alpha}; & L^*/\Delta_1 \!<\! L \!<\! L^* \\ (L/L^*)^{-\beta}; & L^* \!<\!
L \!<\! \Delta_2 L^*
\end{array}
\right.$$ where $C_0$ is a normalization constant. This is the “isotropic-equivalent" LF, i.e. it does not include a correction factor due to beaming. Following Schmidt ([@S01]) we approximate the effective spectral index in the observed range of 20 or 50keV to 300 keV as $-1.1$ ($N(E)\propto E^{-1.1}$), and we use $\Delta_{1,2}=(30,100)$ (GP05). Both values are chosen such that even if there are bursts less luminous than $L^*/\Delta_1$ or more luminous than $\Delta_2 L^*$ they will be only very few (less than about 1%) of the observed bursts outside the range $(L^{*}/\Delta_1,L^*
\Delta_2)$. We find $\alpha=0.6$ and $\beta=2$, and then constrain $L^{*}$ by comparing the predicted peak flux distribution with the one observed by BATSE (see Guetta et al. \[[@GPW04]\] and GP05 for details). The best fit values of $L^*$ are reported in Table (\[t:fit\]).
[lllll]{} Model & $L^*$ & P$_{\rm lim}$ &KS test$^a$& KS test$^a$\
& $[10^{51}{\,{\rm erg}}{\,{\rm s^{-1}}}]$ &$[{\rm ph\, cm^{-2}\, s^{-1}}]$ &(with z=0.7)&(no z=0.7)\
SF2-$1/\tau$ & $2 $ & 1 & 0.05 & 0.01\
RR-$1/\tau$ & $1.5$ & 1 & 0.09 & 0.03\
SF2-$(dp/d\tau)_{\rm dyn}$ & $0.3$ & 1 & 0.7 & 0.8\
SF2-$1/\tau$ & $2 $ & 2.5 & 0.15 & 0.05\
RR-$1/\tau$ & $1.5$ & 2.5 & 0.25 & 0.09\
SF2-$(dp/d\tau)_{\rm dyn}$ & $0.3$ & 2.5 & 0.4 & 0.6\
\[t:fit\]
The observed $z$ distribution
-----------------------------
The expected $z$-DF of the observed bursts is $$\label{redshift} N_{\rm obs}(z)= \frac{N_{\rm intr}(z)}{1+z} \frac{dV(z)}{dz}
\int_{L_{\rm min}(P_{\rm lim},z)}^{L_{\rm max}} \Phi_o(L)d\log L \
,$$ where $L_{\rm max}=\Delta_2 L^*=100L^*$ and $L_{\rm min}$ is the luminosity at $z$ corresponding to the minimum peak flux $P_{\rm lim}$ required for detection. We estimate that $P_{\rm lim}$ for SWIFT is similar to that for BATSE, $P_{\rm lim}\sim 1\,{\rm ph\, cm^{-2}\,
s^{-1}}$, based on the observation that the detection rate of GRBs by SWIFT ($\sim100{\,{\rm yr^{-1}}}$) is similar to that of BATSE (taking into account the different field of views and the fact that BATSE was triggering only 1/3 of the time), and that the fraction of SHBs is similar for both BATSE and SWIFT. Fig. (\[cdf\]) shows a comparison between the observed and the expected integrated $z$-DF of SHBs for the different models listed in table ( \[t:fit\]). Our model for dynamical DNS mergers fits the observed SHBs much better than the model for primordial DNSs (see the results of Kolmogorov-Smirnoff \[KS\] tests in table \[\[t:fit\]\]).
We note, however, that ruling out a primordial binary $z$-DF may be premature based on current data. For SWIFT, only 1/3 of detected SHBs have secure $z$ determinations. If there is a bias against obtaining a secure redshift for higher $z$ bursts, the observed distribution would be shifted, compared to the expected distribution, to low $z$. Assuming, for example, that obtaining a secure $z$ requires a higher $P_{\rm lim}$ compared to that required for detection, the lower rate of detection of SHBs with secure $z$ is accounted for by choosing $P_{\rm lim}=2.5\,{\rm ph\,cm^{-2}\, s^{-1}}$ (which reduces the detection rate by a factor of 3 compared to that obtained for $P_{\rm lim}=1\,{\rm ph\,cm^{-2}\, s^{-1}}$). Indeed, inspection of the 15 to 150 keV peak fluxes of SWIFT SHBs shows that the average peak flux of SHBs with $z$ is higher than that of SHBs without $z$ ($\langle f_z\rangle=5.7\pm 5.7\,{\rm ph\, cm^{-2}\, s^{-1}}$ for GRBs 050509, 050724 and 051221 compared to $\langle f_{{\rm no}\,
z}\rangle=1.2\pm 0.5\,{\rm ph\, cm^{-2}\,s^{-1}}$ for GRBs 050911, 051105, 051114, 051210 and 051227; Barthelmy et al. [@Bar05]; Gehrels et al. [@Gea05]; Bloom et al. [@Bea05]; Berger et al. [@BPea05]; Fox et al. [@Fea05]; Villasenor et al. [@Vil05]).
Choosing $P_{\rm lim}=2.5\,{\rm ph\,cm^{-2}\, s^{-1}}$, the expected $z$-DF of primordial binary mergers is marginally compatible with observations (see table \[t:fit\] and fig. \[cdf\]). The time delay distribution $(dp/d\tau)_{\rm dyn}$ expected for dynamical DNSs yields a $z$-DF compatible with observations for both choices of $P_{\rm lim}$.
{width="1.0\columnwidth"}
Summary and Discussion {#s:discussion}
======================
We have shown (fig \[tdelay\]) that the $z$-DF of SHBs expected from [*dynamical*]{} formation and subsequent merger of DNSs is markedly different from that expected from primordial DNS mergers (assuming $(dp/d\tau)_{\rm prim}\propto 1/\tau$). The large time for core collapse shifts the DF of dynamically formed DNS mergers to low $z$. The observed $z$-DF of SHBs strongly favors that expected for mergers of dynamically formed DNSs, as compared to that expected for primordial DNS mergers (fig \[cdf\] and table 1; see also GP05 and Nakar et al. [@NGF06]).
However, current data do not allow to rule out a $z$-DF consistent with that expected for primordial DNS mergers, since redshifts were obtained only for a minority of the detected SHBs. This may be due to a bias against obtaining redshift information for high $z$ (faint) SHBs (fig \[cdf\], table 1 and discussion at the end of §3.3). Future observations should allow to better constrain the $z$-DF of SHBs, and thus to differentiate between models. For example, detection of only a few high redshift ($z>2$) SHBs would severely constrain the contribution of dynamically formed DNSs.
If the formation rate of primordial and dynamically formed DNSs are comparable (GPZM06), there will be an anti-correlation between $z$ and off-set of the SHB from the center of its host galaxy, because DNSs formed in GCs will be closer in $z$, but farther away from their host center, since GCs reside in the halos of galaxies. The large time delay also implies that more DNS mergers will be observed by LIGO and VIRGO (Nakar et al. [@NGF06])
An alternative method for constraining the progenitors of SHBs is to consider their demography (Gal-Yam et al [@Gal06]; Zheng & Ramirez-Ruiz [@Zhe06]). Zheng & Ramirez-Ruiz ([@Zhe06]) showed that in order to account for the preponderance of SHBs in elliptical galaxies, the delay times should be large, following $dp/d\tau\propto \tau^\delta$ with $\delta\approx1.5$. This method, which does not rely on the observed $z$-DF, also indicates delay times longer than expected for primordial binaries. We note that the specific frequency of GCs (number of GCs per unit luminosity in the V-band) is larger for elliptical galaxies than for spirals (e.g. Harris [@Har91]) by an order of magnitude.
This work was supported by an AEC grant (EW), The Netherlands Academy of Arts and Sciences (KNAW), The Netherlands Organization for Scientific Research (NWO), and the The Netherlands School for Astronomy (NOVA) (SPZ). DG thanks S. Covino, D. Malesani and C. Guidorzi for useful information, and the Weizmann Inst. for its hospitality. The research of DG is partially supported by MIUR-Cofin grant.
[ 2004b]{}
Barthelmy, S. D., et al., 2005, Nature, 438, 994 Baumgardt, H., 2001, MNRAS, 325, 1323 Belczynski, K., Bulik, T., Rudak, B., 2002, ApJ, 571, 394 Belczynski, K., Perna, R., Bulik, T., Kalogera, V., Ivanova, N., & Lamb, D. Q., 2006, preprint: astro-ph/0601458
Berger, E., et al., 2005, astro-ph/0508115 Bloom, J. S., Sigurdsson, S., & Pols, O. R., 1999, MNRAS, 305, 763 Bloom, J. S., et al., 2005, pre-print: astro-ph/0505480 Campana, S. et al. 2006, Nature, submitted Dewi, J. D., Podsiadlowski, Ph., & Pols, O. R., 2005, MNRAS, 363, L71 Efremov, Yu. N., 2000, Astronomy Letters, 26, 558 Eichler, D., Livio, M., Piran, T., & Schramm, D. N., Nature, 340, 126
Fox, D. B., et al., 2005, Nature, 437, 845 Fregeau, J. M., Gürkan, M. A., Joshi, K. J., & Rasio, F. A., 2003, ApJ, 593, 772 Galama, T. J., et al., 1998, Nature, 395, 670 Gal-Yam, A., et al., 2006,pre-print: astro-ph/0509891 Gehrels, N., et al., 2005, Nature, 437, 845 Gladders, M., Berger, E., Morrell, N., & Roth, M., 2005, GCN, 3798, 1. Goodman, J. 1986, , 308, L47 Grindlay, J., Portegies Zwart, S. F., & McMillan, S., 2006, Nat. Phys. 2, 116 (GPZM06) Guetta, D., Piran, T., & Waxman, E., 2004, ApJ, 619, 419 Guetta, D., & Piran, T., 2005, A&A, 435, 421 (GP05) Guetta, D., & Piran, T., 2006, A&A in press (astro-ph/0511239) Gürkan, M. A., Freitag, M., & Rasio, F. A., 2004, ApJ, 604, 632 Harris, W. E., 1991, ARA&A, 29, 543 Harris, W. E., 1996, AJ, 112 Hjorth, J. et al. 2003, , 423, 847 Joshi, K. J., Nave, C. P., & Rasio, F. A., 2001, ApJ, 550, 691
Kouveliotou, C., et al., 1993, ApJ, 413, L101 MacFadyen, A. I. & Woosley, S., 1999, , 524, 262 Makino, J., 1996, ApJ, 471, 796 Malesani, D. et al. 2004, , 609, L5 Nakar, E., Gal-Yam, A., & Fox, D., 2006, pre-print:astro-ph/0511254 Narayan, R., Paczynski, B., & Piran, T., 1992, ApJ, 395, L83 Paczyński, B. 1986, , 308, L43 Paczyński, B. 1998, , 494, L45 Porciani, C., & Madau, P., 2001, ApJ, 548, 522 Portegies Zwart, S. F., & Yungelson, L., 1998, A&A, 332, 173 Portegies Zwart, S. F., McMillan, S. L. W., Hut, P., Makino, J., 2001, MNRAS, 321, 199 Quinlan, G. D., New Astron., Vol. 1, No. 3, p. 255 Rowan-Robinson, M. 1999, Ap&SS, 266, 291 Schmidt, M., 2001, ApJ, 559, L79 Sigurdsson, S. Phinney, E. S., 1995, ApJS, 99, 609 Stanek, K. Z. et al. 2003, , 591, L17 Sugimoto, D., & Bettwieser, E., 1983, MNRAS, 204, 19 van Paradijs, J., et al., 1997, Nature, 386, 686 Villasenor, J. S., et al., Nature, 437, 855 Woosley, S., 1993, , 405, 273 Zheng, Z., & Ramirez-Ruiz, E., 2006, pre-print: astro-ph/0601622
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'Marin Ballu[^1]'
- 'Quentin Berthet[^2]'
- 'Francis Bach[^3]'
bibliography:
- 'references.bib'
title: |
Stochastic Optimization for\
Regularized Wasserstein Estimators
---
=1
[^1]: Department of Pure Mathematics and Mathematical Statistics, University of Cambridge, `[email protected]`
[^2]: Google Research, Brain team, Paris, `[email protected]`
[^3]: INRIA - DI, ENS, PSL Research University Paris, `[email protected]`
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The profiles of local density of states (LDOS) around single vortices and vortex pairs in type-II superconductors are studied taking account of the interference of quasiparticle waves experiencing Andreev reflection within the vortex cores, hybridization of vortex core states and normal reflection at the boundaries or defects. For subgap energy levels these interference effects reveal themselves in a nontrivial dependence of the positions of the LDOS peaks on the intervortex distance and sample size: the peak positions generally do not coincide with the superconducting phase singularity points. The LDOS profiles are calculated for three generic examples which can be realized, e.g., in mesoscopic superconductors: (i) vortex-vortex pair; (ii) vortex positioned near a flat boundary; (iii) vortex positioned in the center of a superconducting disk. The resulting evolution of the Andreev interference patterns could be observable by scanning tunneling spectroscopy techniques in mesoscopic superconductors or disordered vortex arrays.'
author:
- 'A. S. Mel’nikov, D. A. Ryzhov, M. A. Silaev'
title: |
Local density of states around single vortices and vortex pairs:\
effect of boundaries and hybridization of vortex core states
---
Introduction
============
Physics of vortex matter in superconductors has been a long standing topic of intensive research for many years. Over a last decade growing interest to this topic has been stimulated by technological advances allowing deeper insight into the structure of vortices as well as investigation of new exotic vortex states. In particular, local density of states (LDOS) measurements with the help of scanning tunneling spectroscopy techniques have been proven to be an effective tool of experimental study of electronic structure of Abrikosov vortices[@STSHess; @STSRoditchev; @STSmore1; @STSmore2; @RMP-2007]. The observation of the zero-bias anomaly of tunneling conductance at the center of each vortex in these experiments clearly confirmed the existence of bound vortex core states predicted by Caroli, de Gennes and Matricon (CdGM)[@CdGM]. The wave functions of the subgap states are localized inside the vortex core because of the Andreev reflection of quasiparticles at the core boundary. For each individual vortex the energy $\varepsilon(\mu)$ of a subgap state varies from $-\Delta_0$ to $+\Delta_0$ as one changes the angular momentum $\mu$ defined with respect to the vortex axis. At small energies $|\varepsilon|\ll\Delta_0$ the spectrum is a linear function of $\mu$: $\varepsilon(\mu)\simeq-\mu\omega$, where $\omega\approx\Delta_0/(k_F\xi)$, $\Delta_0$ is the superconducting gap value far from the vortex axis, $k_F$ is the Fermi momentum, $\xi=\hbar
v_F/\Delta_0$ is the coherence length, $v_F$ is the Fermi velocity, and $\mu$ is half an odd integer. Note that hereafter we assume the Fermi surface to be a cylinder and neglect the dependence of the quasiparticle energy on the momentum component along the vortex axis $z$ considering a motion of quasiparticles only in $xy$ plane. Introducing a cylindrical coordinate system $(r,\theta,z)$ and defining an impact parameter as $b=-\mu/k_F=[{\bf r},{\bf k}_F]\cdot{\bf z}_0/k_F$ the quasiclassical LDOS inside the isolated vortex core ($r\ll\xi$) can be found as follows: $N=(k_F/2\pi\xi)\int_0^{2\pi}\delta(\varepsilon-\omega
k_Fr\sin(\theta-\theta_p))d\theta_p$. Here we evaluate the LDOS summing up over the quasiparticle states at the trajectories characterized by the orientation angle $\theta_p$. This expression yields a singularity of zero energy LDOS at the vortex center[@Ullah; @Maki; @IchiokaStar]: $N=1/(2\pi\omega r\xi)\approx N_0 \xi/r$, where $N_0=(1/2\pi)m/\hbar^2$ is a normal metal LDOS. Smearing of energy levels due to scattering effects leads to a reduction of LDOS peak amplitude at the vortex center. However, the peak in the LDOS distribution survives even in “dirty” limit when a mean free path is smaller than a coherence length $l<\xi$ (see Refs. ).
The increase in the external magnetic field results in the decrease in the intervortex distance which should be accompanied by the overlapping of wave functions describing the quasiparticle states bound to the neighboring vortex cores. This hybridization of vortex core states can perturb the CdGM spectra of isolated vortices and the circular symmetry of the LDOS peaks (see, e.g., Ref. ). Note that the LDOS peak anisotropy may also reflect possible superconducting gap and Fermi surface anisotropy, and formation of the charge density wave order[@STSHess; @STSmore2; @IchiokaStar].
The goal of the present manuscript is to show by the example of mesoscopic superconductor that the interference of quasiparticle waves trapped in neighboring vortex cores can result even in more dramatic consequences for the LDOS profiles around the vortex configurations. These consequences are not limited just to the distortion of the LDOS peaks at individual vortex positions caused by the concrete symmetry of the system and/or vortex configuration. The peak positions themselves appear to be shifted from the superconducting phase singularity points which are usually considered as vortex centers. Changing the external magnetic field we can control the vortex configuration in the sample and switch between different Andreev interference patterns in LDOS. The distinctive features of the electronic structure of vortices in mesoscopic samples with sizes of several coherence lengths are controlled by two key factors: (i) quasiparticle tunneling between the neighboring vortex cores which becomes extremely important when the individual vortices merge into the giant (multiquantum) vortex[@Mesovortices] with a winding number larger than unity $M>1$; (ii) normal scattering of quasiparticles at the sample boundary.
The multiquantum vortex provides a simple illustration of the above statement regarding the difference in the positions of the LDOS peaks and phase singularity points. Indeed, the subgap spectrum of such vortex contains $M$ anomalous spectral branches (per spin projection)[@VolovikBranches]. Each anomalous branch intersects the Fermi level and contributes to the low-energy DOS. At low energies the spectrum has the following form[@VolovikBranches; @multi-spectrum-num; @Melnikov-Vinokur-2002; @Janko]: $$\label{Volovik-spectr}
\varepsilon_j(\mu)\sim-\frac{\Delta_0}{k_F\xi}(\mu-\mu_{j})\,,$$ where the index $j$ enumerates different spectral branches ($1<j<M$), $-k_F\xi\lesssim\mu_{j}\lesssim k_F\xi$. The LDOS profile corresponding to the spectrum (\[Volovik-spectr\]) consists of a set of axially symmetric ring structures[@Melnikov-Vinokur-2002; @Janko]. Note that for an even winding number the anomalous branch crossing the Fermi level at $\mu=0$ (i.e. at zero impact parameter) is absent and, as a result, the LDOS peak at the vortex center disappears. The splitting of a multiquantum vortex into the individual vortices should lead to the transformation of LDOS rings into a set of peaks positioned in the cores of individual vortices. The initial stage of this LDOS transformation for small intervortex distances was studied in Ref. within a perturbation approach. A nonperturbative approach which allows to describe the spectrum transformation accompanying the decay of the multiquantum vortex has been suggested in Refs. [@MelnikovSilaev; @MelnikovPrb2008].
The second factor which is crucial in mesoscopic superconducting samples is a quasiparticle scattering at the sample boundaries which comes into play when vortices approach a superconductor surface or a size of superconducting sample is small enough. For vortices positioned rather close to the sample surface the effect of quasiparticle reflection at the boundary on the spectrum and total DOS was investigated recently in Refs. . The early stage of transformation of the LDOS profiles for vortices approaching a flat boundary has been studied numerically in Ref. .
In the present paper we analyze the effect of both the hybridization of vortex core states and boundary scattering on the LDOS structures and for this purpose consider three generic examples: (i) vortex-vortex pair; (ii) vortex positioned near a flat boundary; (iii) vortex positioned in the center of a superconducting disk. We will show that the LDOS profiles can be obtained qualitatively from the spectra found in Refs. . Our qualitative considerations will be confirmed by the detailed numerical calculations.
This paper is organized as follows. In Sec. \[sec:theory\] we give an overview of the theoretical framework which is employed in this work, namely the Bogoliubov–de Gennes theory and the quasiclassical Eilenberger approach. We discuss the spectrum and LDOS patterns for multi vortex configurations in Sec. \[subsec:molecules\] and address the case of a vortex in a small mesoscopic cylinder in Sec. \[subsec:cylinder\]. We give our conclusions in Sec. \[sec:summary\].
Model and basic equations {#sec:theory}
=========================
Our further consideration is based on two approaches: (i) Bogoliubov–de Gennes (BdG) equations; (ii) quasiclassical Eilenberger theory. The first one appears to be more transparent for the qualitative analysis of the interference effects and convenient for numerical calculation of LDOS profiles around the vortex placed in a superconducting disk. The Bogoliubov–de Gennes (BdG) equations for particle– ($u$) and hole–like ($v$) parts of the wave function have the following form: $$\begin{aligned}
\label{BdG}
\nonumber
-\frac{1}{2m}\left(\hat {\bf p}-\frac{e}{c}{\bf A}\right)^2u+\Delta v
&=&(\varepsilon+\varepsilon_F) u\,,\\
\frac{1}{2m}\left(\hat {\bf p}+\frac{e}{c}{\bf A}\right)^2v
+\Delta^* u&
=&(\varepsilon-\varepsilon_F) v\,.\end{aligned}$$ Here $\Delta$ is a gap function, ${\bf A}$ is a vector potential, $\hat
{\bf p}=-i\hbar(\partial/\partial x,\partial/\partial y)$, and ${\bf
r}=(x,y)$ is a radius vector in the plane perpendicular to the vortex axis. The LDOS can be expressed through the eigenfunctions of the BdG equation (\[BdG\]) in the following form (see, e.g., Ref.): $$\label{DOS}
N({\bf r},\varepsilon)=\sum_n
|u_n({\bf{r}})|^2\delta(\varepsilon-\varepsilon_n)\,,$$ where $u_n({\bf{r}})$ is electron component of quasiparticle eigenfunction corresponding to an energy level $\varepsilon_n$ (we sum over both positive and negative values $\varepsilon_n$). The eigenfunction has to be normalized: $$\int\left(|u_n ({\bf r})|^2+|v_n ({\bf r})|^2\right)d^2{\bf r}=1\,.$$
In general case BdG equations are rather complicated. A simplification can be obtained if one considers a quasiclassical approximation, assuming that the wavelength of quasiparticles is much smaller than the superconducting coherence length (see, e.g., Ref. ). Within such an approximation, quasiparticles move along linear trajectories, i.e. straight lines along the direction of the quasiparticle momentum ${\bf
k}_F=k_F(\cos\theta_p,\sin\theta_p)$. Generally, the quasiclassical form of the wave function can be constructed as follows: $(u,v)=e^{i{\bf
k}_F{\bf r}}(\tilde U,\tilde V)$, where $(\tilde U({\bf r}),\tilde V({\bf
r}))$ is a slowly varying envelope function. Then the system (\[BdG\]) is reduced to the quasiclassical Andreev equations: $$\begin{aligned}
\label{quasiclass}
\frac{\hbar\mathbf{k}_F}{m}\left(-i\hbar\nabla+
\frac{e}{c}\mathbf{A}\right)\tilde U+\Delta\tilde V
=\varepsilon\tilde U\,,
\notag\\
\frac{\hbar\mathbf{k}_F}{m}\left(i\hbar\nabla+
\frac{e}{c}\mathbf{A}\right)\tilde V+\Delta^*\tilde U
=\varepsilon\tilde V\,,\end{aligned}$$ which are defined at the linear trajectories determined by the direction of the quasiparticle momentum ${\bf k}_F$.
The quasiclassical approximation allows to develop a powerful method for calculation of the LDOS based on the solution of Eilenberger equations for quasiclassical propagator along the trajectories[@Eilenberger]. For numerical treatment of these equations we follow the Refs. and introduce a Ricatti parametrization for the Green function. The essence of this method is a mathematical trick which allows to solve two first order Ricatti equations instead of fourth-order system of Eilenberger equations. Following Refs. one can obtain: $$\begin{aligned}
\label{eq:eilenberger}
\hbar v_F\frac{\partial}{\partial x'}a(x')
&+[2\tilde{\omega}_n+\Delta^*a(x')]a(x') - \Delta = 0 ,\notag\\
\hbar v_F\frac{\partial}{\partial x'}b(x') &-
[2\tilde{\omega}_n + \Delta b(x')]b(x') + \Delta^* = 0,\end{aligned}$$ where $x'=({\bf k}_F{\bf r})/k_F=r\cos(\theta_p-\theta)$ is the coordinate along trajectory, $\Delta({\bf r})=|\Delta|e^{i\Phi}$, $\Phi({\bf r})$ is a superconducting phase, $i\tilde{\omega}_n=i\omega_n+m{\bf v}_F{\bf v}_s$ is a Doppler-shifted Matsubara frequency, $\omega_n=\pi T(2n+1)$ and $${\bf v}_s=\frac{1}{2m}\left(\hbar\nabla\Phi-\frac{2e}{c}{\bf A}\right)$$ is a gauge-invariant superfluid velocity. The LDOS may be expressed through the scalar coherence functions $a$ and $b$ as follows[@Schopohl] $$\begin{aligned}
\label{eq:LDOS}
N({\bf r},\varepsilon) = \int^{2\pi}_0 \frac{d\theta}{2\pi}
\text{Re}\left\{\frac{1-ab}{1+ab}\right\}_{i\omega_n \to \varepsilon+i\nu},\end{aligned}$$ where $\varepsilon$ is the quasiparticle energy measured from Fermi level and $\nu$ is a parameter which accounts for an elastic scattering. Throughout this paper we will assume the simplest model[@clem] for the gap function distribution around an isolated vortex positioned at the origin $\Delta({\bf r})=\Delta_0f_1({\bf r})$, where $$\label{Clem-profile}
f_1({\bf r})=\frac{x+iy}{\sqrt{x^2+y^2+\xi_v^2}}
=\frac{r e^{i\theta}}{\sqrt{r^2+\xi_v^2}}$$ with the core size $\xi_v=\xi$. To describe the system of two singly–quantized vortices positioned at ${\bf r}=\pm{\bf a}/2=\pm(a/2,0)$ we fit the gap function as follows: $$\begin{aligned}
\label{2VortMolecGap}
\nonumber
\Delta({\bf r})=\Delta_0f_1({\bf r}-{\bf a}/2)f_1({\bf r}+{\bf a}/2)\,.\end{aligned}$$
The image method allows us to reduce the problem of a vortex positioned at the point ${\bf r}=(a/2,0)$ near a flat boundary $x=0$ to the one describing a vortex-antivortex pair with the antivortex situated at ${\bf
r}=-(a/2,0)$. For the latter case we choose the gap function in the form: $$\begin{aligned}
\label{VaVVortMolecGap}
\nonumber
\Delta({\bf r})=\Delta_0f_1({\bf r}-{\bf a}/2)f_1^*({\bf r}+{\bf a}/2)\,.\end{aligned}$$
Certainly, the above simplest approximation for the gap function might not be enough to describe closely spaced vortex–vortex or vortex–antivortex pairs. However, the resulting LDOS patterns obtained below are mainly controlled only by the distance between the phase singularity points and appear to be weakly sensitive to the concrete profiles of the order parameter absolute value. The key features of the LDOS patterns appear to hold even in the limit of the zero $\xi_v$ value, i.e., even for phase vortices without any suppression of the gap value inside the cores.
Considering hereafter only the vortex configurations of the finite size comparable with several coherence lengths we neglect the vector potential which is known to give only a moderate renormalization of the interlevel spacing $\omega$ (see Ref. ).
Spectrum and local density of states {#sec:results}
====================================
Vortex molecules {#subsec:molecules}
----------------
We now proceed with the calculation of LDOS profiles for two systems: (i) vortex–vortex pair and (ii) vortex positioned near the flat superconducting boundary. The image method discussed in Ref. allows us to show that the latter case is equivalent to the vortex-antivortex pair in the bulk provided we choose the distance between the vortex and antivortex to be two times larger than the distance from the vortex to the flat surface in the original problem. To elucidate our main results we start from a qualitative description of the quasiparticle spectra following Ref. .
Let us consider two vortices oriented along the $z$ axis. In the plane $(xy)$ the corresponding phase singularity points (or vortex centers) are positioned at ${\bf r}_{1,2}=(\pm a/2,0)$. For quasiparticles propagating along the classical trajectories parallel to ${\bf k}_F$ we introduce the angular momenta $\mu=[{\bf r},{\bf k}_F]\cdot{\bf z_0}=k_F
r\sin(\theta_p-\theta)$ and $\tilde\mu_i=\mu-[{\bf r}_i,{\bf
k}_F]\cdot{\bf z}_0$ defined with respect to the $z$–axis passing through the origin and with respect to the $i$-th vortex axis passing through the point ${\bf r}_i$, correspondingly. Neglecting the quasiparticle tunneling between the vortex cores the wave function can be found as a superposition of two states localized at different vortices and having close energies: $\varepsilon_{v1}=-\omega[\mu-(k_F a/2)\sin\theta_p]$ and $\varepsilon_{v2}=-\omega[\mu+(k_F a/2)\sin\theta_p]$. The transformation of the quasiclassical spectrum occurs due to the overlapping of the corresponding wave functions and can be analysed using a standard quantum mechanical approach describing a two–level system[@LL-III], which yields the secular equation $$\label{secularVV}
(\varepsilon-\varepsilon_{v1})(\varepsilon-\varepsilon_{v2})
=(\delta\varepsilon)^2\,,$$ and the resulting splitting of isoenergetic lines near the degeneracy point in the $\mu-\theta$ plane ($\theta_p=0$ for our case) $$\label{orbit1}
\omega\mu=-\varepsilon\pm\sqrt{\omega^2(k_F a/2)^2\theta_p^2
+(\delta\varepsilon)^2}\,.$$ The tunneling of quasiparticles between the vortex cores is determined by the exponentially small overlapping of wave functions localized near the cores and results in the splitting of the energy levels: $\delta\varepsilon\sim\Delta_0\exp(-a/\xi)$. Thus, the estimate for the splitting $\delta\mu\simeq\delta\varepsilon/\omega$ of isoenergetic lines in the $(\mu,\theta_p)$–plane reads (see Eq. (\[orbit1\])): $$\label{delta-mu}
\delta \mu(a)\sim k_F\xi\exp\left(-\frac{a}{\xi}\right)\,.$$
In the case of a vortex-antivortex pair the non-interacting states localized at different vortices have the energies $\varepsilon_v=-\omega[\mu+(k_F a/2)\sin\theta_p]$ and $\varepsilon_{av}=\omega[\mu-(k_F a/2)\sin\theta_p]$. Taking into account the overlapping of the corresponding wave functions states we obtain the secular equation (\[secularVV\]) with $\varepsilon_{v1}=\varepsilon_v$ and $\varepsilon_{v2}=\varepsilon_{av}$. Therefore, the quasiclassical orbits near the degeneracy point $\theta_p=0$ are defined by the following equation: $$\label{orbit2}
\varepsilon=\pm\sqrt{(\omega\mu)^2+(\delta \varepsilon)^2}
-\omega(k_F a/2)\theta_p\,.$$ The classically forbidden angular domain at $\varepsilon=0$ has the width $\delta\theta_p=4\delta\varepsilon/(\omega k_F a)$. One can assume that the appearance of such classically forbidden domain explains the deep structure in the local DOS profile observed numerically in Ref. for a vortex near the flat boundary of an $s$–wave superconductor. The classically forbidden angular domain results in the suppression of the overall DOS[@MelnikovPrb2008] and we show below that this mechanism is responsible for the anomalous spectrum branch disappearance when the vortex exits the sample.
To study the transformation of spectrum in the entire range of intervortex distances we solve numerically the quasiclassical Andreev equations (\[quasiclass\]) to find isoenergetic lines in $\mu -\theta_p$ plane. The resulting quasiclassical orbits for a vortex–vortex pair with the intervortex distances $a=2.5\xi$ and $a=1.5\xi$ are shown in Fig. \[Orbit\]a.
![\[Orbit\] Quasiclassical orbits in the $\mu-\theta_p$ plane corresponding to zero energy level. Upper panel (a): two vortices at distances $a=2.5\xi$ (solid line) and $a=1.5\xi$ (dashed line). Dotted lines correspond to the case of a doubly quantized vortex, i.e. to $a=0$. Lower panel (b): vortex-antivortex system with $a=2.5\xi$ (solid line), $a=1.7\xi$ (dashed line) and $a=1.6\xi$ (dashdot line).](Orbit.eps){width="1.0\linewidth"}
The dotted lines show the orbits for $a=0$, i.e., for a doubly quantized vortex. The splitting of the quasiclassical orbits grows with the decrease in the size of the vortex molecule both for the two-vortex and vortex-antivortex configurations. When the intervortex distance is of the order $\xi$ the energy splitting becomes so large ($\delta\varepsilon\sim\Delta_0$) that the simplified equations (\[orbit1\]),(\[orbit2\]) are no longer valid and the spectrum is disturbed in the whole angular domain $0<\theta_p<2\pi$. One can clearly observe that with the decrease in the intervortex distance the CdGM orbits $b=\pm (a/2)\sin\theta_p$ corresponding to zero energy in isolated vortices transform into two lines $\mu=\pm\mu_0$ which correspond to the spectrum of a doubly quantized vortex. Comparing the isoenergetic lines with the unperturbed CdGM ones one can see that the zero energy quasiclassical trajectories do not pass through the vortex center. It means that the LDOS peaks are necessary shifted from the vortex centers as the intervortex distance becomes small enough. This phenomenon can be understood considering the following qualitative arguments. Indeed, when the intervortex distance is large $a\gg\xi$ the equilibrium size of the vortex pair is determined by the stability condition yielding the zero value of Lorentz force acting on each vortex: $ {\bf F}_L\propto\phi_0
[{\bf z}_0,{\bf v}_s]=0$, where $\phi_0$ is a magnetic flux piercing the vortex and ${\bf v}_s$ is a a superfluid velocity induced in a given vortex core by other vortices and screening currents flowing along the boundaries. Thus, the local Doppler shift of the quasiparticle energy levels $\varepsilon_d=\hbar{\bf k}_F{\bf v}_s$ appears to vanish for trajectories passing through the vortex center. As a result, the position of the LDOS peaks coincide with the phase singularity points which are usually defined as vortex centers. However when the intervortex distance is comparable with (or less than) $\xi$ the part of the superfluid velocity induced by the neighboring vortex becomes strongly inhomogeneous inside the core region and the ${\bf v}_s$ value at a given vortex center diverges as $v_s\sim1/a$. As a consequence, the equilibrium size of the pair is no more determined by the condition of zero ${\bf v}_s$ at the vortex centers. The resulting nonvanishing Doppler shift for the trajectories passing through the vortex centers suppresses the LDOS at these points. Thus, for small $a$ the LDOS peak positions do not coincide with the vortex centers.
The Fig. \[Orbit\]b corresponds to the vortex-antivortex pair with the intervortex distances $a=2.5\xi$, $a=1.7\xi$, and $a=1.6\xi$. One can see that contrary to the case of the vortex-vortex pair the decrease in the intervortex distance in this system leads to a rapid shrinking of the isoenergetic orbits. When the distance $a$ is small enough all the trajectories corresponding to the zero energy are characterized by the $\mu$ and $\theta_p$ values close to $\mu=0$ and $\theta_p=\pi (n+1/2)$, where $n$ is integer. For the intervortex distance $a/\xi=1.5$ the quasiclassical orbits completely shrink to these points in $\mu-\theta_p$ plane. In the real space the corresponding trajectories pass through the point in the middle between the vortex and antivortex and are perpendicular to the line connecting the phase singularity points. Thus, the two peaks of LDOS should shift from the vortex centers to the middle point and merge into one peak as the vortex and antivortex approach each other.
For a detailed study of these effects we calculate the LDOS profiles for two–vortex and vortex–antivortex configurations by solving numerically the Eilengerger equations (\[eq:eilenberger\]) and applying the expression for the LDOS (\[eq:LDOS\]). The results are shown in Fig. \[LDOSvv\] for the vortex–vortex pair with $a=1.5\xi$ (upper panel) and $a=\xi$ (lower panel) and in Fig. \[LDOSvav\] for a vortex–antivortex pair with $a/\xi=2$ (upper panel) and $a=1.5\xi$ (lower panel).
![\[LDOSvv\] (Color online) The normalized LDOS profile in a two–vortex system for $a/\xi=1.5$ (a) and $a/\xi=1$ (b) calculated for $\nu=0.06\Delta_0$. Positions of vortices are marked by red crosses.](vv.eps){width="1.0\linewidth"}
![\[LDOSvav\] (Color online) The normalized LDOS profile in a vortex–antivortex system for $a/\xi=2$ (a) and $a/\xi=1.5$ (b) calculated for $\nu=0.06\Delta_0$. Positions of vortices are marked by red crosses.](vav.eps){width="1.0\linewidth"}
In Fig. \[LDOSvv\] it is shown how the LDOS profile gradually transforms from a two-peaked structure corresponding to two isolated vortices (upper panel) to the axially symmetric ring for a doubly quantized vortex. The positions of the phase singularity points are shown on the contour plot by red crosses. It can be seen that even when the distance between the vortices exceeds the size of the vortex core the positions of the LDOS peaks do not coincide with the vortex centers. The distance between the LDOS peaks reduces slower than the size $a$ of the vortex molecule and finally the peaks are smeared into the anisotropic ring structure (see Fig. \[LDOSvv\]b). The anisotropy vanishes and the ring becomes axially symmetric when vortices merge into a doubly quantized vortex.
The evolution of the LDOS profile in a vortex-antivortex pair follows a different scenario (see Fig. \[LDOSvav\]). When the distance between vortices is large enough there appear deep structures at the LDOS peaks (see Fig. \[LDOSvav\]a). The peaks themselves are slightly shifted towards the middle point at the line connecting the vortex and antivortex centers. Below the critical intervortex distance $a_c/\xi\sim 1.7$ the two peaks merge into one peak at the middle point (see Fig. \[LDOSvav\]b). As we discussed above this phenomenon is explained by the shrinkage of the quasiclassical orbits in the $\mu - \theta_p$ plane. When the vortex and antivortex come closer the amplitude of the LDOS peak is reduced and finally it disappears completely when the vortex and antivortex merge and annihilate. Mapping the solution of the vortex–antivortex problem on the one for a vortex near the flat surface we conclude that the zero bias LDOS peak appears to be positioned exactly at the surface provided the distance from the vortex to the surface becomes less than $0.85\xi$.
Vortex in mesoscopic cylinder {#subsec:cylinder}
-----------------------------
In this subsection we study another generic example illustrating the effect of the normal quasiparticle reflection at the sample boundary on the LDOS profiles in the vortex state. We consider a singly quantized vortex situated at the center of a superconducting disk of rather small radius $R$ comparable with the coherence length. In this case a constructive interference of quasiparticle waves reflected from the boundary is known to result in mesoscopic oscillations of energy levels[@MelnikovPrl2005]: $$\label{SpectrumDisc}
\varepsilon(\mu)=-\omega\mu -\delta\sin(2k_FR-\pi \mu)\,,$$ where $\delta\sim \Delta_0\exp(-2R/\xi)$ and $|\varepsilon(\mu)|\ll\Delta_0$. The remarkable fact is that at low energies these levels can be combined into two groups corresponding to odd and even values of the integer index $\mu+1/2$. A spacing between the energy levels belonging to one group is $2\omega$ while the energy shift between the different groups is $2\delta|\cos(2k_F R)|$. Analogous shift of vortex core levels defined by the factor $\exp(-R/\xi)$ has been observed recently for a vortex-antivortex pair on a sphere with a finite radius $R$ in $p_x+ip_y$ superconductors[@Fertig]. Note that for a small superconducting disc with $R<R_c\sim (\xi/2)\ln (k_F\xi)$ we have $\delta\gg\omega$. Therefore the two groups of energy levels can be considered as continuous branches if the energy discreteness with the scale $\omega$ is neglected. In this approximation one can use a quasiclassical expression $\mu=-k_Fr\sin(\theta-\theta_p)$ and obtain the low energy branches in the form: $$\label{SpectrumDisc-q}
\varepsilon_{1,2}
=\omega k_Fr\sin(\theta-\theta_p)\pm\delta\cos(2k_F R)\,.$$ Thus, for low energies $\varepsilon\ll\Delta_0$ the quasiclassical expression for the LDOS near the vortex center reads: $$\label{LDOSDisc0}
N(r,\varepsilon)=\frac{k_F}{2\pi\xi}\sum_{j=1,2}\int_0^{2\pi}\delta
(\varepsilon-\varepsilon_{j})d\theta_p\,.$$ The Eq.(\[LDOSDisc0\]) can be derived from the general expression for LDOS (\[DOS\]) by setting the amplitude of the wave function $|\tilde
U_n|^2={\rm const}$ which is a good approximation at the small distances $r\ll\xi$ from the vortex center. Evaluating this expression for the LDOS we obtain: $$\begin{aligned}
\label{LDOSDisc}
N(r,\varepsilon=0)=N_0\frac{\xi}{\sqrt{r^2-r_0^2}}\,,
\quad {\rm for}\ \ r>r_0\,,
\nonumber\\
N(r,\varepsilon=0)=0\,, \quad {\rm for}\ \ r<r_0\,,\end{aligned}$$ where $$r_0=\frac{\delta|\cos(2k_F R)|}{\omega k_F}
\sim\xi|\cos(2k_F R)|e^{-2R/\xi}\,.$$ Thus, the zero-bias peak of the LDOS at the vortex center which exists in an isolated vortex transforms into a ring structure which is similar to the LDOS pattern for a doubly quantized vortex. The only difference is that the radius of this ring is exponentially small comparing to the one in a doubly quantized vortex: $r_0/\xi\propto\exp(-R/\xi)\ll1$. Therefore, this splitting of the LDOS peak in tunneling spectroscopy experiments should be strongly affected by different smearing effects like disorder scattering or finite temperature. In particular, for the observation of the above effect we should assume the elastic mean free path to exceed the characteristic size of the LDOS ring: $l\geq\xi\exp(-R/\xi)$. Note that this condition is not very restrictive because it can be fulfilled even in “dirty” superconductors with $l\ll\xi$ because typically $\exp(-R/\xi)\ll1$. The effect of smearing due to the finite temperature is controlled by the parameter $T/\delta$.
To take into account the finite temperature effects and investigate the LDOS profiles in the whole range of energies and distances from the vortex core we solve numerically the BdG system (\[BdG\]) for a vortex in mesoscopic cylinder. We use the same method which was successfully applied to study the mesoscopic oscillations of vortex core levels[@MelnikovPrl2005], the spectrum of multiquanta vortices in mesoscopic superconductors[@MelnikovPrb2008] and the heat transport along vortex lines[@MelnikovPrb2007]. The resulting spectrum as a function of the angular momentum is shown in Fig. \[FigSpectrumDisc\] which clearly demonstrates the splitting of the anomalous CdGM branch at low energies caused by the boundary effects.
![\[FigSpectrumDisc\] (Color online) Quasiparticle spectrum vs $\mu$ for a singly quantized vortex in a disk of the radius $R=2\xi$. The CdGM spectrum is shown by the solid line. We choose here $k_F\xi=200$.](Spectrum){width="1.0\linewidth"}
For large energies $\varepsilon\sim\Delta_0$ the spectrum is not well described by the Eq.(\[SpectrumDisc\]).
Using the spectrum and the expression (\[DOS\]) for the LDOS we can find the local differential conductance as a function of voltage $V$: $$\frac{dI}{dV}=\left(\frac{dI}{dV}\right)_{N}
\int_{-\infty}^{+\infty}\frac{N({\bf r},\varepsilon)}{N_0}
\frac{\partial f(\varepsilon-eV)}{\partial V}\,d\varepsilon\,,$$ where $(dI/dV)_{N}$ is a conductance of the normal metal junction and $f(E)=[1+\exp(E/T)]^{-1}$ is a Fermi distribution function. The typical plots of the differential conductance are shown in Fig. \[dIdV\_M=1\] for different disk radii.
![\[dIdV\_M=1\] (Color online) Three-dimensional plot of the normalized local differential conductance $dI/dV$ as a function of voltage ($eV$) and distance from the vortex center ($r$) in disks with a large radius $R=7\xi$ (a) and small radius $R=2\xi$ (b) for $T/\Delta_0=0.005$. We choose here $k_F\xi=200$.](dIdV.eps){width="1.1\linewidth"}
The zero bias peak at the vortex center is clearly seen for a large disk radius when the normal scattering at the surface can be neglected. The decrease in the disk radius results in the shift of this conductance peak to higher voltages and formation of ring structure of local zero-bias conductance with finite radius $r\sim r_0$.
Note, that normal scattering of quasiparticles at the boundaries can be also important for multiquanta vortex configurations, e.g., vortex pairs, trapped in rather small samples. In our consideration of the LDOS profiles for a vortex-vortex pair in the previous subsection we have neglected the normal scattering at the sample surface assuming the vortex centers to be situated rather far from the boundaries. However, from our consideration of the vortex – antivortex problem one can expect that the effect of boundary scattering on the LDOS becomes important only when the distance from the vortex center to the sample edge is comparable with $\xi$. In this case the LDOS profiles in a vortex pair are affected by both the hybridization of the CdGM states and normal scattering. Considering the particular case of a doubly quantized vortex one can expect the boundary scattering to result in the spectrum oscillations (see Fig. 6) similar to the ones in a singly quantized vortex.
![\[FigSpectrumDisc2\] (Color online) Quasiparticle spectrum vs $\mu$ for a doubly quantized vortex in a disk of the small radius $R=2\xi$. The spectrum for a doubly quantized vortex in a disk of the large radius $R=7\xi$ is shown by the solid line. We choose here $k_F\xi=200$.](SpectrumD){width="1.0\linewidth"}
These oscillations should cause the broadening of the ring in the LDOS pattern with the typical broadening scale $r_0$. The same effect should result in the broadening of the anisotropic ring structure for a vortex pair of a finite size.
Conclusion {#sec:summary}
==========
To summarize, we have investigated the distinctive features of the LDOS profiles specific both for the vortex states in mesoscopic superconducting samples and for vortex arrays in bulk superconductors. These profiles are shown to be strongly affected by the Andreev interference effects and the normal scattering at the sample boundaries. In particular, we find that in contrast to the case of isolated vortices far from the boundaries the positions of peaks in the LDOS profiles in vortex configurations in small size samples are shifted from the phase singularity points. Taking a generic example of the two vortex system, we have considered the evolution of the LDOS profile which accompanies the merging of two vortices and appearance of the doubly quantized vortex. In this case the distance between LDOS peaks reduces slower than the intervortex distance $a$. Finally, when $a$ is smaller than the vortex core size the two-peaked structure of LDOS transforms into an anisotropic ring. The anisotropy vanishes when the vortex positions coincide and a doubly quantized vortex is formed.
Earlier (see Ref. ) it was shown that the spectrum of a single vortex placed near the flat surface is transformed analogously to the vortex-antivortex system and when the distance is of the order of the vortex core size, the interlevel spacing in the vortex spectrum becomes larger than the CdGM value. This effect was argued to lead to the disappearance of the anomalous spectrum branch when the vortex approaches the surface. In the present paper we have confirmed this prediction by calculating LDOS profile for the entire range of distances from the vortex to surface. We have found an amazing effect for such system: as the vortex approaches the surface the zero-bias peak of LDOS shifts from the vortex center to the surface point positioned at the minimal distance to the vortex center. For the particular vortex core model that we have used in this paper the shift of the LDOS peak occurs when the distance from vortex to surface is $0.85\xi$. Experimentally the LDOS peaks positioned very close to the superconductor boundary have been recently observed in vortex state of tungsten (W) based thin films by scanning tunneling microscopy/spectroscopy techniques[@Suderow-2009]. Note that in the latter experiment LDOS vortex peaks have been studied near the superconductor/normal metal (Au) interface. We suppose that the normal scattering of quasiparticles at such boundary caused by either the Fermi velocity mismatch or some surface barrier can result in the transformation of the LDOS profiles analogous to the one discussed above for a superconductor/insulator boundary.
Also we have investigated a measurable consequence of the giant mesoscopic oscillations of vortex core levels in finite size superconducting systems[@MelnikovPrl2005]. Considering the simplest case of a vortex positioned at the center of a superconducting cylinder with a radius $R$ we have shown that the zero-bias peak of the local differential conductance at the vortex center transforms into a ring structure of the radius of the order of $r_0\sim\xi e^{-2R/\xi}$. We expect that the unusual behavior of the LDOS profiles which we have investigated could be observable in scanning tunneling microscopy/spectroscopy experiments in mesoscopic superconductors which are now in the focus of interest (see, e.g., Refs. ).
Certainly, the Andreev interference patterns discussed above should be smeared by the disorder effects which are controlled by $\nu$ parameter within the Eilenberger theory. The resulting smearing of the LDOS profiles can be extremely important, e.g., for amorphous superconducting samples with rather small mean free path. Still we expect that the shift of the LDOS peaks from vortex singularity points caused by the Doppler effect will be observable even in this limit of strong disorder.
Acknowledgements
================
It is our pleasure to thank H. Suderow and N. B. Kopnin for stimulating discussions, and D. Roditchev for correspondence. This work was supported, in part, by Russian Foundation for Basic Research, by the Program “Quantum Physics of Condensed Matter” of RAS, and by the “Dynasty” Foundation.
[99]{}
H. F. Hess, R. B. Robinson, R. C. Dynes, J. M. Valles, Jr., and J. V. Waszczak, Phys. Rev. Lett. [**62**]{}, 214 (1989); H. F. Hess, R. B. Robinson, and J. V. Waszczak, Phys. Rev. Lett. [**64**]{}, 2711 (1990).
B. W. Hoogenboom, M. Kugler, B. Revaz, I. Maggio- Aprile, O.Fischer, and Ch. Renner, Phys. Rev. B [**62**]{}, 9179 (2000).
A. Kohen, Th. Proslier, T. Cren, Y. Noat, W. Sacks, H. Berger, and D. Roditchev, Phys. Rev. Lett. [**97**]{}, 027001 (2006).
I. Guillamon, H. Suderow, S. Vieira, L. Cario, P. Diener, and P.Rodiere, Phys. Rev. Lett. [**101**]{}, 166407 (2008).
O. Fischer, M. Kugler, I. Maggio-Aprile, C. Berthod, and C. Renner, Rev. Mod. Phys. [**79**]{}, 353 (2007).
C. Caroli, P. G. de Gennes, J. Matricon, Phys. Lett. [**9**]{}, 307 (1964).
S. Ullah, A. T. Dorsey and L. J. Buchholtz, Phys. Rev. B [**42**]{}, 9950 (1990).
N. Schopohl and K. Maki, Phys. Rev. B [**52**]{}, 490 (1995).
N. Hayashi, M. Ichioka, and K. Machida, Phys. Rev. Lett. [**77**]{}, 4074 (1996); Phys. Rev. B [**56**]{}, 9052 (1997).
A. A. Golubov and U. Hartmann, Phys. Rev. Lett. [**72**]{}, 3602 (1994).
P. Miranovic, M. Ichioka, and K. Machida, Phys. Rev. B [**70**]{}, 104510 (2004).
G. Boato, G. Gallinaro, and C. Rizzuto, Solid State Commun. [**3**]{}, 173 (1965); D. S. McLachlan, Solid State Commun. [**8**]{}, 1589 (1970); V. A. Schweigert, F. M. Peeters, and P. S. Deo, Phys. Rev. Lett. [**81**]{}, 2783 (1998); A. K. Geim, S. V. Dubonos, J. J. Palacios, I. V. Grigorieva, M. Henini, and J. J. Schermer, Phys. Rev. Lett. [**85**]{}, 1528 (2000); L. F. Chibotaru, A. Ceulemans, V. Bruyndoncx, V. V. Moshchalkov, Nature (London) [**408**]{}, 833 (2000).
G. E. Volovik, Pis’ma Zh. Eksp. Teor. Fiz. [**57**]{}, 233 (1993) \[JETP Lett. [**57**]{}, 244 (1993)\].
Y. Tanaka, A. Hasegawa, and H. Takayanagi, Solid State Commun. [**85**]{}, 321 (1993); Y. Tanaka, S. Kashiwaya, and H. Takayanagi, Jpn. J.Appl. Phys. [**34**]{}, 4566 (1995); D. Rainer, J. A. Sauls, and D.Waxman, Phys. Rev. B [**54**]{}, 10094 (1996); S. M. M. Virtanen and M. M. Salomaa, Phys. Rev. B [**60**]{}, 14581 (1999).
A. S. Mel’nikov and V. M. Vinokur, Nature, [**415**]{}, 60 (2002); Phys. Rev. B [**65**]{}, 224514 (2002).
K. Tanaka, I. Robel, and B. Janko, Proc. Natl. Acad. Sci. U.S.A. [**99**]{}, 5233 (2002).
A. S. Melnikov and M. A. Silaev, Pis’ma Zh. Eksp. Teor. Fiz. [**83**]{}, 675 (2006) \[JETP Lett. [**83**]{}, 578 (2006)\].
A. S. Melnikov, D. A. Ryzhov, and M. A. Silaev Phys. Rev. B [**78**]{}, 064513 (2008).
N. B. Kopnin, A. S. Mel’nikov, D. A. Ryzhov, I. A. Shereshevskii, V. I. Pozdnyakova, and V. M. Vinokur, Phys. Rev. Lett. [**95**]{}, 197002 (2005).
N. B. Kopnin, A. S. Mel’nikov, V. I. Pozdnyakova, D. A. Ryzhov, I. A. Shereshevskii, and V. M. Vinokur, Phys. Rev. B [**75**]{}, 024514 (2007).
S. Graser, C. Iniotakis, T. Dahm, and N. Schopohl, Phys. Rev. Lett. [**93**]{}, 247001 (2004).
J. D. Shore, Ming Huang, A. T. Dorsey, and J. P. Sethna, Phys.Rev. Lett. [**62**]{}, 3089 (1989).
J. Bardeen, R. Kummel, A. E. Jacobs, and L. Tewordt, Phys. Rev. [**187**]{}, 556 (1969).
G. Eilenberger, Z. Phys. [**214**]{}, 195 (1968); A. I. Larkin and Y.N. Ovchinnikov, Zh. Eksp. Teor. Fiz., [**55**]{}, 2262 (1968); \[Sov.Phys. JETP, [**28**]{}, 1200 (1969)\]; J. W. Serene and D. Rainer, Phys.Rep. [**101**]{}, 22 (1983).
N. Schopohl, arXiv:9804064.
J. R. Clem, J. Low Temp. Phys. [**18**]{}, 427 (1975).
Brun E. Hansen, Phys. Lett. A [**27**]{}, 576 (1968).
L. D. Landau and E. M. Lifshitz [*Quantum mechanics. Non-relativistic theory*]{}, (Pergamon, New York, 1991).
Y. E. Kraus, A. Auerbach, H. A. Fertig, and S. H. Simon, Phys.Rev. Lett. [**101**]{}, 267002 (2008).
I. Guillamon, H. Suderow, S. Vieira, A. Fernandez-Pacheco, J. Sese, R. Cordoba, J. M. De Teresa, and M. R. Ibarra, Journal of Physics: Conference Series (to be published).
M. Zalalutdinov, H. Fujioka, Y. Hashimoto, S. Katsumoto, and Y. Iye, Physica B [**284**]{}-[**288**]{}, 817 (2000).
T. Cren, D. Fokin, F. Debontridder, V. Dubost, and D. Roditchev, (unpublished).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: '**Artificial neural networks (ANNs) suffer from catastrophic forgetting when trained on a sequence of tasks. While this phenomenon was studied in the past, there is only very limited recent research on this phenomenon. We propose a method for determining the contribution of individual parameters in an ANN to catastrophic forgetting. The method is used to analyze an ANNs response to three different continual learning scenarios.**'
address: 'Institute of Signal Processing and System Theory, University of Stuttgart, Germany'
bibliography:
- 'refs.bib'
title: Localizing Catastrophic Forgetting in Neural Networks
---
Continual Learning, Catastrophic Forgetting, Path Integral, Localization
Introduction {#sec:intro}
============
Artificial neural networks suffer from a phenomenon called catastrophic forgetting, which is characterized by a rapid decrease in performance on a learned task when trained on a new task [@ratcliff1990connectionist; @mccloskey1989catastrophic]. For example an ANN trained on machine translation between English and German will essentially “forget” everything it has learned when it is trained on translating between German and French. This is in contrast to human learning, where a human typically will remember at least something he or she has learned on a past task. Solving this problem of catastrophic forgetting, i.e. enabling continual learning of ANNs, is of great interest, because it can enable the accumulation of knowledge over possibly long periods of time without requiring training examples from all but the most recent task. This comes with a number of benefits compared with the current standard of jointly training an ANN on all tasks simultaneously. First: Since training would not require examples from all previously learned tasks, data for a task that has already been learned is no longer needed and can be discarded, reducing the memory required for training. Second: After a ANN was trained to solve some tasks, it is not static but can be adjusted to solve newly and potentially unforeseen tasks. Third: The overall time required for training an ANN on a sequence of tasks could be reduced, since it only needs to be trained on the new task without retraining with the data of previously learned tasks.\
\
While catastrophic forgetting is known in the literature since 1989 and has been studied quite intensively in the past [@robins1996consolidation; @yamaguchi2004reassessment; @french1999catastrophic; @hetherington1993catastrophic], interest in this phenomenon has decayed over the years. Only recently there has been a renewed interest in solving this problem. Many new methods for overcoming catastrophic forgetting like *Elastic Weight Consolidation* (EWC), *Synaptic Intelligence* (SI), *Deep Generative Replay* (DGR), *Variational Continual Learning* (VCL) and more have been proposed [@kirkpatrick2017overcoming; @zenke2017continual; @shin2017continual; @v.2018variational]. Although these works propose new ways of mitigating catastrophic forgetting there is only a very limited research on the phenomenon itself. One example of this is an empirical study on catastrophic forgetting by Goodfellow et al. [@goodfellow2013empirical]. In their work the authors compare different activation functions and their effect on mitigating catastrophic forgetting. The choice of activation function is a vital part of designing a neural architecture and its resilience to catastrophic forgetting is import, but it does not give an insight into the internal mechanisms of an ANN.\
\
In this paper, we study catastrophic forgetting in ANNs by quantifying which part of the network contributed with what extend to forget a previously learned task. While catastrophic forgetting in previous works is measured as a scalar value, e.g. the increase of loss or decrease of accuracy on a previously learned task, we propose a method to quantify catastrophic forgetting separately for every parameter in an ANN. This not only allows for a coarse analysis, i.e. if a ANN experiences catastrophic forgetting or not, but it also localizes which part of a neural architecture contributes to which extend of forgetting a previously learned task. We think that a deeper understanding of catastrophic forgetting in ANNs, enabled through this work, can lead to better methods for overcoming it.
Methods {#sec:metho}
=======
In this section we describe the notation, define different scenarios and methods used in this work.
Notation {#ssec:notat}
--------
In order to clearly define a sequence of tasks on which a ANN is trained, we borrow the notation used in the remainder of this paper from the closely related field of transfer learning with a slight modification[^1] in order to avoid confusion [@weiss2016survey]. We start by defining a domain $\mathcal{D}$ which consists of two parts, a feature space $\mathcal{X}$ and a marginal data generating distribution $P(\mathbf{X})$, where $\mathbf{X}=\lbrace\mathbf{x}_{1},\ldots,\mathbf{x}_{N}\rbrace\in\mathcal{X}$ is a set of training examples. In image classification, the feature space is given by $\mathcal{X}=\lbrace 0,1,\ldots,255\rbrace^{N\times C}$, where $N$ and $C$ are the number of pixels and channels an image contains. An assignment $\mathcal{A}$ for a given domain $\mathcal{D}$ is again defined by two parts, a label space $\mathcal{Y}$ and a function $f:\mathcal{X}\rightarrow\mathcal{Y}$, which represents the mapping from feature to label space. The function $f$ is learned from pairs $\lbrace x_{i},y_{i}\rbrace$, where $x_{i}\in\mathcal{X}$ is a training example and $y_{i}\in\mathcal{Y}$ is the corresponding label. In image classification, this function maps an image to its label. With this notation we can define the phenomenon of catastrophic forgetting more precisely as a rapid decrease in performance of an ANN on Task $A$, defined by $\mathcal{D}_{A}$ and $\mathcal{A}_{A}$, as it is trained on Task $B$, defined by $\mathcal{D}_{B}$ and $\mathcal{A}_{B}$, if $\mathcal{D}_{B}\neq\mathcal{D}_{A}$ and/or $\mathcal{A}_{B}\neq\mathcal{A}_{A}$.
Continual Learning Scenarios {#ssec:CLsce}
----------------------------
Although the recent work on continual learning shares the same goal of mitigating catastrophic forgetting, different experimental setups are used to evaluate the proposed methods. These differ significantly and pose different challenges to the algorithms and methods which are evaluated on them. In order to make the research in this area more comparable, three different scenarios of continual learning were recently proposed [@Hsu2018ReevaluatingCL; @Ven2018GenerativeRW]. These categories are defined in the following subsections.
### Incremental Domain Learning {#sssec:IncDo}
Incremental domain learning (IDL) is characterized by a change in at least one part of the domain $\mathcal{D}$, either the feature space $\mathcal{X}$, the data generating distribution $P(\mathbf{X})$ or both change. This scenario is similar but not identical to domain adaptation in the field of transfer learning [@weiss2016survey]. The difference between domain adaptation and IDL is given by the fact, that in domain adaption one is only interested in transferring an ANNs knowledge from domain $\mathcal{D}_{A}$ to $\mathcal{D}_{B}$. After this transfer the ANNs performance on Domain $\mathcal{D}_{A}$ is typically irrelevant. In IDL this is not the case. Here one is interested learning to solve a task on domain $\mathcal{D}_{A}$ and on $\mathcal{D}_{B}$ without catastrophic forgetting and possibly a transfer of knowledge between the two domains. Another important property of IDL is that the assignment remains unchanged, i.e. the label space $\mathcal{Y}$ and the function $f$ remain unchanged. This can be represented more formally with $f(\mathbf{x}_{i})=f(\hat{\mathbf{x}}_{i})=\mathbf{y}_{i}$, where $\hat{\mathbf{x}}_{i}$ is the representation of $\mathbf{x}_{i}$ in a different domain. In practice this means, that we can share the same output layer of an ANN over different domains. A widely used example for this scenario is permutation MNIST [@kirkpatrick2017overcoming; @zenke2017continual; @shin2017continual; @v.2018variational]. In order to generate different domains, random pixel permutations of images in the MNIST are used, where one realization of a permutation is applied to all images. Although this does not change the feature space $\mathcal{X}$, it changes the data generating distribution $P(\mathbf{X})$ and hence the domain $\mathcal{D}$. The random permutations used in this example generate uncorrelated domains, which is not very realistic and has caused some criticism [@zenke2017continual; @Hsu2018ReevaluatingCL]. A very simple example of IDL with highly correlated, and therefore more realistic, domains based on the MNIST data set can be generated by just inverting the pixel intensities.
### Incremental Class Learning {#sssec:IncCl}
In incremental class learning (ICL) each task adds one or possibly more new classes to classify by an ANN. In each task the ANN is presented with a data set containing only examples of at least one new class to learn. This means that not only the domain $\mathcal{D}$ but also the assignment $\mathcal{A}$ changes between tasks. Formally, the feature space $\mathcal{X}$ and/or the data generating distribution $P(\mathbf{X})$ and the function $f$ change. The label space $\mathcal{Y}$ remains unchanged and therefore the output layer of the ANN can also be shared between tasks. A widely used example for this scenario is split MNIST [@shin2017continual]. In order to generate a sequence of tasks, the MNIST data set is split in such a way that each split contains all the examples from at least one class. A typical way to split the MNIST data set is to separate it into $5$ disjoint sets each containing two classes, e.g. $(0,1)$, $(2,3)$, $(4,5)$, $(6,7)$ and $(8,9)$. In practice this means that in the first task only examples of the classes $0$ and $1$ are used to train an ANN while its output layer contains $10$ neurons and is therefore able to distinct between $10$ different classes.
### Incremental Task Learning {#sssec:IncTa}
The last scenario, incremental task learning (ITL), also allows for changes in both the domain $\mathcal{D}$ and the assignment $\mathcal{A}$ between tasks. In contrast to ICL the label space $\mathcal{Y}$ also changes in ITL, i.e. a ANN can first learn a classification and then a regression task. Since these tasks typically require different activation functions for the last layer of an ANN, a new output layer is used for every task. Such an ANN is known as a multi-headed ANN in the continual learning literature [@v.2018variational; @Hsu2018ReevaluatingCL]. This implies that during inference the identity of a task, which needs to be solved by an ANN, is known in order to select the corresponding head. Requiring such a prior knowledge about the task identity is in stark contrast to ICL, where the ANN not only solves the task at hand but also infers which task is to be solved. A typical example for a sequence of tasks for ITL can be generate from split MNIST or a set of different data sets, where each split or data set has to be learned in each task, by using a different output layer for each task.\
\
These three scenarios for continual learning differ not only in their setup but also in the challenges they pose to an ANN. While ICL requires the model not only to solve a task but also to recognize which task, from all the previously learned tasks, it has to solve. IDL on the other hand requires an ANN to solve the same task across different domains. The last scenario ITL just requires a model to solve individual tasks while sharing only a subset of the network across them.
Proposed Method {#ssec:PropM}
---------------
![Proposed method for localizing catastrophic forgetting[]{data-label="fig:method"}](method){width="1.0\linewidth"}
In order to quantify catastrophic forgetting we need a measure for the performance of an ANN on a given task. Since ANNs are typically trained using stochastic gradient descent, or some variant of it, to minimize a defined loss function $\mathcal{L}(\boldsymbol{\theta}, \mathcal{D}, \mathcal{A})$ with respect to its parameters $\boldsymbol{\theta}$, it is a natural choice to use this loss function for quantifying catastrophic forgetting. To simplify the notation and reduce clutter we will use $\mathcal{L}_{A}(\boldsymbol{\theta})$ for a loss function $\mathcal{L}(\boldsymbol{\theta}, \mathcal{D}_{A}, \mathcal{A}_{A})$, which is defined on domain $\mathcal{D}_{A}$ with assignment $\mathcal{A}_{A}$.\
\
Similar to the inspiring work of Zenke et al. [@zenke2017continual] we interpret the training process of an ANN as a trajectory in parameter space defined by $\boldsymbol{\theta}(n)$, where $n\in\mathbb{N}^{+}$ is the current training step. Moving the parameters $\boldsymbol{\theta}$ along this trajectory causes a change in the loss function $\Delta\mathcal{L}$. If we compute the gradient $\boldsymbol{\nabla}_{\boldsymbol{\theta}}\mathcal{L}$ at each point in parameter space, we can compute $\Delta\mathcal{L}$ either through the difference in loss between the start- and endpoint or through the path integral of $\boldsymbol{\nabla}_{\boldsymbol{\theta}}\mathcal{L}$ along the trajectory as $$\begin{aligned}
\label{eq:deltaL}
\Delta\mathcal{L}=\mathcal{L}(\boldsymbol{\theta}(N))-\mathcal{L}(\boldsymbol{\theta}(1))=\int_{\mathcal{C}}\boldsymbol{\nabla}_{\boldsymbol{\theta}}\mathcal{L}(\boldsymbol{\theta})\mathrm{d}\boldsymbol{\theta},\end{aligned}$$ where $N$ is the number of training steps and $\mathcal{C}$ is the trajectory of $\boldsymbol{\theta}$ through parameter space during training. This equivalence holds, since the gradient vector field is a conservative field. Although both methods for computing this change yield the same result, they differ in their complexity and the insight they can provide. While evaluating $\Delta\mathcal{L}$ via the difference in loss at the start- and endpoint is fast and simple, it can only provide information about the ANN as a whole. Using the path integral is computationally more expensive and therefore slower, but it enables us to determine the contribution of individual parameters. In order to calculate a parameter specific contribution to the change in loss, we decompose the path integral and approximate it with a sum as $$\begin{aligned}
\Delta\mathcal{L}&=\int_{\mathcal{C}}\sum_{i}\boldsymbol{\nabla}_{\theta_{i}}\mathcal{L}(\boldsymbol{\theta})\mathrm{d}\theta_{i}\nonumber\\&\approx\sum_{j=1}^{N}\sum_{i}\boldsymbol{\nabla}_{\theta_{i}}\mathcal{L}(\boldsymbol{\theta}_{j})\Delta\theta_{ij}=\sum_{i}\Delta\mathcal{L}_{i},\end{aligned}$$ where $\Delta\theta_{ij}$ is a small change in the $i$th parameter at training step $j$. With this approximation we can determine the individual contribution of the $i$th parameter $\Delta\mathcal{L}_{i}$ to the overall change in loss $\Delta\mathcal{L}$. In order to check if the approximation is accurate, we can use equation \[eq:deltaL\] to compute $\Delta\mathcal{L}$ exactly and compare it with our approximation. In general the accuracy will depend on the curvature of the loss surface and the step size used for parameter updates. If there is an unacceptable difference between the exact change and the proposed approximation, one can improve the accuracy by inserting some intermediate steps for evaluation of the path integral between two parameter updates.\
\
Catastrophic forgetting occurs when we transition from training and ANN on minimizing $\mathcal{L}_{A}(\boldsymbol{\theta})$ to it being trained to minimize $\mathcal{L}_{B}(\boldsymbol{\theta})$ and is characterized by a rapid increase of the former right after the transition. This period of rapid change is of particular interest, since it represents the period over which the ANN forgets a previously learned task. Determining the contributions of individual parameters is therefore most useful right after the transition and has to be done for the loss function $\mathcal{L}_{A}(\boldsymbol{\theta})$. This process is illustrated in Fig. \[fig:method\].\
\
Although we limit our study of catastrophic forgetting to the scenarios described in section \[ssec:CLsce\], where every task represents a supervised classification problem, the proposed method can be applied to many other settings. For any continual learning task, which involves training an ANN to minimize a loss functions over a sequence of tasks, the change in loss can be approximated as described above. Examples for such sequences of task include, but are not limited to, learning of representations over different domains, a sequence of regression tasks or training generative models continually to capture different data generating distributions.
Experiments {#sec:exper}
===========
In this section we will introduce the model used for the following experiments and describe the exact realization of the different continual learning scenarios introduced in section \[ssec:CLsce\]. We use the same architecture on all three scenarios in order to allow for a comparison of the results. Since a ANNs architecture can have a significant influence on its resilience to catastrophic forgetting [@goodfellow2013empirical] and we are interested in analyzing what challenges the different scenarios pose to an ANN, changing the models structure between evaluations is avoided. The architecture used in this work is a small convolutional neural network (CNN) with four hidden layers according to table \[tbl:CNN\]. Dropout is applied to the input of respective layers and a stride of $2$ is used in all convolutions.
[cccc]{}\
Layer & Act. Size & Act. Func. & Dropout\
Input & $28\times 28\times 1$ & - & -\
Conv. $32\times 3\times 3$& $14\times 14\times 32$ & ReLU & -\
Conv. $32\times 3\times 3$& $7\times 7\times 32$ & ReLU & -\
Dense & $64$ & ReLU & $0.2$\
Dense & $32$ & ReLU & $0.2$\
Dense & $10$ & Soft max & $0.2$\
\[tbl:CNN\]
Incremental Task Learning {#ssec:IncTa}
-------------------------
As described in section \[ssec:CLsce\], ITL is characterized by a change in the domain $\mathcal{D}$ and the assignment $\mathcal{A}$. In order to generate two tasks, which can be used to localize catastrophic forgetting in this scenario, we utilize two popular data sets, MNIST [@lecun-mnisthandwrittendigit-2010] and FashionMNIST [@xiao2017/online]. While MNIST is a data set for handwritten digit classification with $60000$ training and $10000$ test samples of size $28\times 28\times 1$, FashionMNIST is a drop in replacement for MNIST containing images of fashion from 10 different categories.\
\
The sequence of tasks is created by first training the ANN on classifying the handwritten digits of MNIST and then the different fashion categories of FashionMNIST. During this sequence the domain and assignment change. Considering the domains of both tasks, $\mathcal{D}_{M}$ and $\mathcal{D}_{F}$, the feature space $\mathcal{X}_{M}=\mathcal{X}_{F}=\lbrace 0,1,\ldots,255\rbrace^{28\times 28\times 1}$ remains unchanged, while the data generating distributions change, i.e. $\mathcal{P}_{M}(\mathbf{X})\neq\mathcal{P}_{F}(\mathbf{X})$. The assignments, $\mathcal{A}_{M}$ and $\mathcal{A}_{F}$, differ in both the label space and the function learned from training examples, i.e. $\mathcal{Y}_{M}\neq\mathcal{Y}_{F}$ and $f_{M}\neq f_{F}$. We realize this by utilizing a separate output layer, with $10$ neurons, of the ANN for each task.
Incremental Domain Learning {#ssec:IncDo}
---------------------------
In IDL the domain $\mathcal{D}$ changes between tasks while the assignment $\mathcal{A}$ is unchanged. This means, although the feature space $\mathcal{X}$ and/or the corresponding data generating distribution $P(\mathbf{X})$ change, the ANN has to solve the same task but based on different inputs. For experiments on IDL we again utilize the MNIST data set.\
\
A common way to generate a sequence of tasks for IDL based on the MNIST data set is to apply a different random permutation of pixels to each image in the data set for every new task. This is known as permutation MNIST in the literature [@kirkpatrick2017overcoming; @zenke2017continual; @shin2017continual; @v.2018variational; @Ven2018GenerativeRW]. Formally we change the domain $\mathcal{D}$ between tasks by changing the data generating distribution $P(\mathbf{X})$, while the feature space $\mathcal{X}=\lbrace 0,1,\ldots,255\rbrace^{28\times 28\times 1}$ remains unchanged. It is possible to create a very large number of different tasks using this method, but most of the generated domains are uncorrelated and do not resemble natural images.
Incremental Class Learning {#ssec:IncCl}
--------------------------
Experiments for ICL are commonly based on learning the classes in one specific data set in an incremental way. This typically corresponds to a change in both the domain $\mathcal{D}$ and the assignment $\mathcal{A}$ between tasks. But, in contrast to ITL, the feature space $\mathcal{X}$ and the label space $\mathcal{Y}$ are shared between the tasks.\
\
In order to generate a sequence of tasks for ICL, a data set is commonly split into disjoint subsets, where each subset contains only examples of one or more classes. A widely used example for this is split MNIST [@zenke2017continual; @shin2017continual; @v.2018variational; @Ven2018GenerativeRW]. In this case the MNIST data set is commonly split into 5 subsets containing two classes each, e.g. $(0,1)$, $(2,3)$, $(4,5)$, $(6,7)$ and $(8,9)$. Learning to classify the classes in each of these subsets is considered a task.
Results {#sec:resul}
=======
![Results of the experiments described in section \[sec:exper\]. This plot shows the absolute contribution of the weight matrices and bias vectors of individual layers. The ordering from left to right mirrors the CNN’s structure from table \[tbl:CNN\].[]{data-label="fig:combinedresultssum"}](combined_results_sum){width="1.0\linewidth"}
![Results of the experiments described in section \[sec:exper\]. This plot shows the average contribution of an element in the weight matrices and bias vectors of individual layers. The ordering from left to right mirrors the CNN’s structure from table \[tbl:CNN\].[]{data-label="fig:combinedresultsmean"}](combined_results_mean){width="1.0\linewidth"}
In this section we present the results obtained during the experiments described in section \[sec:exper\]. We use the CNN architecture depicted in table \[tbl:CNN\] and train with the Adam optimizer and a learning rate of $0.001$ and a batch size of $128$ for $10$ epochs on every task. No learning rate schedules or early stopping was used. Since the extend of catastrophic forgetting depends not only on the architecture used but also the random initialization of the model we run every experiment $10$ times and report average values with their standard deviation. Although an interpretation of the results is difficult, since it is highly dependent on many different factors like the model architecture, weight initialization, the optimizer and many other hyperparameters, we can at least compare the same configuration across the three different continual learning scenarios introduced in section \[ssec:CLsce\].\
\
Figure \[fig:combinedresultssum\] shows the absolute change in loss aggregated over the weight matrices/tensors and bias vectors of every layer. The ordering from left to right corresponds with the architecture shown in table \[tbl:CNN\] from top to bottom. Comparing the overall distribution of change in loss over the different continual learning scenarios, we can observe distinct patterns for these. The absolute contributions of the convolutional layers on all scenarios is lower than those of the dense layers. Also the absolute contribution of the bias vectors is small when compared to the weight matrices. We can even identify an average decrease in loss for the weight tensor of the second convolutional layer. On ICL the variance of this change in loss is however very high when compared to the other scenarios. While the convolutional layers show a more or less homogeneous change over the different continual learning scenarios, we can observe an interesting difference in the dense layers across them. On ITL and IDL the absolute contribution of the layers decreases from left to right. This is expected since the overall number of neurons in these layers also decreases from left to right. On ICL however we can observe an opposite behavior. Although the number of neurons decreases, the over all contribution to the change in loss increases. This observation is in line with Farguhar & Gal [@Farquhar2018TowardsRE] who observe more catastrophic forgetting on split MNIST than on permutation MNIST and reason that this is caused by gradients with higher magnitude while training the last layer due to more similar looking images in ICL when compared with IDL. Although our observations also indicate that in ICL the last layers are mostly responsible for the change in loss and therefore catastrophic forgetting, we can not give a general explanation for this when considering the limited scope of our experiments. But we can at least support Farguhar & Gal observation that the last layers are mostly responsible for catastrophic forgetting in ICL. This becomes even more evident when we consider the average contribution of a neuron/filter over the layers as shown in figure \[fig:combinedresultsmean\]. Here we have averaged the contributions of neurons, filters without their bias elements, which are plotted separately. Comparing ITL, IDL and ICL we can again observe that the average contribution of a neuron/filter increases when going from ITL to IDL and reaches its maximum for ICL. But we can also again observe that while on ITL and IDL the average contribution of a neuron/filter is approx constant over the dense layers while on ICL it increases from the first dense layer to the output layer.\
\
Overall we can observe different responses of the studied architecture when exposed to the three continual learning scenarios. While ITL causes the least catastrophic forgetting, the evaluation of IDL shows very similar behavior but increased catastrophic forgetting. Our evaluation on ICL not only shows the overall highest change in loss but also a very different pattern than the other two scenarios.
Conclusion {#sec:concl}
==========
Catastrophic forgetting is a fundamental problem in the training process of ANNs. Although it was studied in the past, there was surprisingly few research on the phenomenon itself published over the recent years. We proposed a method for determining the contribution of individual parameters in an ANN to a change in loss, which can be linked to catastrophic forgetting. This method allows a more detailed analysis of this phenomenon through a localization of parts in an ANN that contribute the most to such a change in loss. We evaluated our method on three different continual learning scenarios on common data sets in the field. We could not only support claims from other researchers based on a different experimental evaluation but also found similarities and differences in the response of a specific ANN, which was exposed to these different scenarios.
[^1]: We use the term “assignment” for what is referred to as a “task” in transfer learning.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Given a profinite group $G$ with finite virtual cohomological dimension, let $\{X_i\}$ be a tower of discrete $G$-spectra, each of which is fibrant as a spectrum, so that $X=\operatorname*{holim}_i X_i$ is a continuous $G$-spectrum, with homotopy fixed point spectrum $X^{hG}$. The $E_2$-term of the descent spectral sequence for $\pi_\ast(X^{hG})$ cannot always be expressed as continuous cohomology. However, we show that the $E_2$-term is always built out of a certain complex of spectra, that, in the context of abelian groups, is used to compute the continuous cochain cohomology of $G$ with coefficients in $\lim_i M_i$, where $\{M_i\}$ is a tower of discrete $G$-modules.'
author:
- 'Daniel G. Davis$\sp 1$'
bibliography:
- 'biblio.bib'
date: 'June 21, 2006'
title: 'The $E_2$-term of the descent spectral sequence for continuous $G$-spectra'
---
[Introduction]{}
In this note, $G$ always denotes a profinite group. Let $H_c^*(G;M)$ denote the continuous cohomology of $G$ with coefficients in the discrete $G$-module $M$. This cohomology is defined as the right derived functors of $G$-fixed points. Then we always assume that $G$ has finite virtual cohomological dimension; that is, there exists an open subgroup $H$ and a non-negative integer $m$, such that $H^s_c(H;M)=0$, for all discrete $H$-modules $M$ and all $s \geq m$.
All of our spectra are Bousfield-Friedlander spectra of simplicial sets. In particular, a [*discrete $G$-spectrum*]{} is a $G$-spectrum such that each simplicial set $X_k$ is a simplicial object in the category of discrete $G$-sets (thus, for any $l \geq 0$, the action map on the $l$-simplices, $G \times (X_k)_l
\rightarrow (X_k)_l$, is continuous when $(X_k)_l$ is regarded as a discrete space). The category of discrete $G$-spectra, with morphisms being $G$-equivariant maps of spectra, is denoted by $\mathrm{Spt}_G$.
Discrete $G$-spectra are considered in more detail in [@cts], which shows (see [@cts Theorem 3.6]) that $\mathrm{Spt}_G$ is a model category, where a morphism $f$ in $\mathrm{Spt}_G$ is a weak equivalence (cofibration) if and only if $f$ is a weak equivalence (cofibration) in $\mathrm{Spt}$, the category of spectra. Given a discrete $G$-spectrum $X$, the [*homotopy fixed point spectrum*]{} $X^{hG}$ is obtained as the total right derived functor of fixed points: $X^{hG} = (X_{f,G})^G,$ where $X \rightarrow X_{f,G}$ is a trivial cofibration and $X_{f,G}$ is fibrant, all in $\mathrm{Spt}_G$.
Let $X_0 \leftarrow
X_1 \leftarrow X_2 \leftarrow \cdots$ be a tower of discrete $G$-spectra, such that each $X_i$ is a fibrant spectrum. As explained in [@cts Lemma 4.4], there exists a tower $\{X_i'\}$ of discrete $G$-spectra, such that there are weak equivalences $$\operatorname*{holim}_i X_i \overset{\simeq}{\longrightarrow} \operatorname*{holim}_i X_i'
\overset{\simeq}{\longleftarrow} \lim_i X_i'.$$ (In this paper, $\mathrm{holim}$ always denotes the version of the homotopy limit of spectra that is constructed levelwise in the category of simplicial sets, as defined in [@Bousfield/Kan] and [@Thomason 5.6].) Since the inverse limit of a tower of discrete $G$-sets is a topological $G$-space, and because $\operatorname*{holim}_i X_i$ can be identified with $\lim_i X_i'$, in the eyes of homotopy, $\operatorname*{holim}_i X_i$ is a [*continuous $G$-spectrum*]{}. Notice that, under this identification, the continuous $G$-action respects the topology of both $G$ and all the $X_i$ together. Continuous $G$-spectra and examples of such in chromatic stable homotopy theory are considered in [@cts].
Given the tower $\{X_i\}$ and the continuous $G$-spectrum $\operatorname*{holim}_i X_i$, its [*homotopy fixed point spectrum*]{}, $(\operatorname*{holim}_i X_i)^{hG}$, is defined to be $\operatorname*{holim}_i (X_i)^{hG}.$ This construction is called homotopy fixed points because it is equivalent to the usual definition when $G$ is a finite group and it is the total right derived functor of fixed points in the appropriate sense (see [@cts Remark 8.4]).
By [@cts Theorem 8.8], thanks to the assumption of finite virtual cohomological dimension, there is a descent spectral sequence $${\addtocounter{Lem}{1}\tag{\theLem}}\label{dss}
E_2^{s,t} \Rightarrow \pi_{t-s}((\operatorname*{holim}_i X_i)^{hG}),$$ where $${\addtocounter{Lem}{1}\tag{\theLem}}\label{e2}
E_2^{s,t} = \pi^s\pi_t(\operatorname*{holim}_i (\Gamma^\bullet_G(X_i)_{f,G})^G)$$ (see the beginning of Section 2 for the meaning of $\Gamma_G^\bullet$), and, if the tower of abelian groups $\{\pi_t(X_i)\}$ satisfies the Mittag-Leffler condition for every integer $t$, then $E_2^{s,t} \cong
H^s_\mathrm{cont}(G;\{\pi_t(X_i)\}),$ which is continuous cohomology in the sense of Jannsen. (This cohomology is obtained by taking the right derived functors of $\lim_i (-)^G$, a functor from towers of discrete $G$-modules to abelian groups; see [@Jannsen].)
In expression (\[e2\]), since $\pi_t(\operatorname*{holim}_i (-))$ is not necessarily $\lim_i \pi_t(-)$, the $E_2$-term of descent spectral sequence (\[dss\]), in general, can not be expressed as continuous cohomology, and, in general, it has no compact algebraic description. For example, as pointed out in [@cts Remark 8.10], when $G=\{e\},$ $$H^s_\mathrm{cont}(\{e\};\{\pi_t(X_i)\}) = \lim_i \pi_t(X_i)$$ and $E_2^{0,t} = \pi_t(\operatorname*{holim}_i X_i)$, and these need not be isomorphic, due to the familiar $\operatorname*{lim^1}_i \pi_{t+1}(X_i)$ obstruction. However, in this note, we show that the $E_2$-term (\[e2\]) can always be described in an interesting way.
In more detail, Theorem \[main1\] gives a particular cochain complex $\mathcal{C}^\ast$ for computing the continuous cochain cohomology of $G$ for a topological $G$-module $\lim_i M_i,$ where $\{M_i\}$ is a tower of discrete $G$-modules. In Corollary \[main3\], we show that the $E_2$-term of (\[e2\]) can always be given by taking the cohomology of the homotopy groups of the complex $\mathcal{C}^\ast$, where $\lim_i M_i$ is replaced by the continuous $G$-spectrum $\operatorname*{holim}_i X_i$, in an appropriate sense. This presentation of the $E_2$-term shows that $E_2^{\ast,\ast}$ always takes into account the topology of the continuous $G$-spectrum, even when it cannot be expressed as continuous cohomology.
Section 3 of this note consists of a discussion of the construction of descent spectral sequence (\[dss\]). Section 5 explains why two other potentially plausible interpretations of (\[e2\]) fail to work.
**Acknowledgements.** I am grateful to the referee for helpful comments and for pointing out that the statement of the main result could be simplified. Also, I thank Paul Goerss and Halvard Fausk for useful remarks.
[The pro-discrete cochain complex and continuous cohomology]{}
We begin this section with some terminology. If $\mathcal{C}$ is a category, then $\mathbf{tow}(\mathcal{C})$ is the category of towers $$C_0 \leftarrow C_1 \leftarrow C_2 \leftarrow \cdots$$ in $\mathcal{C}$. The morphisms $\{f_i\}$ are natural transformations such that each $f_i$ is a morphism in $\mathcal{C}$. In this note, we will be working with $\mathbf{tow}(\mathrm{DMod}(G))$, where $\mathrm{DMod}(G)$ is the category of discrete $G$-modules, and $\mathbf{tow}(\mathrm{Spt}_G)$.
If $A$ is an abelian group with the discrete topology, let $\mathrm{Map}_c(G,A)$ be the abelian group of continuous maps from $G$ to $A$. If $X$ is a spectrum, one can also define $\mathrm{Map}_c(G,X),$ where the $l$-simplices of the $k$th simplicial set $(\mathrm{Map}_c(G,X)_k)_l$ are given by $\mathrm{Map}_c(G,(X_k)_l),$ where $(X_k)_l$ is given the discrete topology.
Consider the functor $$\Gamma_G \colon \mathrm{Spt}_G \rightarrow \mathrm{Spt}_G, \ \ \
X \mapsto \Gamma_G(X) = \mathrm{Map}_c(G,X),$$ where the action of $G$ on $\mathrm{Map}_c(G,X)$ is induced on the level of sets by $(g \cdot f)(g') = f(g'g),$ for $g, g' \in G$ and $f \in \mathrm{Map}_c(G,(X_k)_l),$ for each $k, l \geq 0.$ As explained in [@cts Definition 7.1], the functor $\Gamma_G$ forms a triple and there is a cosimplicial discrete $G$-spectrum $\Gamma^\bullet_G X.$ Also, it is clear that $\Gamma_G \: \mathrm{DMod}(G) \rightarrow \mathrm{DMod}(G)$ can be defined as above, so that, given a discrete $G$-module $M$, $\Gamma^\bullet_G M$ is a cosimplicial discrete $G$-module.
We do not claim any originality in the following definition.
Let $\{X_i\}$ be an object in $\mathbf{tow}(\mathrm{DMod}(G))$ or in $\mathbf{tow}(\mathrm{Spt}_G)$. Then the [*pro-discrete cochain complex*]{} is defined to be the complex $$\mathcal{C}^\ast(G;\{X_i\}) =
\lim_i (\Gamma^\ast_G X_i)^G,$$ where $\Gamma^\ast_G X_i$ is the canonical complex associated to $\Gamma^\bullet_G X_i,$ where, if $\{X_i\}$ is in $\mathbf{tow}(\mathrm{Spt}_G)$, then the complex lives in the stable homotopy category. The pro-discrete cochain complex is a complex of abelian groups or spectra, respectively, and the limit and colimit are both formed in abelian groups or in spectra (not the stable homotopy category), respectively.
Let $M$ be any topological $G$-module (that is, an abelian Hausdorff topological group that is a $G$-module, with a continuous $G$-action). Then the [ *continuous cochain cohomology of $G$ with coefficients in $M$*]{}, $H^*_\mathrm{cts}(G;M)$, is the cohomology of a cochain complex that has the form $${\addtocounter{Lem}{1}\tag{\theLem}}\label{cplx}
M \rightarrow \mathrm{Map}_c(G,M) \rightarrow \mathrm{Map}_c(G^2,M)
\rightarrow \cdots$$ (see [@Neuk pg. 106] for details). We note that, by [@Jannsen Theorem (2.2)], if $\{M_i\}$ is a tower of discrete $G$-modules that satisfies the Mittag-Leffler condition, then $$H^s_\mathrm{cts}(G;\lim_i M_i) \cong H^s_\mathrm{cont}(G;\{M_i\}),$$ for all $s \geq 0$, but, in general, these two versions of continuous cohomology need not be isomorphic. Also, if $M$ is a discrete $G$-module, then $H^s_\mathrm{cts}(G;M) = H^s_c(G;M).$
Now we show that the pro-discrete cochain complex can be used to compute continuous cochain cohomology.
\[main1\] If $\{M_i\}$ is a tower of discrete $G$-modules, then $$H^s_\mathrm{cts}(G;\lim_i M_i) \cong H^s[\mathcal{C}^\ast(G;\{M_i\})].$$
As explained in [@Neuk pg. 106], for a topological $G$-module $M$, the chain complex in (\[cplx\]) is defined by taking the $G$-fixed points of the complex $${\addtocounter{Lem}{1}\tag{\theLem}}\label{cplx2}
X^\ast(G;M) = [\mathrm{Map}_c(G,M) \rightarrow \mathrm{Map}_c(G^2,M)
\rightarrow \cdots],$$ where $X^n(G;M) = \mathrm{Map}_c(G^{n+1},M)$ has a $G$-action that is defined by $$(g\cdot f)(g_1,...,g_{n+1}) = g \cdot f(g^{-1}g_1, ..., g^{-1}g_{n+1}).$$
Now let $M$ be a discrete $G$-module. Then it is a standard fact that the cochain complex $(X^\ast(G,M))^G$ is naturally isomorphic as a complex to the cochain complex $(\Gamma^\ast_G M)^G.$ This isomorphism uses the fact that the abelian group of $n$-cochains of $(\Gamma^\ast_G M)^G$ is isomorphic to $\mathrm{Map}_c(G^{n+1},M)^G,$ where $\mathrm{Map}_c(G^{n+1},M)$ has a $G$-action that is given by $$(g\cdot f)(g_1,g_2,g_3,...,g_{n+1})
= f(g_1g,g_2,g_3,...,g_{n+1}).$$
Since $$(X^n(G;\lim_i M_i))^G \cong
\lim_i ((X^n(G;M_i))^G) \cong \lim_i (\Gamma^{n+1}_G M_i)^G,$$ we have: $$H^s_\mathrm{cts}(G;\lim_i M_i) = H^s[(X^\ast(G;\lim_i M_i))^G]
= H^s[\lim_i (\Gamma^\ast_G M_i)^G],$$ where we used the aforementioned fact that $(X^\ast(G,M_i))^G$ and $(\Gamma^\ast_G M_i)^G$ are naturally isomorphic cochain complexes.
[Constructing the descent spectral sequence]{}
In this section, we review how descent spectral sequence (\[dss\]) is constructed and we compare it with a spectral sequence whose $E_2$-term is always Jannsen’s continuous cohomology.
Given a tower $\{X_i\}$ of discrete $G$-spectra, such that each $X_i$ is a fibrant spectrum, by [@cts Remark 7.8, Definition 8.1], $$(\operatorname*{holim}_i X_i)^{hG} =
\operatorname*{holim}_i \operatorname*{holim}_\Delta (\Gamma^\bullet_G (X_i)_{f,G})^G.$$ Thus, $$(\operatorname*{holim}_i X_i)^{hG} \cong
\operatorname*{holim}_\Delta (\operatorname*{holim}_i (\Gamma^\bullet_G (X_i)_{f,G})^G),$$ and descent spectral sequence (\[dss\]) is the conditionally convergent homotopy spectral sequence $$\operatorname*{lim^s}_\Delta \pi_t(\operatorname*{holim}_i (\Gamma^\bullet_G (X_i)_{f,G})^G)
\Rightarrow
\pi_{t-s}(\operatorname*{holim}_\Delta (\operatorname*{holim}_i (\Gamma^\bullet_G (X_i)_{f,G})^G))$$ (see [@cts Theorem 8.8]).
In the above context, there is another spectral sequence that is natural to consider. Since $$(\operatorname*{holim}_i X_i)^{hG} \cong
\operatorname*{holim}_{\Delta \times \{i\}} (\Gamma^\bullet_G (X_i)_{f,G})^G,$$ there is a conditionally convergent homotopy spectral sequence $$\operatorname*{lim^s}_{\Delta \times \{i\}} \pi_t((\Gamma^\bullet_G(X_i)_{f,G})^G)
\Rightarrow
\pi_{t-s}(\operatorname*{holim}_{\Delta \times \{i\}} (\Gamma^\bullet_G (X_i)_{f,G})^G)
\cong \pi_{t-s}((\operatorname*{holim}_i X_i)^{hG}),$$ such that $$\operatorname*{lim^s}_{\Delta \times \{i\}} \pi_t((\Gamma^\bullet_G(X_i)_{f,G})^G)
\cong H^s_\mathrm{cont}(G;\{\pi_t(X_i)\})$$ (see [@Geisser Proposition 3.1.2]). This spectral sequence is closely related to the $\ell$-adic descent spectral sequence of algebraic $K$-theory (see [@Thomason], [@Mitchell]).
We see that we have two spectral sequences with abutment $\pi_\ast((\operatorname*{holim}_i X_i)^{hG})$. As pointed out in the Introduction, if the tower $\{\pi_t(X_i)\}$ satisfies the Mittag-Leffler condition for every integer $t$, then the spectral sequences have isomorphic $E_2$-terms. However, as stated in the Introduction, the case $G=\{e\}$ shows that these two spectral sequences can have different $E_2$-terms, so that the spectral sequences can be different from each other.
Since the second spectral sequence described above has an $E_2$-term that always has a nice algebraic description, it is natural to ask what is the value of descent spectral sequence (\[dss\]). We will see that, because the descent spectral sequence is the homotopy spectral sequence of a cosimplicial spectrum, in certain cases it can be compared with an Adams-type spectral sequence that is strongly convergent.
Let $n \geq 1$ and let $p$ be a prime. Let $K(n)$ be the $n$th Morava $K$-theory spectrum with $K(n)_\ast =
\mathbb{F}_p[v_n^{\pm 1}],$ where the degree of $v_n$ is $2(p^n-1)$. Also, let $E_n$ denote the Lubin-Tate spectrum, where $E_{n \ast} =
W(\mathbb{F}_{p^n})[[u_1, ..., u_{n-1}]][u^{\pm 1}],$ where the degree of $u$ is $-2$, and the complete power series ring over the Witt vectors is in degree zero.
Let $Z$ be a $K(n)$-local spectrum and suppose that there is an augmentation $Z \rightarrow \operatorname*{holim}_i (\Gamma^\bullet_G (X_i)_{f,G})^G$. If the associated complex of spectra $$\ast \rightarrow Z \rightarrow \operatorname*{holim}_i \mathrm{Map}_c(G,(X_i)_{f,G})^G
\rightarrow \operatorname*{holim}_i \mathrm{Map}_c(G,\mathrm{Map}_c(G,(X_i)_{f,G}))^G
\rightarrow \cdots$$ is a $K(n)$-local $E_n$-resolution of $Z$ (for the definition of this, see [@DH Appendix A]), then descent spectral sequence (\[dss\]) is isomorphic to the strongly convergent $K(n)$-local $E_n$-Adams spectral sequence with abutment $\pi_\ast(Z)$ (see [@DH Proposition A.5, Corollary A.8]). Thus, the descent spectral sequence is strongly convergent and the map $Z \rightarrow \operatorname*{holim}_\Delta
\operatorname*{holim}_i (\Gamma^\bullet_G (X_i)_{f,G})^G \cong (\operatorname*{holim}_i X_i)^{hG}$ is a weak equivalence ([@DH Corollary A.8]).
In this way, in [@thesis Chapter 10], the author showed that, given a finite spectrum $X$, the descent spectral sequence for $\pi_\ast((E_n \wedge X)^{hG})$ is strongly convergent and isomorphic to the $K(n)$-local $E_n$-Adams spectral sequence with abutment $\pi_\ast(E_n^{dhG} \wedge X)$, where $G$ is a closed subgroup of the extended Morava stabilizer group $G_n$ and $E_n^{dhG}$ is the spectrum constructed by Devinatz and Hopkins in [@DH] ($E_n^{dhG}$ is denoted by $E_n^{hG}$ in [@DH]).
[The $E_2$-term and the pro-discrete cochain complex]{}
In this section, we show that the $E_2$-term of (\[e2\]) can be built out of the same complex that computes continuous cochain cohomology. More precisely, given a continuous $G$-spectrum $\operatorname*{holim}_i X_i$, there exists a tower $\{X_i'\}$ of discrete $G$-spectra, such that $${\addtocounter{Lem}{1}\tag{\theLem}}\label{neato}
E_2^{s,t} \cong H^s[\pi_t(\mathcal{C}^\ast(G;\{X_i'\}))].$$
We find the expression on the right-hand side in (\[neato\]) interesting for the following reason. The homotopy fixed point spectrum is defined with respect to a continuous action of $G$ on the spectrum. Thus, homotopy fixed points take into account the topology of the spectrum. Similarly, since the $E_2$-term is built out of the pro-discrete cochain complex of spectra, the $E_2$-term is always taking into account the topology of the spectrum.
By [@GJ VI, Proposition 1.3], $\mathbf{tow}(\mathrm{Spt}_G)$ is a model category, where $\{f_i\}$ is a weak equivalence (cofibration) if and only if each $f_i$ is a weak equivalence (cofibration) in $\mathrm{Spt}_G$.
\[main2\] The $E_2$-term (\[e2\]) of descent spectral sequence (\[dss\]) has the form $${\addtocounter{Lem}{1}\tag{\theLem}}\label{nice}
E_2^{s,t} \cong \pi^s\pi_t(\lim_i (\Gamma^\bullet_G X_i')^G),$$ where $\{X_i\} \rightarrow
\{X_i'\}$ is a trivial cofibration with $\{X_i'\}$ fibrant, all in $\mathbf{tow}(\mathrm{Spt}_G)$.
Let $\{X_i'\}$ be as stated in the theorem. By [@GJ VI, Remark 1.5], each $X_i'$ is fibrant and each map $X_i' \rightarrow X_{i-1}'$ is a fibration, all in $\mathrm{Spt}_G$.
For any $k \geq 0$, we consider the expression $$\operatorname*{holim}_i ((\Gamma^\bullet_G X_i')^G)^k =
\operatorname*{holim}_i (\mathrm{Map}_c(G,\mathrm{Map}_c(G, \cdots
,\mathrm{Map}_c(G, X_i') \cdots )))^G,$$ where $\mathrm{Map}_c(G,-)$ appears $k+1$ times. By [@cts Section 3], the forgetful functor $U \: \mathrm{Spt}_G
\rightarrow \mathrm{Spt},$ $\mathrm{Map}_c(G,-) \: \mathrm{Spt} \rightarrow \mathrm{Spt}_G,$ where $\mathrm{Map}_c(G,X) = \Gamma_G(X)$, and the functor $(-)^G \: \mathrm{Spt}_G \rightarrow
\mathrm{Spt}$ all preserve fibrations. Thus, $\{X_i'\}$ is a tower of fibrations of fibrant spectra, all in $\mathrm{Spt}.$ This implies that $\{\mathrm{Map}_c(G,X_i')\}$ is a tower of fibrations of fibrant spectra, in $\mathrm{Spt}_G$, and hence, in $\mathrm{Spt}.$ By iteration, $$\{\mathrm{Map}_c(G,\mathrm{Map}_c(G, \cdots,
\mathrm{Map}_c(G, X_i') \cdots ))\}$$ is a tower of fibrations of fibrant spectra, in $\mathrm{Spt}_G$, so that $$\{(\mathrm{Map}_c(G,\mathrm{Map}_c(G, \cdots,
\mathrm{Map}_c(G, X_i') \cdots )))^G\}$$ is a tower of fibrations of fibrant spectra in $\mathrm{Spt}.$ Therefore, the canonical map $$\lim_i
((\Gamma^\bullet_G X_i')^G)^k \rightarrow \operatorname*{holim}_i
((\Gamma^\bullet_G X_i')^G)^k$$ is a weak equivalence.
Since $\{((\Gamma^\bullet_G X_i)^G)^k\}$ and $\{((\Gamma^\bullet_G (X_i)_{f,G})^G)^k\}$ are towers of fibrant spectra, there is a zigzag of weak equivalences $$\lim_i ((\Gamma^\bullet X_i')^G)^k \rightarrow \operatorname*{holim}_i
((\Gamma^\bullet X_i')^G)^k \leftarrow \operatorname*{holim}_i
((\Gamma^\bullet X_i)^G)^k \rightarrow \operatorname*{holim}_i
((\Gamma^\bullet (X_i)_{f,G})^G)^k,$$ where $\Gamma = \Gamma_G$. This zigzag of weak equivalences implies that $$\pi^s\pi_t(\lim_i
((\Gamma^\bullet_G X_i')^G)) \cong \pi^s\pi_t(\operatorname*{holim}_i
((\Gamma^\bullet_G (X_i)_{f,G})^G)).$$
\[main3\] Let $\{X_i'\}$ be as in Theorem \[main2\]. Then there is an isomorphism $$E_2^{s,t} \cong H^s[\pi_t(\mathcal{C}^\ast(G;\{X_i'\}))],$$ where $E_2^{s,t}$ is the $E_2$-term of (\[e2\]).
By Theorem \[main1\], $H^s[\mathcal{C}^\ast(G;\{\pi_t(X_i')\})]
\cong H^s_\mathrm{cts}(G;\lim_i \pi_t(X_i)).$
[The failure of other possible descriptions of the $E_2$-term]{} After studying the expression in (\[nice\]) further, one recalls that $\lim_i(-)^G$ is the functor used to define $H^s_\mathrm{cont}(G;-)$, and, if $M$ is any discrete $G$-module, then $$0 \rightarrow M \rightarrow \Gamma^\ast_G M$$ is a $(-)^G$-acyclic resolution of $M$, so that $H^s[(\Gamma^\ast_G M)^G] = H^s_c(G;M).$
Let $\{M_i\}$ be a tower of discrete $G$-modules. If $$\{0\} \rightarrow \{M_i\} \rightarrow \{\Gamma^\ast_G M_i\}$$ is a $\lim_i(-)^G$-acyclic resolution of $\{M_i\}$ in $\mathbf{tow}(\mathrm{DMod}(G))$, then $$H^s[(\lim_i(-)^G)(\{\Gamma^\ast_G M_i\})] = H^s_\mathrm{cont}(G;\{M_i\}).$$ This would imply that $E_2^{s,t} \cong
H^s[\pi_t(\lim_i (\Gamma^\ast_G X_i')^G)]$ is computed by taking the cohomology of the homotopy groups of a complex of spectra in the stable homotopy category, that, in the context of abelian groups, computes continuous cohomology. This would be an interesting presentation of the $E_2$-term.
However, it is not hard to show that $$\{0\} \rightarrow \{M_i\} \rightarrow \{\Gamma^\ast_G M_i\}$$ need not be a $\lim_i(-)^G$-acyclic resolution of $\{M_i\}$ in $\mathbf{tow}(\mathrm{DMod}(G))$, so that the above interpretation of the $E_2$-term does not work out. For example, by [@Jannsen (2.1)], there is a short exact sequence $$0 \rightarrow \operatorname*{lim^1}_i H^{s-1}_c(G;\Gamma_G M_i) \rightarrow
H^s_\mathrm{cont}(G;\{\Gamma_G M_i\})
\rightarrow \lim_i H^s_c(G;\Gamma_G M_i) \rightarrow 0,$$ for each $s \geq 0$, where $H^{-1}_c(G;-)=0$. Therefore, when $s \geq 1$, $H^s_c(G;\Gamma_G M_i)=0$, so that, for all $s
\geq 2$, $H^s_\mathrm{cont}(G;\{\Gamma_G M_i\}) = 0.$ But, the short exact sequence also implies that $$H^1_\mathrm{cont}(G;\{\Gamma_G M_i\})
\cong \operatorname*{lim^1}_i M_i,$$ which need not vanish. Thus, $\{\Gamma_G M_i\}$, the first object in the complex $\{\Gamma^\ast_G M_i \}$, need not be $\lim_i(-)^G$-acyclic in $\mathbf{tow}(\mathrm{DMod}(G))$.
Upon further consideration of the expression in (\[nice\]), one notices that, for any $k,l,m \geq 0$, $$((\lim_i(\Gamma^{m+1}_G X_i')^G)_k)_l =
\lim_i(\Gamma^{m+1}_G((X_i')_k)_l)^G \cong
\mathrm{Map}_c(G^m, \lim_i((X_i')_k)_l)$$ is an isomorphism of sets. If one could promote this isomorphism to $${\addtocounter{Lem}{1}\tag{\theLem}}\label{iso}
\lim_i(\Gamma^{m+1}_G X_i')^G \cong
\mathrm{Map}_c(G^m, \lim_i X_i'),$$ then one could use this to interpret the expression in (\[nice\]) as being the cohomology of homotopy groups applied to the complex of continuous cochains with target (“coefficients”) the continuous $G$-spectrum $\lim_i X_i'$.
But notice that, in this interpretation, the expression $\mathrm{Map}_c(G^m, \lim_i X_i')$ does not have the desired meaning. For isomorphism (\[iso\]) to hold, $\lim_i X_i'$ must be a spectrum whose simplicial sets have simplices with the pro-discrete topology. But, as a Bousfield-Friedlander spectrum, in the construction $\mathrm{Map}_c(G^m, \lim_i X_i')$, $\lim_i X_i'$ consists of simplicial sets whose simplices all have the discrete topology, by default. This conflict means that this interpretation also fails to work.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Neural network models recently proposed for question answering (QA) primarily focus on capturing the passage-question relation. However, they have minimal capability to link relevant facts distributed across multiple sentences which is crucial in achieving deeper understanding, such as performing multi-sentence reasoning, co-reference resolution, etc. They also do not explicitly focus on the question and answer type which often plays a critical role in QA. In this paper, we propose a novel end-to-end question-focused multi-factor attention network for answer extraction. Multi-factor attentive encoding using tensor-based transformation aggregates meaningful facts even when they are located in multiple sentences. To implicitly infer the answer type, we also propose a max-attentional question aggregation mechanism to encode a question vector based on the important words in a question. During prediction, we incorporate sequence-level encoding of the first wh-word and its immediately following word as an additional source of question type information. Our proposed model achieves significant improvements over the best prior state-of-the-art results on three large-scale challenging QA datasets, namely NewsQA, TriviaQA, and SearchQA.'
author:
- Souvik Kundu
- |
Hwee Tou Ng\
Department of Computer Science\
National University of Singapore\
{souvik, nght}@comp.nus.edu.sg\
bibliography:
- 'aaai2018.bib'
title: 'A Question-Focused Multi-Factor Attention Network for Question Answering'
---
Introduction {#sec:intro}
============
In machine comprehension-based (MC) question answering (QA), a machine is expected to provide an answer for a given question by understanding texts. In recent years, several MC datasets have been released. @MCTestdata ([-@MCTestdata]) released a multiple-choice question answering dataset. @HermannKGEKSB15 ([-@HermannKGEKSB15]) created a large cloze-style dataset using CNN and Daily Mail news articles. Several models [@HermannKGEKSB15; @chen2016thorough; @KadlecSBK16; @Kobayashi2016; @attn_over_attn; @GAR] based on neural attentional and pointer networks [@pointer_net] have been proposed since then. @RajpurkarZLL16 ([-@RajpurkarZLL16]) released the SQuAD dataset where the answers are free-form unlike in the previous MC datasets.
Most of the previously released datasets are closed-world, i.e., the questions and answers are formulated given the text passages. As such, the answer spans can often be extracted by simple word and context matching. @newsqa ([-@newsqa]) attempted to alleviate this issue by proposing the NewsQA dataset where the questions are formed only using the CNN article summaries without accessing the full text. As a result, a significant proportion of questions require reasoning beyond simple word matching. Two even more challenging open-world QA datasets, TriviaQA [@triviaqa] and SearchQA [@searchqa], have recently been released. TriviaQA consists of question-answer pairs authored by trivia enthusiasts and independently gathered evidence documents from Wikipedia as well as Bing Web search. In SearchQA, the question-answer pairs are crawled from the Jeopardy archive and are augmented with text snippets retrieved from Google search. Recently, many neural models have been proposed [@mpcm_squad; @memen; @allenai_squad; @smu_squad; @fastqa_squad; @salesforce_squad; @cmu_squad], which mostly focus on passage-question interaction to capture the context similarity for extracting a text span as the answer. However, most of the models do not focus on synthesizing evidence from multiple sentences and fail to perform well on challenging open-world QA tasks such as NewsQA and TriviaQA. Moreover, none of the models explicitly focus on question/answer type information for predicting the answer. In practice, fine-grained understanding of question/answer type plays an important role in QA.
In this work, we propose n end-to-end question-focused ulti-factor ttention etwork for ocument-based question nswering (AMANDA), which learns to aggregate evidence distributed across multiple sentences and identifies the important question words to help extract the answer. Intuitively, AMANDA extracts the answer not only by synthesizing relevant facts from the passage but also by implicitly determining the suitable answer type during prediction. The key contributions of this paper are:
- We propose a multi-factor attentive encoding approach based on tensor transformation to synthesize meaningful evidence across multiple sentences. It is particularly effective when answering a question requires deeper understanding such as multi-sentence reasoning, co-reference resolution, etc.
- To subsume fine-grained answer type information, we propose a max-attentional question aggregation mechanism which learns to identify the meaningful portions of a question. We also incorporate sequence-level representations of the first wh-word and its immediately following word in a question as an additional source of question type information.
Problem Definition {#sec:prob_def}
==================
Given a pair of passage and question, an MC system needs to extract a text span from the passage as the answer. We formulate the answer as two pointers in the passage, which represent the beginning and ending tokens of the answer. Let $\mathcal{P}$ be a passage with tokens $(\mathcal{P}_1, \mathcal{P}_2, \ldots, \mathcal{P}_T)$ and $\mathcal{Q}$ be a question with tokens $(\mathcal{Q}_1,\mathcal{Q}_2, \ldots, \mathcal{Q}_U)$, where $T$ and $U$ are the length of the passage and question respectively. To answer the question, a system needs to determine two pointers in the passage, $b$ and $e$, such that $1 \leq b \leq e \leq T$. The resulting answer tokens will be $(\mathcal{P}_b, \mathcal{P}_{b+1}, \ldots, \mathcal{P}_e)$.
Network Architecture {#sec:network}
====================
![Architecture of the proposed model. Hidden unit representations of Bi-LSTMs, **B** and **E**, are shown to illustrate the answer pointers. Blue and red arrows represent the start and end answer pointers respectively.[]{data-label="fig:squad_system_block"}](amandasystem.png){width="45.00000%"}
The architecture of the proposed question-focused multi-factor attention network[^1] is given in Figure \[fig:squad\_system\_block\].
Word-level Embedding
--------------------
Word-level embeddings are formed by two components: pre-trained word embedding vectors from GloVe [@pennington2014glove] and convolutional neural network-based (CNN) character embeddings [@cnn_char_emb_kim2014]. Character embeddings have proven to be very useful for out-of-vocabulary (OOV) words. We use a character-level CNN followed by max-pooling over an entire word to get the embedding vector for each word. Prior to that, a character-based lookup table is used to generate the embedding for every character and the lookup table weights are learned during training. We concatenate these two embedding vectors for every word to generate word-level embeddings.
Sequence-level Encoding
-----------------------
We apply sequence-level encoding to incorporate contextual information. Let $\mathbf{e}^p_t$ and $\mathbf{e}^q_t$ be the $t$th embedding vectors of the passage and the question respectively. The embedding vectors are fed to a bi-directional LSTM (BiLSTM) [@LSTM]. Considering that the outputs of the BiLSTMs are unfolded across time, we represent the outputs as $\mathbf{P} \in \mathbb{R} ^{T \times H}$ and $\mathbf{Q} \in \mathbb{R} ^{U \times H}$ for passage and question respectively. $H$ is the number of hidden units for the BiLSTMs. At every time step, the hidden unit representation of the BiLSTMs is obtained by concatenating the hidden unit representations of the corresponding forward and backward LSTMs. For the passage, at time step $t$, the forward and backward LSTM hidden unit representations can be written as: $$\begin{aligned}
\label{eq:for_back_lstm}
\overrightarrow{\mathbf{h}}^p_t &=& \overrightarrow{\textnormal{LSTM}}(\overrightarrow{\mathbf{h}}^p_{t-1} ,~ \mathbf{e}^p_t) \nonumber\\
\overleftarrow{\mathbf{h}}^p_t &=& \overleftarrow{\textnormal{LSTM}}(\overleftarrow{\mathbf{h}}^p_{t+1} ,~ \mathbf{e}^p_t) \end{aligned}$$ The $t$th row of $\mathbf{P}$ is represented as $\mathbf{p}_t = \overrightarrow{\mathbf{h}}^p_t ~||~ \overleftarrow{\mathbf{h}}^p_t$, where $||$ represents the concatenation of two vectors. Similarly, the sequence level encoding for a question is $\mathbf{q}_t = \overrightarrow{\mathbf{h}}^q_t ~||~ \overleftarrow{\mathbf{h}}^q_t$, where $\mathbf{q}_t$ is the $t$th row of $\mathbf{Q}$.
Cartesian Similarity-based Attention Layer
------------------------------------------
The attention matrix is calculated by taking dot products between all possible combinations of sequence-level encoding vectors for a passage and a question. Note that for calculating the attention matrix, we do not introduce any additional learnable parameters. The attention matrix $\mathbf{A} \in \mathbb{R} ^{T \times U}$ can be expressed as: $$\mathbf{A} = \mathbf{P} ~\mathbf{Q}^\top ~~$$ Intuitively, $A_{i,j}$ is a measure of the similarity between the sequence-level encoding vectors of the $i$th passage word and the $j$th question word.
Question-dependent Passage Encoding
-----------------------------------
In this step, we jointly encode the passage and question. We apply a row-wise softmax function on the attention matrix:
$$\mathbf{R} = \textnormal{row-wise softmax} (\mathbf{A})$$
If $\mathbf{r}_t \in \mathbb{R} ^{U}$ is the $t$th row of $\mathbf{R} \in \mathbb{R} ^{T \times U}$, then $\sum_{j=1}^{U} r_{t,j} = 1$. Each row of $\mathbf{R}$ measures how relevant every question word is with respect to a given passage word. Next, an aggregated question vector is computed corresponding to each sequence-level passage word encoding vector. The aggregated question vector $\mathbf{g}_t \in \mathbb{R} ^{H}$ corresponding to the $t$th passage word is computed as $\mathbf{g}_t = \mathbf{r}_t \mathbf{Q}$. The aggregated question vectors corresponding to all the passage words can be computed as $\mathbf{G} = \mathbf{R} ~ \mathbf{Q}$, where $\mathbf{g}_t$ is the $t$th row of $\mathbf{G} \in \mathbb{R} ^{T \times H}$.
The aggregated question vectors corresponding to the passage words are then concatenated with the sequence-level passage word encoding vectors. If the question-dependent passage encoding is denoted as $\mathbf{S} \in \mathbb{R} ^{T \times 2H}$ and $\mathbf{s}_t$ is the $t$th row of $\mathbf{S}$, then $\mathbf{s}_t = \mathbf{c}_t ~||~ \mathbf{g}_t$, where $\mathbf{c}_t$ is the sequence-level encoding vector of the $t$th passage word ($t$th row of $\mathbf{P}$). Then a BiLSTM is applied on $\mathbf{S}$ to obtain $\mathbf{V} \in \mathbb{R} ^{T \times H}$.
Multi-factor Attentive Encoding
-------------------------------
Tensor-based neural network approaches have been used in a variety of natural language processing tasks [@tensor_neural_net; @tensor_discourse]. We propose a multi-factor attentive encoding approach using tensor-based transformation. In practice, recurrent neural networks fail to remember information when the context is long. Our proposed multi-factor attentive encoding approach helps to aggregate meaningful information from a long context with fine-grained inference due to the use of multiple factors while calculating attention.
Let $\mathbf{v}_{i} \in \mathbb{R} ^{H}$ and $\mathbf{v}_{j} \in \mathbb{R} ^{H}$ represent the question-dependent passage vectors of the $i$th and $j$th word, i.e., the $i$th and $j$th row of $\mathbf{V}$. Tensor-based transformation for multi-factor attention is formulated as follows: $$\mathbf{f}^m_{i,j} = \mathbf{v}_i ~ \mathbf{W}^{[1:m]}_f ~ \mathbf{v}^\top_j ~~,$$ where $\mathbf{W}^{[1:m]}_f \in \mathbb{R} ^{H \times m \times H}$ is a 3-way tensor and $m$ is the number of factors. The output of the tensor product $\mathbf{f}^m_{i,j} \in \mathbb{R} ^{m}$ is a vector where each element $f^m_{i,j,k}$ is a result of the bilinear form defined by each tensor slice $\mathbf{W}^{[k]}_f \in \mathbb{R} ^{H \times H}$: $$f^m_{i,j,k} = \mathbf{v}_i ~ \mathbf{W}^{[k]}_f ~ \mathbf{v}^\top_j = \sum_{a,b} v_{i,a} W^{[k]}_{f_{a,b}} v_{j,b}$$ $\forall i,j \in [1,T]$, the multi-factor attention tensor can be given as $\mathbf{F}^{[1:m]} \in \mathbb{R} ^{m \times T \times T}$. For every vector $\mathbf{f}^m_{i,j}$ of $\mathbf{F}^{[1:m]}$, we perform a max pooling operation over all the elements to obtain the resulting attention value: $$F_{i,j} = \textnormal{max}(\mathbf{f}^m_{i,j}) ~~,$$ where $F_{i,j}$ represents the element in the $i$th row and $j$th column of $\mathbf{F} \in \mathbb{R} ^{T \times T}$. Each row of $\mathbf{F}$ measures how relevant every passage word is with respect to a given question-dependent passage encoding of a word. We apply a row-wise softmax function on $\mathbf{F}$ to normalize the attention weights, obtaining $\mathbf{\tilde{F}} \in \mathbb{R} ^{T \times T}$. Next, an aggregated multi-factor attentive encoding vector is computed corresponding to each question-dependent passage word encoding vector. The aggregated vectors corresponding to all the passage words, $\mathbf{M} \in \mathbb{R} ^{T \times H}$, can be given as $\mathbf{M} = \mathbf{\tilde{F}} ~ \mathbf{V}$. The aggregated multi-factor attentive encoding vectors are concatenated with the question-dependent passage word encoding vectors to obtain $\mathbf{\tilde{M}} \in \mathbb{R} ^{T \times 2H}$. To control the impact of $\mathbf{\tilde{M}}$, we apply a feed-forward neural network-based gating method to obtain $\mathbf{Y} \in \mathbb{R} ^{T \times 2H}$. If the $t$th row of $\mathbf{\tilde{M}}$ is $\mathbf{\tilde{m}}_t$, then the $t$th row of $\mathbf{Y}$ is: $$\mathbf{y}_t = \mathbf{\tilde{m}}_t \odot \textnormal{sigmoid}(\mathbf{\tilde{m}}_t \mathbf{W}^g + \mathbf{b}^g) ~~,$$ where $\odot$ represents element-wise multiplication. $\mathbf{W}^g \in \mathbb{R} ^{2H \times 2H}$ and $\mathbf{b}^g \in \mathbb{R} ^{2H}$ are the transformation matrix and bias vector respectively.
We use another pair of stacked BiLSTMs on top of $\mathbf{Y}$ to determine the beginning and ending pointers. Let the hidden unit representations of these two BiLSTMs be $\mathbf{B} \in \mathbb{R} ^{T \times H}$ and $\mathbf{E} \in \mathbb{R} ^{T \times H}$. To incorporate the dependency of the ending pointer on the beginning pointer, the hidden unit representation of $\mathbf{B}$ is used as input to $\mathbf{E}$.
Question-focused Attentional Pointing
-------------------------------------
Unlike previous approaches, our proposed model does not predict the answer pointers directly from contextual passage encoding or use another decoder for generating the pointers. We formulate a question representation based on two parts:
- max-attentional question aggregation ($\mathbf{q}_{ma}$)
- question type representation ($\mathbf{q}_{f}$)
$\mathbf{q}_{ma}$ is formulated by using the attention matrix $\mathbf{A}$ and the sequence-level question encoding $\mathbf{Q}$. We apply a [maxcol]{.nodecor} operation on $\mathbf{A}$ which forms a row vector whose elements are the maximum of the corresponding columns of $\mathbf{A}$. We define $\mathbf{k} \in \mathbb{R} ^{U}$ as the normalized max-attentional weights: $$\label{eq:maxcolA}
\mathbf{k} = \textnormal{softmax} (\textnormal{maxcol}(\mathbf{A}))$$ where [softmax]{.nodecor} is used for normalization. The max-attentional question representation $\mathbf{q}_{ma} \in \mathbb{R} ^{H}$ is: $$\label{eq:q_ma}
\mathbf{q}_{ma} = \mathbf{k} ~ \mathbf{Q}$$ Intuitively, $\mathbf{q}_{ma}$ aggregates the most relevant parts of the question with respect to all the words in the passage.
$\mathbf{q}_{f}$ is the vector concatenation of the representations of the first wh-word and its following word from the sequence-level question encoding $\mathbf{Q}$. The set of wh-words we used is {[*what, who, how, when, which, where, why*]{}}. If $\mathbf{q}_{t_{wh}}$ and $\mathbf{q}_{t_{wh}+1}$ represent the first wh-word and its following word (i.e., the $t_{wh}$th and $(t_{wh}+1)$th rows of $\mathbf{Q}$), $\mathbf{q}_{f} \in \mathbb{R} ^{2H}$ is expressed as: $$\centering
\label{eq:q_type}
\mathbf{q}_{f} ~=~ \mathbf{q}_{t_{wh}} ~||~ \mathbf{q}_{t_{wh}+1}$$
The final question representation $\mathbf{\tilde{q}} \in \mathbb{R} ^{H}$ is expressed as: $$\label{eq:q_tilde}
\mathbf{\tilde{q}} = \textnormal{tanh} ((\mathbf{q}_{ma} ~||~ \mathbf{q}_{f}) \mathbf{W}_q ~+~ \mathbf{b}_q )$$ where $\mathbf{W}_q \in \mathbb{R} ^{3H \times H}$ and $\mathbf{b}_q \in \mathbb{R} ^{H}$ are the weight matrix and bias vector respectively. If no wh-word is present in a question, we use the first two sequence-level question word representations for calculating $\mathbf{\tilde{q}}$.
We measure the similarity between $\mathbf{\tilde{q}}$ and the contextual encoding vectors in $\mathbf{B}$ and $\mathbf{E}$ to determine the beginning and ending answer pointers. Corresponding similarity vectors $\mathbf{s}_b \in \mathbb{R} ^{T}$ and $\mathbf{s}_e \in \mathbb{R} ^{T}$ are computed as: $$\mathbf{s}_b = \mathbf{\tilde{q}} ~ \mathbf{B}^\top ~~,~~ \mathbf{s}_e = \mathbf{\tilde{q}} ~ \mathbf{E}^\top$$ The probability distributions for the beginning pointer $b$ and the ending pointer $e$ for a given passage $\mathcal{P}$ and a question $\mathcal{Q}$ can be given as: $$\begin{aligned}
\textnormal{Pr}(b ~|~ \mathcal{P},\mathcal{Q}) &=& \textnormal{softmax}(\mathbf{s}_b) \nonumber \\
\textnormal{Pr}(e ~|~ \mathcal{P},\mathcal{Q}, b) &=& \textnormal{softmax}(\mathbf{s}_e)\end{aligned}$$ The joint probability distribution for obtaining the answer $a$ is given as: $$\label{eq:joint_prob}
\textnormal{Pr}(a ~|~ \mathcal{P},\mathcal{Q}) = \textnormal{Pr}(b ~|~ \mathcal{P},\mathcal{Q}) ~ \textnormal{Pr}(e ~|~ \mathcal{P},\mathcal{Q}, b)$$ To train our model, we minimize the cross entropy loss: $$\textnormal{loss} = - \sum \textnormal{log} ~~ \textnormal{Pr}(a ~|~ \mathcal{P},\mathcal{Q})$$ summing over all training instances. During prediction, we select the locations in the passage for which the product of $\textnormal{Pr}(b)$ and $\textnormal{Pr}(e)$ is maximum keeping the constraint $1 \leq b \leq e \leq T$.
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
**Passage**: ... The family of a Korean-American missionary believed held in North Korea said Tuesday they are working with U.S. officials to get him returned home. Robert Park told relatives before Christmas that he was trying to sneak into the isolated communist state to bring a message of “Christ’s love and forgiveness” to North Korean leader Kim ...
**Question**: What is the name of the Korean-American missionary?
**Reference Answer**: Robert Park
---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
: Example of a (passage, question, answer)[]{data-label="tab:ex2"}
Visualization {#sec:visualization}
=============
![Multi-factor attention weights (darker regions signify higher weights).[]{data-label="fig:ex_mfa"}](short_mfa.png){width="30.00000%"}
![Max-attentional weights for question (the origin is set to $-0.1$ for clarity).[]{data-label="fig:ex_qma"}](qma.png){width="5cm" height="3cm"}
To understand how the proposed model works, for the example given in Table \[tab:ex2\], we visualize the normalized multi-factor attention weights $\mathbf{\tilde{F}}$ and the attention weights $\mathbf{k}$ which are used for max-attentional question aggregation.
In Figure \[fig:ex\_mfa\], a small portion of $\mathbf{\tilde{F}}$ has been shown, in which the answer words [*Robert*]{} and [*Park*]{} are both assigned higher weights when paired with the context word [*Korean-American*]{}. Due to the use of multi-factor attention, the answer segment pays more attention to the important keyword although it is quite far in the context passage and thus effectively infers the correct answer by deeper understanding. In Figure \[fig:ex\_qma\], it is clear that the important question word [*name*]{} is getting a higher weight than the other question words. This helps to infer the fine-grained answer type during prediction, i.e., a person’s name in this example.
Experiments {#sec:experiments}
===========
[l c c c c]{} & &\
& **EM** & **F1** & **EM** & **F1**\
\
Match-LSTM & 34.4 & 49.6 & 34.9 & 50.0\
BARB & 36.1 & 49.6 & 34.1 & 48.2\
---------------------
[@two_stage_synnet]
BiDAF on NewsQA
---------------------
: Results on the NewsQA dataset. $^\dagger$ denotes the models of [@fastqa_squad].[]{data-label="tab:newsqa"}
& - & - & 37.1 & 52.3\
$^\dagger$Neural BoW Baseline & 25.8 & 37.6 & 24.1 & 36.6\
$^\dagger$FastQA & 43.7 & 56.4 & 41.9 & 55.7\
$^\dagger$FastQAExt & 43.7 & 56.1 & 42.8 & 56.1\
------------------
[@reading_twice]
R2-BiLSTM
------------------
: Results on the NewsQA dataset. $^\dagger$ denotes the models of [@fastqa_squad].[]{data-label="tab:newsqa"}
& - & - & 43.7 & 56.7\
AMANDA & **48.4** & **63.3** & **48.4** & **63.7**\
----------------------- -- ----------- ----------- ----------- ----------- ----------- ----------- ----------- -----------
**EM** **F1** **EM** **F1** **EM** **F1** **EM** **F1**
$^\ddagger$Random 12.72 22.91 12.74 22.35 14.81 23.31 15.41 25.44
$^\ddagger$Classifier 23.42 27.68 22.45 26.52 24.91 29.43 27.23 31.37
$^\ddagger$BiDAF 40.26 45.74 40.32 45.91 47.47 53.70 44.86 50.71
$^\star$MEMEN 43.16 46.90 - - 49.28 55.83 - -
**46.95** **52.51** **46.67** **52.22** **52.86** **58.74** **50.51** **55.93**
$^\ddagger$Classifier 24.64 29.08 24.00 28.38 27.38 31.91 30.17 34.67
$^\ddagger$BiDAF 41.08 47.40 40.74 47.05 51.38 55.47 49.54 55.80
$^\star$MEMEN 44.25 48.34 - - 53.27 57.64 - -
**46.68** **53.27** **46.58** **53.13** **60.31** **64.90** **55.14** **62.88**
----------------------- -- ----------- ----------- ----------- ----------- ----------- ----------- ----------- -----------
------------- ------ -------------- ------------
**Unigram** **N-gram**
**Accuracy** **F1**
[@searchqa]
Dev 13.0 -
Test 12.7 -
Dev 43.9 24.2
Test 41.3 22.8
Dev **48.6** **57.7**
Test **46.8** **56.6**
------------- ------ -------------- ------------
: Results on the SearchQA dataset.[]{data-label="tab:results_searchqa"}
We evaluated AMANDA on three challenging QA datasets: NewsQA, TriviaQA, and SearchQA. Using the NewsQA development set as a benchmark, we perform rigorous analysis for better understanding of how our proposed model works.
Datasets {#sec:dataset}
--------
The NewsQA dataset [@newsqa] consists of around 100K answerable questions in total. Similar to [@newsqa; @fastqa_squad], we do not consider the unanswerable questions in our experiments. NewsQA is more challenging compared to the previously released datasets as a significant proportion of questions requires reasoning beyond simple word- and context-matching. This is due to the fact that the questions in NewsQA were formulated only based on summaries without accessing the main text of the articles. Moreover, NewsQA passages are significantly longer (average length of 616 words) and cover a wider range of topics. TriviaQA [@triviaqa] consists of question-answer pairs authored by trivia enthusiasts and independently gathered evidence documents from Wikipedia and Bing Web search. This makes the task more similar to real-life IR-style QA. In total, the dataset consists of over 650K question-answer-evidence triples. Due to the high redundancy in Web search results (around 6 documents per question), each question-answer-evidence triple is treated as a separate sample and evaluation is performed at document level. However, in Wikipedia, questions are not repeated (each question has 1.8 evidence documents) and evaluation is performed over questions. In addition to distant supervision, TriviaQA also has a verified human-annotated question-evidence collection. Compared to previous datasets, TriviaQA has more complex compositional questions which require greater multi-sentence reasoning.
SearchQA [@searchqa] is also constructed to more closely reflect IR-style QA. They first collected existing question-answer pairs from a Jeopardy archive and augmented them with text snippets retrieved by Google. One difference with TriviaQA is that the evidence passages in SearchQA are Google snippets instead of Wikipedia or Web search documents. This makes reasoning more challenging as the snippets are often very noisy. SearchQA consists of 140,461 question-answer pairs, where each pair has 49.6 snippets on average and each snippet has 37.3 tokens on average.
Experimental Settings {#sec:exp_set}
---------------------
We tokenize the corpora with NLTK[^2]. We use the 300-dimension pre-trained word vectors from GloVe [@pennington2014glove] and we do not update them during training. The out-of-vocabulary words are initialized with zero vectors. We use 50-dimension character-level embedding vectors. The number of hidden units in all the LSTMs is 150. We use dropout [@dropout] with probability 0.3 for every learnable layer. For multi-factor attentive encoding, we choose 4 factors ($m$) based on our experimental findings (refer to Table \[tab:variation\_num\_factors\]). During training, the minibatch size is fixed at 60. We use the Adam optimizer [@adam] with learning rate of 0.001 and clipnorm of 5. During testing, we enforce the constraint that the ending pointer will always be equal to or greater than the beginning pointer. We use exact match (EM) and F1 scores as the evaluation metrics.
Results {#sec:results}
-------
Table \[tab:newsqa\] shows that AMANDA outperforms all the state-of-the-art models by a significant margin on the NewsQA dataset. Table \[tab:triviaqa\] shows the results on the TriviaQA dataset. In Table \[tab:triviaqa\], the model named [Classifier]{} based on feature engineering was proposed by @triviaqa ([-@triviaqa]). They also reported the performance of BiDAF [@allenai_squad]. A memory network-based approach, [MEMEN]{}, was recently proposed by [@memen]. Note that in the Wikipedia domain, we choose the answer which provides the highest maximum joint probability (according to Eq. (\[eq:joint\_prob\])) for any document. Table \[tab:triviaqa\] shows that AMANDA achieves state-of-the-art results in both Wikipedia and Web domain on distantly supervised and verified data.
Results on the SearchQA dataset are shown in Table \[tab:results\_searchqa\]. In addition to a TF-IDF approach, @searchqa ([-@searchqa]) modified and reported the performance of attention sum reader (ASR) which was originally proposed by @KadlecSBK16 ([-@KadlecSBK16]). We consider a maximum of 150 words surrounding the answer from the concatenated ranked list of snippets as a passage to more quickly train the model and to reduce the amount of noisy information. During prediction, we choose the first 200 words (about 5 snippets) from the concatenated ranked list of snippets as an evidence passage. These are chosen based on performance on the development set. Based on question patterns, question types are always represented by the first two sequence-level representations of question words. To make the results comparable, we also report accuracy for single-word-answer (unigram) questions and F1 score for multi-word-answer (n-gram) questions. AMANDA outperforms both systems, especially for multi-word-answer questions by a huge margin. This indicates that AMANDA can learn to make inference reasonably well even if the evidence passages are noisy.
Effectiveness of the Model Components {#sec:ablation}
-------------------------------------
**Model** **EM** **F1**
---------------------------------------------- ---------- ----------
minus multi factor attn. 46.4 61.2
minus $\mathbf{q}_{ma}$ and $\mathbf{q}_{f}$ 46.2 60.5
minus $\mathbf{q}_{ma}$ 46.6 61.3
minus $\mathbf{q}_{f}$ 46.8 61.8
[AMANDA]{} **48.4** **63.3**
: Ablation of proposed components on the NewsQA development set.[]{data-label="tab:ablations_comps"}
**Model** **EM** **F1**
--------------------------------------- ---------- ----------
minus char embedding 47.5 61.4
minus question-dependent passage enc. 32.1 45.0
minus 2nd LSTM during prediction 46.5 61.6
[AMANDA]{} **48.4** **63.3**
: Ablation of other components on the NewsQA development set[]{data-label="tab:other_ablations"}
Table \[tab:ablations\_comps\] shows that AMANDA performs better than any of the ablated models which include the ablation of multi-factor attentive encoding, max-attentional question aggregation ($\mathbf{q}_{ma}$), and question type representation ($\mathbf{q}_{f}$). We also perform statistical significance test using paired t-test and bootstrap resampling. Performance of AMANDA (both in terms of EM and F1) is significantly better $(p < 0.01)$ than the ablated models.
One of the key contributions of this paper is multi-factor attentive encoding which aggregates information from the relevant passage words by using a tensor-based attention mechanism. The use of multiple factors helps to fine-tune answer inference by synthesizing information distributed across multiple sentences. The number of factors is the granularity to which the model is allowed to refine the evidence. The effect of multi-factor attentive encoding is illustrated by the following example taken from the NewsQA development set:
*What will allow storage on remote servers?*
*...The **iCloud service** will now be integrated into the iOS 5 operating system. It will work with and allow content to be stored on remote servers instead of the users’ iPod, iPhone or other device...*
When multi-factor attentive encoding is ablated, the model could not figure out the cross-sentence co-reference and wrongly predicted the answer as [*apps*]{}. On the contrary, with multi-factor attentive encoding, AMANDA could correctly infer the answer as [*iCloud service*]{}.
Another contribution of this work is to include the question focus during prediction. It is performed by adding two components: $\mathbf{q}_{ma}$ (max-attentional question aggregation) and $\mathbf{q}_f$ (question type representation). $\mathbf{q}_{ma}$ and $\mathbf{q}_f$ implicitly infer the answer type during prediction by focusing on the important question words. Impact of the question focus components is illustrated by the following example taken from the NewsQA development set:
*who speaks on Holocaust remembrance day?* *... Israel’s vice prime minister **Silvan Shalom** said Tuesday “Israel can never ... people just 65 years ago” ... He was speaking as observes its Holocaust memorial day, remembering the roughly...* Without the $\mathbf{q}_{ma}$ and $\mathbf{q}_f$ components, the answer was wrongly predicted as [*Israel*]{}, whereas with $\mathbf{q}_{ma}$ and $\mathbf{q}_{f}$, AMANDA could correctly infer the answer type (i.e., a person’s name) and predict [*Silvan Shalom*]{} as the answer.
Ablation studies of other components such as character embedding, question-dependent passage encoding, and the second LSTM during prediction are given in Table \[tab:other\_ablations\]. When the second LSTM ($\mathbf{E}$) is ablated, a feed-forward layer is used instead. Table \[tab:other\_ablations\] shows that question-dependent passage encoding has the highest impact on performance.
Variation on the number of factors ($m$) and $\mathbf{q}_{ma}$ {#sec:variation}
--------------------------------------------------------------
**Value of $m$** **1** **2** **3** **4** **5**
------------------ ------- ------- ---------- ---------- -------
**EM** 45.8 47.4 **48.7** 48.4 48.0
**F1** 61.2 61.9 62.9 **63.3** 62.5
: Variation of $m$ on the NewsQA development set.[]{data-label="tab:variation_num_factors"}
Table \[tab:variation\_num\_factors\] shows the performance of AMANDA for different values of $m$. We use 4 factors for all the experiments as it gives the highest F1 score. Note that $m = 1$ is equivalent to standard bilinear attention.
**Aggregation** **EM** **F1**
----------------- ---------- ----------
Mean 46.6 61.3
Sum 47.9 62.2
Max (AMANDA) **48.4** **63.3**
: Variation of question aggregation formulation on the NewsQA development set.[]{data-label="tab:var_q"}
Table \[tab:var\_q\] shows the variation of question aggregation formulation. For mean aggregation, the attentional weight vector $\mathbf{k}$ is formulated by applying column-wise averaging on the attention matrix $\mathbf{A}$. Intuitively, it is giving equal priority to all the passage words to determine a particular question word attention. Similarly, in the case of sum aggregation, we apply a column-wise sum operation. Table \[tab:var\_q\] shows that the best performance is obtained when $\mathbf{q}_{ma}$ is obtained with a column-wise maximum operation on $\mathbf{A}$. Effectively, it is helping to give higher weights to the more important question words based on the most relevant passage words.
Quantitative Error Analysis {#sec:quant_error_analysis}
---------------------------
We analyzed the performance of AMANDA across different question types and different predicted answer lengths. Figure \[fig:error\_analysis\](a) shows that it performs poorly on [*why*]{} and [*other*]{} questions whose answers are usually longer. Figure \[fig:error\_analysis\](b) supports this fact as well. When the predicted answer length increases, both F1 and EM start to degrade. The gap between F1 and EM also increases for longer answers. This is because for longer answers, the model is not able to decide the exact boundaries (low EM score) but manages to predict some correct words which partially overlap with the reference answer (relatively higher F1 score).
Qualitative Error Analysis {#sec:qual_error_analysis}
--------------------------
----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
[Answer ambiguity (42%)]{}
Q: What happens to the power supply?
... customers possible.” The outages were due mostly to power **lines downed** by Saturday’s hurricane-force winds, which . At ...
[Context mismatch (22%)]{}
Q: Who was Kandi Burrus’s fiance?
Kandi Burruss, the ” ... fiance, who died ... fiance, 34-year-old **Ashley “A.J.” Jewell**, also...
[Complex inference (10%)]{}
Q: When did the Delta Queen first serve?
... the Delta Queen steamboat, a floating National ... scheduled voyage ... The Delta Queen will go ... Supporters of the boat, which has roamed the nation’s waterways since **1927** and helped the Navy ...
[Paraphrasing issues (6%)]{}
Q: What was Ralph Lauren’s first job?
Ralph Lauren has ... Japan. For four ... than the former **tie salesman** from the Bronx. “Those ties ... Lauren originally named his because ...\
----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
: Examples of different error types and their percentages. Ground truth answers are bold-faced and predicted answers are underlined.[]{data-label="tab:qual_analysis"}
On the NewsQA development set, AMANDA predicted completely wrong answers on 25.1% of the questions. We randomly picked 50 such questions for analysis. The observed types of errors are given in Table \[tab:qual\_analysis\] with examples. 42% of the errors are due to answer ambiguities, i.e., no unique answer is present. 22% of the errors are due to mismatch between question and context words. 10% of the errors are due to the need for highly complex inference. 6% of the errors occur due to paraphrasing, i.e., the question is posed with different words which do not appear in the passage context. The remaining 20% of the errors are due to insufficient evidence, incorrect tokenization, wrong co-reference resolution, etc.
Related Work {#sec:related_works}
============
Recently, several neural network-based models have been proposed for QA. Models based on the idea of chunking and ranking include @ibm_squad ([-@ibm_squad]) and @google_squad ([-@google_squad]). @cmu_squad ([-@cmu_squad]) used a fine-grained gating mechanism to capture the correlation between a passage and a question. @smu_squad ([-@smu_squad]) used a Match-LSTM to encode the question and passage together and a boundary model determined the beginning and ending boundary of an answer. @newsqa ([-@newsqa]) reimplemented Match-LSTM for the NewsQA dataset and proposed a faster version of it. @salesforce_squad ([-@salesforce_squad]) used a co-attentive encoder followed by a dynamic decoder for iteratively estimating the boundary pointers. @allenai_squad ([-@allenai_squad]) proposed a bi-directional attention flow approach to capture the interactions between passages and questions. @fastqa_squad ([-@fastqa_squad]) proposed a simple context matching-based neural encoder and incorporated word overlap and term frequency features to estimate the start and end pointers. @rnet_squad ([-@rnet_squad]) proposed a gated self-matching approach which encodes the passage and question together using a self-matching attention mechanism. @memen ([-@memen]) proposed a memory network-based multi-layer embedding model and reported results on the TriviaQA dataset.
Different from all prior approaches, our proposed multi-factor attentive encoding helps to aggregate relevant evidence by using a tensor-based multi-factor attention mechanism. This in turn helps to infer the answer by synthesizing information from multiple sentences. AMANDA also learns to focus on the important question words to encode the aggregated question vector for predicting the answer with suitable answer type.
Conclusion
==========
In this paper, we have proposed a question-focused multi-factor attention network (AMANDA), which learns to aggregate meaningful evidence from multiple sentences with deeper understanding and to focus on the important words in a question for extracting an answer span from the passage with suitable answer type. AMANDA achieves state-of-the-art performance on NewsQA, TriviaQA, and SearchQA datasets, outperforming all prior published models by significant margins. Ablation results show the importance of the proposed components.
Acknowledgement
===============
This research was supported by research grant R-252-000-634-592.
[^1]: Our code is available at <https://github.com/nusnlp/amanda>
[^2]: http://www.nltk.org/
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We study the ratio of the energy and particle currents ($j^E/j^N$) in an integrable one dimensional system of interacting fermions. Both currents are driven by a finite (nonzero) dc electric field. In doped insulators, where the local conserved quantities saturate the so called Mazur bound on the charge stiffness, $j^E/j^N$ agrees with the linear–response theory, even though such agreement may be violated for each current alone. However, in the metallic regime with a non-saturated Mazur bound, the ratio $j^E/j^N$ in a driven system is shown to be much larger than predicted by the linear–response theory.'
author:
- 'D. Crivelli'
- 'P. Prelovšek'
- 'M. Mierzejewski'
bibliography:
- 'bibliography.bib'
title: Energy and particle currents in a driven integrable system
---
Introduction and Motivation.
============================
The physics beyond the linear response (LR) regime is interesting for basic research and potentially important for the future applications. The underlying phenomena have recently become accessible to novel experimental techniques like ultrafast pump–probe spectroscopy of solid state systems or measurements of the relaxation processes in ultracold atoms driven far from equilibrium. Significant progress has also been achieved in the theoretical description of solids driven by a finite (nonzero) electric field [@Amaricci2012; @Arrachea2002; @Aron2012; @einhellinger2012; @Joura2008; @meisner; @lev2011; @lev2011_1; @janez1; @Boulat2008]. Recently developed numerical approaches allow to study response to the electric field of (almost) arbitrary strength. In particular, applicability of the LR theory has been tested for a weak–to–moderate driving [@Russomanno2013; @Mierzejewski2011; @Steinigeweg2012; @Karrasch2013], while for extremely strong fields one has studied the Bloch oscillations in systems of strongly interacting carriers.[@Buchleitner2003; @Eckstein2011; @Freericks2008; @Mierzejewski2010; @covaci] At the same time, the combined transport of energy and charge, which determines the thermoelectric properties, has been studied mostly in the LR regime with only a few attempts to the nonequilibrium regime [@leijnse; @sanchez2013; @Ajisaka2012; @ourcouple]. Promising results concerning enhanced the thermoelectric performance have been reported for low–dimensional systems [@Benenti2013; @Dresselhaus2007; @Dubi2011; @Kim2009] for systems with ballistic (coherent) charge carriers [@Hlubek2010; @Benenti2013a; @Karlstroem2011] as well as for systems with strongly interacting electrons [@Arsenault2013; @Zlatic2012; @Zlatic2014; @Peterson2007; @Shastry2009; @Zemljic2005; @Paul2003; @Kargarian2013].
We first note that not all currents which are well established in the LR theory remain uniquely defined also in a generic nonequilibrium situation. Related to conservation laws and continuity equations, the energy and particle currents are well defined also beyond LR, while e.g. the heat current is not. The main objective of our research is to establish the ratio of the energy current $ j^E $ and the particle current $ j^N $ $$\begin{aligned}
R(t) = \frac{j^E(t)}{j^N(t)},
\label{eq:time-ratio}\end{aligned}$$ in a homogeneous [*integrable*]{} system which at time $t=0$ is in equilibrium while for $t>0$ is driven by a finite electric field $F$. Generic (nonintegrable) systems show a dissipative transport, hence a steady driving induces steady currents $j^{N(E)}(t \rightarrow \infty)=\mathrm{const}$ and the dc ratio $R(t \rightarrow \infty)$ is well defined. One would wish to discuss directly the heat current usually expressed as $j^Q=j^E-\mu j^N$, but the chemical potential $\mu$ is essentially an equilibrium concept. While $j^E$ is still not the heat current, at least under close–to–equilibrium conditions $R(t \rightarrow \infty)$ can be related to various thermoelectric properties,[@Goupil2011] e.g. the Peltier coefficient $\Pi=R(t\rightarrow \infty)-\mu$.
Integrable systems display unusual relaxation [@gge; @Eckstein2012; @Cassidy2011; @Steinigeweg2012] and transport properties [@ZotosIdeal; @Mierzejewski2010; @Mierzejewski2011; @Tomaz2011; @Marko2011; @Sirker2009; @Steinigeweg2011]. In particular they show a [*ballistic*]{} transport quantified by a nonzero charge stiffness leading to singular response functions.[@ZotosIdeal; @Naef97; @Heidrich-Meisner2003; @Mukerjee2008; @Sirker2009; @Orignac2003; @Steinigeweg2014] On the one hand, the basic understanding of the ballistic transport is that a steady driving induces a steadily growing currents $j^{N(E)}(t ) \propto t$. On the other hand, in the tight–binding models the expectation values of currents cannot become arbitrarily large. This poses limits on the time–window in which currents may indeed vary linearly in time. It has recently been shown for driven integrable systems that the particle current undergoes the Bloch oscillations [@Mierzejewski2010] and it is straightforward to expect the same also for $j^E$. Note also that finite electric field acts as a source of the currents, but doubles as an integrability breaking mechanism. Despite singular response functions, the dc ratio of energy and particle currents has been expected to remain well defined and finite [@Zemljic2005] at least in the LR regime ($F\rightarrow 0$). However, the above arguments indicate that it is by far not obvious whether/when/why it may actually take place under a finite driving. It is the main problem which we address in this paper.
The manuscript is organized as follows. In the subsequent section we introduce a model and the specify the details of driving. Then, as a test of our approach we study a generic system and show how the LR results for $R(t)$ can be extracted from time evolved observables. Next we turn our approach on integrable systems, where strictly equilibrium predictions for the ratio of currents are ambiguous due to singularities of the response functions. First, we investigate a doped insulator, for which the ballistic transport can be explained as originating from local conserved quantities.[@Herbrych2011] Finally, we present conjectural results for a metallic system in which a relation between charge stiffness and local conservation laws has not been established.
Setup and Methods
=================
The system under study is a closed, homogeneous one–dimensional ring of charged, spinless, but interacting fermions. The Hamiltonian is that of the *t-V-W* model, arranged on a periodic ring of $L$ sites: $$\begin{aligned}
H(t)=&-t_0 \sum_i \left[ {\mathrm e}^{i \phi(t)}\; c^{\dagger}_{i+1} \, c_i +{\mathrm H.c.} \right] \nonumber \\
& + V \sum_i \tilde{n}_i \, \tilde{n}_{i+1} + W \sum_i \tilde{n}_i \, \tilde n_{i+2},
\label{hamiltonian}\end{aligned}$$ where $n_i = c^{\dagger}_{i} \, c_i$, $\tilde n_i=n_i -1/2$ and $t_0$ is the hopping integral. $V$ and $W$ are repulsive interactions between first and second nearest neighbors, respectively. The latter interaction is introduced to break integrability in a controlled manner and to allow for the normal diffusion (at least at weak driving). Below we use units in which $\hbar=k_B=t_0=1$.
The dynamics is studied by explicitly solving the time-dependent Schrödinger equation for a pure quantum state. The initial equilibrium state $|\Psi(t=0) \rangle$ is determined from the Microcanonical Lanczos Method (MCLM) [@Long2003] for the energy $E_0= \langle \Psi(0)| H(0) |\Psi(0) \rangle$ corresponding to a target inverse temperature $\beta$. If not specified otherwise we take $\beta \simeq 0.4$, while a typical energy uncertainty is $\delta E_0=\langle \Psi(0)| [H(0)-E_0]^2|\Psi(0) \rangle^{1/2} \simeq 0.003$. For time $t>0$ the system is driven by a constant electrical field $F$ [@ring1; @ring2; @ring3; @ring4; @Arrachea2002], induced by linearly varying magnetic flux $\phi(t) = - F t$. The evolution under driving $|\Psi(0)\rangle \to |\Psi(t)\rangle$ is obtained by means of a fourth order expansion [@Alvermann2012] of the time ordered exponential, with Chebyshev approximation of the unitary propagators [@chebytime] on successive small time intervals. The evolution is thus unitary and numerically accurate, allowing long timescales up to $t \alt 1 / \delta E_0$.
![(Color online) Results for a generic system with $V=3, W=1, L=24, N=10$. (a) Time dependence of energy and particle currents for $F=0.2$. (b) Ratio $j^E/F$ for different fields $F$ as function of the instantaneous energy $E$ relative to the energy at infinite temperature $E_{\infty}$. Point marks the initial state and arrow shows the direction of time–evolution. []{data-label="fig1"}](fig1){width="45.00000%"}
We study the particle (charge) current $j^N = \langle J^N \rangle$ and the energy current $j^E = \langle J^E \rangle$, both induced by the same field $F$. The currents follow uniquely from the continuity relations for the local charge and energy densities [@Mierzejewski2013; @Naef97; @Mukerjee2008] and have the form: $$\begin{aligned}
J^N(t)=& \frac{1}{L} \sum_i J^N_i = \frac{1}{L} \sum_i \left[i \mathrm{e}^{i\phi} c^{\dagger}_{i+1}\, c_{i}+ \mathrm{H.c.}\right],
\label{jn} \\
J^E(t) =& \frac{1}{L} \sum_i J^E_i = \,-\frac{1}{L}\sum_i \Big\lbrace[i \mathrm{e}^{2i\phi} c^{\dagger}_{i+1}\, c_{i-1}+{\mathrm H.c.}] \nonumber \\
+ \frac{J^N_i}{2} \big[ 3W & (\tilde{n}_{i+3}\!+\!\tilde{n}_{i-2})+ (2V\!-\!W)(\tilde{n}_{i+2} \!
+ \! \tilde{n}_{i-1}) \big] \Big\rbrace.
\label{je}\end{aligned}$$ The equilibrium continuity equation for charge holds true also in driven systems, because driving does not influence the conservation of particles. However, the energy of a driven system is not conserved. Therefore, the relevant continuity equation contains also the source terms which for systems driven by electric field represent the effects of the Joule heating: $$\frac{\mathrm d}{{\mathrm d} t}
\langle H_i \rangle +
\nabla \langle J^E_i \rangle = F \langle J^N_i \rangle.$$ Here, $H_i$ is the energy density operator, $H=\sum_i H_i$. In the LR regime the currents can be equivalently derived from the polarization operators [@Shastry2009; @Louis2003; @Paul2003].
Departure from half–filling ($N \ne L/2$) is necessary to obtain nonzero $j^E$ since at half filling the Hamiltonian is invariant under the particle–hole transformation $c_i \rightarrow (-1)^i c^{\dagger}_i$ while $J^E \rightarrow -J^E$ under this transformation. Furtheron, the number of charged fermions is taken to be $N=10$ for the $L=24$ site ring or $N=9$ for $L=26$ sites both slightly below half–filling. We investigate systems with $V=1.5$ and $V=3$ which for $W=0$ correspond, respectively, to doped metals and insulators [@Rigol-tVW]. The latter insulating phase is induced by a short range fermion–fermion interaction ($V$) and is charge ordered. Hence, it shares common properties with Mott insulators as well as with charge density wave insulators.
Generic response of nonintegrable systems {#Sec:Nonintegrable}
=========================================
Figure \[fig1\](a) shows the time–dependence of both currents in a driven generic system. Shortly after turning on the electric field, $j^{E(N)}$ can be easily determined from the equations of motion [@Mierzejewski2010] $$\frac{d}{dt} j^{N(E)} = -\tau^{N(E)} \dot{\phi} + i \langle [H,J^{N(E)}] \rangle,
\label{eq_motion}$$ where $ \tau^{N} = -\langle \partial_{\phi} J^{N} \rangle $ and $\tau^{E} = -\langle \partial_{\phi} J^{E} \rangle $ are stress coefficients (tensors in general) determining the short–time LR to the flux change. The last term in Eq. (\[eq\_motion\]) vanishes for the initial equilibrium state, hence the short–time ratio of the energy and particle currents $$\begin{aligned}
R(t\rightarrow 0^{+}) = \frac{\tau^{E}}{\tau^{N}},
\label{eq:initial-ratio}\end{aligned}$$ is field–independent and always consistent with the LR theory [@Louis2003; @Zemljic2005].
![(Color online) Results for a generic system with $V=3$, $L=24$, $N=10$. (a) Ratio $R=j^E/ J^N$ for $W=1$ as a function of the instantaneous energy compared to the LR results. (b) Long–time ratio $R(t\rightarrow \infty)$. The LR results have been computed at the average temperature of evolution ($\beta\simeq 0.05$). (c) Time dependence of $R(t)$ for $V=3, W=1, F=0.4$ with a single initial state (dashed line), averaged over 16 initial states (solid line), and for a bigger system ($L=26, N=11$) with similar concentration of fermions (dot dashed line). []{data-label="fig2"}](fig2){width="45.00000%"}
In a closed tight–binding model, constant $F$ cannot induce strictly time–independent current since such d.c. response would cause a steady and unlimited in time increase of the energy[@Russomanno2013], while the energy spectrum is bounded from above. However, the long–time dependence can still be reconciled with LR theory[@Mierzejewski2010; @Mierzejewski2011; @Eckstein2011] provided these nonlinear effects of heating are properly filtered out. For a weak but finite $F$ the system undergoes a quasiequilibrium evolution, when the instantaneous expectation values of observable are uniquely determined only by $F$ and the instantaneous energy E(t) or (equivalently) by the instantaneous effective temperature.[@Mierzejewski2013] Consequently one should also consider the dc response functions as quantities which depend on $E(t)$. An extended form of LR, $j^{N(E)}(t) \simeq \sigma^{N(E)}[\omega\to 0, E(t)] F$ holds true in the quasi–equilibrium regime.[@Mierzejewski2010] In this regime the ratios $j^{N(E)}/F$ weakly depend on $F$ and vanish [@Eckstein2011; @Mierzejewski2010] when the system’s energy approaches its value at the infinite temperature, $E_{\infty}$. Both results are explicitly shown in Fig. \[fig1\](b) for the case of $j^{E}$. For $T \rightarrow \infty $ the energy dependence of the response functions cancels out and $R(t\rightarrow \infty)$ should be a well defined and finite. Figure \[fig2\](a) shows that it is actually the case. Moreover, the results obtained from the time–evolution remain in a good agreement with the LR results for the high–temperature regime: $$\begin{aligned}
R(t\rightarrow \infty) \simeq
\left.
\frac{\sigma^{E}(\omega \to 0, E) }
{\sigma^{N}(\omega \to 0, E) } \right|_{E\rightarrow E_{\infty}} .
\label{eq:lr-ratio}\end{aligned}$$ For a nonintegrable systems at nonzero temperature $\sigma^{N}(\omega,E)$ and $\sigma^{E}(\omega,E)$ are regular. In the MCLM they are proportional to the current-current correlator on the state $|\Psi_{\beta}\rangle$ for the energy $E_{\beta}$ corresponding to the inverse temperature $\beta$: $$\begin{aligned}
\sigma^{N(E)}_{\text{reg}}(\omega) =&
L \frac{1 - e^{-\beta\omega}}{\omega} \; \mathrm{Im}\, C^{>}_{N(E)}(\omega), \label{eq:conductivity}\\
C^{>}_{N(E)}(\omega) =& \langle \Psi_\beta | J^{N(E)} (\omega^{+} +E_{\beta} - H)^{-1} J^N | \Psi_\beta\rangle. \label{eq:correlator}
$$ We use a Lanczos expansion with Lorentzian broadening $\omega^{+} = \omega + i 0 ^{+}$. In a driven system, the estimate of $R$ can be obtained robustly from the least-squares scaling of $j^N$ against $j^E$ for long times ($t>t_0 \simeq 50$) $$\begin{aligned}
\frac{d}{d R}\sum_{t_i > t_0} \left[j^E(t_i) - R\, j^N(t_i)\right]^2 = 0.
\label{eq:least-squares-ratio}\end{aligned}$$ The results are shown in Fig. \[fig2\](b) for $V=3$ and various $W$. The data obtained for driven systems nicely recover the equilibrium results from the standard LR approach. Extracting the LR limit is thus possible from the time-dependent quantities, since the ratio of Eq. (\[eq:time-ratio\]) is a well behaved monotonic function of $F$. This holds true as long as the driving is not as strong as to induce the Bloch Oscillations (BO) of the currents [@Freericks2008; @Eckstein2011; @Mierzejewski2011], which eventually occur also in generic nonintegrable systems.
The tiny oscillations of currents around their average values (see Fig. \[fig1\](b)) originate from the fact that we carry out calculations for a finite quantum system and for a single initial state. However small are these oscillations they eventually dominate when the system approaches $\beta\to 0$ and the smooth components of the currents vanish. Then, the numerical results for $R(t)$ being the ratio of two vanishing quantities unavoidably becomes noisy (see Fig. \[fig2\](a)). These oscillations have no physical meaning and can be reduced by either increasing the system size or by averaging over many initial states. Both cases are shown in Fig. \[fig2\](c).
Doped integrable insulator
==========================
After showing that our method reliably applies to the generic case, furtheron we restrict the scope to driven integrable systems and set $W = 0$. In equilibrium the real part of the dynamical conductivity has two separate contributions: $$\begin{aligned}
\sigma^{N}(\omega) = 2\pi D^{N} \delta(\omega) + \sigma^{N}_{\text{reg}}(\omega).
\label{eq:total-conductivity}\end{aligned}$$ The regular part $\sigma^{N}_{\text{reg}}$ is connected with normal (diffusive) behavior while the singular one is weighted by the stiffness $D^{N}$ and implies anomalous (ballistic) transport as well as non-decaying currents. The sum rule $\int_{-\infty}^{\infty} \sigma^{N}(\omega) d\,\omega = \pi \tau^{N}$ allows to normalize and weight the different contributions with the previously defined $\tau^{N}$ operator, thus linking initial-time \[see Eq. (\[eq\_motion\])\] with the dynamical response. Since $j^E$ is conserved, the regular part of $\sigma^{E}(\omega)$ vanishes and the LR response of the energy current is purely singular $$\sigma^{E}(\omega)=2\pi D^{E} \delta(\omega)= \pi \tau^{E} \delta(\omega).
\label{sne}$$
In the case of doped insulators ($V>2$) the Drude weight $D^{N}$ can be well estimated from the Mazur bound by taking the overlap of $J^N$ with a single conserved quantity - the energy current [@Herbrych2011; @Mukerjee2008; @Naef97; @ZotosIdeal]: $$\begin{aligned}
D^{N} \approx D^{N}_{\text{Mazur}} = \frac{\beta L}{2}\; \frac{\langle J^N\, J^E \rangle^2}{\langle J^E\, J^E \rangle^{\phantom{2}}}
\label{eq-mazur}.\end{aligned}$$
![(Color online) Currents for $L=24,N=10$, $V=3$ and $W=0$. (a) $j^E(t)$ (solid lines) together with $R j^N(t)$ (dashed lines) for $R$ shown in the legend. (b) $R(t)$ vs instantaneous energy compared to $R_{\text{Mazur}}$ \[see Eq. (\[eq-mazur-bound\])\].[]{data-label="fig3"}](fig3){width="45.00000%"}
![ (Color online) Parametric plots $j^E(t)$ vs. $j^N(t)$ for $W=0$. (a) Results for $L=24,N=10$ and extremely weak interaction $V=0.2$ when $j^E(t)$ oscillates with frequency twice larger than $j^{N}(t)$. (b) The same as in (a) but for $V=3$ while $L=26,N=9$ and $V=3$ are used in (c). In the two latter panels the straight lines show $\tau^E/\tau^N$ for the initial $\beta$ \[see Eq. (\[eq:initial-ratio\])\] and $R_{\text{Mazur}}$ for $\beta \rightarrow 0$ \[see Eq. (\[eq-mazur-bound\])\].[]{data-label="fig4"}](fig4){width="45.00000%"}
According to the LR theory, $j^N$ and $j^E$ should grow linearly in time for a dc driving. However, this linear growth cannot be unlimited in time under a finite driving as argued in the preceding sections. Then, the currents may develop either into BO [@Freericks2008; @Eckstein2011] or into quasistatic current as observed for generic systems. The latter is also possible since finite $F$ breaks the integrability. Figure \[fig3\](a) shows that the strength of driving determines the scenario which prevails. We observe oscillatory response in the limits of very weak and very strong driving, and quasisteady currents for the intermediate $F$.
The relation between $j^N$ and $j^E$ can be inferred from Fig. \[fig3\] as well as from the parametric plots shown in Fig. \[fig4\]. For a weak–to–moderate driving both currents are roughly proportional to each other. It holds true independently of whether these currents are quasistatic as shown in Figs. \[fig3\](b) and \[fig4\](b) or undergo the BO (Fig. \[fig4\](c)). Hence in this regime the ratio $R(t)$ is indeed well defined and meaningful despite the singular LR of the integrable system. The proportionality between oscillating currents $j^N$ and $j^E$ for $F \rightarrow 0 $ is rather unexpected. Such proportionality is evidently broken for BO under large $F$ (see Fig. \[fig3\](a) for $F=2$) and/or for very weak $V$. Due to an exact doubling of the frequency of their oscillations in the latter case \[see Eqs. (\[jn\]),(\[je\])\] the currents form a damped Lissajous figures in the parametric plane ($j^E$,$j^N$) as shown in Fig. \[fig4\](a).
![(Color online) Results for $V=3$, $L=24$ and $N=10$. (a) Phenomenological scattering rate $\Gamma$ as a function of the instantaneous energy for $W=0$. (b) $R(t\rightarrow \infty)$ for small but finite integrability-breaking interaction $W$. The value of $R_{\text{Mazur}}$ (at $\beta \rightarrow 0$) for $W=0$ is shown for comparison.[]{data-label="fig5"}](fig5){width="45.00000%"}
In order to explain the numerical results we first focus on the regime of intermediate driving, when currents show the same steady behavior as in generic systems under quasiequilibrium evolution. Hence, we apply a similar phenomenological modification of LR which turned out to be successful in the case of generic systems.[@Mierzejewski2010; @Mierzejewski2011] Since the driving itself is sufficient to damp oscillations of the energy current, the main effects must be the broadening of the singular response functions [@Alvarez2002]. A phenomenological attempt would be to modify Eq. (\[sne\]) using a Lorentzian ansatz with an effective scattering rate $\Gamma$ $$\begin{aligned}
\delta(\omega) \longrightarrow \frac{1}{\pi} \frac{\Gamma}{(\omega^2 + \Gamma^2)}.\end{aligned}$$ It leads to an effective dc response $\sigma^{E}(\omega\to0) = \tau^{E} / \Gamma$ and a quasistatic energy current $$\begin{aligned}
j^E = \frac{\tau^{E}}{\Gamma} F.
\label{modlr}\end{aligned}$$ We have used this formula together with the numerical data for $j^E(t)$ and determined the (phenomenological) effective scattering rate shown in Fig. \[fig5\](a). One may observe that $\Gamma$ increases with $F$ and after the initial transient it becomes independent of the instantaneous energy. Therefore the heating effect (dependence on the energy) is included entirely in the sum rule $\tau^{E}$, while $\Gamma$ describes solely the broadening of the response–function by external driving.
It is also interesting that the numerical values of $\Gamma$ are very close to $F$. Hence the effective scattering (damping) rate is close to the frequency of the BO ($\omega_B=F$). Therefore, within this phenomenological picture the regime of the quasistatic current is just at the boundary of overdamped BO.
The same reasoning should also hold for the particle current, however the numerical analysis would be much more demanding since close to half-filling ($ \langle n \rangle \sim 1/2$) the stiffness $D^{N} \ll \tau^{N}/2$ in contrast to $D^{E} = \tau^{E}/2$. However, assuming that a single scattering rate gives the broadening of both response functions, one may estimate the ratio $R(t\rightarrow \infty)$ in the quasiequilibrium regime $$\begin{aligned}
R(t \rightarrow \infty) = \frac{D^{E}}{D^{N}} \simeq
R_{\text{Mazur}}= \frac{\tau^{E}}{\beta L}\frac{\langle J^E\, J^E \rangle^{\phantom{2}}}{\langle J^N\, J^E \rangle^2}.
\label{eq-mazur-bound}\end{aligned}$$ Results in Fig. \[fig3\](b) and \[fig5\](b) show that $R(t \rightarrow \infty)$ is reasonably close to $R_{\text{Mazur}}$, provided $F$ is small enough. The averages at the rhs of Eq. (\[eq-mazur-bound\]) were computed by means of the kernel polynomial method[@Weiße2006] in the canonical ensemble at the temperature determined by the instantaneous energy during the evolution. The deviations between the results from the real–time dynamics and Eq. (\[eq-mazur-bound\]) in Fig. \[fig5\](b) are overestimated since the real–time currents are determined at $E(t) < E_{\infty}$ while $R_{\text{Mazur}}$ for $E \rightarrow E_{\infty}$.
Quite surprisingly, the prediction (\[eq-mazur-bound\]) is accurately fulfilled also for weaker driving when both currents oscillate. In Fig. \[fig4\](c) such behavior is shown for a different filling factor, providing an independent test. After a short transient, the currents oscillate perfectly in phase with a relative amplitude $R$ satisfying the Mazur bound of Eq. (\[eq-mazur-bound\]), regardless of $F$. This agreement makes a clear connection between the BO under finite but weak $F$ and the stiffnesses within the LR theory. Note also that this relation is broken for large $F$, when BO are independent of integrability and occur also in generic systems.
![(Color online) $R(t)$ as a function of instantaneous energy for $V=1.5$, $W=0$. (a) Results for $L=24$ and $N=10$. (b) $L=26$ and $N=9$. Dashed lines show $R_{\text{Mazur}}$ \[see Eq. (\[eq-mazur-bound\])\] and the LR ratio $\frac{D^{E}}{D^{N}}$ \[see Eq. (\[dtrue\])\] both at $\beta \rightarrow 0$.[]{data-label="fig6"}](fig6){width="45.00000%"}
![(Color online) Results for $V=1.5$, $L=24,N=10$. (a) $R(t\rightarrow \infty)$ for decreasing $W$ compared with the LR result \[Eq. (\[eq:lr-ratio\])\]. For the case $W=0$ we also show $R_{\text{Mazur}}$ \[Eq. (\[eq-mazur-bound\])\] and LR ratio $\frac{D^{E}}{D^{N}}$ \[see Eq. (\[dtrue\])\] both at $\beta \rightarrow 0$. (b) The stiffness $D^{N}$, $D^{N}_{\text{Mazur}}$, $D^{N}_{\text{driving}}$ and $D^{N}_{\text{quench}}$ normalized to $\tau^{N}/2$ for $W=0$ as detailed in the text. []{data-label="fig7"}](fig7){width="45.00000%"}
Integrable metals close to half-filling
=======================================
We now turn to the case $V<2$ when the system is metallic at arbitrary filling factor. For moderate fields, currents again display only modest oscillations, so the ratio $R(t)$ can be determined directly (see Fig. \[fig6\]).
It has been shown for integrable metals at half–filling ($\langle n \rangle =1/2$) that the Mazur bound formulated in terms of strictly local conserved operators fails, in particular $D^{N}_{\text{Mazur}}=0$ while $D^N$ stays nonzero. In order to saturate the Mazur bound, one (probably) needs to introduce quasi–local conserved operators.[@quasi11; @quasi13; @Ilievski2013] For slightly smaller concentration of fermions[@Herbrych2011] ($\langle n \rangle < 1/2$), $D^{N}$ is still much larger than $D^{N}_{\text{Mazur}} $, hence the ratio $R(t \rightarrow \infty)$ was expected to be consistently lower than $R_{\text{Mazur}}$ given by (\[eq-mazur-bound\]). However, the numerical data in Fig. \[fig6\] show that $R(t)$ departures from LR and approaches $R_{\text{Mazur}}$, as if the energy current were the relevant conserved quantity. Fig. \[fig7\](a) shows $R(t \rightarrow \infty)$ calculated for small but nonzero $W$ in comparison to the LR results obtained directly from the response functions as well as with $R_{\text{Mazur}} (W=0)$ given by Eq. (\[eq-mazur-bound\]). Upon decreasing $W$ one again observes that results for driven system departure from the predictions of LR theory towards $R_{\text{Mazur}}$ for $W=0$. We expect that breaking the integrability by finite $F$ is responsible for the observed departure from LR regime. In order to verify this expectation we have compared the response of the system driven by $F>0$ with its nonequilibrium relaxation at $F=0$. In particular, we have calculated $D^{N}_{\text{Mazur}}$ given by Eq. (\[eq-mazur\]) as well as the actual charge stiffness calculated from the sum rule in Eq. (\[eq:total-conductivity\]) taking the regular conductivity \[Eq. (\[eq:conductivity\])\] in the initial MCLM state $$\begin{aligned}
D^{N} = \frac{\tau^{N}}{2}-\frac{1}{2\pi}\int_{-\infty}^{\infty} \sigma_{\text{reg}}^{N}(\omega)\, d\omega.
\label{dtrue}\end{aligned}$$ These equilibrium results have been compared with two nonequilibrium cases. For a system evolving under finite $F$ one can estimate the charge stiffness from $R(t\rightarrow \infty)$ assuming that $j^E/j^N \simeq D^{E}/D^{N}_{\text{driving}}$ holds in long–time regime similarly to the case of doped insulators. Then, $$\begin{aligned}
D^{N}_{\text{driving}} = \frac{\tau^{E}}{2R(t\to\infty)}.\end{aligned}$$ Finally, we have studied an instantaneous change of the magnetic flux which should also be consisted with LR. At $t=0$ we quench the flux $\phi(t) = \Delta \phi\, \theta(t) $ inducing an electric field $F(t) = -\Delta \phi\, \delta(t)$. To the first order in $\Delta \phi$ the time–dependent particle current reads $$\begin{aligned}
j^N(t) \!=-\tau^{N} \, \Delta \phi - i L \int_{0}^{t} \langle [J^N(t'),J^N(t)] \rangle \Delta \phi\, dt'\end{aligned}$$ which gives the peak value $j^N(t \rightarrow 0^+) = -\tau^{N} \Delta \phi$ since the integrand is smooth. The real-time LR current is given by $j^N(t) = \int \,d\omega\,F(\omega) \boldsymbol{\sigma}^{N}(\omega) e^{-i \omega^{+} t}$ where $F(\omega) = -\frac{\Delta \phi}{2 \pi}$ and $\boldsymbol{\sigma}^{N}$ is the complex conductivity. The regular part of $\boldsymbol{\sigma}^{N}$ is smooth and gives no contribution to $j^N(t)$ for $t \rightarrow \infty$. With the complex singular part $\boldsymbol{\sigma}^{N}_{\text{sing}}(\omega) = \frac{2iD^{N}}{\omega + i 0^{+}}$, the current after the quench stabilizes to $$\begin{aligned}
j^N(t \to \infty)\! =\! - \int_{-\infty}^{\infty} \frac{\Delta \phi}{2 \pi} \frac{2i D^{N}}{\omega + i 0^{+}} e^{-i \omega t} \,d\omega
\overset{Res}{=} - 2 D^{N} \Delta \phi.\end{aligned}$$ We have calculated the ratio of the peak to long time currents also for finite $\Delta \phi$ and estimate the ratio of the Drude weight intervening in the quench to the sum-rule expectation value: $$\begin{aligned}
\frac{j^{N}(t\to\infty)}{j^{N}(t\to 0^{+})} = \frac{2 D^{N}_{\text{quench}}}{\tau^{N}}.
\label{eq:quench-ratio}\end{aligned}$$ We stress that the actual stiffness is defined within LR by Eq.(\[dtrue\]). The results for $D^{N}_{\text{driving}}$ and $D^{N}_{\text{quench}}$ are expected to merge with $D^{N}$ when LR is applicable respectively to a system driven by a nonzero field and a system that relaxed after a nonzero quench of the magnetic flux. All these estimates of the stiffness are compared in Fig. \[fig7\](b). For vanishing electric field $D^{N}_{\text{driving}}$ approaches $D^{N}_{\text{Mazur}} \ll D^{N}$, whereas $D^{N}_{\text{quench}}$ nicely reproduces the LR result $D^{N}$. The latter agreements holds also for strong quenches $\Delta \phi$, i.e. for relaxation from far–from–equilibrium states. The deeper understanding of the contrasting result for driving and relaxation remains an open problem and requires further studies. In particular, it remains to be checked whether $j^E(t)/j^N(t)$ approaches $R_{\text{Mazur}}$ also for other driven integrable systems.
Summary
=======
We have studied an integrable one–dimensional system of interacting spinless fermions and established the long–time ratio of the energy current ($j^E$) and the particle current ($j^N$) under dc driving by nonzero electric field $F$. The equilibrium LR theory predicts singular (ballistic) responses of both currents, as quantified by the stiffnesses $D^E$ and $D^N$, respectively. Since $j^E$ is a conserved quantity (at $F=0$), $D^E$ represents simply the stress coefficient. However, $j^N$ is not conserved and the physical origin of a finite $D^N$ is more complex. We have first considered a system (doped insulator) where the local conserved quantities saturate the Mazur bound on $D^N$. In this case the long–time results for $j^E(t)/j^N(t)$ agree with the LR ratio $D^E/D^N$, despite the currents themselves are steady or oscillating in contrast to the LR prediction $j^{N(E)} \propto t$. We have then studied a system (doped metal close to half–filling) where large $D^N$ cannot be explained by the Mazur bound formulated in terms of local conserved quantities. On the one hand, the ratio $j^E(t)/j^N(t)$ obtained for a system which relaxes after a flux–quench ($\delta$–like pulse of electric field) nicely agrees with the LR theory. On the other hand, $j^E(t)/j^N(t)$ obtained for a steady driving becomes much larger than the LR value $D^E/D^N$. While the deviation from the LR theory in the latter case is evident, we are not aware of any qualitative explanation for this discrepancy.
*Acknowledgments.* This work has been carried out within the project DEC-2013/09/B/ST3/01659 financed by the Polish National Science Center (NCN). P.P. acknowledges the support by the Program P1-0044 and project J1-4244 of the Slovenian Research Agency.
| {
"pile_set_name": "ArXiv"
} |
epsf
$^1$[*IPNL, CNRS/IN2P3, 4 rue E. Fermi, 69622 Villeurbanne cedex, France; Université Lyon 1, Villeurbanne; Université de Lyon, F-69622, Lyon, France* ]{}\
> An Effective Field Theory for dark matter at a TeV-scale hadron collider should include contact interactions of dark matter with the partons, the Higgs and the $Z$. This note estimates the impact of including dark matter-$Z$ interactions on the complementarity of spin dependent direct detection and LHC monojet searches for dark matter. The effect of the $Z$ is small, because it interacts with quarks via small electroweak couplings, and the contact interaction self-consistency condition $C/\Lambda^2 < 4\pi/\hat{s}$ restricts the coupling to dark matter. In this note, the contact interactions between the $Z$ and dark matter are parametrised by derivative operators; this is convenient at colliders because such interactions do not match onto low energy quark-dark matter contact interactions.
Introduction
============
Various experiments attempt to detect the particle making up the “dark matter”[@DM] of our Universe. For instance, direct detection(DD) experiments [@EdelCDMS; @Xenon; @SC; @autres], search for $\sim $ MeV energy deposits due to scattering of dark matter particles from the galactic halo on detector nuclei. And the Large Hadron Collider (LHC) searches[@CMS; @ATLAS] for dark matter pairs produced in multi-TeV $pp$ collisions, which would materialise as an excess of events with missing energy and jets. The LHC and DD searches are at very different energy scales, so different Standard Model (SM) particles are present, and also the quantum interferences are different[@PST]. The expected rates can be compared in specific dark matter models [@R], or, in recent years, several studies[@CMS; @tevatron; @toutlmonde; @Z; @unitarity1; @monoH] have compared the LHC and DD sensitivities using a contact interaction parametrisation of the dark matter interactions with the standard model particles.
The LHC bounds obtained in this way are restrictive, and probe smaller couplings than direct detection experiments searching for “spin dependent” interactions between partons and dark matter [@SC]. These contact interaction studies are refered to as “Effective Field Theory” (EFT), and considered to be relatively model independent. However, the particle content is an input in EFT, and the restrictive LHC limits assume that the dark matter particle is the only new particle accessible at the LHC. Relaxing this assumption can significantly modify the experimental sensitivities[@st; @pvz; @toni2]. This has motivated various simplified models for dark matter searches at the LHC [@marcusmodel; @DMLQ; @deSGS]. Retaining this assumption, as will be done in this note, is only marginally consistent, because the contact interactions to which the LHC is sensitive would have to be mediated by strongly coupled particles. As recalled in the next section, this implies that colliders can exclude contact interactions of order their sensitivity, but not much larger.
Effective Field Theory (EFT) is a recipe to get the correct answer in a simple way[@Georgi]. So this note attempts to compare LHC and DD constraints on dark matter, according to the prescriptions of [@Georgi]. An EFT for dark matter at the LHC should parametrise all possible SM-gauge invariant interactions of the dark matter with other on-shell particles. So first, contact interactions between the dark matter and the Higgs or $Z$ should be included at the LHC. These can interfere with the contact interactions studied in previous analyses, but contribute differently at colliders from in direct detection, so the linear combination of operator coefficients constrained at high and low energy will be different. Secondly, an EFT contains in principle a tower of operators[@Porod] organised in increasing powers of the inverse cutoff scale $1/\Lambda$, and higher orders can only be neglected if there is a sufficient hierarchy of scales: $\Lambda_{NP}\gg v$. This hierarchy is absent in dark matter production at the LHC. Addressing the importance of higher dimensional operators will be left to a subsequent publication [^1].
This note focuses on including the $Z$ in the EFT for dark matter at the LHC, and estimates analytically the consequences of including the lowest dimension operators allowing dark matter interactions with the $Z$ [^2]. Section \[sec:EFT\] outlines a peculiar choice of operators for the $Z$ vertex; they are proportional to the momentum-transfer-squared. This choice appears convenient, because the effects of the $Z$ are therefore absent in direct detection. Section \[sec:LHC\] estimates the impact of cancellations between $Z$ exchange and dark matter contact interactions with quarks at the LHC, and section \[sec:DD\] recalls the direct detection bounds.
EFT, assumptions and operators {#sec:EFT}
===============================
The low energy consequences of New Physics from above a scale $\Lambda$ can be parametrised by contact interactions of coefficient $C/\Lambda^n$. Unitarity [@unitarity1; @unitarity2] approximately implies that $C< 4\pi$, and the contact interaction approximation implies that the momentum exchange should be less than $\Lambda$. This means that an experiment can [*exclude*]{} > > [sensitivity]{} , \[plage\] where $\hat{s}$ is the four-momentum-squared of the process. Low energy experiments, where $\hat{s} \to 0$, therefore can be taken to exclude everything above their sensitivity. However, the upper limit of eqn (\[plage\]) is relevant for collider searches, where $\hat{s}$ is the invariant mass of the invisibles. This upper limit is rarely taken into account in the literature.
The first step in the EFT recipe to parametrise New Physics from beyond the scale $\Lambda$ it to add to the Lagrangian (at the scale $\Lambda$), all the non-renormalisable operators which can be constructed out of the fields present, consistently with the symmetries of the theory[@Georgi]. The coefficients $C^{(n)}_O$ of these operators are unknown “coupling constants” which evolve with scale via Renormalisation Group Equations. This infinite set of operators would be unmanageable, so EFT is useful when there is a hierachy between the experimental and NP scales. Then only the lowest dimension operators need be considered.
In this note, the dark matter is assumed to be the only new New Physics particle lighter than a TeV, and is taken to be a SM gauge singlet dirac fermion $\chi$ with a conserved parity, and of mass $m_\chi \geq m_Z/2$ (maybe $\geq m_h/2$), to avoid bounds on the coupling to the $Z$ from the invisible width[^3] of the $Z$ (and Higgs). So the particle content of the EFT for $\chi$ at the LHC should be $\chi$, plus all relevant particles of the SM, which I take to be the partons, the Higgs, and the $Z$.
The operators should be SM gauge invariant, to profit from our knowledge of the SM gauge sector. They are of dimension $>$ 4, and should attach a $\chi \overline{\chi}$ pair to partons, to the Higgs, or to the $Z$. The quark operators are taken generation diagonal; flavour-changing operators were considered in [@CK11]. The quarks are chiral because the operators are SM gauge invariant, and also because opposite chiralities do not interfere at the LHC. The dark matter currents are taken in a vector, axial vector, etc basis because these do not interfere in direct detection, nor at the LHC in the limit where the $\chi$ mass is neglected, as done here.
I focus on operators of lowest dimension, that is six and seven. This is an arbitrary simplification, because $\Lambda \sim$ TeV, which is the energy scale probed at the LHC. The contact interactions considered here therefore do not provide a “model-independent” parametrisation of the interactions of $\chi$ with the SM. This problem is left for a later publication. Concretely, $\Lambda$ will be taken as 1- $2 $ TeV, for reasons discussed above eqn (\[lim3\]). Experimental limits on contact interactions will therefore be presented as limits on the dimensionless coefficient $C^{(n)}_O$.
At dimension six, there are vector and axial vector $\chi$ currents coupled to quarks: \_ \^P\_X Q\_i , - \_\_5 \^P\_X Q\_i \[qDMd6\] where the quarks $Q_i$ are first generation SM multiplets $ \{q_L,u_R,d_R\}$, and $P_X$ is the appropriate chiral projector. The contact interactions between the dark matter and the $Z$ boson are taken as - D\^B\_ \_ s\_[w]{} p\_Z\^2 Z\^ \_\
D\^B\_ \_\_5 -s\_[w]{} p\_Z\^2 Z\^ \_\_5 \[inteff\] where to the right of the arrow is the resulting vertex, $B^\mu$ is the hypercharge gauge boson with coupling $g'=e\tan\theta_W \equiv e s_{\rm w}/c_{\rm w}$, $B^{\mu\nu} = \partial ^\mu B^\nu - \partial ^\nu B^\mu$, and a term $\propto p_Z\cdot Z$ was dropped after the arrow in the axial current operator, assuming the $Z$ was produced by light quarks. There is in addition a “dipole moment” operator $B^{\mu \nu}\overline{\chi} \sigma_{\mu \nu}\chi $, which is neglected here because it also induces dark matter interactions with the photon [@magmoDM] which are more interesting.
Then at dimension seven, there are four-fermion operators: H d + \[ H d \]\^ , \_5 H d + \[ H d \]\^\
\^ H \_ d + \[ H \_ d \]\^ (and similarly for $u$ quarks, but with a charge conjugate Higgs field), interactions with the gluons: $$\frac{C^{(7)}_{gg,S} }{\Lambda^3}
\overline{\chi} \chi G_{\mu \nu}^A G^{\mu \nu, A}
~~~,~~~
\frac{C^{(7)}_{g\tilde{g},P} }{\Lambda^3}
\overline{\chi}\g_5 \chi G_{\mu \nu}^A \tilde{G}^{\mu \nu, A}~~~,$$ and double-derivative interactions between dark matter and the Higgs: H\^D\^D\_ H - m\_W\^2 W\_\^+W\^[-]{} - m\_Z\^2 Z\_Z\^[ ]{} + h\
H\^D\^D\_ H \_5 - m\_W\^2 W\_\^+W\^[- ]{} \_5 - m\_Z\^2 Z\_Z\^[ ]{} \_5 + h \_5\[higgs\] where $\langle H \rangle = v = 174$ GeV, $p_h$ is the four-momentum of the physical Higgs particle $h$, and after the arrow are the interactions induced by the operator.
The $Z$ and Higgs operators are choson $\propto p^2$ so that they are relevant at the LHC where the $Z$ and Higgs are external legs in the EFT, but do not contribute in the low-energy scattering of DD. This choice should be acceptable, because the operator basis can always be reduced by using the equations of motion[@Simma]. Focussing for simplicity on the hypercharge boson $B$, and neglecting gauge-fixing terms, the equations of motion are [@polonais] D\_B\^= g’y\_H (H\^D\^H - (D\^H)\^H) + g’\_y\_\^\[EoMZ\] where $\psi$ is a SM fermion of hypercharge $y_\psi$. Usually[@polonais], operators containing the double derivative on the left of eqn(\[EoMZ\]) are dropped, and the operators containing the Higgs v.e.v. squared $ \langle H^\dagger \Dlr H \rangle$ are retained. In this usual basis, $\chi-Z$ interactions could be parametrised by $(\overline{\chi}\gamma^\mu \chi) H^\dagger \Dlr H$, in which case the matrix element for $Z$ exchange at the LHC is $\propto m_Z^2/(p_Z^2 -m_Z^2)$, so negligeable for $p_Z^2 \gg m_Z^2$. But $Z$ exchange should be included in the quark-$\chi$ contact interaction used in direct detection, so the coefficient of the operators of eqn (\[qDMd6\]) would not be the same in direct detection as at the LHC. To avoid this discrepancy, I retain the derivative operators of eqn (\[inteff\]), and use eq. (\[EoMZ\]) to remove the operator $ \propto \langle H^\dagger \Dlr H \rangle$. This means that the $Z$ couples significantly to $\chi$ at the LHC, but negligeably in DD, and the operator coefficients do not change when the $Z$ is matched out of the theory.
In the case of the Higgs, the equation of motion is $$D_\mu D^\mu H = \mu^2 H -\lambda H^\dagger H H -
\overline{e}Y_e^\dagger P_L
\ell - \overline{d}Y_d^\dagger q_L +
\varepsilon \overline{q_L}Y_u u$$ where $Y_f$ are Yukawa matrices. This has been used to exchange the more usual $ (H^\dagger H)^2 \overline{\chi} \chi$, and $ (H^\dagger H) \overline{\chi} \chi$ operators for the double-derivative interactions between dark matter and the Higgs given in eqn (\[higgs\]). Notice that it is possible to use the equations of motion to replace two operators ($\overline{\chi} \chi H^\dagger H$ and $\overline{\chi} \chi (H^\dagger H)^2$) with one (involving the DM and $H^\dagger D^2 H$), because I am only interested in the $h$-$\chi$-$\bar{\chi}$ interaction induced by these operators. The linear combination of operators $[\mu^2 H^\dagger H -\lambda (H^\dagger H)^2]\overline{\chi}\chi$, which is orthogonal to the combination in the Equations of Motion, gives a vanishing $h$-$\chi$-$\bar{\chi}$ interaction, due to the minimisation condition of the Higgs potential. As in the case of the $Z$, the derivative operators of eqn (\[higgs\]) are interesting, because they give a higgs coupling to dark matter $\propto p_h^2$, which has the desirable feature of being relevant at the LHC where the Higgs is in the effective theory, but not contributing at low energy.
This note focusses on the interactions of $\chi$ with the $Z$ (eqn \[inteff\]), and with the quark currents of eqn (\[qDMd6\]) which can interfere with $Z$ exchange. So the dimension seven operators will be neglected in the following sections. However, it is interesting to first review the sensitivity to the coefficients of the operators of eqn (\[higgs\]). The dark matter interactions to $W$ and $Z$ pairs were studied in [@HR], who used $U(1)_{em} \times
SU(3)$ invariant operators such that these contact interactions have dimension five with coupling $1/\Lambda_{CHLR}$. They find that the 8 TeV LHC with luminosity 25 fb$^{-1}$ could probe $\Lambda_{CHLR} \lsim $ TeV. This constrains the coefficients of the operators of eqn (\[higgs\]) to be $ \lsim 1/({\rm TeV}m_W^2)$, which is not restrictive. For $m_\chi< m_h/2$, a more significant limit of 10 TeV$^{-3}$ arises from requiring $\Gamma(h\to \chi \overline{\chi})
\lsim \Gamma(h\to b \overline{b})$. This restriction should be reasonable[@BMM] because the Higgs is observed to decay to $b\bar{b}$.
Estimated limits from the LHC {#sec:LHC}
=============================
Dark matter particles are invisible to the LHC detectors, so pair production of $\chi$s can be searched for in events with missing transverse energy ($\ETm)$, which can be identified by jet(s) radiated from the incident partons. The principle Standard Model background for such “monojet” searches is $Z +$ jet production, followed by $Z\to \bar{\nu}\nu$. The 8 TeV LHC is sensitive to dark matter contact interactions with $C/\Lambda^2 \sim $ TeV$^{-2}$.
Given the operators of eqns (\[qDMd6\]) and (\[inteff\]) at the LHC, the axial vector dark matter current can interact with quarks $Q$ via the diagrams of figure \[figop\], which can be written as a four-fermion interaction of coefficient c\_[QX,A]{} & = & C\_[QX,A]{}+ g\_X\^Q\
& &C\_[QX,A]{}+ g\_X\^Q C\_[Z,A]{} , \[annuler\] where $g_X^Q =\{ 1- \frac{4}{3}s_{\rm w}^2,- \frac{4}{3}s_{\rm w}^2,
-1+ \frac{2}{3}s_{\rm w}^2, \frac{2}{3}s_{\rm w}^2\}$ for $\{u_L,u_R,d_L,d_R\}$ [@PDB]. A similiar expression can be obtained for the vector $\chi$ current. The $Z$ exchange looks like a contact interactions for large $p_Z^2 = M^2_{inv} \gg m_Z^2$, where $M^2_{inv}$ is the invariant mass-squared of the dark matter pair. This is a useful approximation, because the arguments below suggests that most $\chi\bar{\chi}$ events arise at larger $M^2_{inv}$.
The aim here is to analytically estimate the invisible four-momentum-squared $ M^2_{inv} $, by comparing the partonic cross-sections for $\nu\bar{\nu}$ and $\chi\bar{\chi}$ production. I assume that the QCD part of the amplitude is identical in both cases, so it does not need to be calculated. This allows for an arbitrary number of jets, which is more difficult to simulate[@Uli2] (the data frequently contains more than one jet[@CMS]). In the matrix element for jets +$\nu \bar{\nu}$ will appear $$g_X^Q\frac{g^2}{4c^2_W}
\frac{1}{p^2 -m_Z^2 +im_Z\Gamma_Z}
(\bar{Q} \g^\a P_X Q )
( \overline{\nu}\g_\a P_L\nu)$$ whereas, for DM production via the $\bar{\chi} \g^\mu\g_5 \chi$ current, this is replaced by: $$\frac{c_{QX,A}}{\Lambda^2}
(\bar{Q} \g^\a P_X Q )
(\overline{\chi}\g_\a\g_5 \chi)~~.$$ Then the full matrix element must be squared, and integrated over the phase space of $N$ jets and two invisible particles. The invisibles can be treated as a single particle of variable mass $p^2 = M^2_{inv}$, using the identity d\_[N+2]{} &=& \^4(P\_[in]{}- q\_i -p) \_[i:1..N]{} ( ) (2)\^3 dp\^2 \^4(p-p\_- p\_[|]{}) . Neglecting spin correlations and the dark matter mass, the invisible phase space integral over the gamma-matrix trace for the invisible fermions gives $M^2_{inv}/(8\pi)$ for $\chi$s, and $3M^2_{inv}/(16\pi)$ for neutrinos. For neutrinos in the final state, $M_{inv}^2 = m_Z^2$ due to the delta-function-like behaviour of the $Z$ propagator-squared. However, for dark matter, the $d M^2_{inv}$ phase space integral will privilege larger values of $M^2_{inv}$. Treating the $N$ jets of the event as a particle of negligeable mass, the upper bound on $M_{inv}^2$ is $ \gsim 4 \ETm^2$, where $\ETm$ is the invisible transverse energy. The CMS study [@CMS] uses the range 400 GeV $\leq \ETm\lsim$ TeV. However, the assumption that the jet emission part of the cross-section is the same as for $\nu$ pairs will fail, if $M_{inv}^2$ is a significant fraction of the energy of the event. With the $M_{inv}$ cutoff ranging from 800 GeV to 2 TeV, requiring that the dark matter contribute $\lsim 1/6$ [@CMS] of the SM background, gives an estimated bound $\Lambda \gsim 880 \to 2200$ GeV, for $c_{uL,A} =c_{uR,A} =c_{dL,A} =c_{dR,A}=1$. This compares favourably to the CMS bound of $\Lambda > 950$ GeV, for $C_{uL,A} =C_{uR,A} =C_{dL,A} =C_{dR,A}=1$. Since the analytical estimate is reasonable, most of the dark matter signal probably comes from $M_{inv}^2 \gg m_Z^2$, and the approximation (\[annuler\]) is consistent. However, the analytic bound is a bit to restrictive (perhaps in part because it includes any number of jets), so in the remainder of the paper, the CMS limit of 950 GeV will be used.
There is also an upper limit on the $C$s which a collider can exclude, eqn (\[plage\]), from requiring that the contact interaction approximation be self-consistent: $C/\Lambda^2 <4\pi/M_{inv}^2$. Since the previous analytic estimate reproduces the CMS bound for $M_{inv}^2 \sim$ TeV$^2$, the consistency condition is taken as $C< 4\pi$. For the axial $\chi$ current with $\Lambda = $ TeV, the CMS limit and eqn (\[plage\]) give 3 independent bounds on $\{ c_{qL,A}, c_{uR,A}, c_{dR,A}\}$: 4\
4 |C\_[uR,A]{} - C\_[Z,A]{}| \
4 |C\_[dR,A]{} + C\_[Z,A]{}| \[lim3\] where the first line is the summed contributions of $u_L$ and $d_L$, the fractions are approximations $gg_X^Q s_{\rm w}/2c_{\rm w}$, and the $d$ to $u$ pdf ratio is taken 1/2. Similar limits apply for the vector operator of eqn (\[qDMd6\]).
It can be seen already from eqn (\[lim3\]), that including the interactions with the $Z$ will make little differences to the LHC limits on the $C_{QX,A}$: for the doublet quarks, the $Z$ contribution cannot cancel simultaneously against the $u_L$ and $d_L$ contributions, and the $Z$ contribution is irrelevant for the singlet quarks, because also $C_{Z,A}$ must be $\lsim 4\pi$. The parameters ruled out by the first and second eqns of (\[lim3\]) are represented as the central regions in figure \[fig2\].
From the TeV to the MeV {#sec:DD}
=======================
In direct detection, the dark matter scatters non-relativistically off nuclei. Therefore, to translate the EFT from the TeV to the MeV, the $Z$ must be removed, the effects of QCD loops in running the operator coefficients should be included, and the quarks must be embedded in the nucleons.
To remove the $Z$, the Greens function for two quarks and two $\chi$s in the effective theory with a $Z$, should be matching to the same Greens function in the theory without a $Z$. Since the matching is performed at zero momentum for the fermion legs, the contact interactions of eqn (\[inteff\]) do not contribute, and the coefficients of the four-fermion operators of eqn (\[qDMd6\]) remain the same after the $Z$ is “matched out”. The $Z$ vertices were taken $\propto p_Z^2$ to obtain this.
The light quark currents $\overline{q} \g^\mu P_X q$ are conserved in QCD, so do not run. Also, since $\chi$ is a SM gauge singlet and the only dark sector particle below the TeV, I suppose that the operators with vector and axial vector $\chi$ currents do not mix below the TeV. See [*e.g.*]{} [@UliSDSI] about loop effects mixing various operators involving dark matter and the SM.
Finally, the quark currents can be embedded in nucleons $N = \{p,n\}$ using identities [@BBPS] such as N| \^ Q\_i |N &=& c\_[V,i]{}\^[N]{} N| \^\_N |N where $ c_{V,u}^{p}= c_{V,d}^{n} = 2$, and $c_{V,d}^{p} = c_{V,u}^{n} = 1$, because this current counts valence quarks in the nucleon. The axial quark current is proportional to the nucleon spin: N| \^\_5 Q\_i |N &=& 2 s\^Q\_i\^N = Q\_i\^N N| \^\_5 \_N |N where the proportionality constants are measured [@DqN] as $ \Delta u^p = \Delta d^n = 0.84$, $\Delta d^p = \Delta u^n = -0.43$. In the zero-momentum-transfer limit of non-relativistic scattering, the dark matter can have spin-dependent interactions via the axial current, or spin-independent interactions via the first component of the vector current.
The spin-independent scattering amplitude for $\chi$ on a nucleon, is a coherent sum of vector and scalar interactions, for quarks of both chiralities and all flavours. The experimental limit on the cross-section per nucleon is $\sigma_{SI}\lsim 10^{-44}$ cm$^2$ for $m_\chi \sim 100$ GeV [@Xenon]. For the proton ($C_{uR}\leftrightarrow C_{dR}$ for the neutron), with $ C_{qR,V} =\frac{1}{3}( C_{dR,V } + 2C_{uR,V})$, this gives [@BBPS] \_[SI]{} \^2 3 10\^[-17]{} [GeV]{}\^[-2]{} where the $+...$ contains scalar contact interactions neglected in this note. For $\Lambda =$ TeV, this gives \[C\_[qL,V]{} + ( C\_[dR,V ]{} + 2C\_[uR,V]{}) + ...\] 10\^[-2]{} (SI). \[SIbd\] The spin dependent cross-section per proton is [@BBPS] $$\sigma_{SD} \!\simeq \! m_p^2 \left[
\frac{.42 (C_{qL,A} \!+ \!C_{uR,A} \!- \ \!
2C_{dR,A} )
}{2\Lambda^2} \right] ^2 \! \! \! \lsim \frac{ 10^{-10 }}{4} {\rm GeV}^{-2}$$ where the experimental bound is for $m_\chi \sim 100$ GeV. For $\Lambda = $ TeV, this gives | (C\_[qL,A]{}+C\_[uR,A]{} - 2C\_[dR,A]{} )| 20 (SD). \[SD\] Comparing to eqn (\[lim3\]) shows that the contact interactions explored by SD direct detection experiments are mediated by physics which is not a contact interaction at the LHC, so are not excluded by the limits given in eqn (\[lim3\]). The limit (\[SD\]) is represented in figure \[fig2\] as the vertical exclusions.
Discussion
==========
From a bottom-up EFT point of view, it is important to include all operators which can interfere, when computing experimental constaints. This is to allow for cancellations. Including several operators which do not interfere improves the bound, but is not otherwise motivated. In this note, operators with vector and axial vector currents for the dark matter fermion $\chi$ were presented as an example, which illustrates two points.
First, the EFT at the LHC contains more particles than the light partons and dark matter that are relevant in direct detection. At the LHC, the Higgs and $Z$ should also be included. Matching the high and low energy EFTs, as done in this note, suggests that the LHC constrains several combinations of operator coefficients that are different from direct detection, as can be seen by comparing eqns (\[lim3\]) and (\[SD\]). However, the contribution of the $Z$ is relatively unimportant, because its couplings to singlet quarks are small, and it interferes with opposite sign with $u_L$ and $d_L$. The LHC limits on the dark matter couplings to quarks and the $Z$ are represented as the central exclusion areas of figure \[fig2\]: the coupling to quarks is more constrained than the coupling to the $Z$, and arbitrary axial current dark matter interactions to quarks cannot be allowed by tuning the dark matter coupling to the $Z$. This is because there is a self-consistency upper bound on contact interaction coefficients at colliders $C/\Lambda^2 < 4\pi/\hat{s}$ (see eqn (\[plage\])). It is important to notice that this upper bound also implies that the LHC limits do not exclude the parameter space probed by spin dependent direct detect experiments.
Second, an interesting difference between direct detection and collider experiments, is that quarks of different chirality and flavour interfere in direct detection, whereas the LHC can constrain the interactions of dark matter with each flavour and chirality of quark individually. This is related to the relative unimportance of the $Z$: it cannot cancel separately against the contributions of $u_L, d_L, u_R$ and $d_R$.
In summary, the rules of bottom-up Effective Field Theory say that one should include all operators up to some specified dimension. So to parametrise at dimension six the axial vector interactions of dark matter with quarks, one should include contact interactions of dark matter with the quarks and with the $Z$. Including interactions with the $Z$ that are $\propto p_Z^2$, as done here, suggests that these are not crucial.
Acknowledgements {#acknowledgements .unnumbered}
----------------
I thank J.P. Chou, S Malik, and S. Perries for useful comments.
[222222]{}
G. Bertone, D. Hooper and J. Silk, “Particle dark matter: Evidence, candidates and constraints,” Phys. Rept. [**405**]{} (2005) 279 \[hep-ph/0404175\]. [[Jungman]{}, G. and [Kamionkowski]{}, M. and [Griest]{}, K.]{}, “[Supersymmetric dark matter]{}”, Phys. Rept. [**267**]{} (1996) 195 \[hep-ph/9506380\].
Z. Ahmed [*et al.*]{} \[CDMS and EDELWEISS Collaborations\], “Combined Limits on WIMPs from the CDMS and EDELWEISS Experiments,” Phys. Rev. D [**84**]{} (2011) 011102 \[arXiv:1105.3377 \[astro-ph.CO\]\].
D. S. Akerib [*et al.*]{} \[LUX Collaboration\], “First results from the LUX dark matter experiment at the Sanford Underground Research Facility,” arXiv:1310.8214 \[astro-ph.CO\]. E. Aprile [*et al.*]{} \[XENON100 Collaboration\], “Dark Matter Results from 225 Live Days of XENON100 Data,” Phys. Rev. Lett. [**109**]{} (2012) 181301 \[arXiv:1207.5988 \[astro-ph.CO\]\]. M. Felizardo, T. A. Girard, T. Morlat, A. C. Fernandes, A. R. Ramos, J. G. Marques, A. Kling and J. Puibasset [*et al.*]{}, “Final Analysis and Results of the Phase II SIMPLE Dark Matter Search,” Phys. Rev. Lett. [**108**]{} (2012) 201302 \[arXiv:1106.3014 \[astro-ph.CO\]\]. E. Vazquez-Jauregui \[COUPP Collaboration\], “COUPP: Bubble chambers for Dark Matter detection,” in Proceedings of the 48th Rencontres de Moriond on Very High Energy Phenomena in the Universe.
V. Zacek, S. Archambault, E. Behnke, J. Behnke, M. Das, A. Davour, F. Debris and N. Dhungana [*et al.*]{}, “Dark matter search with PICASSO,” J. Phys. Conf. Ser. [**375**]{} (2012) 012023.
CMS Collaboration, “Search for new physics in monojet events in pp collisions at sqrt(s)= 8 TeV”, CMS-PAS-EXO-12-048.
ATLAS Collaboration, “Search for New Phenomena in Monojet plus Missing Transverse Momentum Final States using 10fb$^{-1}$ of pp Collisions at $\sqrt{s}=8$ TeV with the ATLAS detector at the LHC”, ATLAS-CONF-2012-147.
S. Profumo, W. Shepherd and T. Tait, “The Pitfalls of Dark Crossings,” arXiv:1307.6277 \[hep-ph\]. see [*e.g.*]{} T. G. Rizzo, “Dark Matter Complementarity in the pMSSM and the ILC,” arXiv:1402.5870 \[hep-ph\], and references therein.
Y. Bai, P. J. Fox and R. Harnik, “The Tevatron at the Frontier of Dark Matter Direct Detection,” JHEP [**1012**]{} (2010) 048 \[arXiv:1005.3797 \[hep-ph\]\].
J. Goodman, M. Ibe, A. Rajaraman, W. Shepherd, T. M. P. Tait and H. B. Yu, “Constraints on Light Majorana dark Matter from Colliders,” Phys. Lett. B [**695**]{} (2011) 185 \[arXiv:1005.1286 \[hep-ph\]\]. P. J. Fox, R. Harnik, J. Kopp and Y. Tsai, “Missing Energy Signatures of Dark Matter at the LHC,” Phys. Rev. D [**85**]{} (2012) 056011 \[arXiv:1109.4398 \[hep-ph\]\]. J. Goodman, M. Ibe, A. Rajaraman, W. Shepherd, T. M. P. Tait and H. -B. Yu, “Constraints on Dark Matter from Colliders,” Phys. Rev. D [**82**]{} (2010) 116010 \[arXiv:1008.1783 \[hep-ph\]\]. P. J. Fox, R. Harnik, R. Primulando and C. T. Yu, “Taking a Razor to Dark Matter Parameter Space at the LHC,” Phys. Rev. D [**86**]{} (2012) 015010 \[arXiv:1203.1662 \[hep-ph\]\]. N. Zhou, D. Berge and D. Whiteson, “Mono-everything: combined limits on dark matter production at colliders from multiple final states,” Phys. Rev. D [**87**]{} (2013) 9, 095013 \[arXiv:1302.3619 \[hep-ex\]\]. L. M. Carpenter, A. Nelson, C. Shimmin, T. M. P. Tait and D. Whiteson, “Collider searches for dark matter in events with a Z boson and missing energy,” Phys. Rev. D [**87**]{} (2013) 7, 074005 \[arXiv:1212.3352\]. I. M. Shoemaker and L. Vecchi, “Unitarity and Monojet Bounds on Models for DAMA, CoGeNT, and CRESST-II,” Phys. Rev. D [**86**]{}, 015023 (2012) \[arXiv:1112.5457 \[hep-ph\]\]. A. A. Petrov and W. Shepherd, “Searching for dark matter at LHC with Mono-Higgs production,” Phys. Lett. B [**730**]{} (2014) 178 \[arXiv:1311.1511 \[hep-ph\]\]. M. T. Frandsen, F. Kahlhoefer, A. Preston, S. Sarkar and K. Schmidt-Hoberg, “LHC and Tevatron Bounds on the Dark Matter Direct Detection Cross-Section for Vector Mediators,” JHEP [**1207**]{} (2012) 123 \[arXiv:1204.3839 \[hep-ph\]\]. M. Papucci, A. Vichi and K. M. Zurek, “Monojet versus rest of the world I: t-channel Models,” arXiv:1402.2285 \[hep-ph\].
G. Busoni, A. De Simone, J. Gramling, E. Morgante and A. Riotto, “On the Validity of the Effective Field Theory for Dark Matter Searches at the LHC, Part II: Complete Analysis for the s-channel,” arXiv:1402.1275 \[hep-ph\].
S. Chang, R. Edezhath, J. Hutchinson and M. Luty, “Effective WIMPs,” Phys. Rev. D [**89**]{} (2014) 015011 \[arXiv:1307.8120 \[hep-ph\]\]. A. DiFranzo, K. I. Nagao, A. Rajaraman and T. M. P. Tait, “Simplified Models for Dark Matter Interacting with Quarks,” JHEP [**1311**]{} (2013) 014 \[arXiv:1308.2679 \[hep-ph\]\]. A. De Simone, G. F. Giudice and A. Strumia, “Benchmarks for Dark Matter Searches at the LHC,” arXiv:1402.6287 \[hep-ph\]. H. Georgi, “Effective field theory,” Ann. Rev. Nucl. Part. Sci. [**43**]{} (1993) 209. H. Georgi, “On-shell effective field theory,” Nucl. Phys. B [**361**]{} (1991) 339. M. B. Krauss, S. Morisi, W. Porod and W. Winter, “Higher Dimensional Effective Operators for Direct Dark Matter Detection,” arXiv:1312.0009 \[hep-ph\].
M. Endo and Y. Yamamoto, “Unitarity Bounds on Dark Matter Effective Interactions at LHC,” arXiv:1403.6610 \[hep-ph\].
J. Beringer [*et al.*]{} \[Particle Data Group Collaboration\], “Review of Particle Physics (RPP),” Phys. Rev. D [**86**]{} (2012) 010001. J. F. Kamenik and C. Smith, “FCNC portals to the dark sector,” JHEP [**1203**]{} (2012) 090 \[arXiv:1111.6402 \[hep-ph\]\]. K. Sigurdson, M. Doran, A. Kurylov, R. R. Caldwell and M. Kamionkowski, “Dark-matter electric and magnetic dipole moments,” Phys. Rev. D [**70**]{} (2004) 083501 \[Erratum-ibid. D [**73**]{} (2006) 089903\] \[astro-ph/0406355\]. V. Barger, W. -Y. Keung, D. Marfatia and P. -Y. Tseng, “Dipole Moment Dark Matter at the LHC,” Phys. Lett. B [**717**]{} (2012) 219 \[arXiv:1206.0640 \[hep-ph\]\].
H. Simma, “Equations of motion for effective Lagrangians and penguins in rare B decays,” Z. Phys. C [**61**]{} (1994) 67 \[hep-ph/9307274\]. B. Grzadkowski, M. Iskrzynski, M. Misiak and J. Rosiek, “Dimension-Six Terms in the Standard Model Lagrangian,” JHEP [**1010**]{} (2010) 085 \[arXiv:1008.4884 \[hep-ph\]\]. R. C. Cotta, J. L. Hewett, M. P. Le and T. G. Rizzo, “Bounds on Dark Matter Interactions with Electroweak Gauge Bosons,” Phys. Rev. D [**88**]{} (2013) 116009 \[arXiv:1210.0525 \[hep-ph\]\].
S. Banerjee, S. Mukhopadhyay and B. Mukhopadhyaya, “New Higgs interactions and recent data from the LHC and the Tevatron,” JHEP [**1210**]{} (2012) 062 \[arXiv:1207.3588 \[hep-ph\]\]. P. P. Giardino, K. Kannike, M. Raidal and A. Strumia, “Reconstructing Higgs boson properties from the LHC and Tevatron data,” JHEP [**1206**]{} (2012) 117 \[arXiv:1203.4254 \[hep-ph\]\].
U. Haisch, F. Kahlhoefer and E. Re, “QCD effects in mono-jet searches for dark matter,” arXiv:1310.4491 \[hep-ph\].
U. Haisch and F. Kahlhoefer, “On the importance of loop-induced spin-independent interactions for dark matter direct detection,” JCAP [**1304**]{} (2013) 050 \[arXiv:1302.4454 \[hep-ph\]\].
G. Belanger, F. Boudjema, A. Pukhov and A. Semenov, “Dark matter direct detection rate in a generic model with micrOMEGAs 2.2,” Comput. Phys. Commun. [**180**]{} (2009) 747 \[arXiv:0803.2360 \[hep-ph\]\].
A. Airapetian [*et al.*]{} \[HERMES Collaboration\], “Precise determination of the spin structure function g(1) of the proton, deuteron and neutron,” Phys. Rev. D [**75**]{} (2007) 012007 \[hep-ex/0609039\].
[^1]: Higher dimensional operators can contain more fields and be suppressed by phase space, or contain Higgs fields and be suppressed by $\langle H \rangle^2/\Lambda^2$, or contain derivatives and be dangerous.
[^2]: Contact interactions between dark matter and the $Z$ have been proposed in [@deSGS] as a benchmark model, assuming other contact interactions to be absent.
[^3]: For $m_\chi <m_Z/2$, the invisible width of the $Z$ (at “2$\sigma$”, so[@PDB] $\Gamma (Z\to \chi
\overline{\chi})\leq 3$ MeV) imposes that $ | C_{Z,B}| < 8.9 (\Lambda/{\rm TeV})^2$, for $B= V,A$.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Observed cusps with density profiles $\rho\propto r^{-1}$ or shallower, in the central regions of galaxies, cannot be reproduced in the standard Cold Dark Matter (CDM) picture of hierarchical clustering. Previous claims to the contrary were based on simulations with relatively few particles, and substantial softening. We present simulations with particle numbers an order of magnitude higher, and essentially no softening, and show that typical central density profiles are clearly steeper than $\rho\propto r^{-1}$. The observed shallower profiles may have formed through the smoothing effect of the spiral-in of central black holes in previous merger phases. In addition, we confirm the presence of a temperature inversion in the inner 5 kpc of massive galactic halos, and illustrate its formation as a natural result of the merging of unequal progenitors.'
author:
- |
Toshiyuki Fukushige and Junichiro Makino\
Department of General Systems Studies,\
College of Arts and Sciences, University of Tokyo,\
3-8-1 Komaba, Meguro-ku, Tokyo 153, Japan\
Email: [email protected]
title: '**On the Origin of Cusps in Dark Matter Halos**'
---
startsection [section]{}[1]{}[@]{}[24pt plus 2pt minus 2pt]{} [12pt plus 2pt minus 2pt]{}[****]{}
=================================================================================================
startsection [subsection]{}[2]{}[@]{}[12pt plus 2pt minus 2pt]{} [12pt plus 2pt minus 2pt]{}[****]{}
----------------------------------------------------------------------------------------------------
Introduction
============
Recent high resolution ground-based and Hubble Space Telescope observations (Lauer et al. 1995) have revealed that elliptical galaxies do not have a constant-density core and the densities continue to rise until the resolution limit. In faint ellipticals, the density profile at the central region increases roughly as $\rho\propto r^{-2}$, and in bright ellipticals it increases as $\rho
\propto r^{-1}$ or shallower (Merritt and Friedman 1996).
Several simulations demonstrated that the dark matter halos formed through hierarchical clustering were not well approximated by isothermal spheres, and better fitted by a Hernquist model (Hernquist 1990) or similar models with $\rho \sim 1/r$ at the center (Dubinski and Carlberg 1991; Navarro, Frenk, and White 1996a, 1996b). In these simulations, halos have a power law index which changes from around $-1$ to $-3$ or $-4$ as the radius increases. Recently, Moore (1994) and Flores and Primack (1995) have argued that the gently rising $\rho(r)$ curves of dwarf galaxies are inconsistent with a $r^{-1}$ central cusp, and therefore challenge the hypothesis of CDM halos.
However, whether the results of numerical simulations are physically valid or are numerical artifacts remains unclear for the following reasons. First, no physical mechanism has been presented so far for the formation of cusps found in simulations (White 1996). Secondly, since the mass resolution in these simulation was rather low, the central structure may have been affected by two-body relaxation effects (Quinlan 1996, Steinmetz and White 1996, Fukushige and Makino 1996). Thirdly, the central structures of halos formed in these simulations are strongly affected by the potential softening used, and show rapid change in the power index of $\rho(r)$ around a radius not much larger than the potential softening length.
In this letter, we have been able to separate physically real effects from numerical artifacts, through $N$-body simulations of hierarchical clustering with a resolution far higher than those in previous work. The number of particles we used ($N=786,400$) is more than 10 times larger than those used in previous simulations ($N\sim
10,000-30,000$), and our softening is significantly smaller (§2).
Our results show that dark halos formed through hierarchical clustering typically exhibit an inwardly decreasing temperature structure in their inner regions, with a density cusp shallower than $\rho\propto r^{-2}$, but steeper than $\rho\propto r^{-1}$ (§3). We illustrate the origin of this behavior through an idealized merger simulation of different types of galaxies (§4), which enables us to give a physical explanation of temperature inversion (§5).
Method
======
Initial conditions were constructed following Dubinski and Carlberg (1991). We assigned initial positions and velocities to particles in a spherical region with a radius of 2Mpc surrounding a density peak selected from a discrete realization of a standard CDM model ($H_o = 50$km/s/Mpc and $\Omega = 1$). The peak was chosen from a realization of the density contrast in an 8Mpc box, using COBE normalization. The peaks were found by smoothing the density with a Gaussian filter of radius 0.75 Mpc.
We followed the evolution of a density peak through direct $N$-body simulation. We added the local Hubble flow and integrated the orbits directly in physical space, with $N=786,400$, an individual particle mass of $4.0 \times 10^6M_{\odot}$, and a Plummer softened potential with a length of $\varepsilon=0.14$kpc. We started the simulation at $z\sim 46$. We did not include tidal effects from outside our 2Mpc sphere, since we are mainly interested in the core properties within 10kpc, where tidal effects from outside 2Mpc scale are negligible.
We used a 4-th order Hermite integration scheme (Makino and Aarseth 1992) with individual (hierarchical) timestep algorithm (McMillan 1986, Makino 1991a).
For the force calculation, we used the GRAPE-4 (Taiji et al. 1996), a special-purpose computer designed to accelerate $N$-body simulations using a Hermite integrator and a hierarchical timestep algorithm. The total system consists of 1692 pipeline processor chips dedicated to gravity calculation and has a peak performance of 1 Tflops. The calculation with $N=786,400$ took about 180 CPU hours ($5.7\times
10^9$ particle steps) using 3/4 of the total system. The sustained speed of computation was 406 Gflops. Since the force calculations on GRAPE-4 are effectively of double-precision accuracy (Makino et al. 1996), our simulations exhibit much higher accuracy both for force calculations and orbit integrations than previous simulations.
Hierarchical Clustering Simulations
===================================
Figure 1 shows the particle distributions in our simulation at the redshift $z=$(a)$8.7$, (b)$5.1$.and (c)$1.8$. Figure 2 shows the density and temperature structure of the halo at $z=1.8$, well after most of the mergings has already taken place (results at $z=0$ are similar, but these earlier stages shows even less effects of two-body relaxation). In figure 2 we can see a clear temperature decrease toward the center within 5kpc. This non-isothermal structure produces a density cusp shallower than $\rho\propto r^{-2}$. In our simulation, the structure outside $1$ kpc is not affected by two-body relaxation after the formation of a large single halo at $z\sim 3$; only the central region within $1$ kpc shows some expansion because of two-body relaxation effects.
The large potential softening employed in previous studies has produced spurious structures, which we have been able to reproduce, using additional simulations with larger softening, leading to a central temperature decrease within a region a few times larger than the softening radius. The potential softening tends to produce a flat core on the scale of the softening length. Thus, unless we use point-mass particles, we always see a tendency for the power index of the density profile to approach zero at a radius comparable to the softening radius.
After most of the merging has taken place, subsequent two-body relaxation effects due to small $N$ continue to lead to spurious changes in the central region. For example, a density cusp might evolve to a flat core through gravothermal expansion (Quinlan 1996, Fukushige and Makino 1996). In our simulation, the local two-body relaxation times at 1 and 5 kpc are $1.2\times 10^{10}$ and $2.3\times
10^{11}$ years, respectively. Therefore, only the structure within 1kpc is somewhat affected by two-body relaxation in our simulation. In simulations with $N\sim 10,000$ such as reported by Dubinski and Carlberg (1991) and Navarro et al. (1996a, 1996b), the local relaxation time at 5kpc was $\sim 3.0\times 10^9$ years, implying that the effect of two-body relaxation completely dominated their results at that distance.
Idealized Merger Simulations
============================
In order to illustrate the formation of temperature inversion after merging, we performed an idealized simulation in which we merged two equal-mass spherical halos with different central densities, a Plummer model and a King model with central potential $W_0=9$. The result is shown in figure 3. The central temperature inversion is striking.
The encounter between the halos is head-on, starting from rest at an initial separation of $R=5$ (in standard units with $M=G=-4E=1$ where $G$ is the gravitational constant, and $M$ and $E$ are the total mass and the initial total energy of a halo; cf. Heggie and Mathieu 1986). The total number of particles is $N=262,144$ and the softening length is 1/256. We used Barnes-Hut tree code on GRAPE-4 (Makino 1991b) for the force calculations. The timestep is shared and constant ($\Delta
t=1/512$).
In hierarchical clustering, temperature inversion takes place in the inner halos in a similar way. Figure 4 shows the evolution of temperature structure for the simulation of halo formation presented in section 3. We can see that the temperature of the outer halo regions increases faster than the temperature of the central halo regions.
Discussion
==========
In this letter, we have shown that dark halos formed through hierarchical clustering have a central density cusp shallower than $\rho\propto r^{-2}$, and a velocity dispersion that has a local minimum in the center, peaks at a distance of $5\sim10$ kpc from the center, and then drops again in the outer halo. We offer the following interpretations, for the temperature and the density structures found.
The occurrence of the striking temperature inversion can be understood as follows. In bottom-up structure formation, as in CDM hierarchical clustering, a typical halo is formed through repeated merging of smaller subclumps. Each time this happens, the less concentrated clump tends to be disrupted by the tidal field of the more centrally concentrated clump. The dense core of the latter survives the merging process more or less intact, and settles down at the center of merger remnant, with locally unchanged temperature. In contrast, the temperature of the merger remnant as a whole increases since the specific binding energy of the merger remnant is almost always larger than that of its progenitors.
As a result, present-day halos tend to carry some memory of the temperature of their densest progenitor clump, which has set the temperature scale in the inner, and densest, regions. Since the bulk of the halo is affected more by subsequent mixing and heating, a central temperature inversion is naturally created. This formation process is evident, not only in our hierarchical clustering simulations, but also in our idealized simulation of a merger between two different halos, one with dense inner region, and another one with a much flatter core.
The occurrence of steep density profiles, significantly steeper than $\rho\propto r^{-1}$, naturally follows from the same physical picture. According to our detailed $N$-body calculations, in a purely CDM scenario, each final halo forms a repository for the small inner cores of the subclumps that have made up the final dark matter aggregate.
Comparing our results with observations, we conclude that the shallow cusps of large ellipticals cannot have formed through the dissipationless processes that we have modeled here. Instead, we interpret the presence of density profiles shallower than $\rho\propto
r^{-1}$ as strong evidence for the existence of central massive black holes, through the following mechanism. As an aftereffect of the merger of two black-hole containing galaxies, the spiral-in of those two massive black holes tends to smear out the central cusp that would have otherwise formed, as shown in detailed calculations of hierarchical merging of galaxies by Makino and Ebisuzaki (1996).
Acknowledgement {#acknowledgement .unnumbered}
===============
We are grateful to Makoto Taiji, for his important contribution in developing the GRAPE-4, and to the referee, Piet Hut, for many helpful comments on the manuscript. T.F. acknowledges financial support by JSPS. To generate initial condition, we used the COSMIC package developed by Edmund Bertschinger, to whom we express our thanks. This research was partially supported by the Grand-in-Aid for Specially Promoted Research (04102002) of The Ministry of Education, Science and Culture.
References {#references .unnumbered}
==========
Dubinski, J., end Carlberg, R., 1991, ApJ, 378, 496
Fukushige, T., and Makino, J., 1996, in preparation.
Heggie, D. C., and Mathieu, R. D., 1986, in The Use of Supercomputer in Stellar Dynamics (ed. Hut, P., McMillan, S.) 233, (Springer-Verlag, Berlin)
Hernquist, L., 1990, ApJ, 356, 359
Lauer, T. R. et al., 1995, AJ, 110, 2622
Makino J. 1991a, PASJ, 43, 859
Makino J. 1991b, PASJ, 43, 621
Makino J. and Aarseth, S. J. 1992, PASJ, 44, 141
Makino, J., and Ebisuzaki, T., 1996, ApJ, 465, 527
Makino, J., Taiji, M, Ebisuzaki, T. and Sugimoto, D., 1996, submitted to ApJ.
McMillan, S. L. W., 1986, in The Use of Supercomputers in Stellar Dynamics, eds. S. L. W. McMillan, P. Hut (Springer, Berlin) p 156
Merritt, D., and Fridman, T., 1996, in Fresh Views of Elliptical Galaxies, eds. Buzzoni, A., Renzini, A. Serrano, A. (ASP conference Series Vol. 86)
Moore, B., 1994, Nature, 370, 629
Navarro, J. F., Frenk, C. S., and White, S. D. M., 1996a, ApJ, 462, 563
Navarro, J. F., Frenk, C. S., and White, S. D. M., 1996b, preprint (astro-ph/9611107)
Quinlan, G, D., 1996, preprint (astro-ph/9606182)
Steinmetz, M. H., and White, S. D. M., 1996, preprint (astro-ph/9609021)
Taiji, M., Makino, J., Fukushige, T., Ebisuzaki, T., and Sugimoto, D., 1996, in IAU Symposium 174, eds. Makino, J., Hut, P. (Kluwer Academic Press)
White, S. D. M. 1996, in Gravitational Dynamics, ed. Lahav, O., Terlevich, E., Terlevich, R. J., 121 (Cambridge University Press)
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'A conjecture by Aharoni and Berger states that every family of $n$ matchings of size $n+1$ in a bipartite multigraph contains a rainbow matching of size $n$. In this paper we prove that matching sizes of $\left(\frac 3 2 + o(1)\right) n$ suffice to guarantee such a rainbow matching, which is asymptotically the same bound as the best known one in case we only aim to find a rainbow matching of size $n-1$. This improves previous results by Aharoni, Charbit and Howard, and Kotlar and Ziv.'
author:
- Dennis Clemens
- Julia Ehrenmüller
bibliography:
- 'rainbowmatchings.bib'
title: An improved bound on the sizes of matchings guaranteeing a rainbow matching
---
=1
Introduction
============
In this paper we are concerned with the question which sizes of $n$ matchings in a bipartite multigraph suffice in order to guarantee a rainbow matching of size $n$.
One motivation for considering these kinds of problems is due to some well known conjectures on Latin squares. A *Latin square* of order $n$ is an $n \times n$ matrix in which each symbol appears exactly once in every row and exactly once in every column. A *partial transversal* in a Latin square is a set of entries with distinct symbols such that from each row and each column at most one entry is contained in this set. We call a partial transversal of size $n$ in a Latin square of order $n$ simply *transversal*. A famous conjecture of Ryser [@ryser1967] states that for every odd integer $n$ any Latin square of order $n$ contains a transversal. The conjecture is known to be true for $n \leq 9$. Omitting the restriction to odd numbers yields a false statement. Brualdi [@brualdi1991; @denes1974] and Stein [@stein1975] independently formulated the following conjecture for all orders $n$.
\[conj:brualdi\] For every $n \geq 1$ any Latin square of order $n$ has a partial transversal of size $n-1$.
A natural way to transfer this problem to graphs is the following. Let $L=(\ell_{i,j})_{i,j\in[n]}$ be a Latin square of order $n$. We define $G_{L}:=(A\cup B, E)$ as the complete bipartite edge-coloured graph with partite sets $A=\{a_1,\ldots,a_n\}$ and $B=\{b_1,\ldots,b_n\}$, where $a_ib_j$ is coloured $\ell_{i,j}$. That is, $A$ and $B$ represent the columns and rows of $L$, respectively. Moreover, a transversal of $L$ corresponds to a perfect matching in $G_L$ that uses each edge colour exactly once, which we call a *rainbow matching* of size $n$. Using this notion, Conjecture \[conj:brualdi\] is equivalent to the following: For every $n\geq 1$ any complete bipartite edge-coloured graph, the colour classes of which are perfect matchings, contains a rainbow matching of size $n-1$. One may wonder whether this might even be true in the more general setting of bipartite edge-coloured multigraphs.
Following Aharoni, Charbit and Howard [@aharoni2015], we define $f(n)$ to be the smallest integer $m$ such that every bipartite edge-coloured multigraph with exactly $n$ colour classes, each being a matching of size at least $m$, contains a rainbow matching of size $n$. Aharoni and Berger [@aharoni2009] conjectured the following generalization of Conjecture \[conj:brualdi\].
\[conj:aharoni\] For every $n\geq 1$ we have $f(n) = n+1$.
The first approaches towards this conjecture are given by the bounds $f(n)\leq \left\lfloor \frac 7 4 n\right\rfloor$ due to Aharoni, Charbit and Howard [@aharoni2015] and $f(n) \leq \left\lfloor \frac 5 3 n\right\rfloor$ due to Kotlar and Ziv [@kotlar2014]. Here, we give an improved bound, which is asymptotically the same as the best known bound on the sizes of the colour classes in case we aim to find a rainbow matchings of size $n-1$ [@kotlar2014]. In particular, we prove the following.
\[thm:main\] For every ${\varepsilon}>0$ there exists an integer $n_0\geq 1$ such that for every $n \geq n_0$ we have $f(n)\leq \left(\frac{3}{2} + {\varepsilon}\right)n$.
Subsequently, we use the following notation. Let $G$ be a bipartite multigraph with partite sets $A$ and $B$ and let $R$ be a matching in $G$. For a set $X \subseteq A$ we denote by $N_G(X|R):= \{y \in B: \exists xy\in R \text{ with } x \in X\}$ the neighbourhood of $X$ with respect to $R$. For the sake of readability, we omit floor and ceiling signs and do not intend to optimize constants in the proofs.
Proof of Theorem \[thm:main\]
=============================
In this section we give a proof of Theorem \[thm:main\] the idea of which can be summarized as follows. We start with assuming for a contradiction that a maximum rainbow matching in the given graph $G=(A\cup B,E)$ is of size $n-1$. A rainbow matching of this size is known to exist [@kotlar2014]. We fix such a matching $R$ and find two sequence $e_1,\ldots,e_k$ and $g_1,\ldots,g_k$ of edges, the first consisting of edges from $R$ and the second consisting of edges outside $R$. We then show that either we can switch between some of the edges from the edge sequences to produce a rainbow matching of size $n$ (see the proofs of the Claims \[claim1\], \[claim2\] and \[claim3\]), or the matchings represented by the edges $e_1,\ldots,e_k$ need to touch at least $n$ vertices in $B$ that are saturated by $R$, both leading to a contradiction. To make the second case more precise we additionally introduce in the proof certain sequences $X_1,\ldots,X_k\subseteq A$ and $Y_1,\ldots,Y_k\subseteq B$.
Let ${\varepsilon}>0$ be given and whenever necessary we may assume that $n$ is large enough. Let ${{\mathcal F}}=\{F_0,\ F_1,\ \ldots,\ F_{n-1}\}$ be a family of $n$ matchings of size at least $(3/2 +\varepsilon)n$ in a bipartite multigraph $G=(A\cup B,E)$ with partite sets $A$ and $B$. We aim to find a rainbow matching of size $n$.
For a contradiction, let us assume that there is no such matching. As shown in [@kotlar2014], there must exist a rainbow matching $R$ of size $n-1$. We may assume without loss of generality that none of the edges of $F_0$ appears in $R$. Let $t$ be the smallest positive integer with $1/(2t-1)\leq \varepsilon$. Moreover, let $X\subseteq A$ and $Y\subseteq B$ be the sets of vertices that are saturated by $R$, i.e. incident with some edge of $R$.
In the following we show that for every $k \in [t]$ we can construct sequences
1. $e_1,\ldots,e_k$ of $k$ distinct edges $e_i=x_iy_i$ in $R$ with $x_i\in X$ and $y_i\in Y$,
2. $g_1,\ldots,g_k$ of $k$ distinct edges $g_i=z_iy_i$ with $z_i\in A\setminus X$,
3. $ X_1, \ldots, X_k$ of subsets of $X$,
4. $ Y_1, \ldots, Y_k$ of subsets of $Y$,
and an injective function $\pi: \{0,1,\ldots,k\} \rightarrow \{0,1,\ldots,n-1\}$ with $\pi(0):=0$ such that the following properties hold:
1. \[color\_e\] for each $ i\in [k]$ we have $e_i\in F_{\pi(i)}$,
2. \[color\_g\] for each $ i\in [k]$ we have $g_i\in \bigcup_{j=0}^{i-1} F_{\pi(j)}$,
3. \[disjoint\] $(e_1\cup\ldots\cup e_k) \cap (X_k\cup Y_k)=\varnothing$,
4. \[sizes\] $|X_k|=|Y_k|=s_k:=2k\varepsilon n + k(7-3k)/2$,
5. \[many\_colors\] for each $ i\in [k]$ and each $j\in\{0,\ldots,n-1\}$ it holds that if $R$ contains an edge of the matching $F_j$ between $X_i$ and $Y_i$, then there is also an edge of $F_j$ between $x_i$ and $B\setminus Y$,
6. \[good\_edges\] for each $i\in [k]$ and each $w\in Y_i\setminus Y_{i-1}$ there exists a vertex $v\in A\setminus (X\cup \{z_1,\ldots,z_{i-1}\})$ such that $vw\in F_{\pi(i-1)}$ (where $Y_0 := \varnothing$), and
7. \[different\_endpoints\] for each $i\in [k]$ and each $j\in [i-1]$ it holds that if $g_i\in F_{\pi(j)}$, then $z_i\in A\setminus (X\cup \{z_1,\ldots,z_j\})$.
Before we start with the construction, let us first observe that by Property \[sizes\] we have a set $Y_t\subseteq Y$ which satisfies $2t\varepsilon n + t(7-3t)/2= |Y_t| \leq |Y| <n$. However, for large enough $n$ and by the choice of $t$ we have that $2t\varepsilon n + t(7-3t)/2 > n$, a contradiction.
In order to find the sequences described above, we proceed by induction on $k$. For the base case, let us argue why we find edges $e_1$, $g_1$, sets $X_1$, $Y_1$, and an injective function $\pi$ with Properties \[color\_e\]-\[different\_endpoints\]. First observe that $F_0$ does not have any edges between $A\setminus X$ and $B \setminus Y$, by assumption on $R$. As $|F_0| \geq(3/2+{\varepsilon})n$, there are at least $(1/2+{\varepsilon})n+1$ edges of $F_0$ between $A\setminus X$ and $Y$. Let $N_0 \subseteq Y$ denote a set of size $(1/2+{\varepsilon})n+1$ such that for every vertex $w \in N_0$ there exists a vertex $v\in A\setminus X$ such that $vw \in F_0$. Furthermore, let $X_1':= N_G(N_0|R)$ and let ${{\mathcal R}}_1 := \{F_j \in {{\mathcal F}}: F_j \cap R[N_0, X_1']\neq \varnothing\}$.
Let $F$ be any matching in ${{\mathcal R}}_1$, let $vw$ be the unique edge in $R[N_0, X_1'] \cap F$ and let $z \in A\setminus X$ be the unique vertex such that $zw \in F_0$. Notice that there cannot be any edge $g$ of $F$ between $A\setminus (X \cup \{z\})$ and $B\setminus Y$, since otherwise $(R\setminus \{vw\})\cup \{zw,g\}$ would give a rainbow matching of size $n$, in contradiction with $R$ being a maximum rainbow matching. Therefore, there are at least $(1/2+{\varepsilon})n + 1$ edges of $F$ between $B \setminus Y$ and $X \cup \{z\}$. Since $|X_1'| = (1/2+{\varepsilon})n+1$, there are at least $2{\varepsilon}n +2$ edges of $F$ between $B\setminus Y$ and $X_1'$. Since this is true for any $F \in {{\mathcal R}}_1$, we know by the pigeonhole principle that there is a vertex $x_1 \in X_1'$ and a subset $X_1 \subseteq X_1'$ of size $2{\varepsilon}n +2$ such that, for every $F_j\in {{\mathcal F}}$, if $F_j \cap R[X_1, N_G(X_1|R)]\neq \varnothing$ then $F_j$ has an edge between $x_1$ and $B\setminus Y$. Note that $x_1\notin X_1$. Let $e_1= x_1y_1$ be the unique edge in $R$ incident with $x_1$ and let $g_1 = z_1y_1$ be the unique edge of $F_0$ incident with $y_1 \in N_0$. Set $\pi(1)$ to the unique index $j\in [k]$ such that $e_1\in F_j$. One can easily verify that $e_1 = x_1y_1$, $g_1 = z_1y_1$, $X_1$, $Y_1 := N_G(X_1|R)$, and $\pi$ satisfy Properties \[color\_e\]-\[different\_endpoints\].
For the induction hypothesis let us assume that for some $k \in [t-1]$ the above sequences are given with Properties \[color\_e\]-\[different\_endpoints\]. We now aim to extend these by edges $e_{k+1}, g_{k+1}$, sets $X_{k+1}, Y_{k+1}$, and a value $\pi(k+1)$ while maintaining Properties \[color\_e\]-\[different\_endpoints\]. We start with some useful claims.
\[claim1\] $F_{\pi(k)}$ has no edge between $A\setminus (X\cup \{z_1,\ldots,z_k\})$ and $B\setminus Y$.
Assume for a contradiction that there exists an edge $g\in F_{\pi(k)}$ between the sets $A\setminus (X\cup \{z_1,\ldots,z_k\})$ and $B\setminus Y$. (See Figure \[fig:cl1\] for an illustration.) By Property \[color\_g\] we find a sequence $k>j_1>j_2>\ldots>j_s=0$ with $1\leq s\leq k$ such that $$\begin{aligned}
g_k & \in F_{\pi(j_1)}\ , \\
g_{j_i} & \in
F_{\pi(j_{i+1})} \text{\ \ for }i< s.\end{aligned}$$ Moreover, according to Property \[different\_endpoints\] we know that $z_k,z_{j_1},\ldots,z_{j_{s-1}}$ are distinct, and thus, also using Property \[color\_e\], we conclude that $$(R\setminus \{e_k,e_{j_1},\ldots,e_{j_{s-1}}\})\cup \{g_k,g_{j_1},\ldots,g_{j_{s-1}},g\}$$ forms a rainbow matching which is larger than $R$, a contradiction.
\[htbp\] \[fig:cl1\]
\[claim2\] $F_{\pi(k)}$ has no edge between $A\setminus (X\cup \{z_1,\ldots,z_k\})$ and $Y_k$.
Assume for a contradiction that there is an edge $g\in F_{\pi(k)}$ between the sets $A\setminus (X\cup \{z_1,\ldots,z_k\})$ and $Y_k$. (See Figure \[fig:cl2\] for an illustration.) Let $e$ be the unique edge in $R$ which is adjacent to $g$. Observe that $e$ lies between $X_k$ and $Y_k$ by assumption. Let $j\in [n-1]$ be such that $e\in F_j$. By Property \[disjoint\] we have $e\notin \{e_1,\ldots,e_k\}$. Thus, using Property \[color\_e\] and the fact that $R$ is a rainbow matching, we can conclude that $j\notin \{\pi(i):1\leq i\leq k\}$. Now, by Property \[many\_colors\] it holds that there is an edge $\overline{e}\in F_j$ between $x_k$ and $B\setminus Y$. Moreover, by Properties \[color\_g\] and \[different\_endpoints\], we find a sequence $k>j_1>j_2>\ldots>j_s=0$ with $1\leq s\leq k$ such that $$\begin{aligned}
g_k & \in F_{\pi(j_1)}\ , \\
g_{j_i} & \in
F_{\pi(j_{i+1})} \text{\ \ for }i< s\end{aligned}$$ and all vertices $z_k,z_{j_1},\ldots,z_{j_{s-1}}$ are distinct. Therefore, using Property \[color\_e\], we conclude that $$(R\setminus \{e_k,e_{j_1},\ldots,e_{j_{s-1}},e\})\cup \{g_k,g_{j_1},\ldots,g_{j_{s-1}},\overline{e},g\}$$ forms a rainbow matching which is larger than $R$, a contradiction.
\[htbp\] \[fig:cl2\]
\[existence\_Nk\] The matching $F_{\pi(k)}$ has at least $\left(\frac{1}{2}+\varepsilon\right)n+1-2k$ edges between $A\setminus (X\cup \{z_1,\ldots,z_k\})$ and $Y\setminus (Y_k\cup \{y_1,\ldots,y_k\})$.
As $|F_{\pi(k)}|\geq (3/2+\varepsilon)n$ and $|X\cup \{z_1,\ldots,z_k\}|\leq n-1+k$, we conclude that at least $(1/2+\varepsilon)n+1-k$ edges of $F_{\pi(k)}$ are incident with vertices in $A\setminus (X\cup \{z_1,\ldots,z_k\})$. Each of these edges intersects $Y\setminus Y_k$ by the previous claims and thus the statement follows.
In the following, let $N_k\subseteq Y\setminus (Y_k\cup \{y_1,\ldots,y_k\})$ be a set of size $1/2+\varepsilon)n+1-2k$ such that for each vertex $w\in N_k$ there is a vertex $v\in A\setminus (X\cup \{z_1,\ldots,z_k\})$ with $vw\in F_{\pi(k)}$. Such a set exists by the previous corollary. Moreover, let $$Y_{k+1}':=Y_k\cup N_k$$ and let $X_{k+1}':= N_G(Y_{k+1}'|R)$ be the neighbourhood of $Y_{k+1}'$ with respect to $R$. By Property \[sizes\], and as $N_k\cap Y_k=\varnothing$, we obtain $$\begin{aligned}
|X_{k+1}'|=|Y_{k+1}'|&=2k\varepsilon n + \frac{k(7-3k)}{2} + \left(\frac{1}{2}+\varepsilon\right)n+1-2k \nonumber \\
& = \frac{1}{2}n + (2k+1)\varepsilon n + \frac{-3k^2+3k+2}{2}\ . \label{sizeXY}\tag{$\ast$}\end{aligned}$$
We now look at all matchings that have an edge in $R$ between $X_{k+1}'$ and $Y_{k+1}'$. Formally, we consider $${{\mathcal R}}_{k+1}:=\big\{F_j\in {{\mathcal F}}: F_j\cap R[X_{k+1}',Y_{k+1}']\neq \varnothing\big\}\ .$$
\[claim3\] Every $F_j\in {{\mathcal R}}_{k+1}$ has at least $s_{k+1}$ edges between $X_{k+1}'$ and $B\setminus Y$.
The main argument is similar to that of Claim \[claim1\] - Corollary \[existence\_Nk\]. For $F_j\in {{\mathcal R}}_{k+1}$ let $f=vw$, with $v\in X_{k+1}',\ w\in Y_{k+1}'$, denote the unique edge in $F_j\cap R[X_{k+1}',Y_{k+1}']$. Since $Y_{k+1}':=Y_k\cup N_k$, we either have $w\in Y_k$ or $w\in N_k$. In particular, by Property \[disjoint\] from the hypothesis and by the definition of $N_k$, we know that $w\notin \{y_1,\ldots,y_k\}$, and therefore $j\notin \{\pi(i): 0\leq i\leq k\}$.
If $w\in Y_k$, then we find an integer $j_1\in [k]$ such that $w\in Y_{j_1}\setminus Y_{j_1-1}$ since $Y_k=\bigcup_{i\in [k]} Y_i\setminus Y_{i-1}$, and by Property \[good\_edges\] there is a vertex $z\in A\setminus (X\cup \{z_1,\ldots,z_{j_1-1}\})$ such that $zw\in F_{\pi(j_1-1)}$.\
If otherwise $w\in N_k$, we find a vertex $z\in A\setminus (X\cup \{z_1,\ldots,z_k\})$ such that $zw\in F_{\pi(k)}$, by construction of $N_k$. In either case, let us fix this particular vertex $z$. We now prove the claim by showing first that (i) $F_j$ has no edge between $A\setminus (X\cup \{z_1,\ldots,z_k,z\})$ and $B\setminus Y$, and then we conclude that (ii) the statement holds for $F_j$.
We start with the discussion of (i). So, assume that $F_j$ has an edge $\overline{f}$ between $A\setminus (X\cup \{z_1,\ldots,z_k,z\})$ and $B\setminus Y$.
If $w\in Y_k$, then by the definition of $z$ we have $zw\in F_{\pi(j_1-1)}$, with $j_1$ being defined above. We can assume that $j_1> 1$, as otherwise $zw\in F_0$ and thus $(R\setminus \{f\})\cup \{\overline{f},zw\}$ forms a full rainbow matching, in contradiction to our main assumption. But then, using Property \[color\_g\], we find a sequence $j_1-1>j_2>\ldots >j_s=0$ with $2\leq s<k$ such that $$\begin{aligned}
g_{j_1-1} & \in F_{\pi(j_2)}\ , \\
g_{j_i} & \in F_{\pi(j_{i+1})} \text{\ \ for }2\leq i\leq s-1\end{aligned}$$ and, by Property \[different\_endpoints\] and since $z\in A\setminus (X\cup \{z_1,\ldots,z_{j_1-1}\})$, all the vertices $z,z_{j_1-1},z_{j_2},\ldots,z_{j_{s-1}}$ are distinct. We thus find the rainbow matching $$(R\setminus \{e_{j_1-1},e_{j_2},\ldots,e_{j_{s-1}},f\})\cup \{g_{j_1-1},g_{j_2},\ldots,g_{j_{s-1}},\overline{f},zw\}$$ which is larger than $R$, a contradiction.
\[htbp\] \[fig:cl3a\]
If otherwise $w\in N_k$, then $zw\in F_{\pi(k)}$. Analogously we find a sequence $k>j_1>j_2>\ldots>j_s=0$ with $1\leq s\leq k$ such that $g_{k} \in F_{\pi(j_1)}$ and $g_{j_i} \in F_{\pi(j_{i+1})}$ for $i<s$, and we obtain a contradiction as $$(R\setminus \{e_{k},e_{j_1},\ldots,e_{j_s},f\})\cup \{g_{k},g_{j_1},\ldots,g_{j_s},\overline{f},zw\}$$ forms a rainbow matching which is larger than $R$. Thus, we are done with part (i).
Let us proceed with (ii): $F_j$ needs to saturate at least $(1/2 + \varepsilon )n+1$ vertices of $B\setminus Y$, as $|F_j|\geq (3/2 + \varepsilon )n$ and $|Y|\leq n-1$. Thus, by part (i), we have at least $(1/2 + \varepsilon )n+1$ edges of $F_j$ between $X\cup \{z_1,\ldots,z_k,z\}$ and $B\setminus Y$. Using (\[sizeXY\]), we further calculate that $$\begin{aligned}
|X\cup \{z_1,\ldots,z_k,z\}|-|X_{k+1}'| & \leq (n+k) - \left(\frac{1}{2}n + (2k+1)\varepsilon n + \frac{-3k^2+3k+2}{2} \right)\\
& = \frac{1}{2}n - (2k+1)\varepsilon n + \frac{3k^2-k-2}{2} \ .\end{aligned}$$ Thus, the number of edges in $F_j$ between $X_{k+1}'$ and $B\setminus Y$ needs to be at least $$\begin{aligned}
\left( \frac{1}{2} + \varepsilon \right)n+1 - \left( \frac{1}{2}n - (2k+1)\varepsilon n + \frac{3k^2-k-2}{2} \right) = s_{k+1}\ ,
$$ as claimed.
We now proceed with the construction of the edges $e_{k+1}, g_{k+1}$ and the sets $X_{k+1}, Y_{k+1}$, and afterwards we show that all required properties are maintained. The next corollary is by the pigeonhole principle an immediate consequence of Claim \[claim3\].
\[sequence\] There exists a vertex $x_{k+1}\in X_{k+1}'$, a set $X_{k+1}\subseteq X_{k+1}'$ of size $s_{k+1}$ and its neighborhood $Y_{k+1}\subseteq Y_{k+1}'$ with respect to $R$ such that the following holds for every $j\in[n-1]$: If $F_j\cap R[X_{k+1},Y_{k+1}]\neq \varnothing$, then $F_j$ has an edge between $x_{k+1}$ and $B\setminus Y$.
To extend the sequences, choose $X_{k+1}$ and $Y_{k+1}$ according to Corollary \[sequence\], and let $e_{k+1}=x_{k+1}y_{k+1}$ be the unique edge in $R$ that is incident with $x_{k+1}$. Note that $x_{k+1}\notin X_{k+1}$, as otherwise $x_{k+1}$ would need to be incident to two edges of the same matching $F_j$.
Observe that $y_{k+1}\notin \{y_1,\ldots,y_k\}$. Indeed, $y_{k+1}\in Y_{k+1}'=Y_k\cup N_k$, and by construction we have $N_k\cap \{y_1,\ldots,y_k\}=\varnothing$, while $Y_k\cap \{y_1,\ldots,y_k\}=\varnothing$ holds by Property \[disjoint\].
Now, let $e_{k+1}\in F_j$. As $e_{k+1}\in R\setminus \{e_1,\ldots,e_k\}$, we have $j\notin \{\pi(i):\ 0\leq i\leq k\}$. We extend the injective function $\pi$ with $\pi(k+1)=j$.
Finally, we choose $g_{k+1}$ as follows: If $y_{k+1}\in N_k$, then by construction of $N_k$ there is a vertex $z_{k+1}\in A\setminus (X\cup \{z_1,\ldots,z_k\})$ with $z_{k+1}y_{k+1}\in F_{\pi(k)}$. Otherwise, if $y_{k+1}\in Y_k$, then there is an $i\in [k]$ with $y_{k+1}\in Y_i\setminus Y_{i-1}$, and by Property \[good\_edges\] there is a vertex $z_{k+1}\in A\setminus (X\cup \{z_1,\ldots, z_{i-1}\})$ such that $z_{k+1}y_{k+1}\in F_{\pi(i-1)}$. In any case, we set $g_{k+1}:=z_{k+1}y_{k+1}$.
\[claim:properties\] The extended sequences satisfy Properties \[color\_e\]-\[different\_endpoints\].
Properties \[color\_e\] and \[color\_g\] follow immediately from the induction hypothesis and from the definition of $\pi(k+1)$ and $g_{k+1}$. By construction, we have $Y_{k+1}\subseteq Y_{k+1}' = Y_k \cup N_k$. By Property \[disjoint\] of the induction hypothesis and by the definition of $N_k$, we have $\{y_1,\ldots,y_k\} \cap Y_{k+1} = \varnothing$. It follows from the construction of $X_{k+1}$ (Corollary \[sequence\]) that $y_{k+1} \notin Y_{k+1}$. By symmetry, we have $\{e_1, \ldots, e_{k+1}\}\cap(X_{k+1}\cup Y_{k+1}) = \varnothing$, which shows Property \[disjoint\]. Properties \[sizes\] and \[many\_colors\] hold by Corollary \[sequence\] and by Property \[many\_colors\] of the induction hypothesis. Recall that $Y_{k+1} \setminus Y_{k} \subseteq N_k$. This means that for every $w\in Y_{k+1}\setminus Y_k$ there exists a vertex $v \in A\setminus (X\cup \{z_1,\ldots, z_{k}\})$ such that $vw \in F_{\pi(k)}$, proving Property \[good\_edges\]. Finally, Property \[different\_endpoints\] holds by the induction hypothesis and since we chose $z_{k+1}$ from a set $A\setminus (X\cup \{z_1,\ldots, z_{i-1}\})$ such that $z_{k+1}y_{k+1}\in F_{\pi(i-1)}$ for the appropriate $i \in [k+1]$. Consequently, all Properties \[color\_e\]-\[different\_endpoints\] are fulfilled by the extended sequences.
Claim \[claim:properties\] concludes the induction and thus the proof of Theorem \[thm:main\].
Open problems and concluding remarks
====================================
In this paper we proved that a collection of $n$ matchings of size $\left(3/2 + o(1)\right)n$ in a bipartite multigraph guarantees a rainbow matching of size $n$. One of the obstacles why our proof does not work for smaller values is that it is not clear what matching sizes are sufficient for guaranteeing a rainbow matching of size $n-1$. More generally, as suggested by Tibor Szabó (private communication), it would be interesting to determine upper bounds on the smallest integer $\mu(n,\ell)$ such that every family of $n$ matchings of size $\mu(n,\ell)$ in a bipartite multigraph guarantees a rainbow matching of size $n-\ell$. One can verify that $\mu(n,l) \leq \frac{l+2}{l+1} n$. Moreover, it holds that $\mu(n, \sqrt{n})\leq n$, which is a generalization (see e.g. [@aharoni2013]) of a result proved in the context of Latin squares by Woolbright [@woolbright], and independently by Brouwer, de Vries and Wieringa [@brouwer1978]. In order to approach Conjecture \[conj:aharoni\], one can also increase the number of matchings and fix their sizes to be equal to $n$ instead of considering families of $n$ matchings of sizes greater than $n$. Drisko [@drisko1998] proved that a collection of $2n-1$ matchings of size $n$ in a bipartite multigraph with partite sets of size $n$ guarantees a rainbow mathching of size $n$. He also showed that this result is sharp. This problem can be further investigated in the following two directions. Does the statement also hold if we omit the restriction on the sizes of the vertex classes? And how many matchings do we need to find a rainbow matching of size $n-\ell$ for every $\ell \geq 1$?
Finally, in case Conjecture \[conj:aharoni\] turns out to be true, it is of interest to see how sharp it is. As shown by Barat and Wanless [@barat2014], one can find constructions of $n$ matchings with $\left\lfloor \frac{n}{2}\right\rfloor -1$ matchings of size $n+1$ and the remaining ones being of size $n$ such that there is no rainbow matching of size $n$. We wonder whether the expression $\left\lfloor \frac{n}{2}\right\rfloor -1$ above could also be replaced by $(1-o(1))n$.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Positive results of dark matter searches in DAMA/NaI and DAMA/LIBRA experiments, being put together with the results of other groups, can imply nontrivial particle physics solutions for cosmological dark matter. Stable particles with charge -2, bound with primordial helium in O-helium “atoms” (OHe), represent a specific Warmer than Cold nuclear-interacting form of dark matter. Slowed down in the terrestrial matter, OHe is elusive for direct methods of underground Dark matter detection used in cryogenic experiments. However radiative capture of OHe by Na and I nuclei can lead to annual variations of energy release in the interval of energy 2-5 keV in DAMA/NaI and DAMA/LIBRA experiments.'
address:
- |
Virtual Institute of Astroparticle physics, APC laboratory\
10, rue Alice Domon et Léonie Duquet 75205 Paris Cedex 13, France,\
Centre for Cosmoparticle Physics “Cosmion” and\
National Research Nuclear University (Moscow Engineering Physics Institute)\
115409 Moscow, Russia\
[email protected]
- |
National Research Nuclear University (Moscow Engineering Physics Institute)\
115409 Moscow, Russia\
[email protected]
- |
National Research Nuclear University (Moscow Engineering Physics Institute)\
115409 Moscow, Russia\
[email protected]
author:
- 'MAXIM YU. KHLOPOV'
- 'ANDREY G. MAYOROV'
- 'EVGENY YU. SOLDATOV'
title: Composite Dark Matter and Puzzles of Dark Matter Searches
---
The widely shared belief is that the dark matter, corresponding to $25\%$ of the total cosmological density, is nonbaryonic and consists of new stable particles. One can formulate the set of conditions under which new particles can be considered as candidates to dark matter (see e.g. Refs. – for review and reference): they should be stable, saturate the measured dark matter density and decouple from plasma and radiation at least before the beginning of matter dominated stage. The easiest way to satisfy these conditions is to involve neutral weakly interacting particles. However it is not the only particle physics solution for the dark matter problem. In the composite dark matter scenarios new stable particles can have electric charge, but escape experimental discovery, because they are hidden in atom-like states maintaining dark matter of the modern Universe.
It offers new solutions for the physical nature of the cosmological dark matter. The main problem for these solutions is to suppress the abundance of positively charged species bound with ordinary electrons, which behave as anomalous isotopes of hydrogen or helium. This problem is unresolvable, if the model predicts stable particles with charge -1, as it is the case for tera-electrons [@Glashow; @Fargion:2005xz]. To avoid anomalous isotopes overproduction, stable particles with charge -1 should be absent, so that stable negatively charged particles should have charge -2 only.
Elementary particle frames for heavy stable -2 charged species are provided by: (a) stable “antibaryons” $\bar U \bar U \bar U$ formed by anti-$U$ quark of fourth generation [@Q; @I; @lom; @Khlopov:2006dk] (b) AC-leptons [@Khlopov:2006dk; @5; @FKS], predicted in the extension [@5] of standard model, based on the approach of almost-commutative geometry [@bookAC]. (c) Technileptons and anti-technibaryons [@KK] in the framework of walking technicolor models (WTC) [@Sannino:2004qp]. (d) Finally, stable charged clusters $\bar u_5 \bar u_5 \bar u_5$ of (anti)quarks $\bar
u_5$ of 5th family can follow from the approach, unifying spins and charges [@Norma].
In the asymmetric case, corresponding to excess of -2 charge species, $X^{--}$, their positively charged antiparticles effectively annihilate in the early Universe. In all the models, in which new stable species belong to non-trivial representations of electroweak SU(2) group sphaleron transitions at high temperatures provide the relationship between baryon asymmetry and excess of -2 charge stable species. After it is formed in the Standard Big Bang Nucleosynthesis (SBBN), $^4He$ screens the $X^{--}$ charged particles in $(^4He^{++}X^{--})$ [*O-helium*]{} “atoms” [@I]. In all the forms of O-helium, $X^{--}$ behaves either as lepton or as specific “heavy quark cluster” with strongly suppressed hadronic interaction. Therefore O-helium interaction with matter is determined by nuclear interaction of $He$. These neutral primordial nuclear interacting objects contribute to the modern dark matter density and play the role of a nontrivial form of strongly interacting dark matter [@Starkman]. Here after a brief review of main features of OHe Universe we concentrate on its effects in underground detectors. We present a quantitative confirmation of the earlier guess [@I; @I2; @KK2; @Bled09; @unesco] that the positive results of dark matter searches in DAMA/NaI (see for review Ref. ) and DAMA/LIBRA [@Bernabei:2008yi] experiments can be explained by effect of O-helium, resolving the controversy between these data and negative results of other experimental groups.
O-helium Universe
=================
Following Refs. –, and consider charge asymmetric case, when excess of $X^{--}$ provides effective suppression of positively charged species.
In the period $100\s \le t \le 300\s$ at $100 \keV\ge T \ge T_o=
I_{o}/27 \approx 60 \keV$, $^4He$ has already been formed in the SBBN and virtually all free $X^{--}$ are trapped by $^4He$ in O-helium “atoms" $(^4He^{++} X^{--})$. Here the O-helium ionization potential is[^1] I\_[o]{} = Z\_[x]{}\^2 Z\_[He]{}\^2 \^2 m\_[He]{}/2 1.6 ,\[IO\]where $\alpha$ is the fine structure constant,$Z_{He}= 2$ and $Z_{x}= 2$ stands for the absolute value of electric charge of $X^{--}$. The size of these “atoms" is [@I; @FKS] R\_[o]{} \~1/(Z\_[x]{} Z\_[He]{}m\_[He]{}) 2 10\^[-13]{} \[REHe\] Here and further, if not specified, we use the system of units $\hbar=c=k=1$.
Due to nuclear interactions with nuclei of cosmic plasma, the O-helium gas is in thermal equilibrium with plasma and radiation on the Radiation Dominance (RD) stage, while the energy and momentum transfer from plasma is effective. The radiation pressure acting on the plasma is then transferred to density fluctuations of the O-helium gas and transforms them in acoustic waves at scales up to the size of the horizon.
At temperature $T < T_{od} \approx 200 S^{2/3}_3\eV$ the energy and momentum transfer from baryons to O-helium is not effective [@I; @KK] because $n_B \sv (m_p/m_o) t < 1,$ where $m_o$ is the mass of the $OHe$ atom, $S_3= m_o/(1 \TeV)$, $m_p$ is the mass of proton, $\sigma
\approx \sigma_{o} \sim \pi R_{o}^2 \approx
10^{-25}\cm^2$ and $v = \sqrt{3T/m_p}$ is the baryon thermal velocity. Then O-helium gas decouples from plasma. It starts to dominate in the Universe after $t \sim 10^{12}\s$ at $T
\le T_{RM} \approx 1 \eV$ and O-helium “atoms" play the main dynamical role in the development of gravitational instability, triggering the large scale structure formation. The composite nature of O-helium determines the specifics of the corresponding dark matter scenario, which has qualitative feature of a Warmer Than Cold Dark Matter model [@unesco].
Being decoupled from baryonic matter, the $OHe$ gas does not follow the formation of baryonic astrophysical objects (stars, planets, molecular clouds...) and forms dark matter halos of galaxies. It can be easily seen that O-helium gas is collisionless for its number density, saturating galactic dark matter. Taking the average density of baryonic matter one can also find that the Galaxy as a whole is transparent for O-helium in spite of its nuclear interaction. Only individual baryonic objects like stars and planets are opaque for it.
O-helium atoms can be destroyed in astrophysical processes, giving rise to acceleration of free $X^{--}$ in the Galaxy.
If the mechanisms of $X^{--}$ acceleration are effective, the anomalous low $Z/A$ component of $-2$ charged $X^{--}$ can be present in cosmic rays at the level [@unesco; @Mayorov] $X/p \sim n_{X}/n_g \sim
10^{-9}S_3^{-1},$ and be within the reach for PAMELA and AMS02 cosmic ray experiments.
In the framework of Walking Technicolor model the excess of both stable $X^{--}$ and $Y^{++}$ is possible [@KK2], the latter being two-three orders of magnitude smaller, than the former. It leads to the two-component composite dark matter scenario with the dominant OHe accompanied by a subdominant WIMP-like component of $(X^{--}Y^{++})$ bound systems. Technibaryons and technileptons can be metastable and decays of $X^{--}$ and $Y^{++}$ can provide explanation for anomalies, observed in high energy cosmic positron spectrum by PAMELA and in high energy electron spectrum by FERMI and ATIC.
O-helium collisions in the galactic bulge can lead to excitation of O-helium. If 2S level is excited, pair production dominates over two-photon channel in the de-excitation by $E0$ transition and positron production with the rate $3 \cdot 10^{42}S_3^{-2} \s^{-1}$ is not accompanied by strong gamma signal. According to Ref. this rate of positron production for $S_3 \sim 1$ is sufficient to explain the excess in positron annihilation line from bulge, measured by INTEGRAL (see Ref. for review and references). If $OHe$ levels with nonzero orbital momentum are excited, gamma lines should be observed from transitions ($ n>m$) $E_{nm}= 1.598 \MeV (1/m^2
-1/n^2)$ (or from the similar transitions corresponding to the case $I_o = 1.287 \MeV $) at the level of $3 \cdot 10^{-4}S_3^{-2}(\cm^2 \s
\MeV sr)^{-1}$.
O-helium in the terrestrial matter
==================================
The evident consequence of the O-helium dark matter is its inevitable presence in the terrestrial matter, which appears opaque to O-helium and stores all its in-falling flux.
The nuclear cross section of the O-helium interaction with matter escapes the severe constraints on strongly interacting dark matter particles (SIMPs) [@Starkman] imposed by the XQC experiment [@XQC]. Therefore, a special strategy of direct O-helium search is needed, as it was proposed in Ref. .
After they fall down terrestrial surface the in-falling $OHe$ particles are effectively slowed down due to elastic collisions with matter. Then they drift, sinking down towards the center of the Earth with velocity V = 80 S\_3 A\_[med]{}\^[1/2]{} /. \[dif\]Here $A_{med} \sim 30$ is the average atomic weight in terrestrial surface matter, $n=2.4 \cdot 10^{24}/A_{med}$ is the number of terrestrial atomic nuclei, $\sigma v$ is the rate of nuclear collisions and $g=980~ \cm/\s^2$.
Then the O-helium abundance the Earth is determined by the equilibrium between the in-falling and down-drifting fluxes.
The in-falling O-helium flux from dark matter halo is $$F=\frac{n_{0}}{8\pi}\cdot |\overline{V_{h}}+\overline{V_{E}}|,$$ where $V_{h}$-speed of Solar System (220 km/s), $V_{E}$-speed of Earth (29.5 km/s) and $n_{0}=3 \cdot 10^{-4} S_3^{-1} \cm^{-3}$ is the local density of O-helium dark matter. Here, for simplicity, we don’t take into account velocity dispersion and distribution of particles in the incoming flux that can lead to significant effect.
At a depth $L$ below the Earth’s surface, the drift timescale is $t_{dr} \sim L/V$, where $V \sim 400 S_3 \cm/\s$ is given by Eq. (\[dif\]). It means that the change of the incoming flux, caused by the motion of the Earth along its orbit, should lead at the depth $L \sim 10^5 \cm$ to the corresponding change in the equilibrium underground concentration of $OHe$ on the timescale $t_{dr} \approx 2.5 \cdot 10^2 S_3^{-1}\s$.
In underground detectors, $OHe$ “atoms” are slowed down to thermal energies and give rise to energy transfer $\sim 2.5 \cdot 10^{-4}
\eV A/S_3$, far below the threshold for direct dark matter detection. It makes this form of dark matter insensitive to the severe CDMS constraints [@Akerib:2005kh]. However, in $OHe$ reactions with the matter of underground detectors can lead to observable effects.
The equilibrium concentration, which is established in the matter of underground detectors, is given by $$n_{oE}=\frac{2\pi \cdot F}{V} = n_{oE}^{(1)}+n_{oE}^{(2)}\cdot sin(\omega (t-t_0)),
\label{noE}$$ where $\omega = 2\pi/T$, $T=1yr$ and $t_0$ is the phase. The averaged concentration is given by $$n_{oE}^{(1)}=\frac{n_o}{320S_3 A_{med}^{1/2}} V_{h}$$ and the annual modulation of concentration is characterized by $$n_{oE}^{(2)}= \frac{n_o}{640S_3 A_{med}^{1/2}} V_E$$ The rate of nuclear reactions of OHe with nuclei is proportional to the local concentration and the energy release in these reactions should lead to observable signal. There are two parts of the signal: the one determined by the constant part and annual modulation, which is concerned by the strategy of dark matter search in DAMA experiment [@Bernabei:2008yi].
Low energy bound state of O-helium with nuclei
----------------------------------------------
Our explanation [@Bled09; @unesco] is based on the idea that OHe, slowed down in the matter of DAMA/NaI or DAMA/LIBRA detector, can form a few keV bound state with nucleus, in which OHe is situated **beyond** the nucleus. Therefore the positive result of this experiment is explained by reaction $$A+(^4He^{++}X^{--}) \rightarrow [A(^4He^{++}X^{--})]+\gamma
\label{HeEAZ}$$ with sodium and/or iodine. In detectors with different chemical content such level may not exist at all, or has other value of energy. The rate of reaction (\[HeEAZ\]) is proportional to temperature and suppressed in cryogenic detectors, making the comparison of their results with DAMA a nontrivial task.
The approach of Refs. and assumes the following picture: at the distances larger, than its size, OHe is neutral and it feels only Yukawa exponential tail of nuclear attraction, due to scalar-isoscalar nuclear potential. It should be noted that scalar-isoscalar nature of He nucleus excludes its nuclear interaction due to $\pi$ or $\rho$ meson exchange, so that the main role in its nuclear interaction outside the nucleus plays $\sigma$ meson exchange, on which nuclear physics data are not very definite. When the distance from the surface of nucleus becomes smaller than the size of OHe, the mutual attraction of nucleus and OHe is changed by dipole Coulomb repulsion. Inside the nucleus strong nuclear attraction takes place. In the result the spherically symmetric potential appears, given by $$U=-\frac{A_{He} A g^2 exp(-\mu r)}{r} + \frac{Z_{He} Z e^2 r_o \cdot F(r)}{r^2}.
\label{epot}$$ Here $A_{He}=4$, $Z_{He}=2$ are atomic weight and charge of helium, $A$ and $Z$ are respectively atomic weight and charge of nucleus, $\mu$ and $g^2$ are the mass and coupling of scalar-isoscalar meson - mediator of nuclear attraction, $r_o$ is the size of OHe and $F(r)$ is its electromagnetic formfactor, which strongly suppresses the strength of dipole electromagnetic interaction outside the OHe “atom”.
To simplify the solution of Schrodinger equation the potential (\[epot\]) was approximated in [@Bled09] by a rectangular potential, presented on Fig. \[pic23\].
Solutions of Schrodinger equation for each of the four regions, indicated on Fig. \[pic23\], are given in textbooks (see e.g.[@LL3]) and their sewing determines the condition, under which a low-energy OHe-nucleus bound state appears in the region III.
The energy of this bound state and its existence strongly depend on the parameters $\mu$ and $g^2$ of nuclear potential (\[epot\]). On the Fig. \[NaI\] the region of these parameters, giving 2-6 keV energy level in OHe bound states with sodium and iodine are presented. In these calculations [@Bled09] the mass of OHe was taken equal to $m_o=1 TeV$.
The rate of radiative capture of OHe by nuclei can be calculated with the use of the analogy with the radiative capture of neutron by proton with the account for: i) absence of M1 transition that follows from conservation of orbital momentum and ii) suppression of E1 transition in the case of OHe. Since OHe is isoscalar, isovector E1 transition can take place in OHe-nucleus system only due to effect of isospin nonconservation, which can be estimated by factor $f \sim 10^{-3}$, corresponding to relative mass difference of neutron and proton. In the result the rate of OHe radiative capture by nucleus with atomic number $A$ and charge $Z$ to the energy level $E$ in the medium with temperature $T$ is given by $$\sigma v=\frac{f \pi \alpha}{m_p^2} \frac{3}{\sqrt{2}} (\frac{Z}{A})^2 \frac{T}{\sqrt{Am_pE}}.
\label{radcap}$$
Formation of OHe-nucleus bound system leads to energy release of its binding energy, detected as ionization signal. In the context of our approach the existence of annual modulations of this signal in the range 2-6 keV and absence of such effect at energies above 6 keV means that binding energy of Na-OHe and I-OHe systems in DAMA experiment should not exceed 6 keV, being in the range 2-4 keV for at least one of these elements. The amplitude of annual modulation of ionization signal (measured in counts per day per kg, cpd/kg) is given by $$\zeta=\frac{3\pi \alpha \cdot n_o N_A V_E t Q}{640\sqrt{2} A_{med}^{1/2} (A_I+A_{Na})} \frac{f}{S_3 m_p^2} (\frac{Z_i}{A_i})^2 \frac{T}{\sqrt{A_i m_p E_i}}=
4.3\cdot10^{10}\frac{f}{S_3^2} (\frac{Z_i}{A_i})^2 \frac{T}{\sqrt{A_i m_p E_i}}.$$ Here $N_A$ is Avogadro number, $i$ denotes Na or I, $Q=10^3$ (corresponding to 1kg of the matter of detector), $t=86400 \s$, $E_i$ is the binding energy of Na-OHe (I-OHe) system and $n_{0}=3 \cdot 10^{-4} S_3^{-1} \cm^{-3}$ is the local density of O-helium dark matter. The value of $\zeta$ should be compared with the integrated over energy bins signals in DAMA/NaI and DAMA/LIBRA experiments and the result of these experiments can be reproduced e.g. for $E_{Na} = 3 \keV$ and $E_{I} = 5 \keV$.
At the corresponding values of $\mu$ and $g^2$ energy of OHe binding with other nuclei can strongly differ from 2-6 keV. In particular, energy release at the formation of OHe bound state with thallium can be larger than 6 keV. However, for the cross section of radiative capture of thallium by OHe, given by Eq. (\[radcap\]) and taking into account that thallium content in DAMA detector is 3 orders of magnitude smaller, than NaI, such signal is below the experimental errors.
It should be noted that the results of DAMA experiment exhibit also absence of annual modulations at the energy of MeV-tens MeV. Energy release in this range should take place, if OHe-nucleus system comes to the deep level inside the nucleus (in the region I of Fig. \[pic23\]). This transition implies tunneling through dipole Coulomb barrier and is suppressed below the experimental limits. The actual rate of these transitions is under our current study. Since OHe capture rate is proportional to the temperature, it is suppressed in cryogenic detectors by a factor of order $10^{-4}$. The predicted effects of OHe radiative capture in different cryogenic detectors at $T=10 \mK$, $f=10^{-3}$ and $S_3=1$ are given in Table \[ta1\].
Conclusions
===========
To conclude, the existence of heavy stable charged particles may not only be compatible with the experimental constraints but even lead to composite dark matter scenario of nuclear interacting Warmer than Cold Dark Matter. This new form of dark matter can provide explanation of excess of positron annihilation line radiation, observed by INTEGRAL in the galactic bulge. The search for stable -2 charge component of cosmic rays is challenging for PAMELA and AMS02 experiments. Decays of heavy charged constituents of composite dark matter can provide explanation for anomalies in spectra of cosmic high energy positrons and electrons, observed by PAMELA, FERMI and ATIC. In the context of our approach search for heavy stable charged quarks and leptons at LHC acquires the significance of experimental probe for components of cosmological composite dark matter.
The results of dark matter search in experiments DAMA/NaI and DAMA/LIBRA can be explained in the framework of our scenario without contradiction with negative results of other groups. Our approach contains distinct features, by which the present explanation can be distinguished from other recent approaches to this problem [@Edward] (see also review and more references in Ref. ).
The proposed explanation is based on the mechanism of low energy binding of OHe with nuclei. Within the uncertainty of nuclear physics parameters there exists a range at which OHe binding energy with sodium and/or iodine is in the interval 2-6 keV. Radiative capture of OHe to this bound state leads to the corresponding energy release observed as an ionization signal in DAMA detector.
OHe concentration in the matter of underground detectors is determined by the equilibrium between the incoming cosmic flux of OHe and diffusion towards the center of Earth. It is rapidly adjusted and follows the change in this flux with the relaxation time of few minutes. Therefore the rate of radiative capture of OHe should experience annual modulations reflected in annual modulations of the ionization signal from these reactions.
An inevitable consequence of the proposed explanation is appearance in the matter of DAMA/NaI or DAMA/LIBRA detector anomalous superheavy isotopes of sodium and/or iodine, having the mass roughly by $m_o$ larger, than ordinary isotopes of these elements.
Our results show that the ionization signal, detected by DAMA, is proportional to the temperature and should be suppressed in cryogenic detectors. Therefore test of results of DAMA/NaI and DAMA/LIBRA experiments by other experimental groups can become a very nontrivial task, especially, in view of their rejection of electromagnetic part of counting rate in the absence of nuclear recoil.
The presented approach sheds new light on the physical nature of dark matter. Specific properties of composite dark matter and its constituents are challenging for their experimental search. OHe interaction with matter is an important aspect of these studies. In this context positive result of DAMA/NaI and DAMA/LIBRA experiments may be a signature for exciting phenomena of O-helium nuclear physics.
Acknowledgments {#acknowledgments .unnumbered}
===============
We would like to thank Pierluigi Belli, Rita Bernabei, Jean Pierre Gazeau and Bernard Sadoulet for discussions and important comments. We express our gratitude to organizers of IWARA09 for cooperation.
[0]{}
M.Yu. Khlopov [*Cosmoparticle physics*]{} (World Scientific, Singapore, 1999).
M.Yu. Khlopov in [*Cosmion-94*]{}, Eds. M.Yu.Khlopov et al. (Editions frontieres, 1996) P. 67; M. Y. Khlopov, . M. Y. Khlopov, ; in arXiv:0711.4681, p. 114; M. Y. Khlopov and N. S. Mankoc-Borstnik, ibid, p. 195.
S. L. Glashow, arXiv:hep-ph/0504287. D. Fargion and M. Khlopov, arXiv:hep-ph/0507087. K.M.Belotsky [*et al*]{}., M.Yu. Khlopov, . K. Belotsky [*et al*]{}., arXiv:astro-ph/0602261. K. Belotsky [*et al*]{}., ; K. Belotsky [*et al*]{}., arXiv:0806.1067 \[astro-ph\]. M. Y. Khlopov, arXiv:astro-ph/0607048.
C. A. Stephan, arXiv:hep-th/0509213. D. Fargion [*et al*]{}., ; M. Y. Khlopov and C. A. Stephan, arXiv:astro-ph/0603187. A. Connes [*Noncommutative Geometry*]{} (Academic Press, London and San Diego, 1994).
M. Y. Khlopov and C. Kouvaris, . F. Sannino and K. Tuominen, ; D. K. Hong [*et al*]{}., ; D. D. Dietrich [*et al*]{}., ; S. B. Gudnason [*et al*]{}., ; S. B. Gudnason [*et al*]{}, . N.S. Mankoč Borštnik, ; N.S. Mankoč Borštnik, ; G. Bregar, M. Breskvar, D. Lukman, N.S. Mankoč Borštnik, .
C. B. Dover [*et al*]{}., ; S. Wolfram, ; G. D. Starkman [*et al*]{}., ; R. Bernabei [*et al*]{}., ; D. Javorsek [*et al*]{}., ; S. Mitra, ; G. D. Mack [*et al*]{}., ; B. D. Wandelt [*et al*]{}., arXiv:astro-ph/0006344; P. C. McGuire and P. J. Steinhardt, arXiv:astro-ph/0105567; G. Zaharijas and G. R. Farrar, . M. Y. Khlopov, arXiv:0806.3581 \[astro-ph\]; M. Y. Khlopov, arXiv:0801.0167 \[astro-ph\]; arXiv:0801.0169 \[astro-ph\]. M. Yu. Khlopov [*et al*]{}., M. Y. Khlopov and C. Kouvaris, M. Yu. Khlopov, A. G. Mayorov, E.Yu. Soldatov, ; arXiv:0911.5606 \[astro-ph\]; M. Yu. Khlopov, The puzzles of dark matter searches, to appear in [*AIP Proceedings of the ’Invisible Universe International Conference’*]{}, arXiv:0911.5685 \[astro-ph.CO\].
R. Bernabei [*et al*]{}., DAMA Collaboration (R. Bernabei [*et al*]{}.), ; R. Bernabei [*et al*]{}., arXiv:1002.1028 \[astro-ph.GA\].
R. N. Cahn and S. L. Glashow, ; M. Pospelov, ; K. Kohri and F. Takayama, .
K. M. Belotsky, A. G. Mayorov and M.Yu.Khlopov, “Charged particles of dark matter in the cosmic rays”, to appear in [*Proc. of the Scientific Session MIFI-2010*]{}
D. P. Finkbeiner and N. Weiner, B. J. Teegarden [*et al*]{}.,
D. McCammon [*et al*]{}., ; D. McCammon [*et al*]{}., .
K. Belotsky [*et al*]{}., arXiv:astro-ph/0606350. CDMS Collaboration (D. S. Akerib [*et al*]{}.), ; CDMS Collaboration (Z. Ahmed [*et al*]{}.), ; The CDMS-II Collaboration (Z. Ahmed [*et al*]{}.), arXiv:0912.3592 \[astro-ph.CO\]; O. Kamaev for the CDMS Collaboration, arXiv:0910.3005 \[hep-ex\].
L. D. Landau, E. M. Lifshitz [*Quantum Mechanics: Non-Relativistic Theory*]{} (Fizmatlit, Moscow, 2004).
F. Petriello and K. M. Zurek, ; R. Foot, ; J. L. Feng, J. Kumar and L. E. Strigari, arXiv:0806.3746 \[hep-ph\]; J. L. Feng, J. Kumar, J. Learned and L. E. Strigari, arXiv:0808.4151 \[hep-ph\]; E. M. Drobyshevski, arXiv:0811.0151 \[astro-ph\].
G. B. Gelmini, arXiv:0810.3733 \[hep-ph\].
[^1]: The account for charge distribution in $He$ nucleus leads to smaller value $I_o \approx 1.3 \MeV$ [@Pospelov].
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In this paper we describe observations acquired by satellite instruments (*Hinode*/SOT and *IRIS*) and ground-based telescopes (ROSA@DST) during two consecutive C7.0 and X1.6 flares occurred in active region NOAA 12205 on 2014 November 7. The analysis of these data show the presence of continuum enhancements during the evolution of the events, observed both in ROSA images and in *IRIS* spectra. Moreover, we analyze the role played by the evolution of the $\delta$ sunspots of the active region in the flare triggering, indicating the disappearance of a large portion of penumbra around these sunspots.'
author:
- 'F. Zuccarello, S.L. Guglielmino, V. Capparelli, M. Mathioudakis, P. Keys, L. Fletcher, S. Criscuoli, M. Falco M. Murabito'
title: |
Continuum emission enhancements and penumbral changes\
observed during flares by *IRIS*, ROSA, and *Hinode*
---
Introduction
============
During solar flares, electromagnetic radiation from radio waves to $\gamma$ rays can be emitted due to the conversion of magnetic energy as a consequence of magnetic reconnection ([@Zuccarello:13], [@Romano:15], [@Guglielmino:16]). Enhancements in the continuum at visible wavelengths (white-light \[WL\] flares, see, e.g., [@neidig83]), as well as in the FUV/NUV passbands may be observed, especially in the case of the most energetic flares. Moreover, the strong energy release occurring during reconnection can give rise to a rearrangement of the magnetic field at the photospheric level, leading in some cases to morphological changes in the penumbrae of sunspots [@Song:16]. In this paper we show some examples of these processes occurring during two consecutive flares.
![*Left*: ROSA *G*-band image showing the entire FOV2, during the occurrence of the X1.6 flare. The white arrow indicates the southern $\delta$ spot of the AR. *Middle*: Zoomed image showing the location of a ribbon observed in the *G*-band (white arrow) at 17:22 UT (i.e., few minutes before the X1.6 flare peak). *Right*: *G*-band difference image showing part of FOV2, with the location of the WL ribbons (red arrows). North is on the top, West is to the right. \[fig:ROSA\]](Map195_FOV2.pdf "fig:") ![*Left*: ROSA *G*-band image showing the entire FOV2, during the occurrence of the X1.6 flare. The white arrow indicates the southern $\delta$ spot of the AR. *Middle*: Zoomed image showing the location of a ribbon observed in the *G*-band (white arrow) at 17:22 UT (i.e., few minutes before the X1.6 flare peak). *Right*: *G*-band difference image showing part of FOV2, with the location of the WL ribbons (red arrows). North is on the top, West is to the right. \[fig:ROSA\]](SubMap915.pdf "fig:") ![*Left*: ROSA *G*-band image showing the entire FOV2, during the occurrence of the X1.6 flare. The white arrow indicates the southern $\delta$ spot of the AR. *Middle*: Zoomed image showing the location of a ribbon observed in the *G*-band (white arrow) at 17:22 UT (i.e., few minutes before the X1.6 flare peak). *Right*: *G*-band difference image showing part of FOV2, with the location of the WL ribbons (red arrows). North is on the top, West is to the right. \[fig:ROSA\]](Diff_ribbon.pdf "fig:")
Observations
============
On 2014 November 7, active region (AR) NOAA 12205 was observed during a coordinated observing campaign carried out using the ground-based ROSA (Rapid Oscillations in the Solar Atmosphere, [@jess10]) instrument at the Dunn Solar Telescope (DST) and the *IRIS* (Interface Region Imaging Spectrograph, [@depontieu]) satellite. ROSA@DST acquired data simultaneously in the Blue continuum at 4170 Å and in the *G* band at 4305.5 Å, with a diffraction limited spatial sampling of $0.069^{\prime\prime}$/pixel in two different fields of view (both $69^{\prime\prime} \times 69^{\prime\prime}$). *IRIS* acquired SJI images from 16:07 to 16:57 UT at 3 wavelengths (C II at 1330 Å, Mg II k at 2796 Å, Mg II wing at 2832 Å) and used the large 4-step coarse raster mode. High-resolution data acquired in the *G* band between 15:15 UT and 19:06 UT by the Solar Optical Telescope ([@Tsuneta:08]) aboard the *Hinode* satellite ([@Kosugi:07]) were also used.
![*Left*: Plots of the average intensity as a function of wavelength in three of the *IRIS* spectral windows in the pixels at raster positions (3,\[537:539\]; blue), (3,\[543:545\]; orange), and (3,\[558:560\]; green) at 16:56:47 UT. For comparison, the black line is the average intensity calculated at the same time in a quiet-Sun region. *Right*: Difference image between *G*-band filtergrams acquired by *Hinode*/SOT at 18:36:40 UT and 15:15:04 UT. White (black) areas indicate regions with penumbral enhancement (decay). Red contours indicate the umbral boundary at 15:15:04 UT. \[fig:iris\_continua\]](profile_iris_3_pro.pdf "fig:") ![*Left*: Plots of the average intensity as a function of wavelength in three of the *IRIS* spectral windows in the pixels at raster positions (3,\[537:539\]; blue), (3,\[543:545\]; orange), and (3,\[558:560\]; green) at 16:56:47 UT. For comparison, the black line is the average intensity calculated at the same time in a quiet-Sun region. *Right*: Difference image between *G*-band filtergrams acquired by *Hinode*/SOT at 18:36:40 UT and 15:15:04 UT. White (black) areas indicate regions with penumbral enhancement (decay). Red contours indicate the umbral boundary at 15:15:04 UT. \[fig:iris\_continua\]](sot_sequence_new.pdf "fig:")
Results
=======
AR NOAA 12205 was characterized by the presence of two $\delta$ spots: one, located on the northern side of the AR, comprised in the ROSA FOV1, showed the presence of penumbral filaments running almost parallel to the umbrae edges; the other, included in the ROSA FOV2 and located in the southern part of the AR, was characterized by a magnetic inversion line almost parallel to the equator (see Figure \[fig:ROSA\], *left*). Two flares took place in the AR during the observations: a C7.0 flare (peak at 16:39 UT) and an X1.6 flare, starting at 16:53 UT, peaking at 17:26 UT, ending at 18:34 UT. Few minutes before the peak of the X1.6 flare (17:22 UT), the analysis of ROSA images relevant to FOV2 indicates that both in the *G* band and in the 4170 Å continuum, it is possible to detect WL ribbons close to the southern $\delta$ spot (see, e.g., Figure \[fig:ROSA\], *middle*), separating with a velocity of $\sim10 \,\mathrm{km\,s}^{-1}$. In particular, the southernmost ribbon is clearly visible in the images and movies obtained in these wavelength ranges, while the northern ribbon can be distinguished only using the difference images (see, e.g., Figure \[fig:ROSA\], *right*).
In Figure \[fig:iris\_continua\] (*left*) we display the radiometric calibrated intensities for selected pixels along the *IRIS* slit at 16:56:47 UT. The blue, orange and green curves show the intensity in correspondence of a ribbon observed in the northern $\delta$ spot. We can see that in all the channels there is an intensity enhancement also in the continuum region. Moreover, the blue pixel exhibits a very prominent bump in the blue wing of the Si IV 1402 Å line, but also in the blue wings of the C II 1334 and 1336, Mg II h and k lines.
Using filtergrams obtained by *Hinode*/SOT in the *G* band, we obtained the difference image displayed in Figure \[fig:iris\_continua\] (*right*), showing areas of permanent penumbral decay (black) and enhancement (white) during the evolution of the observed flares.
The authors acknowledge support by the Università degli Studi di Catania (Piano per la Ricerca 2016-2018 – Linee di intervento 1-2) and by the European Union’s H2020 programme under grant agreement no. 739500 (PRE-EST project).
[0]{} . . . . .
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The rate of Schwinger pair production due to an external electric field can be derived heuristically from the uncertainty principle. In the presence of a cosmological constant, it has been argued in the literature that the uncertainty principle receives a correction due to the background curvature, which is known as the “extended uncertainty principle” (EUP). We show that EUP does indeed lead to the correct result for Schwinger pair production rate in anti-de Sitter spacetime (the case for de Sitter spacetime is similar), provided that the EUP correction term is *negative* (positive for the de Sitter case). We compare the results with previous works in the EUP literature.'
author:
- Yen Chin
title: Schwinger Pair Production and the Extended Uncertainty Principle
---
Schwinger Pair Production from the Uncertainty Principle {#1}
========================================================
The quantum vacuum is teeming with virtual particles, whose fleeting existence is governed by the uncertainty principle. On the other hand, if we apply a sufficiently strong external electric field, we can “boil the vacuum” [@pc] and create real particle pairs from the virtual ones. This is the well known Schwinger effect [@Schwinger]. There are many ways to derive the Schwinger critical field and the corresponding pair production rate. However, a heuristic derivation can already give us some insights as to why such an effect should occur (in the Appendix we briefly discuss the Euclidean method).
Consider a virtual electron-positron pair in a constant electric field of strength[^1] $E$. Suppose the particles move apart from each other by a distance $\ell$, then the amount of energy they receive from the electric field is $eE\ell$. The pair will become real if $eE\ell > 2m_e$, i.e., if the energy exceeds the rest mass of the two particles. The typical separation of the virtual pair is of the order of the Compton wavelenth $2\pi \hbar/m_e$. This can be derived from the Heisenberg uncertainty principle as follows. First, denote the characteristic length scale $\ell \sim \Delta x$, and[^2] $\Delta p \sim m_e$. Then the uncertainty relation $\Delta x \Delta p \sim \hbar/2$ implies that $\ell \sim \hbar/(2m_e)$. This is the Compton wavelength $\ell_C=2\pi \hbar/m_e$ up to a dimensionless constant $4\pi$. Thus the condition that the virtual pair becomes real is the inequality $$\label{condition}
e E\ell_C = 4\pi \ell eE > 2m_e,$$ which implies that the Schwinger critical field $E_S$ should satisfy (up to a constant $1/\pi$ factor), the relation $$\frac{m_e^2}{\hbar e E} \sim 1,$$ This is indeed the case. In conventional SI units, we have $$E_s = \frac{m_e^2 c^3}{e\hbar} \approx 1.32 \times 10^8 \text{V/m}.$$ The Schwinger pair production rate, which we will denote as $\Gamma$, is proportional to $\exp\left[-S(E)\right]$, where $$S(E)= \frac{\pi m_e^2}{\hbar eE},$$ which is a constant multiple of the left hand side expression in Eq.(2). So far, the derivation is a textbook material [@textbook]. In the following we will generalize this argument to derive the Schwinger effect in anti-de Sitter spacetime.
Schwinger Pair Production in Anti-de Sitter Spacetime {#2}
=====================================================
The Schwinger pair production rate receives a correction in the presence of a nonzero cosmological constant, $\Lambda$. In this work we will focus on the anti-de Sitter (AdS) case, which corresponds to $\Lambda <0$ (the case for $\Lambda > 0$, i.e. in de Sitter (dS) spacetime, is similar, and will be discussed later). With $L$ denoting the curvature length scale of the AdS spacetime, the pair production rate is known from the literature to be $\exp\left[-S(E,L)\right]$, where[^3]
S(E,L)&:=2 L\^2 \^[-1]{} (eE - )\
&, \[ads\]
up to $1/L^2$ order in the large $L$ series expansion [@0501169; @0803.2555; @1804.04140]. A derivation using Euclidean method (Wick rotation) is provided in the Appendix. Note that the pair production is suppressed compared to the Minkowski case. On the other hand, the rate will be enhanced in de Sitter spacetime (heuristically, positive cosmological constant that drives the expansion of the Universe also makes separating particle pairs easier; a negative cosmological constant acts in an opposite manner.)
The question we are interested in is this: *can we derive Eq.(\[ads\]) with a suitable correction to the uncertainty principle?* As we shall see, the answer is yes, but not without leaving a puzzle behind concerning the sign of the correction parameter.
Such a correction to the uncertainty principle is known as the “extended uncertainty principle” (EUP), which takes the form $$\label{EUP}
\Delta x \Delta p \sim \frac{\hbar}{2}\left[1+\frac{\beta(\Delta x)^2}{L^2}\right].$$ The parameter $\beta$ is often taken to be of order unity. There have been some debates concerning the *sign* of $\beta$, an unresolved issue that we will discuss in the next section. For now, let us take Eq.(\[EUP\]) for granted and repeat the calculation in Sec.(\[1\]).
Eq.(\[EUP\]) is a quadratic equation in $\Delta x$ and thus gives two possible solutions $$\Delta x_{\pm} = \frac{L(L\Delta p \pm \sqrt{L^2 (\Delta p)^2 - \beta \hbar^2})}{\beta \hbar}.$$ However, $\Delta x_+ \sim 2\Delta p L^2/(\beta \hbar) - \hbar/(2\Delta p) + \mathcal{O}(\beta/L^2)$ in large $L$ limit, which is divergent. Therefore $\Delta x_-$ is the only sensible solution that yields the correct limiting behavior: $\Delta x_- \sim \hbar/(2\Delta p) + \mathcal{O}(\beta/L^2)$. Thus, with $\Delta p \sim m_e$, we have $$\ell \sim \Delta x_- \sim \frac{L(m_e L - \sqrt{L^2 m_e^2 - \beta \hbar^2})}{\beta \hbar}.$$
From Eq.(\[condition\]), one can obtain the modified Schwinger critical field condition:
1 & \~(m\_e L - )\^[-1]{}\
& =\
& = + ().
Dropping the constant prefactor, the expression in the square bracket should be compared to the expression in the square bracket of Eq.(\[ads\]).
Since the characteristic field strength is $E \sim m_e/(2\pi \ell e)$, we also have
- & \~-(m\_e L - )\
& = - + ().
Therefore, up to the same order of the series expansion, $\hbar = m_e^2/(\pi e E)$. Consequently, we have $$-\frac{\pi \hbar^2}{4}\frac{\beta}{e EL^2} = - \frac{m_e^4 \beta}{4\pi e^3 E^3L^2}.$$ Comparing this with Eq.(\[ads\]), we conclude that $$\beta = -\pi^2.$$ While the exact numerical value is probably not important in such a heuristic treatment anyway, we note that the sign of the EUP correction is *negative*. This is a surprising curiosity. Let us now compare this result with other works in the literature.
The Sign of Extended Uncertainty Principle Parameter {#3}
====================================================
Initially, EUP was motivated by Park from the point of view that such a form of the uncertainty principle would allow a heuristic derivation of the Hawking temperature of black holes in AdS or dS spacetimes [@0709.2307] (further analysis of black hole thermodynamics in this context was carried out in [@0411086v1]). For example, the Hawking temperature of a Schwarzschild-AdS spacetime in $d$-dimension is given by $$T=\frac{\hbar\left[ (d-3)L^2+(d-1)r_+^2\right] }{4\pi L^2 r_+}.$$ Consider a temperature of a typical photon emitted by the black hole (in the unit that the Boltzmann constant $k_B=1$), $T = E = pc$. From the EUP, we can indeed heuristically derive the correct form $$T\sim \Delta p \sim \frac{\hbar}{2}\left[\frac{1}{\Delta x} + \frac{\beta (\Delta x)}{L^2}\right].$$ Such technique is a direct generalization of the discussion in asymptotically flat case [@0106080], in which $\Delta x \sim r_+$ is the horizon scale (a typical Hawking quanta can materialize in a “quantum atmosphere” that extends quite far away from the horizon, as emphasized by Giddings [@1511.08221]). This “derivation” implies that $\beta$ is positive in AdS spacetime, and negative in dS case[^4].
Shortly after, Bambi and Urban [@Bambi] argued that contrary to Park’s proposal, the sign for $\beta$ in de Sitter spacetime should be positive. More recently, Lake et al. proposed a derivation of EUP from superpositions of geometries [@1812.10045], in which the sign of EUP parameter is the same as the sign of the cosmological constant. That is to say, our result in this work agrees with Lake et al. [@1812.10045], and is also consistent with Bambi [@Bambi], but not with Park [@0709.2307] (also not with the anti-Snyder-de Sitter model of Mignemi [@1110.0201]). In this section, we will attempt to further strengthen the argument for the case $\text{sgn}(\beta)=\text{sgn}(\Lambda)$.
First we note that there have been recent attempts to give EUP a more rigorous foundation from other points of view, see, e.g. [@F; @0306080; @1804.02551; @Mignemi; @0911.5695]. Notably, EUP correction can be viewed as a classical curvature correction due to the underlying geometry [@1804.02551; @9405067]. This is different from GUP correction (see Eq.(\[GUP\]) below) which is quantum gravitational in nature. In fact, in the 1960s, Judge essentially showed that on a unit circle $S^1$, the uncertainty principle should take the form[^5] $$\Delta x \Delta p \geqslant \frac{\hbar}{2}\left[1-C(\Delta x)^2\right],$$ where $C$ is a constant, argued to be $3/\pi^2$ [@judge1; @judge2]; see also [@1512.05716]. Hence, along with [@1804.02551], these give good arguments that if the underlying spatial geometry is positively curved, then the corresponding EUP should have a negative correction term and conversely, a negatively curved spatial geometry should give rise to a positive correction term, at least when the geometry is of constant sectional curvature. This would suggest that de Sitter spacetime, whose global spatial section is $S^3$, should correspond to negative EUP parameter. Nevertheless, one has to keep in mind that both de Sitter and anti-de Sitter spacetimes are maximally symmetric, so one could always choose a foliation such that the spatial slices are either positively curved, flat, or negatively curved, so this argument is suggestive at best.
In fact, for locally asymptotically AdS spacetimes, it is well-known that there are topological black hole solutions. Their Hawking temperature takes the form [@9808032] $$T=\frac{ \hbar\left[k(d-3)L^2+(d-1)r_+^2 \right]}{4\pi L^2 r_+},$$ where $k=+1,0,-1$ correspond to horizons that are positively curved, flat, and negatively curved, respectively. The heuristic argument to derive Hawking temperature discussed above therefore only works for $k=1$ case, and even then such subtleties mean that it becomes rather doubtful whether the heuristic argument works as intended. Note that for the $k=0$ case, AdS toral or planar black hole has temperature that is directly proportional to $r_+$, not inversely proportional to it as in asymptotically flat spacetime. If some form of modified uncertainty principle exists that would allow us to derive Hawking temperature in the manner discussed above, then it must take the form $\Delta x /\Delta p = \text{const.}$, which is not the usual Heisenberg form plus a correction term. This would be rather surprising indeed as one can take both $\Delta x$ and $\Delta p$ to be arbitrarily small, while keeping their ratio constant. In other words, Hawking radiation of AdS black holes depend on the underlying topology, which does not seem easily encoded by just a single form of EUP. For a different criticism of [@0709.2307], see [@0607010].
The Schwinger effect, on the other hand, is independent of $k$. This can be readily shown, for example, by deriving the particle production rate from Euclidean method (Wick rotation), as we show in the Appendix. Our heuristic derivation thus fixes the sign for EUP parameter in a more concrete, straightforward, manner.
Our work is, in any case, not the first to employ EUP to derive Schwinger effect in the presence of a cosmological constant. Hamil and Merad had previously derived Schwinger effect in *de Sitter* spacetime by employing a much more rigorous method than ours [@Hamil]. They solved EUP-modified Klein-Gordon equation and obtained the pair production rate, which is known from earlier literature [@0803.2555; @1401.4137] to be, up to the first correction term[^6], $$\label{dseup}
\Gamma_\text{dS-EUP}=\frac{\pi m^2}{eE} - \frac{1}{4}\frac{\pi m^4}{e^3 E^3 L^2},$$ which corresponds to $\beta > 0$ in our work, as expected. This seems strange at first since Hamil and Merad actually assumed from the beginning that EUP in de Sitter spacetime corresponds to $\beta < 0$ (in our notation). However, there appears to be a typo of a sign (going from Eq.(56) and Eq.(57) to Eq.(59) in their paper), which seems to indicate that in order to match Eq.(\[dseup\]), they should have $\beta > 0$ instead. Nevertheless, much of the calculations in [@Hamil] needs to be repeated with $\beta > 0$ to see if this gives consistent results, as the corresponding equations are not readily obtained just by reversing a few signs.
We shall also remark that various authors have employed EUP with positive $\beta$ without specifying whether it corresponds to either dS or AdS (see, e.g., [@9311147; @9412167]), but based their motivations on the ground that this recovers the symmetry with the “generalized uncertainty principle” (GUP), which is a quantum gravitational correction to the Heisenberg uncertainty principle [@9301067; @9305163; @9904025; @9904026; @6; @7; @8]: $$\label{GUP}
\Delta x \Delta p \sim \frac{\hbar}{2}\left[1+\frac{\alpha(\Delta p)^2}{\hbar}\right],$$ in which $\alpha$ is often taken to be positive. Indeed GUP with positive $\alpha$ can be derived from various means, including various quantum gravitational arguments (see also, [@1001.1205]). Curiously, even for the case of GUP, there are still some indications that $\alpha$ might be negative. For example, a lattice “spacetime crystal” gives rise to such a GUP [@0912.2253]. Negative GUP parameter is also needed if one accepts that Wick-rotation can be applied to obtain GUP-corrected black hole temperature from a Schwarzschild-like black hole with higher order terms [@1407.0113]. More recently, non-commutative geometry [@KLVY] and corpuscular gravity, were also shown to give rise to negative $\alpha$ [@1903.01382]. See [@1809.00442; @1809.06348; @1812.03136], as well as the recent review [@1905.00287], for more discussions.
Incidentally, the method used in Sec.(\[2\]) can be used to compute GUP correction to the Schwinger effect as well. Since the steps are nearly identical, we only state the result here: the pair production rate goes like $$\Gamma_\text{GUP-dS} = \exp\left[-\left(\frac{\pi \hbar m_e^2}{eE} - \alpha \pi^3 e E \right)\right],$$ which agrees – up to a constant numerical factor in the second term linear in $E$ – with the the result in [@1310.6966] obtained using a more rigorous method[^7]. This gives another support to the validity of our heuristic method. (However, to be fair, it is inconsistent with [@MWY], in which the sign of the second term is opposite, although both [@1310.6966] and [@MWY] involve a positive GUP parameter.)
Conclusion {#4}
==========
Schwinger particle production by external electric field can be heuristically derived using the Heisenberg’s uncertainty principle [@textbook]. In this work, we provided a heuristic derivation of the Schwinger effect in anti-de Sitter spacetime (similarly for the de Sitter case) using the so-called extended uncertainty principle (EUP). We found that in order to obtain the known correct result, the EUP parameter must be negative in AdS spacetime, and positive in dS spacetime. This result is consistent with some works in the literature, but not with others. We have further discussed why using the known result for Schwinger pair production rate to determine the sign of EUP parameter is more reasonable than using Hawking radiation, though both derivations are heuristic. Nevertheless, the sign of EUP parameter – like that of GUP – still requires further studies, as different considerations and methods seem to yield different results. This issue requires a better understanding so that EGUP can be better employed as a phenomenological tool for us to investigate the interface of quantum mechanics and gravity.
Appendix: Euclidean Derivation of the Schwinger Effect in AdS Spacetime {#appendix-euclidean-derivation-of-the-schwinger-effect-in-ads-spacetime .unnumbered}
=======================================================================
First, let us review the Euclidean method that allows us to compute Schwinger pair production rate in Minkowski spacetime. Upon Wick rotation $t \mapsto \tau = it$, Minkowski space (now Euclidean space) in the $\tau$-$r$ plane can be written in the polar form (the problem is essentially 2-dimensional): $$\text{d} s^2 = \text{d} R^2 + R^2 \text{d} \psi^2.$$ One construct an effective action $S_\text{eff}=m\ell-eEA$, where $\ell$ and $A$ are the length (circumference) and the area of a circle of radius $R$ centered at an arbitrary fixed point. (We prefer not to include $\hbar$ in this effective action because it is a quantity constructed from classical geometry; the actual action is then $S=S_\text{eff}/\hbar$, which is rightfully dimensionless.) Then $$S_\text{eff}=mL-eEA= 2\pi m R - \pi q ER^2.$$ Solving $\partial S_\text{eff}/\partial R =0 $ gives the extremal value $R_\text{ext}=m/(eE)$, which upon substituting back into the action gives $S_\text{eff}=\pi m^2/(eE)$. The pair production rate is $\exp(-S_\text{eff}/\hbar)$. The method is well-known, and was mentioned in, e.g., [@0110178v3].
This method is readily generalized to anti-de Sitter spacetime. Circumference and area are best computed in the generalized version of polar coordinates – the geodesic polar coordinates. Around an arbitrary fixed point, the metric of a space of constant negative Gaussian curvature $K$ in 2-dimensions has the following form (see Corollary 7.2.1 of [@kuhnel]): $$\label{k}
\text{d}s^2 = \text{d}r^2 + \frac{1}{-K}\sinh^2(\sqrt{-K}r) \text{d}\psi^2.$$ Given the 2-dimensional AdS metric in static coordinates, $$\text{d}s^2 =\left(1+\frac{\tilde{r}^2}{L^2}\right)\text{d}\tau^2+\left(1+\frac{\tilde{r}^2}{L^2}\right)^{-1}\text{d}\tilde{r}^2,$$ we have $K=-1$, and so $$\text{d}s^2 = \text{d}r^2 + L^2\sinh^2(r/L) \text{d}\psi^2.$$ The effective action $S=m\ell-eEA$ has circumference $$\ell=\int_0^{2\pi} L \sinh(R/L) ~\text{d} \psi = 2\pi L \sinh(R/L),$$ and area $$A=\int_0^{2\pi} \int_0^{R} L \sinh(r/L) ~\text{d} r\text{d} \psi = 2\pi L^2[\cosh(R/L)-1].$$ Solving for $\partial S_\text{eff}/\partial R =0 $ gives the extremal value $$R_\text{ext}=\frac{L}{2}\ln\left(\frac{eEL+m}{eEL-m}\right).$$ Consequently, $$S_\text{eff-AdS}=2\pi L^2\left(eE-\sqrt{e^2E^2-\frac{m^2}{L^2}}\right).$$ This calculation only depends on the Gaussian curvature of the Euclidean manifold. It can be shown that AdS metric with different foliations such that $$\text{d}s^2 =\left(k+\frac{\tilde{r}^2}{L^2}\right)\text{d}\tau^2+\left(k+\frac{\tilde{r}^2}{L^2}\right)^{-1}\text{d}\tilde{r}^2,$$ also gives $K=-1/L^2$, and so the result is independent of $k$.
Note that alternatively, if we are only interested in the first few correction terms of the pair production rate, we can simply take a geodesic disk and compute with the well-known formula from differential geometry (see, e.g. Theorem 3.1 of [@grey], which gives results for higher dimensions as well) $$A=\pi R^2 \left(1-\frac{KR^2}{12}+\cdots \right),$$ and $$\ell=2\pi R\left(1-\frac{KR^2}{6}+\cdots \right),$$ so that $$S_\text{eff-AdS} \approx 2\pi m R\left(1+\frac{R^2}{6L^2}\right)-\pi e E R^2 \left(1+\frac{R^2}{12 L^2}\right).$$ Again we can solve for $\partial S_\text{eff}/\partial R =0 $ and subsitute the extremal value $R_\text{ext}$ into the effective action. This gives, after some cumbersome algebraic manipulations, the final result: $$S_\text{eff-AdS} \approx \frac{\pi m^2}{eE} + \frac{1}{4}\frac{\pi m^4}{e^3E^3L^2}.$$ Note that this method also does not depend on $k$.
YCO thanks the National Natural Science Foundation of China (No.11705162, No.11922508) and the Natural Science Foundation of Jiangsu Province (No.BK20170479) for funding support. He also thanks Brett McInnes for useful suggestions.
[99]{}
Pisin Chen, Claudio Pellegrini, “Boiling the Vacuum with Ultra Intense Lasers”, in *Quantum Aspects of Beam Physics*, P. Chen, ed. p.571, World Scientific (1999).
Julian Schwinger, “On Gauge Invariance and Vacuum Polarization", [[Phys. Rev. **82** (1951) 664](https://journals.aps.org/pr/abstract/10.1103/PhysRev.82.664)]{}.
Viatcheslav Mukhanov, Sergei Winitzki, *Introduction to Quantum Fields in Classical Backgrounds*, p.11, Cambridge University Press (2007).
Boris Pioline, Jan Troost, “Schwinger Pair Production in $\text{AdS}_2$”, [[JHEP **0503** (2005) 043](https://iopscience.iop.org/article/10.1088/1126-6708/2005/03/043)]{}, [\[arXiv:hep-th/0501169\]](https://arxiv.org/abs/hep-th/0501169).
Sang Pyo Kim, Don N. Page, “Schwinger Pair Production in $\text{dS}_2$ and $\text{AdS}_2$”, [[Phys. Rev. D **78** (2008) 103517](https://journals.aps.org/prd/abstract/10.1103/PhysRevD.78.103517)]{}, [\[arXiv:0803.2555 \[hep-th\]\]](https://arxiv.org/abs/0803.2555).
Rong-Gen Cai, Sang Pyo Kim, “One-Loop Effective Action and Schwinger Effect in (Anti-)de Sitter Space”, [[JHEP **09** (2014) 072](https://link.springer.com/article/10.1007%2FJHEP09%282014%29072)]{}, [\[arXiv:1407.4569 \[hep-th\]\]](https://arxiv.org/abs/1407.4569v3).
Peter Breitenlohner, Daniel Z. Freedman, “Positive Energy in Anti-de Sitter Backgrounds and Gauged Extended Supergravity”, [[Phys. Lett. B **115** (1982) 197](https://www.sciencedirect.com/science/article/abs/pii/0370269382906438?via%3Dihub)]{}.
Prasant Samantray, Suprit Singh, “Schwinger Pair Production in Hot Anti-de Sitter Space”, [[Phys. Rev. D **99** (2019) 085006](https://journals.aps.org/prd/abstract/10.1103/PhysRevD.99.085006)]{}, [\[arXiv:1804.04140 \[hep-th\]\]](https://arxiv.org/abs/1804.04140).
Mu-in Park, “The Generalized Uncertainty Principle in (A)dS Space and the Modification of Hawking Temperature from the Minimal Length”, [[Phys. Lett. B **659** (2008) 698](https://www.sciencedirect.com/science/article/pii/S0370269307014980?via%3Dihub)]{}, [\[arXiv:0709.2307 \[hep-th\]\]](https://arxiv.org/abs/0709.2307).
Brett Bolen, Marco Cavaglia, “(Anti-)de Sitter Black Hole Thermodynamics and the Generalized Uncertainty Principle”, [[Gen. Rel. Grav. **37** (2005) 1255](https://link.springer.com/article/10.1007/s10714-005-0108-x)]{}, [\[arXiv:gr-qc/0411086\]](https://arxiv.org/abs/gr-qc/0411086v1).
Ronald J. Adler, Pisin Chen, David I. Santiago, “The Generalized Uncertainty Principle and Black Hole Remnants”, [[Gen. Rel. Grav. **33** (2001) 2101](https://link.springer.com/10.1023/A:1015281430411)]{}, [\[arXiv:gr-qc/0106080\]](https://arxiv.org/abs/gr-qc/0106080).
Steven B. Giddings, “Hawking Radiation, the Stefan-Boltzmann law, and Unitarization”, [[Phys. Lett. B **754** (2016) 39](https://www.sciencedirect.com/science/article/pii/S0370269316000022?via%3Dihub)]{}, [\[arXiv:1511.08221 \[hep-th\]\]](https://arxiv.org/abs/1511.08221).
Cosimo Bambi, Federico R. Urban, “Natural Extension of the Generalised Uncertainty Principle”, [[Class. Quant. Grav. **25** (2008) 095006](http://iopscience.iop.org/article/10.1088/0264-9381/25/9/095006/meta)]{}, [\[arXiv:0709.1965 \[gr-qc\]\]](https://arxiv.org/abs/0709.1965).
Matthew J. Lake, Marek Miller, Ray F. Ganardi, Zheng Liu, Shi-Dong Liang, Tomasz Paterek, “Generalised Uncertainty Relations From Superpositions of Geometries”, [[Class. Quant. Grav. **36** (2019) 155012](https://iopscience.iop.org/article/10.1088/1361-6382/ab2160)]{}, [\[arXiv:1812.10045 \[quant-ph\]\]](https://arxiv.org/abs/1812.10045).
Salvatore Mignemi, “Classical and Quantum Mechanics of the Nonrelativistic Snyder Model in Curved Space”, [[Class. Quant. Grav. **29** (2012) 215019](https://iopscience.iop.org/article/10.1088/0264-9381/29/21/215019)]{}, [\[arXiv:1110.0201 \[hep-th\]\]](https://arxiv.org/abs/1110.0201).
Raimundo N. Costa Filho, João P.M. Braga, Jorge H.S. Lira, José S. Andrade, “Extended Uncertainty From First Principles”, [[Phys. Lett. B **755** (2016) 367](https://www.sciencedirect.com/science/article/pii/S0370269316001313?via%3Dihub)]{}.
Alexey Golovnev, Lev Vasil’evich Prokhorov, “Uncertainty Relations in Curved Spaces”, [[J. Phys. A **37** (2004) 2765](http://iopscience.iop.org/article/10.1088/0305-4470/37/7/017/meta)]{}, [\[arXiv:quant-ph/0306080\]](https://arxiv.org/abs/quant-ph/0306080).
Thomas Schürmann, “Uncertainty Principle on 3-Dimensional Manifolds of Constant Curvature”, [[Found. Phys. **48** (2018) 716](https://link.springer.com/article/10.1007%2Fs10701-018-0173-0)]{}, [\[arXiv:1804.02551 \[quant-ph\]\]](https://arxiv.org/abs/1804.02551).
Salvatore Mignemi, “Extended Uncertainty Principle and the Geometry of (Anti)-de Sitter Space”, [[Mod. Phys. Lett. A **25** (2010) 1697](https://www.worldscientific.com/doi/abs/10.1142/S0217732310033426)]{}, [\[arXiv:0909.1202 \[gr-qc\]\]](https://arxiv.org/abs/0909.1202v2).
Subir Ghosh, Salvatore Mignemi, “Quantum Mechanics in de Sitter Space”, [[Int. J. Theor. Phys. **50** (2011) 1803](https://link.springer.com/article/10.1007%2Fs10773-011-0692-3)]{}, [\[arXiv:0911.5695 \[hep-th\]\]](https://arxiv.org/abs/0911.5695).
Achim Kempf, “Quantum Field Theory with Nonzero Minimal Uncertainties in Positions and Momenta”, [[Czechoslovak Journal of Physics **44** (1994) 1041](https://link.springer.com/article/10.1007%2FBF01690456)]{}, [\[arXiv:hep-th/9405067\]](https://arxiv.org/abs/hep-th/9405067).
D. Judge, “On the Uncertainty Relation for $L_z$ and $\phi$”, [[Phys. Lett. **5** (1963) 189](https://linkinghub.elsevier.com/retrieve/pii/S0375960163962832)]{}.
D. Judge, “On the Uncertainty Relation for Angle Variables”, [[II Nuovo Cimento **31 (2)** (1964) 332](https://link.springer.com/article/10.1007/BF02733639)]{}.
Adam R. Brown, “Schwinger Pair Production at Nonzero Temperatures or in Compact Directions”, [[Phys. Rev. D **98** (2018) 036008](https://journals.aps.org/prd/abstract/10.1103/PhysRevD.98.036008)]{}, [\[arXiv:1512.05716 \[hep-th\]\]](https://arxiv.org/abs/1512.05716).
Danny Birmingham, “Topological Black Holes in Anti-de Sitter Space”, [[Class. Quant. Grav. **16** (1999) 1197](https://iopscience.iop.org/article/10.1088/0264-9381/16/4/009)]{}, [\[arXiv:hep-th/9808032\]](https://arxiv.org/abs/hep-th/9808032).
Fabio Scardigli, “Hawking Temperature for Various Kinds of Black Holes From Heisenberg Uncertainty Principle”, [\[arXiv:gr-qc/0607010\]](https://arxiv.org/abs/gr-qc/0607010).
B. Hamil, M. Merad, “Schwinger Mechanism on de Sitter Background”, [[Int. J. Mod. Phys. A **33** (2018) 1850177](https://www.worldscientific.com/doi/abs/10.1142/S0217751X18501774)]{}.
Markus B. Fröb, Jaume Garriga, Sugumi Kanno, Misao Sasaki, Jiro Soda, Takahiro Tanaka, Alexander Vilenkin, “Schwinger Effect in de Sitter Space”, [[JCAP **1404** (2014) 009](https://iopscience.iop.org/article/10.1088/1475-7516/2014/04/009)]{}, [\[arXiv:1401.4137 \[hep-th\]\]](https://arxiv.org/abs/1401.4137).
Achim Kempf, “Uncertainty Relation in Quantum Mechanics with Quantum Group Symmetry”, [[J. Math. Phys. **35** (1994) 4483](https://aip.scitation.org/doi/10.1063/1.530798)]{}, [\[arXiv:hep-th/9311147\]](https://arxiv.org/abs/hep-th/9311147).
Achim Kempf, Gianpiero Mangano, Robert B. Mann, “Hilbert Space Representation of the Minimal Length Uncertainty Relation”, [[Phys. Rev. D **52** (1995) 1108](https://journals.aps.org/prd/abstract/10.1103/PhysRevD.52.1108)]{}, [\[arXiv:hep-th/9412167\]](https://arxiv.org/abs/hep-th/9412167).
Michele Maggiore, “A Generalized Uncertainty Principle in Quantum Gravity”, [[Phys. Lett. B **304** (1993) 65](https://www.sciencedirect.com/science/article/pii/0370269393914018?via%3Dihub)]{}, [\[arXiv:hep-th/9301067\]](https://arxiv.org/abs/hep-th/9301067).
Michele Maggiore, “Quantum Groups, Gravity, and the Generalized Uncertainty Principle”, [[Phys. Rev. D **49** (1994) 5182](https://journals.aps.org/prd/abstract/10.1103/PhysRevD.49.5182)]{}, [\[arXiv:hep-th/9305163\]](https://arxiv.org/abs/hep-th/9305163).
Fabio Scardigli, “Generalized Uncertainty Principle in Quantum Gravity from Micro-Black Hole Gedanken Experiment”, [[Phys. Lett. B **452** (1999) 39](https://www.sciencedirect.com/science/article/pii/S0370269399001677?via%3Dihub)]{}, [\[arXiv:hep-th/9904025\]](https://arxiv.org/abs/hep-th/9904025).
Ronald J. Adler, David I. Santiago, “On Gravity and the Uncertainty Principle”, [[Mod. Phys. Lett. A **14** (1999) 1371](https://www.worldscientific.com/doi/abs/10.1142/S0217732399001462)]{}, [\[arXiv:gr-qc/9904026\]](https://arxiv.org/abs/gr-qc/9904026).
David J. Gross, Paul F. Mende, “String Theory Beyond the Planck Scale”, [[Nucl. Phys. B **303** (1988) 407](https://www.sciencedirect.com/science/article/pii/0550321388903902)]{}.
Daniele Amati, Marcello Ciafolini, Gabriele Veneziano, “Can Spacetime be Probed Below the String Size?”, [[Phys. Lett. B **216** (1989) 41](https://www.sciencedirect.com/science/article/pii/037026938991366X)]{}.
Kenichi Konishi, Giampiero Paffuti, Paolo Provero , “Minimum Physical Length and the Generalized Uncertainty Principle in String Theory”, [[Phys. Lett. B **234** (1990) 276](https://www.sciencedirect.com/science/article/pii/0370269390919274)]{}.
Ronald J. Adler, “Six Easy Roads to the Planck Scale”, [[Am. J. Phys. **78** (2010) 925](https://aapt.scitation.org/doi/10.1119/1.3439650)]{}, [\[arXiv:1001.1205 \[gr-qc\]\]](https://arxiv.org/abs/1001.1205).
Petr Jizba, Hagen Kleinert, Fabio Scardigli, “Uncertainty Relation on World Crystal and its Applications to Micro Black Holes”, [[Phys. Rev. D **81** (2010) 084030](https://journals.aps.org/prd/abstract/10.1103/PhysRevD.81.084030)]{}, [\[arXiv:0912.2253 \[hep-th\]\]](https://arxiv.org/abs/0912.2253v2).
Fabio Scardigli, Roberto Casadio, “Gravitational tests of the Generalized Uncertainty Principle”, [[Eur. Phys. J. C **75** (2015) 425](https://link.springer.com/article/10.1007%2FJHEP07%282015%29052)]{}, [\[arXiv:1407.0113 \[hep-th\]\]](https://arxiv.org/abs/1504.07637).
T. Kanazawa, G. Lambiase, G. Vilasi, A. Yoshioka, “Noncommutative Schwarzschild Geometry and Generalized Uncertainty Principle”, [[Eur. Phys. J. C **79** (2019) 95](https://link.springer.com/article/10.1140%2Fepjc%2Fs10052-019-6610-1)]{}.
Luca Buoninfante, Giuseppe Gaetano Luciano, Luciano Petruzziello, “Generalized Uncertainty Principle and Corpuscular Gravity”, [[Eur. Phys. J. C **79** (2019) 663](https://link.springer.com/article/10.1140%2Fepjc%2Fs10052-019-7164-y)]{}, [\[arXiv:1903.01382 \[gr-qc\]\]](https://arxiv.org/abs/1903.01382).
Yen Chin Ong, “GUP-Corrected Black Hole Thermodynamics and the Maximum Force Conjecture", [[Phys. Lett. B **785** (2018) 217](https://linkinghub.elsevier.com/retrieve/pii/S0370269318306828)]{}, [\[arXiv:1809.00442 \[gr-qc\]\]](https://arxiv.org/abs/1809.00442).
Yen Chin Ong, Yuan Yao, “Generalized Uncertainty Principle and White Dwarfs Redux: How Cosmological Constant Protects Chandrasekhar Limit", [[Phys. Rev. D **98** (2018) 12018](https://journals.aps.org/prd/abstract/10.1103/PhysRevD.98.126018)]{}, [\[arXiv:1809.06348 \[gr-qc\]\]](https://arxiv.org/abs/1809.06348).
Yuan Yao, Meng-Shi Hou, Yen Chin Ong, “A Complementary Third Law for Black Hole Thermodynamics”, [[Eur. Phys. J. C **79** (2019) 513](https://link.springer.com/article/10.1140%2Fepjc%2Fs10052-019-7003-1)]{}, [\[arXiv:1812.03136 \[gr-qc\]\]](https://arxiv.org/abs/1812.03136).
Fabio Scardigli, “The Deformation Parameter of the Generalized Uncertainty Principle”, [[J. Phys. Conf. Ser. **1275** (2019) 012004](https://iopscience.iop.org/article/10.1088/1742-6596/1275/1/012004)]{}, [\[arXiv:1905.00287 \[hep-th\]\]](https://arxiv.org/abs/1905.00287).’
Salah Haouat, Khireddine Nouicer, “Influence of a Minimal Length on the Creation of Scalar Particles”, [[Phys. Rev. D **89** (2014) 105030](https://journals.aps.org/prd/abstract/10.1103/PhysRevD.89.105030)]{}, [\[arXiv:1310.6966 \[hep-th\]\]](https://arxiv.org/abs/1310.6966).
Ben-Rong Mu, Peng Wang, Hai-Tang Yang, “Minimal Length Effects on Schwinger Mechanism”, [[Commun. Theor. Phys. **63** (2015) 715](https://doi.org/10.1088/0253-6102/63/6/715)]{}, [\[arXiv:1501.06020 \[gr-qc\]\]](https://arxiv.org/abs/1501.06020).
A. S. Gorsky, K. A. Saraikin, K. G. Selivanov, “Schwinger Type Processes via Branes and Their Gravity Duals”, [[Nucl. Phys. B **628** (2002) 270](https://linkinghub.elsevier.com/retrieve/pii/S0550321302000950)]{}, [\[arXiv:hep-th/0110178\]](https://arxiv.org/abs/hep-th/0110178v3).
Wolfgang Kühnel, *Differential Geometry: Curves-Surfaces-Manifolds*, Second Ed., Student Mathematical Library Vol. 16, American Mathematical Society, 2006.
Alfred Gray, “The Volume of a Small Geodesic Ball of a Riemannian Manifold”, [[Michigan Math. Jour. **20** (1974) 329](https://projecteuclid.org/euclid.mmj/1029001150)]{}.
[^1]: We shall work with the units in which $c=G=4\pi \epsilon_0 =1$ but $\hbar \neq 1$, so $\hbar$ has the dimension of area, while mass and charge have the dimension of length. The electric field has dimension of inverse length.
[^2]: We assume that the speed $v$ is not too large to require relativistic correction for the momentum. In any case, for $v$ not too close to 1, the $\gamma$-factor is of order unity which can be neglected in our heuristic approach.
[^3]: In the square root sign there appears an additional term inversely proportional to $L^4$ [@0501169; @0803.2555; @1407.4569], which is related to the famous Breitenlohner-Freedman bound [@BF] in AdS spacetime, if one considers one-loop vacuum amplitude effect. This term does not appear in our work. Our expression is the same as, e.g. Eq.(2.28) of [@0501169].
[^4]: In [@0709.2307], the absolute value $\beta$ is also dimensional dependent, but because of the heuristic nature of the Hawking temperature derivation, it is not clear that the constant numerical coefficients involved should be taken too seriously. Therefore we shall just focus our discussion on the sign of $\beta$.
[^5]: Judge was actually discussing an equivalent problem: the uncertainty principle between angular momentum $L_z$ and angle $\varphi$.
[^6]: The full expression of $S(E,L)$ in de-Sitter spacetime is $$S(E,L)_\text{dS-EUP}=2\pi L^2 \hbar^{-1}\left[\sqrt{(eE)^2+\left(m/L\right)^2}-eE\right],$$ having ignored the term that corresponds to the one-loop effect mentioned in Footnote 2.
[^7]: Again, modulo the term that corresponds to the one-loop effect discussed in Footnote 2.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We describe a hybrid meta-architecture – the CHAMELEON – for session-based news recommendation that is able to leverage a variety of information types using Recurrent Neural Networks. We evaluated our approach on two public datasets, using a temporal evaluation protocol that simulates the dynamics of a news portal in a realistic way. Our results confirm the benefits of modeling the sequence of session clicks with RNNs and leveraging side information about users and articles, resulting in significantly higher recommendation accuracy and catalog coverage than other session-based algorithms.'
bibliography:
- 'references.bib'
---
Introduction
============
Recommender systems help users to deal with information overload by providing tailored item suggestions to them. One of the earliest application domains is the recommendation of online *news* [@karimi2018news]. News recommendation is sometimes considered as being particularly difficult, as it has a number of distinctive characteristics [@Zheng:2018:DDR:3178876.3185994]. Among other challenges, news recommenders have to deal with a constant stream of news articles being published, which at the same time can become outdated very quickly. Another challenge is that the system often cannot rely on long-term user preference profiles. Typically, most users are not logged in and their short-term reading interests must be estimated from only a few logged interactions, leading to a *session-based recommendation problem* [@QuadranaetalCSUR2018].
In recent years, we observed an increased interest in the problem of session-based recommendation, where the task is to recommend relevant items given an ongoing user session. Recurrent Neural Networks (RNN) represent a natural choice for sequence prediction tasks, as they can learn models from sequential data. *GRU4Rec* [@hidasi2016] was one of the first neural session-based recommendation techniques, and a number of other approaches were proposed in recent years that rely on deep learning architectures, as in [@Liu2018stamp; @Li2017narm].
However, as shown in [@jannach2017recurrent; @ludewig2018evaluation; @LudewigMauro2019], neural approaches that only rely on logged item interactions have certain limitations and they can, depending on the experimental setting, be outperformed by simpler approaches based, e.g., nearest-neighbor techniques. Differently from previous works, we therefore leverage multiple types of side information with RNNs, including textual article embeddings, as well as the context of users and articles. Furthermore, we propose a meta-architecture to address the aforementioned challenges of recommending in the news domain.
Technical Contribution
======================
Our approach is based upon CHAMELEON [@moreira2018chameleon; @moreira2018news; @moreira2019inra; @moreira2019contextual], which is a Deep Learning Meta-Architecture for News Recommendation. It supports session-based news recommendation scenarios, modeling the sequence of user clicks using Recurrent Neural Networks. The resulting system is a hybrid recommender system, which addresses the permanent user and item cold-start problem in the news domain by leveraging the textual content of news articles, the article context (e.g., recent popularity and recency) and the user context (e.g., time, location, device, previous session clicks).
Figure \[figure:chameleon\_instantiation\] shows our instantiation of the *CHAMELEON* framework with its two main modules: the *ACR* module on the left creates distributed representations of articles’ textual content. The *NAR* module on the right is responsible to generate next-click predictions. The *NAR* module is trained on a ranking loss function based on similarities, which is designed to recommend fresh articles without retraining. As proposed for the *DSSM* loss function [@huang2013learning], it is trained to maximize the likelihood of correctly predicting the next clicked article given a user session.
![An architecture instantiation of *CHAMELEON*[]{data-label="figure:chameleon_instantiation"}](figures/chameleon_instantiation_journal.jpg){width="1.0\linewidth"}
Evaluation Protocol
===================
The evaluation was performed as follows: (1) Recommenders are continuously trained on users’ sessions ordered by time and grouped by hours. Each five hours, the recommenders are evaluated on sessions from the next hour; (2) For each session in the evaluation set, we incrementally revealed one click after the other to the recommender; and (3) For each click to be predicted, we created a set containing 50 negative samples articles (not clicked by the user in her session) and compute top-N metrics about accuracy, item coverage, and novelty.
![Illustration of the evaluation protocol. After training for 5 hours, we evaluate using the sessions of the next hour.[]{data-label="figure:eval_protocol"}](figures/chameleon_eval_protocol.jpg){width="0.99\linewidth"}
Experiments were performed with two public real-world datasets from the *G1* [@moreira2018news] and *Adressa* [@gulla2017adressa] news portals, described in Table \[tab:datasets\].
We evaluated the following recommendation quality factors for the top-N ranked items: accuracy – Hit Rate (*HR@n*) and Mean Reciprocal Rank (MRR@n); item coverage – (*COV*) (i.e., the number of distinct articles that appeared in any top-N list divided by the number of recommendable articles); and novelty – *ESI-R*, which is based on item popularity, returning higher values when recommending long-tail items.
As baseline algorithms for session-based recommendation, we have used: two neural approaches (*GRU4Rec* [@hidasi2016] and *SR-GNN* [@wu2019session]); association rules-based methods (*Co-Occurrence (CO)* and *Sequential Rules (SR)* [@ludewig2018evaluation]); neighborhood-based methods (*Item-kNN* [@hidasi2016] and *Vector Multiplication Session-Based kNN (V-SkNN)* [@jannach2017recurrent]); and two other classical methods (*Recently Popular (RP)* [@ludmann2017recommending] and *Content-Based (CB)*.
*Globo.com (G1)* *Adressa*
------------------------------- ------------------ -----------
Language Portuguese Norwegian
Period (days) 16 16
\# users 322,897 314,661
\# sessions 1,048,594 982,210
\# clicks 2,988,181 2,648,999
\# articles 46,033 13,820
Avg. Sessions length (clicks) 2.84 2.70
: Statistics of the datasets used for the experiments.[]{data-label="tab:datasets"}
Results
=======
The evaluation results are presented in Table \[tab:metrics\_results\], as originally reported in [@moreira2019contextual]. The best results for a metric are printed in bold face and marked with \* if they are significantly different [^1] from all other algorithms.
[p[1.8cm]{}p[1.0cm]{}p[1.0cm]{}p[1.1cm]{}p[1.4cm]{}]{} *Recommender* & *HR@10* & *MRR@10* & *COV@10* & *ESI-R@10*\
\
*CHAMELEON* & **0.6738**\* & **0.3458**\* & 0.6373 & 6.4177\
*SR* & 0.5900 & 0.2889 & 0.2763 & 5.9747\
*Item-kNN* & 0.5707 & 0.2801 & 0.3913 & 6.5909\
*CO* & 0.5689 & 0.2626 & 0.2499 & 5.5728\
*V-SkNN* & 0.5467 & 0.2494 & 0.1355 & 5.1760\
*SR-GNN* & 0.5144 & 0.2467 & 0.3196 & 5.4280\
*GRU4Rec* & 0.4669 & 0.2092 & 0.6333 & 5.2332\
*RP* & 0.4577 & 0.1993 & 0.0218 & 4.4904\
*CB* & 0.3643 & 0.1676 & **0.6774** & **8.1531**\*\
\
*CHAMELEON* & **0.7018**\* & **0.3421**\* & 0.7926 & 5.3410\
*SR* & 0.6288 & 0.3022 & 0.4604 & 5.4443\
*Item-kNN* & 0.6179 & 0.2819 & 0.5314 & 5.4675\
*CO* & 0.6131 & 0.2768 & 0.4220 & 5.0789\
*V-SkNN* & 0.6140 & 0.2723 & 0.1997 & 4.6018\
*SR-GNN* & 0.6122 & 0.2991 & 0.5197 & 5.1013\
*GRU4Rec* & 0.4958 & 0.2200 & 0.5143 & 5.0571\
*RP* & 0.5648 & 0.2481 & 0.0542 & 4.1465\
*CB* & 0.3307 & 0.1253 & **0.8875**\* & **7.6715**\*\
Conclusion
==========
CHAMELEON was specifically designed to address news recommendation challenges such as (a) the short lifetime of the recommendable items and (b) the lack of longer-term preference profiles of the users. In the extensive experiments performed, CHAMELEON was able to provide recommendations with much higher accuracy than all other evaluated algorithms, and it led to the second best item coverage.
[^1]: As errors around the reported averages were normally distributed, we used paired Student’s t-tests with Bonferroni correction at $p<0.001$ for significance tests.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We present measurements of the in-plane ($\kappa_{ab}$) and out-of-plane ($\kappa_{c}$) thermal conductivity of Pr$_2$CuO$_4$ and Gd$_2$CuO$_4$ single crystals. The anisotropy gives strong evidence for a large contribution of magnetic excitations to $\kappa_{ab}$, i.e. for a heat current within the CuO$_2$ planes. However, the absolute values of $\kappa_{\rm mag}$ are lower than previous results on La$_2$CuO$_4$. These differences probably arise from deviations from the nominal oxygen stoichiometry. This has a drastic influence on $\kappa_{\rm mag}$, which is shown by an investigation of a La$_2$CuO$_{4+\delta}$ polycrystal.'
address:
- 'II. Physikalisches Institut, Universität zu Köln, 50937 Köln, Germany'
- 'Institute of Solid State $\&$ Semiconductor Physics, Belarussian Academy of Sciences, Minsk 220072, Belarus'
author:
- 'K. Berggold'
- 'T. Lorenz'
- 'J. Baier'
- 'M. Kriener'
- 'D. Senff'
- 'S. Barilo'
- 'and A. Freimuth'
title: 'Thermal conductivity of $R_2$CuO$_4$, with $R$ = La, Pr and Gd'
---
,
,
,
,
,
Thermal conductivity ,low-dimensional quantum spin systems 74.72.-h, 66.70.+f, 75.10.Jm
One peculiarity of low-dimensional spin systems is the possibility of large contributions to the heat transport by magnetic excitations. This has been found in 1D systems like spin ladders,[@sologubenko00a] whereas the situation in 2D is less clear. The CuO$_2$ planes in La$_2$CuO$_4$ are a good realization of a 2D antiferromagnetic Heisenberg square lattice with a large exchange constant $J\approx 1400$K.[@kastner98a] The thermal conductivity of La$_2$CuO$_4$ shows an unusual behavior.[@nakamura91a] For a heat current along the $c$ direction one low-temperature maximum is present, as it is expected for an insulator. The heat is carried by acoustic phonons, and for high temperatures Umklapp scattering yields an approximate $1/T$ behavior of $\kappa_c$. However, for a heat current parallel to the CuO$_2$ planes a second broad maximum arises in addition to the low-temperature maximum. This can be interpreted in terms of an additional contribution to the heat transport by magnetic excitations. Depending on the crystals, values of $\kappa_{\rm mag}^{(300\,{\rm K})}\approx15\dots 24$W/Km are reported.[@nakamura91a; @yan03a; @hess03a] However, a double peak of $\kappa$ could also arise from a suppression acting in a certain temperature window.[@hofmann01a] This requires an additional scattering mechanism, e.g. due to a structural instability, which is present in La$_2$CuO$_4$.[@braden94b] Such a structural instability is not present in the nearly iso-structural compound Sr$_2$CuO$_2$Cl$_2$, but still there is a second maximum of $\kappa_{ab}$.[@hofmann03a] This gives evidence for a magnetic origin of the second maximum and
![Panel a) and b): $\kappa_{ab}$ and $\kappa_c$ for Pr$_2$CuO$_4$ and Gd$_2$CuO$_4$. The phononic behavior of the out-of-plane measurements at high temperatures is confirmed with $1/T$ fits (lines). Panel c): $\kappa$ of a La$_2$CuO$_{4+\delta}$ polycrystal, see text.[]{data-label="fig1"}](berggold-198-f1.eps){width="45.00000%"}
suggests that a magnetic contribution to $\kappa_{ab}$ could be a common feature of all single-layered insulating cuprates. In order to investigate this, we studied $\kappa$ of [*R*]{}$_2$CuO$_4$, with [*R*]{} = Pr, Gd. Both compounds have a very similar CuO$_2$ square lattice and exchange constant as La$_2$CuO$_4$.[@johnston97a] However, for $R$=Pr the structure is stable down to lowest temperatures, whereas for $R$=Gd a structural phase transition occurs at $685$K.[@braden94a]
Fig. \[fig1\]a shows $\kappa_{ab}$ and $\kappa_c$ for Pr$_2$CuO$_4$. For $\kappa_c$ we observe one low-temperature maximum and a $1/T$ behavior (solid line) above about $70$K. For $\kappa_{ab}$, however, a pronounced shoulder occurs at high temperatures, which leads to an anisotropy $\kappa_{ab}/\kappa_c \approx 2$ at room temperature, whereas the heights of the low-temperature maxima differ only by $20$%. For Gd$_2$CuO$_4$ (Fig. \[fig1\]b), $\kappa_c$ can also be described well by phonons. For $\kappa_{ab}$ a second broad maximum arises at high temperatures. The anisotropy at room temperature $\kappa_{ab}/\kappa_{c}\approx3$ is even larger than for Pr$_2$CuO$_4$. The low-temperature maxima differ by about $25$% in height. They are shifted in temperature, which is likely to be caused by sheet defects, which are more efficient for $\kappa_{c}$. This anisotropic behavior, which occurs in both compounds independently from the presence of a structural phase transition, gives further evidence that a magnetic contribution to $\kappa_{ab}$ is a general feature of these cuprates. We estimate the magnetic contribution at room temperature from the difference $\kappa_{\rm mag}=\kappa_ {ab}-\kappa_c$. This assumes that the phononic anisotropy is not too big, what is reasonable, since the low-temperature maxima for the different directions differ much less than the room-temperature values. In this way, we get $\kappa_{\rm mag}^{(300\,{\rm K})}\approx6$W/Km for Pr$_2$CuO$_4$ and $\approx11$W/Km for Gd$_2$CuO$_4$.
The different values of $\kappa_{\rm mag}^{(300\,{\rm K})}$ raise the question, what determines the size of $\kappa_{\rm mag}$. Because $J$ does not differ too much, it is reasonable to assume that the differences mainly arise from different scattering of the heat-carrying excitations. Possible scattering mechanisms are magnon-magnon, magnon-phonon and magnon-hole scattering. Because of the similar magnetic subsystem for all these compounds, we do not expec that magnon-magnon scattering is very different. Whether a different magnon-phonon scattering is present, is difficult to judge at the present stage of the available data. Thus, we concentrate on the study of magnon-hole scattering. The antiferromagnetic order in La$_2$CuO$_{4+\delta}$ is suppressed drastically by hole doping. To investigate the increase of the magnon-hole scattering, we measured $\kappa$ of a La$_2$CuO$_{4+\delta}$ polycrystal (Fig. \[fig1\]c), which was annealed either in oxygen ($T_N=250$K) or in vacuum ($T_N=317$K). For $T_N=317$K a pronounced high-temperature maximum is present. The magnitude of this maximum is lower than observed in single crystals, which partly arises from the averaging $\kappa=1/3\kappa_c+2/3\kappa_{ab}$ in a polycrystal. Moreover, $T_N$ is still lower than the maximum value $T_N^{\rm max}\approx 325$K.[@chen91a] Weak oxygen doping is also the most probable reason for the different $\kappa_{\rm mag}^{(300\,{\rm K})}$ in single crystals.[@nakamura91a; @yan03a; @hess03a] On our Pr$_2$CuO$_4$ and Gd$_2$CuO$_4$ crystals we determined $T_N=250$K and $295$K from neutron scattering and magnetic susceptibility data, respectively. For Gd$_2$CuO$_4$ we are not aware of higher $T_N$ values, but for Pr$_2$CuO$_4$ significantly larger values up to $T_N\approx 270$K are reported.[@johnston97a] This may explain the low value of $\kappa_{\rm mag}$ for Pr$_2$CuO$_4$. However, $T_N$ is not solely determined by the oxygen content, but also depends e.g. on the magnetic anisotropy. Thus, no direct conclusion concerning the hole content can be drawn from a comparison of $T_N$ for $R_2$CuO$_4$ with different $R$.
In conclusion, we have measured $\kappa_{ab}$ and $\kappa_c$ of Pr$_2$CuO$_4$ and Gd$_2$CuO$_4$. The temperature dependence of $\kappa_{ab}$ and the anisotropy $\kappa_{ab}/\kappa_{c}$ give evidence for a magnetic contribution to $\kappa_{ab}$ and suggest that this behavior is a common feature of single-layered cuprates. The magnetic contribution is suppressed drastically by charge-carrier doping, which was shown on a La$_2$CuO$_{4+\delta}$ polycrystal.
This work was supported by the Deutsche Forschungsgemeinschaft through SFB 608.
[10]{}
A. V. Sologubenko [*et al.*]{}, Phys. Rev. Lett. [**84**]{}, 2714 (2000).
M. A. Kastner [*et al.*]{}, Rev. Mod. Phys. [**70**]{}, 897 (1998).
Y. Nakamura [*et al.*]{}, Physica C [**185-189**]{}, 1409 (1991).
J.-Q. Yan [*et al.*]{}, Phys. Rev. B [**68**]{}, 104520 (2003).
C. Hess [*et al.*]{}, Phys. Rev. Lett. [**90**]{}, 197002 (2003).
M. Hofmann [*et al.*]{}, Phys. Rev. Lett. [**87**]{}, 047202 (2001).
M. Braden [*et al.*]{}, Z. Physik B [**94**]{}, 29 (1994).
M. Hofmann [*et al.*]{}, Phys. Rev. B [**67**]{}, 184502 (2003).
D. Johnston, [*Normal-state magnetic properties of single-layer cuprate high-temperature superconductors and related materials*]{} (Elsevier Science B.V., 1997).
M. Braden [*et al.*]{}, Europhys. Lett. [**25**]{}, 625 (1994).
C. Y. Chen [*et al.*]{}, Phys. Rev. B [**43**]{}, 392 (1991).
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'I review the potential for observing cosmic reionization using the HI 21cm line of neutral hydrogren. Studies include observations of the evolution of large scale structure of the IGM (density, excitation temperature, and neutral fraction), through HI 21cm emission, as well as observations of small to intermediate scale structure through absorption toward the first discrete radio sources. I summarize predictions for the HI signals, then consider capabilities of facilities being built, or planned, to detect these signals. I also discuss the significant observational challenges.'
address: |
$^1$National Radio Astronomy Observatory, Socorro, NM, USA, 87801\
E-mail: [email protected]
author:
- 'C.L. Carilli$^1$'
title: 'HI 21cm probes of reionization, and beyond'
---
cosmology ,radio ,lines
Introduction
============
The 21cm line of neutral hydrogren presents a unique probe of the evolution of the neutral intergalactic medium (IGM), and cosmic reionization. Furlanetto & Briggs [@FB04] point out some of the advantages of using the HI line in this regard: (i) unlike Ly$\alpha$ (ie. the Gunn-Peterson effect), the 21cm line does not saturate, and the IGM remains ’translucent’ at large neutral fractions [@CGO02]. And (ii) unlike CMB polarization studies, the HI line provides full three dimensional (3D) information on the evolution of cosmic structure, and the technique involves imaging the neutral IGM directly, and hence can easily discern different reionization models [@BL04; @FSH03]. HI 21cm observations can be used to study evolution of cosmic structure from the linear regime at high redshift (ie. density-only evolution), through the non-linear, ’messy astrophysics’ regime associated with luminous source formation. As such, HI measurements are sensitive to structures ranging from very large scales down to the source scale set by the cosmological Jeans mass, thereby “making 21cm the richest of all cosmological data sets” [@BL04].
Early calculations of HI 21cm signal from a neutral IGM discuss a broad range of models for large scale HI structure, including Zeldovich ’pancakes’ arising in HDM models, explosive structure formation models, and CDM [@SZ72; @HR79; @SR90]. Early (unsuccessful) searches for large scale structure in HI based on this wide range of predictions include searches for pancakes with masses $>
10^{14}$ M$_\odot$ at $z \sim 3$ [@uson; @oort84; @davies78; @debruyn88], and for even more massive structures ($10^{15}$ M$_\odot$) at $z = 8.4$ [@bebbington86]. Fortunately, the cosmological parameter space has been greatly reduced due to the advent of the concordance model of $\Lambda$CDM structure formation, and this paper will assume the standard precision cosmology parameters [@spergel03].
In this review I will summarize the HI 21cm probes of cosmic reionization and the neutral IGM. I will focus on the latest predictions for the expected signals, and the observational capabilities of telescopes being built, or planned, to detect these signals. I begin with a short basic review of the physical processes involved in reionization that relate directly to the HI 21cm signal. Previous reviews that dealt with, at least in part, the HI 21cm signal from cosmic reionization, include [@BL00; @LB00; @meiksin99], and in the context of the Square Kilometer Array (SKA) [@Carilli04a; @FB04]. See also articles by Furlanetto, Morales, and Peterson in this volume.
The physics of the neutral IGM
==============================
The physics and equations of radiative transfer of the HI 21cm line through the neutral IGM have been considered in detail by many authors [@SR90; @BA04; @FB04; @ZFL04; @SCK05; @tozzi99; @MMR97; @Morales04], and I review only the basic results here.
In analogy to the Gunn-Peterson effect for Ly$\alpha$ absorption by the neutral IGM, the optical depth, $\tau$, of the neutral hydrogen to 21cm absorption for our adopted values of the cosmological parameters is: $$\tau
= {{3c^3 h_p A_{10} n_{HI}}\over{32 \pi k_B \nu_{21}^2 T_S H(z)}} \\
\sim 0.0074 \frac{x_{HI}}{T_S} (1+\delta)(1+z)^{3/2},
\label{tauofz}$$ This equation shows immediately the rich physics involved in studying the HI 21cm line during reionization, with $\tau$ depending on the evolution of cosmic over-densities, $\delta$ (predominantly in the linear regime), the neutral fraction, $x_{\rm HI}$ (ie. reionization), and the HI excitation, or spin, temperature, $T_S$.
In the Raleigh-Jeans limit, the observed brightness temperature (relative to the CMB) due to the HI 21cm line at a frequency $\nu = \nu_0/(1+z)$, where $\nu_0=1420.40575$ MHz, is given by: $$T_B \approx ~ \frac{T_S - T_{\rm CMB}}{1+z} \, \tau
\label{eq:dtb} \\
\approx ~ 7 (1+\delta) x_{HI} (1 - \frac{T_{CMB}}{T_S})
(1+z)^{1/2} ~ \rm{mK},$$ The conversion factor from brightness temperature to specific intensity, I$_\nu$, is given by: $I_\nu = \frac{2 k_B}{(\lambda_{21} (1+z))^2} T_B$, or in more common units: $I_\nu = 7.1\times 10^4 (1+z)^{-2} T_B$ Jy sr$^{-1}$. Equation 2 shows that for $T_S \sim T_{CMB}$ one expects no 21cm signal. When $T_S >> T_{CMB}$, the brightness temperature becomes independent of spin temperature. When $T_S << T_{CMB}$, we expect a strong negative (ie. absorption) signal against the CMB.
An important point to keep in mind is that, for the 21cm experiments being considered, the signal being observed corresponds to large scale structure, not individual galaxies. This point is demonstrated in Fig 1, which shows the expected signal in Jy versus redshift for different HI masses, along with the expected sensitivity of current and future radio telescopes. For simplicity, I have assumed a single line width of 300 km s$^{-1}$, which would correspond to individual virialized galaxies, or to large scale structures just separating from the Hubble flow. Fig 1 shows that current instruments, such as the GMRT, can detect large galaxies ($\sim 10^{10}$ M$_\odot$ in gas) out to only modest redshifts ($z\le 0.3$), even in long integration times. At redshifts corresponding to reionization ($z > 6$), future large area low frequency radio telescopes, such as LOFAR, and eventually the SKA, will still be limited to studying large scale structure (HI masses $>
10^{11}$ to 10$^{12}$ M$_\odot$). Fortunately, the entire IGM is made up of neutral hydrogen prior to reionization, and the large scale structure detected is not just density enhancements (ie. protoclusters), but also structure induced by reionization itself (the ’bubble machine’ of HII regions), and possibly spatial variations in the spin temperature.
[@tozzi99] show that the HI excitation temperature will equilibrate with the CMB on a timescale $\sim
\frac{3\times10^5}{(1+z)}$ year, in absence of other effects. However, collisions and resonant scattering of Ly$\alpha$ photons can drive $T_S$ to the gas kinetic temperature, $T_K$ [@field59; @wout]: $$T_S = \frac{T_{\rm CMB} + y_c T_K +
y_{{\rm Ly}\alpha} T_K}{1 + y_c + y_{{\rm Ly}\alpha}}.
\label{eq:hItspin}$$ In this equation, $y_c$ represents collisional excitation of the hyperfine transition, which couples $T_S$ to the gas kinetic temperature $T_K$. The coupling coefficient $y_c \propto n_H$, and [@zygelman05] shows that, for the mean IGM density, collisional coupling between $T_S$ and $T_K$ becomes significant for $z \ge 30$. The third term in equ 3 corresponds to the Wouthuysen-Field effect, in which resonant scattering of Ly$\alpha$ photons couples the spin temperature to $T_K$ [@wout; @field59; @CM03; @hirata04; @MMR97; @meiksin99; @PF05]. [@MMR97] show that this latter mechanism will be important when the radiation background at the Ly$\alpha$ frequency satisfies $J_\alpha > 9\times 10^{-23} (1+z)$ erg cm$^{-2}$ s$^{-1}$ Hz$^{-1}$ sr$^{-1}$, or about one Ly$\alpha$ photon per every two baryons at $z=8$ [@CM03].
The interplay between the CMB temperature, the kinetic temperature, and the spin temperature, coupled with radiative transfer, lead to a number of interesting physical regimes for the HI 21cm signal. [@Ali05; @BL04] suggest the following plausible regimes:
At $z> 200$ residual free electrons couple $T_{CMB}$ and $T_K$ through Thompson scattering and subsequent gas collisions, while the density is high enough to equilibrate $T_K$ and $T_S$. In this case $T_S =
T_{CMB}$ and there is no 21cm signal.
At $z \sim 30$ to 200, the residual ionization fraction and density is too low to couple $T_K$ to $T_{CMB}$, so the gas cools adiabatically, with temperature falling as $(1+z)^2$, ie. faster than the (1+z) for the CMB. Hence the gas becomes colder than the CMB. However, the mean density in this redshift range is still high enough to provide some coupling between $T_S$ and $T_K$ through collisions, and the HI 21cm signal might be seen in absorption against the CMB. In this regime the HI fluctuations are still evolving linearly, essentially following the dark matter [@BL04; @LZ04; @sethi05].
At $z \sim 20$ to 30, the situation starts to become complex. Collisions can no longer couple $T_K$ to $T_S$, and $T_S$ again approaches $T_{CMB}$. However, we might also expect the first luminous structures (Pop III stars or mini-quasars), at least near the end of this redshift range. The Ly$\alpha$ photons from these objects would induce local coupling of $T_K$ and $T_S$, thereby leading to some 21cm absorption regions. On the other hand, [@BL04; @MMR97] point out that these same photons, and more importantly, any Xrays from the first luminous sources [@ChenME03], could lead to local IGM warming above $T_{CMB}$ well before reionization. Energetically it takes 13.6 eV per baryon to ionize the IGM, but only $0.005
\frac{(1+z)}{20}$ eV to warm the IGM about $T_{CMB}$. Hence one might expect a ’patch work’ of regions with no signal, absorption, and perhaps emission, in the 21cm line.
At $z \sim 6$ to 20 all the physical processes come to play. The IGM is being warmed by the resonant scattering of Ly$\alpha$ photons [@MMR97], and penetrating, ionizing hard Xrays, from the first galaxies and black holes [@LZ04; @BL04; @CM03], as well as by weak shocks associated with structure formation [@FSH03; @GS03; @SCK05; @ChenME03], such that $T_K$ is likely larger than $T_{CMB}$ globally [@FSH03]. Likewise, these objects are reionizing the universe, leading to a fundamental topological change in the IGM, from the linear evolution of large scale structure, to a bubble dominated era of HII regions [@FMH05]. It is this regime that most (although not all), theoretical work has been exploring, and the expected HI 21cm signal is rich. Finally, after reionization, $z < 6$ to 10, the IGM is fully ionized ($x_{HI} < 10^{-4}$), and the 21cm signal is gone.
HI 21cm probes of the evolution of the IGM
==========================================
Global HI signature
-------------------
In an idealized sense, if the universe reionized very rapidly everywhere, at the same time, one would expect a global (ie. full sky) step in the background temperature at the frequency corresponding to the redshifted 21cm line [@Shaver99]. However, if nothing else, cosmic variance will lead to different ionization redshifts in different regions, and we now know that reionization is likely an extended process in time, thereby smoothing out the expected signal in space and time. On the other hand, since this is an all sky signal, the sensitivity of the experiment is independent of telescope collecting area, and the experiment can be done using small area telescopes at low frequency, with well controlled frequency response.
-2.65in
The most recent modeling of the expected global HI signal is presented in [@GS03] (see also [@FSH03]), who employ the numerical simulations of [@Gnedin00]. They include both a fast reionization model, plus more complicated, [*ad hoc*]{} reionization histories.
Results from these calculations are reproduced in Fig 2a. The expected signal peaks at roughly 20 mK above the foreground. In their fiducial (fast reionization) model, the peak occurs at $z \sim 10$. They also predict a negative signal, meaning absorption against the CMB, at higher redshift, prior to IGM warming, but allowing for Ly$\alpha$ resonant scattering (ie. the era of ’Ly$\alpha$ coupling [@BL04]). The shaded region shows the potential system thermal noise of a well calibrated low frequency experiment.
[@GS03] point out that detecting this signal against the mean non-thermal foreground radiation may be difficult. The foreground is the sum of relatively smooth Galactic synchrotron emission and discrete distant radio galaxies, with a typical spectral index, $\alpha
\sim -0.8$, where $\alpha$ is defined as: $S_\nu \propto \nu^\alpha$. The foreground temperature behaves roughly as $T_{FG} \sim 100
(\frac{\nu}{200 \rm{MHz}})^{-2.8}$, in the coldest regions of the sky, and can be an order of magnitude higher in the Galactic plane. Hence, the expect change in $T_B$ is at most $\sim 10^{-4}$ that of the mean foreground signal over the frequency range $\sim 100$ to 200 MHz.
Large scale structure
---------------------
### Power spectra and tomography
The majority of theoretical studies have focused on predicting the HI fluctuations in the IGM on scales of arcmins to a few degrees both during [@ZS04; @MMR97; @tozzi99; @FZH04; @FSH03; @BL04; @Ali05; @BA04; @GS03; @SCK05; @FMH05; @Wang05], and prior to [@CM03; @BL05; @LZ04], reionization. Most studies have focused on the diffuse IGM, although a few studies have also considered the signal expected due to the clustering of collapsed structures [@Iliev03].
The HI 21cm signal from the IGM during reionization can be predicted both analytically, using a standard Press-Schechter type of analysis of linear structure formation, plus some recipes to approximate non-linear evolution [@SCK05; @ZFH04; @Ali05; @BA04; @GS03], or through the use of numerical simulations [@CM03; @FSH03].
Fig 3 shows the expected evolution of the HI 21cm signal during reionization based on the numerical simulations of [@FSH03]. They find that the mean HI signal is about $T_B \sim 25$ mK prior to reionization, with fluctuations of only a few mK on arcmin scales due to linear density evolution. In this simulation, the HII regions caused by galaxy formation during reionization are seen in the redshift range $z \sim 8$ to 10, reaching scales up to 2$'$ (frequency widths $\sim 0.3$ MHz $=> 0.5$ Mpc physical size). These regions have (negative) brightness temperatures up to 20 mK relative to the mean HI signal. This corresponds to 5$\mu$Jy beam$^{-1}$ in a 2$'$ beam at 140 MHz.
We compare this signal to the sensitivity of a radio array. The point source rms sensitivity (dual polarization) in an image from a synthesis radio telescope is given by: $${\rm rms} = ~
(\frac{70}{(\Delta\nu_{kHz} t_{hr})^{0.5}}) (\frac{T_{sys}}{300 {\rm K}})
(\frac{0.60}{\epsilon_{eff}}) (\frac{500 {\rm m}^2}{A_{ant}})
(\frac{27}{N_{ant}}) ~~ \rm{mJy ~ beam^{-1}}$$ where $\Delta \nu$ is the channel width in kHz, $t$ is the integration time in hours, $T_{sys}$ is the total system temperature, $A$ is the collecting area of each element in the array, $\epsilon$ is the aperture efficiency, and $N$ is the number of elements ($\epsilon
A N$ = total effective collecting area of the array). At low frequency the value of $T_{sys}$ is dominated by the non-thermal foreground, and again behaves as $T_{sys} \sim 100 (\frac{\nu}{200 \rm{MHz}})^{-2.8}$ K. The beam FWHM is given by the Fourier transform of the uv coverage, assuming equal weight for each visibility.
Consider an array with an effective collecting area of 1 square kilometer at 140 MHz, distributed over 4 km, a system temperature of 300 K, a channel width of 0.3 MHz, and integrating for one month. The rms sensitivity is then 1.3$\mu$Jy beam$^{-1}$, with a beam FWHM $\sim
2'$. This square kilometer array will be just adequate to perform true three dimensional imaging of the average structure of the IGM during reionization.
Unfortunately, the nearer term low frequency ’path finder’ arrays will have $\le 10\%$ the collecting of the SKA, and will likely not be able to perform such direct 3D imaging (Table 1). However, these near term experiments should have enough sensitivity to perform power spectral analyses of the HI 21cm fluctuations. For power spectral analyses the sensitivity is greatly enhanced relative to direct imaging due to the fact that the universe is isotropic, and hence one can average the measurements in annuli in the uv-plane, ie. the statistics of fluctuations along a uv annulus are equivalent. The width of independent uv annuli is set by the primary beam diameter, D, or $\Delta l = 2 \pi \frac{D}{\lambda_{obs}}$. The situation is analogous to COBE and WMAP. COBE lacked the sensitivity to make a true image of the CMB fluctuations, but was able to make a robust determination of the power spectrum of fluctuations. WMAP was able to produce a proper image of the CMB.
Many authors have considered the power spectrum of brightness temperature fluctuations in the HI 21cm line [@SCK05; @ZFH04; @Ali05; @BA04; @Morales04; @SKES98; @BMH05]. It can be shown that the noise power spectrum, in mK$^2$, in standard spherical harmonic units, $\frac{l(l+1)}{2 \pi} C_l^N$, for an array with uniform coverage of the uv plane, is given by:
$$C_l^N = \frac{T_{sys}^2 (2\pi)^3}{\Delta \nu t f_c^2 l_{max}^2}$$
where $f_c$ = areal covering factor of the array $= N_{ant}
\frac{A_{ant}}{A_{tot}}$, and $A_{tot}$ is the total area of the area defined by the longest baseline. For example, for the logarithmic antenna spacing of the VLA in the smallest configuration, the VLA-VHF system has $f_c = 0.017$ on baselines out to 1 km, and $f_c = 0.17$ out to 0.25 km, for a full synthesis observation. For comparison, the SKA will have 30$\%$ of it’s collecting area inside 1 km, or $f_c =
0.3$ out to 1 km.
Most predictions of the HI 21cm power spectrum during reionization published thus far parallel the 2D calculations applied to the CMB. Ultimately, higher sensitivity, and physical insight, will come through full 3D analysis of the HI 21cm fluctuations during reionization [@Morales04]. Of course, at the smallest $l$, approaching the primary beam size $\Delta l$, the signal-to-noise is limited by cosmic variance to essentially the number of synthesized beams per primary beam.
As an example, Fig 4 shows the predicted evolution of the HI 21cm power spectrum using an analytic calculation of structure formation, including a model of (uncorrelated) HII region formation during reionization [@FSH03; @ZFH04] (see also [@BA04; @SCK05; @Wang05] for analytic solutions, and [@FSH03; @CM03] for predictions based on numerical simulations). Fig 4 includes the power spectrum arising from linear density fluctuations following the dark matter, plus the effect of reionization. The signature of reionization can be seen as a bump in the power spectrum above the density-only curve due to the formation of HII regions. In this calculation, the rms fluctuations at $z = 10$ peak at about 10 mK rms (observing frequency = 130 MHz) on scales $l \sim 3000$[^1]. Note that in this calculation the neutral fraction as a function of redshift is a free parameter, and is assumed to be 0.5 at $z=10$. [@BL04] point out that there is no very large scale information contained in the HI 21cm fluctuations, and that power spectral analyses are best done on scales ranging from $l = 100$ to a few thousand, ie. on angular scales $\le$ few degrees.
Also included are the noise power spectra for near and longer term radio arrays. Typical noise values for near-term arrays are $\sim 1$ to 10 mK rms in the range $l = 10^3$ to $10^4$, for long integrations. This should be adequate to determine the HI 21cm power spectrum during reionization. The SKA will go an order of magnitude deeper, providing an accurate measure of the HI power spectrum, and its evolution with redshift.
[@BL05; @Ali05] have shown that gas peculiar velocities, ie. infall into superclusters and along filaments (analogous to the Kaiser effect for galaxy clustering), increases the expected rms fluctuations by about a factor two. Also, clustering of luminous sources (ie. biased galaxy formation) implies a characteristic scale for the HII regions of 1 to 10 Mpc (comoving), and leads to another factor two in fluctuation strength [@FMH05; @SCK05] over the random distribution of HII regions assumed by [@ZFL04]. Specifically, this later effect moves the fluctuation power to larger scales.
[@ZFL04; @BA04; @FSH03] consider the effect of spectral channel width on the power spectral analysis. Bandwidth enters via the finite thickness of the HII regions and the density fluctuations. They conclude that a bandwidth of about 0.2 MHz is optimal for power spectral studies, although the rms of the signal changes by only a factor of about two going from 0.1 to 0.5 MHz channel width.
[@FSH03] show how the power spectra change with different reionization models, including outside-in versus inside-out models, uniform versus HII region dominated reionization, and double reionization. They point out that HI 21cm power spectral studies should provide powerful constraints on the process of reionization, with different reionization scenarios clearly discernable in their power spectral signature. Note that this is unlike CMB large scale polarization measurements, which provide reasonable constraints on the total Thompson scattering optical depth through the universe back to recombination, but only marginal constraints on the reionization history [@HuHolder03].
Lastly, [@BA04; @nusser04] point out that if full 3D tomography of the evolution of IGM structures could be made, then one might also constrain the geometry of universe through the standard Alcock-Pacinsky effect, ie. the difference in structure evolution observed in angle versus along the line-of-sight, due to the non-Euclidean geometry of the univese.
### Clustering of minihalos
A point of debate has been the fraction of HI in collapsed objects, as opposed to the diffuse IGM [@Iliev03]. This fraction has a complex depence on structure formation history, and [@OM03; @Gnedin04] conclude that the majority of the HI during reionization will remain in the diffuse phase. However, at times when $T_S$ approaches $T_{CMB}$ in the diffuse IGM, one might expect HI 21cm emission due to clustering of minihalos[^2], in which $T_S > T_{CMB}$ simply due to the virialization process [@FSH03].
[@Iliev03] have considered the HI 21cm power spectrum due to clustering of minihalos at high redshift. For a beam size of 1$'$, and channel width of 0.2 MHz, they predict 3$\sigma$ brightness temperature fluctuations due to clustering of minihalos $\sim 7$mK at $z = 8.5$, decreasing to 2mK at $z=20$.
### Other tests
[@CF05] discuss circular polarization of the 21cm line due to zeeman splitting during reionization. Such polarization would be detectable with the SKA if the IGM field strength were 100 $\mu$G. [@Salvaterra05; @Cooray04] consider the anti-correlation between HI 21cm fluctuations and sub-degree scale CMB secondary anisotropies due to Thompson scattering in ionized regions during reionization at $l >
1000$ (kinetic SZ effect). A natural anticorrelation would be expected between the neutral and ionized regions. And lastly, [@SC05] discuss delensing of the CMB using the HI 21cm fluctuations to pin-point density inhomogenaeities along the line of sight ($z \sim 10$ to 200). This could prove important for decoupling the E and B mode mixing that occurs due to lensing, possibly recovering the intrinsic (inflationary) B polarization.
### Cosmic Stromgren spheres
While direct detection of the typical structure of HI and HII regions may be out of reach of the near-term 21cm telescopes, there is a chance that even this first generation of telescopes will be able to detect the rare, very large scale HII regions associated with luminous quasars near the end of reionization.
The presence of cosmic Stromgren spheres around the highest redshift SDSS QSOs has been deduced from the observed difference between the redshift of the onset of the GP effect and the systemic redshift of the host galaxy [@White03; @WL04a; @WLC05; @Walter03] (although cf. [@OF04]). The physical size of these spheres is $\sim 5$ Mpc, or an order of magnitude larger than the typical spheres expected from clustered galaxy formation, due to the extreme luminosity of the sources ($\sim 10^{14}$ L$_\odot$). If the neutral fraction remains substantial to relatively low redshift, f(HI) $> 0.1$ [@MH03; @WL04a; @WLC05], then it is plausible to search for these regions as ’holes in the sky’, or regions of negative brightness in the redshift 21cm line frequency. The expected signal is $\sim 20$mK $\times$ f(HI) on a scale $\sim 10'$ to 15$'$, with a line width of $\sim 1$ to 2 MHz [@WL04b]. This corresponds to 0.5 $\times$ f(HI) mJy beam$^{-1}$, for a 15$'$ beam.
Searching for HI signals around known high redshift QSOs has a number of major observational advantages: (i) the exact location in frequecy, and RA and DEC, are well determined, thereby limiting the search space dramatically, and (ii) there is already evidence for the existence of the features based on the Ly$\alpha$ spectra. The VLA-VHF system has been specifically design to search for these spheres around known SDSS quasars at $z \sim 6$ to 6.4 [@VHF]. Fig 5 shows the simulated image and spectrum of such a Stromgren sphere around an SDSS QSO assuming f(HI) = 1, as observed with the VLA-VHF system. While it is unlikely the mean neutral fraction is this high at such a low redshift, in 250 hours the VLA-VHF system should set the first direct constraints on neutral fraction of the IGM, at the level f(HI) $\ge
0.2$, as well as easily rule out more extreme models, such as $T_S <<
T_{CMB}$. [@WLB05] present similar simulations for the MWA and the SKA.
-2.4in
[@Kohler05] also discuss spectral dips due to large HII regions around luminous quasars ($> 2\times 10^{10}$ L$_\odot$) during the EoR. Their Fig 2 shows a spectrum in a typical 10$'$ beam from 100 to 180 MHz. They predict on average one relatively deep (-2 to -4 mK) dip per LoS on this scale. They emphasize that these spectral features may be easier to detect than spatial fluctuations of a similar magnitude, due to the fact that spatial confusion is highly structured on arcsec to arcmin scales, while the spectral confusion should be smooth over 10’s of MHz.
[@WLB05] perform a similar calculation using the evolution of the bright QSO luminosity function to predict the number of HII regions around active QSOs at $z > 6$. They conclude that there should be roughly one SDSS-type HII region around an active QSO (physical radius $>$ 4Mpc) per 400 deg$^2$ field per 16 Mhz bandwidth at $z \sim 6$, and one $R \ge 2$ Mpc region at $z \sim 8$. They also point out that the recombination time may be longer than the Hubble time, so that fossil HII regions may be observed around non-active AGN. Assuming a duty cycle $\sim \frac{<t_{qso}>}{t_{H}(z)} \sim \frac{10^7}{10^9}
\sim 0.01$, where $<t_{qso}>$ is the fiducial lifetime of QSO activity [@YT99], and $t_{H}(z)$ is the age of the universe, leads to a factor 100 more fossil HII regions per FoV.
### Beyond reionization
A number of authors have considered the brightness temperature fluctuations due to the 21cm line from the neutral IGM at redshifts prior to reionization, $z > 20$. [@BL04; @BL05] predict the power spectrum of fluctuations during the era when the Ly$\alpha$ photons from the first luminous objects couple $T_S$ to $T_K$ locally, but the universe remains substantially neutral ($z \sim 20$ to 30), and $T_K < T_{CMB}$ in the diffuse IGM. Brightness temperature fluctuations can be due to emission from clustered minihalos, plus enhanced by absorption against the CMB by the diffuse IGM, with an rms $\sim$ 10mK for $l \sim 10^5$, due to a combination of linear density fluctuations, plus poisson (’shot’) noise and biasing in the Ly$\alpha$ source (ie. galaxy) distribution.
[@LZ04] go even further in redshift, to $z > 50$ to 200. In this regime the HI generally follows linear density fluctuations, and hence the experiments are as ’clean’ as CMB studies, and $T_K < T_{CMB}$, so a relatively strong absorption signal might be expected. [@LZ04] also point out that Silk damping, or photon diffusion, erases structures on scales $l > 2000$ in the CMB at recombination, corresponding to comoving scales = 22 Mpc. The HI 21cm measurements can explore this physical regime at $z \sim 50$ to 300. The predicted rms fluctuations are 1 to 10 mK on scales $l= 10^3$ to 10$^6$ (0.2$^o$ to 1$''$). These authors point out that, due to sensitivity to large $l$, and the 3D nature of the information, HI 21cm power spectral studies in this epoch “contain an amount of information orders of magnitude larger than any other cosmological probe”. These data could provide the best tests of non-Gaussianity of density fluctuations, and for constrainting the tilt, or running power law index, of mass fluctutions to large $l$, providing important tests of inflationary structure formation. [@sethi05] also suggests that a large global signal, up to -0.1 K, might be expected for this redshift range.
Two experiments are being planned to cover this low frequency range, the long wavelength array (LWA) in New Mexico, and in the longer term, LUDAR is being considered for the back side of the moon.
Absorption toward discrete radio sources
----------------------------------------
Observing HI 21cm emission from the EoR implies studying large scale structure (cluster scales and larger). A number of groups have recently considered the possibility of studying smaller scale structure in the neutral IGM by looking for HI 21cm absorption toward the first radio-loud objects (AGN, star forming galaxies, GRBs) [@Carilli04a].
[@CGO02] use numerical simulations to predict the HI 21cm absorption profile of the ’cosmic web’ prior to reionization. For example, for a source at $z = 10$, they predict an average optical depth due to 21cm absorption of about 1$\%$, corresponding to the ’radio Gunn-Peterson effect’. They also find about five narrow (few km/s) absorption lines per MHz with optical depths of a few to 10$\%$ (Fig 2b). These latter lines are equivalent to the Ly $\alpha$ forest seen after reionization, and correspond to over-densities evolving in the linear regime ($\delta
\le 10$). While significant questions remain about simulating the thermal state of the IGM to such detail, the simple point remains that, while the Ly$\alpha$ lines from such structures are highly saturated, the (much) lower 21cm oscillator strength makes the IGM translucent prior to reionization.
[@FL02; @OH03] predict a similar HI 21cm absorption line density due to gas in minihalos as that expected for the 21cm forest. [@Cen02] shows that the presence or absence of HI absorption by minihalos, or the cosmic web, would be a telling diagnostic of early IGM heating mechanisms. [@FL02] also consider the expected 21cm absorption profiles for proto-disk galaxies. While such absorption lines will be rare ($10^4$ times less frequent than the 21cm forest lines), the optical depths may be large enough that the lines could be observed toward faint radio sources, in particular, gamma ray burst radio after-glows within the host galaxy [@IM05].
An important caveat in these calculations is the assumption of radio loud sources during the EoR. This question has been considered in detail by [@CGO02; @haiman04; @JR05]. They show that current models of radio-loud AGN evolution predict between 0.05 and 1 radio sources per square degree at $z > 6$ with $\rm S_{150MHz} \ge 6$ mJy, adequate for EoR HI 21cm absorption studies with the SKA.
Observational challenges
========================
Foregrounds
-----------
It has long been recognized that the HI 21cm signal from reionization must be detected on top of a much larger non-thermal (synchrotron) signal from foreground emission. This foreground includes discrete radio galaxies, and large scale emission from our own Galaxy. The expected HI signal is about $10^{-4}$ of the foreground emission.
Many groups have considered the effects of the foregrounds on HI 21cm EoR imaging and power spectral studies. [@dMat01] show that, even if point sources can be removed to the level of 1 $\mu$Jy, the rms fluctutions on spatial scales $\le 10'$ ($l \ge 1000$) due to residual radio point sources will be $\ge 10$ mK just do to Poisson noise, increasing by a factor 100 if the sources are strongly clustered. Clearly this conclusion depends on the extrapolation of radio source populations to $\le 1\mu$Jy, but under reasonable assumptions, even in the Poisson case the residual source fluctuations will be comparable to the HI signal.
This calculation has led many groups to consider removal of foregrounds in the spectral domain. The important point is that the foregrounds should be relatively smooth in frequency, predominantly the sum of power-law, low frequency non-thermal spectra, or perhaps gently curving on spectral scales $\ge 10$’s MHz. The HI signal should show significant structure on sub-MHz scales, corresponding to the typical size scale for features during reionization.
A number of complimentary approaches have been presented for foreground removal. [@GS03; @Wang05] consider removal of the foregrounds through fitting of smooth spectral models (power laws or low order polynomials in log space) to the observed visibilities or images. [@MH03; @Morales04] present a 3D Fourier analysis of the measured visibilities, where the third dimension is frequency. They show that the different symmetries in this 3D space for the signal arising from the noise-like HI emission, versus the smooth (in frequency) foreground emission, can be a powerful means of differentiating between foreground emission and the EoR line signal. [@SCK05; @BA04; @ZFH04] perform a similar analysis, only in the complementary Fourier space, meaning cross correlation of spectral channels. They show that the 21cm signal will effectively decorrelate for channel separations $> 1$ MHz, while the foregrounds do not, and this too can be a means of separating the foregrounds from the desired 21cm fluctuations. The overall conclusion of these methods is that spectral decomposition should be adequate to separate non-thermal foregrounds from the HI 21cm signal from reionization at the mK level.
However, there are two potentially more insideous observational challenges relating to non-thermal foregrounds beyond simple in-beam confusion. First, there can be frequency dependent sidelobes from confusing sources in the very wide fields being considered. A particularly problematic effect has been the recent observation that telescope spectral response (ie. bandpass) may depend on position of a given source in the primary beam [@Oosterloo05]. The origin of this effect remains uncertain (perhaps relating to scattering and blocking structures), but it will lead to frequency dependent sidelobes that potentially affect widefields and are not removed through normal (on-axis) bandpass calibration, and may require a spatially dependent bandpass calibration.
And second is polarized structure in the diffuse emission in the field. Wide field polarization has been seen with low frequency observations with the WSRT at 330 MHz and below, and is thought to arise due to differential Faraday depth through the ISM [@haverkorn]. Hence, great care is required to control, and calibrate, the wide field polarization response of the primary elements.
Ionosphere
----------
A second potential challenge to low frequency imaging over wide fields is phase fluctuations caused by the ionosphere. These fluctuations are due to index of refraction fluctuations in the ionized plasma, and behave as $\Delta \phi \propto \nu^{-2}$. Morever, the typical ’isoplanatic patch’, or angle over which a single phase error applies, is a few to 10 degrees (physical scales of 10’s km in the ionosphere), depending on frequency [@Cotton04; @lane]. Fields larger than the isoplanatic patch will have multiple phase errors across the field, and hence cannot be corrected through standard (ie. single solution) phase self-calibration techniques.
New wide field self-calibration techniques, involving multiple phase solutions over the field, or a ’rubber screen’ phase model [@Cotton04; @hopkins], are being developed that should allow for self-calibration over wide fields. However, even self-calibration techniques will be insufficient to overcome the very rapid variations caused by the occasional ionospheric storm, or traveling ionospheric disturbances.
At very low frequency ($\le$ a few MHz), one approaches the plasma frequency of the ionosphere, and the optical depth increases dramatically, precluding observations from the ground.
Interference
------------
Perhaps the most difficult problem facing low frequency radio astronomy is terrestrial (man-made) interference. The relevant frequency range corresponds to 7 to 200 MHz ($z= 200$ to 6). These are not protected frequency bands, and commercial allocations include everything from broadcast radio and television, to fixed and mobile communications.
Many groups are pursuing methods for RFI mitigation and excision (see extensive references at [@RFI]). These include: (i) using a reference horn, or one beam of a phased array, for constant monitoring of known, strong, RFI signals, (ii) conversely, arranging interferometric phases to produce a null at the position of the RFI source, and (iii) real-time RFI excision using advanced filtering techniques in time and frequency, of digitized signals both pre- and post-correlation. This requires very high dynamic range (many bit sampling), and very high frequency resolution.
One obvious advantage of HI absorption experiments (section 3.3) is that they can be done using long baselines (10’s to 100’s of km). Long baselines lead to decorrelation of the terrestrial signal due to fringe tracking of the celestial source. However, the fringe rates on the short baselines required for the HI emission experiments ($\le$ few km) are low, such that decorrelation due to fringe winding will not be a very effective RFI filter on the shorter baseline.
In the end, the most effective means of reducing interference is to go to the remotest sites. [@MWA; @PAST] have selected sites in remote regions of Western Australia, and China, respectively, because of known low RFI environments. Of course, the ultimate location would be the back side of the moon.
Telescopes
==========
Many programs have been initiated to study the HI 21cm signal from cosmic reionization, and beyond. These are summarized in Table 1. A number of different approaches are being taken, both techinically and in order to address different aspects of the problem [@Morales04].
The largest near-term efforts are the Mileura Wide Field Array (MWA), the Primeval Structure Telescope (PAST), and the Low Frequency Array (LOFAR). These are being optimized to study the power spectrum of the HI 21cm fluctuations, although in principle they will be able to image the larger HII regions during reionization. The VLA-VHF system is designed specifically to set limits on the HII regions around $z\sim 6$ to 6.4 SDSS QSOs, although it should also constrain the late-time power spectrum. In the long term the Square Kilometer Array should have the sensitivity to perform true three dimensional imaging of the neutral IGM in the 21cm line during reionization. And at the lowest frequencies ($< 50$ MHz), the Long Wavelength Array (LWA), and eventually the lunar array (LUDAR [@LUDAR; @maccone]), are being designed for the higher $z$ signal, prior to reionization.
As discussed above, the technical challenges are many. Use of spectral decomposition to remove the foregrounds requires careful control of the synthesize beam as a function of frequency, with the optimal (although difficult) solution being a telescope design where the synthesized beam is invariant as a function of frequency [@MarkI]. High dynamic range front ends are required to avoid saturation in cases of strong interference, while fine spectral sampling is required to avoid Gibbs ringing in the spectral response. The polarization response must be stable and well calibrated. Calibation in the presence of a significant ionospheric phase screen requires new wide field calibration techniques. The very high data rate expected for many element ($\ge 10^3$) arrays requires new methods for data transmission, cross correlation, and storage [@MWA]. At the lowest frequencies, $\le 20$MHz or so, where we hope to study the pre-reionization IGM, phase fluctuations and the opacity of the ionosphere becomes problematic, leading to the proposed LUDAR project on the far side of the moon. The far side of the moon is also the best location in order to completely avoid terrestrial interference.
[cccccccccc]{} & Freq & Area & B$_{max}$ & Site & Type & FoV & Date & Goal & Ref\
& MHz & 10$^4$ m$^2$ & km & & & deg & & & \
GMRT & 150 – 165 & 3.7 & 10 & India & Parabola & 4 & 2000 & StromSph & [@GMRT]\
PAST & 50 – 200 & 7 & 2 & China & Dipole & 10 & 2006 & PowSpec & [@PAST]\
VLA-VHF & 180 – 200 & 1.3 & 1 & USA & Parabola & 4 & 2006 & StromSph & [@VHF]\
Mark I & 100 – 200 & & 0 & Aus & Spiral & 180 & 2006 & Global & [@MarkI]\
MWA-LFD$^1$ & 80 – 300 & 3 & 1.5 & Aus & Dipole & 20 & 2007 & PowSpec & [@MWA]\
LOFAR$^2$ & 115 – 240 & 10 & 2km=40% & NL & Tiles & 20 & 2007 & PowSpec & [@LOFAR]\
LWA & 10 – 88 & 10 & 5km=30% & USA & Dipole & 20 & 2008 & Pre-reion & [@LWA]\
SKA & 100 – 200 & 100 & 5km=50% & ? & Dipole & 20 & 2015 & Imaging & [@SKA]\
LUDAR & 5 – 50 & 1000 & 100 & Moon & Dipole & 60 & 2020 & pre-reion & [@LUDAR]\
$^1$An order of magnitude increase in area is planned after testing the Low Frequency Demonstrator.
$^2$LOFAR will also have a low frequency component between 30 and 80 MHz.
Acknowledgements: Thanks to the Max-Planck Society and the Alexander von Humboldt Foundation for support through the Max-Planck Research Prize, and S. Furlanetto and N. Gnedin for permission to use figures.
[99]{}
Ali, Sk. 2005, MNRAS, in press, (astroph-0503237)
Barkana, R., Loeb, A. 2005, ApJ, 626, 1-11
Barkana, R., Loeb, A. 2005, ApJ, 624, L65-68
Barkana, R., Loeb, A. 2001, Phys. Rep. 349, 125-238
Bebbington, D. 1986, MNRAS, 218, 577-585
Bharadwaj, S. & Ali, Sk. 2005, MNRAS, 356, 1519-1428
Bowman, J, Morales, M., Hewitt, J. 2005, ApJ, in press (astroph-0507357)
Ciardi, B., Madau, P. 2003, ApJ, 596, 1-8
Ciardi, B., Stoehr, F., White, S. MNRAS, 343, 1101-1109
Carilli, C., Gnedin, N., Furlanetto, S., & Owen, F. 2004, NewAR, 48, 1053-1061
Carilli, C., Furlanetto, S., Briggs, F., Jarvis, M., Rawlings, S., Falcke, H. 2004, NewAR, 48, 1029-1038
Carilli, C., Gnedin, N., Owen, F. 2002, ApJ, 577, 22-30
Cen, R. 2003, ApJ, 591, 12-37
Chen, X. & Miralda-Escude, J. 2004, ApJ, 602, 1-11
Cooray, A. 2004, Phys. Rev. D, 70, 3509-3513
Cooray, A., Furlanetto, S. 2005, MNRAS, 395, L47-52
Cotton, W., Condon, J., Perley, R. et al. 2004, SPIE, 5489, 180-189
Davies, R., Pedlar, A., Mirabel, I. 1978, MNRAS, 182, 727-733
de Bruyn, A., Wieringa, M., Katgert, P., Sancisi, R. 1988, IAUS 130, 211-214
Di Matteo, T., Rosalba, P., Abel, T., Rees, M. 2002, MNRAS, 564, 576-580
Field, G.B. 1959, ApJ, 129, 551-565
Furlanetto, S. & Briggs, F. 2004, New AR, 48, 1039-1052
Furlanetto, S, Loeb, A. 2002, ApJ, 579, 1-9
Furlanetto, S., Zaldarriaga, M. Hernquist, L. 2004, ApJ, 613, 16-22
Furlanetto, S., Sokasian, A., Hernquist, L. 2004, MNRAS 347, 187-195
Furlanetto, S., McQuinn, D., Hernquist, L. 2005, ApJ, in press (astroph-0507524)
Gnedin, N. & Shaver, P. 2004, ApJ, 608, 611-621
Gnedin, N. 2004, ApJ, 610, 9-13
Gnedin, N. 2000, ApJ, 535, 530-554
Haiman, Z., Quartaert, E., Bower, G. 2004, ApJ, 612, 698-705
Haverkorn, M., Katgert, P., de Bruyn, A. 2004, A& A, 427, 549-559
Hopkins, P., Doeleman, S., Lonsdale, C. 2003, AAS, 203, 4005
Hirata, C.M. 2005, ApJ, in press (astroph-0507102)
Hogan, C., Rees, M. 1979, MNRAS, 188, 791-798
Hu, W. & Holder, G. Phys. Rev. D, 68, 3001-3004
Iliev, I., Scannapieco, E., Martel, H., Schapiro, P. 2003, MNRAS, 341, 81-90
Ioka, K. & Meszaros, P. 2005, ApJ 619, 684-696
Jarvis, M. & Rawlings, S. 2005, New AR, 48, 1173-1185
Kohler, K., Gnedin, N., Miralda-Escude, J., Shaver, P. 2005, ApJ, in press (astroph-0501086)
Lane, W., Cohen, A., Cotton, W., et al. 2004, SPIE, 5489, 354-361
Loeb, A., Barkana, R. 2001, ARAA, 39, 19-66
Loeb, A. & Zaldarriaga, M. 2004, Phys.Rev. Lett. 92, 1301-1304
Madau, P., Meiksin, A., Rees, M. 1997, ApJ, 475, 429-444
Meiksin, A. 2000, “Science with Large Antenna Arrays,” (ASTRON: Netherlands), ed. M. van Haarlem p. 37-45
Miralda-Escude, J. 2003, Science, 300, 1904-1909
Mesinger, A., Haiman, Z. 2004, ApJ, 611, L69-72
Morales, M. & Hewitt, J. 2004, ApJ, 615, 7-18
Morales, M. 2005, ApJ, 619, 678-683
Nusser, A. 2005, MNRAS, 359, 183-190
Oh, S.P. & Furlanetto, S. 2005, ApJ, 620, L9-L12
Oh, S.P. & Haiman, Z. 2004, MNRAS, 346, 456-472
Oh, S.P. & Mack, K. 2003, MNRAS, 346, 871-877
Oort, M., Steemers, W., Windhorst, R. 1988, A & A Supp, 73, 103-123
Oosterloo, T. 2005, in [*SKA Workshop, Wide Field Imaging*]{}, (ASTRON: Netherlands), in press
Pritchard, J. & Furlanetto, S. 2005, MNRAS, in press (astroph-0508381)
Salvaterra, R., Ciardi, B., Ferrara, A., Baccigalupi, C. 2005, MNRAS, 360, 1063-1068
Santos, M., Cooray, A., Knox, L. 2005, ApJ, 625, 575-587
Scott, D., Rees, M.J. 1990, MNRAS, 247, 510-516
Sethi, S. 2005, MNRAS, in press (astroph-0508172)
Shaver, P., Windhorst, R., Madau, P., de Bruyn A. 1999, A& A, 345, 380-390
Sigurdson, B. & Cooray, A. 2005, ApJ, in press, astroph-0502549
Spergel, D., Verde, L., Peiris, H. et al. ApJS, 148, 175-194
Subrahmanyan, R., Kesteven, M., Ekers, R., Sinclair, M., Silk, J. 1998 MNRAS 298, 1189-1197
Sunyaev, R., Zeldovich, Ya. 1972, A&A, 20, 189-200
Tozzi, P., Madau, P., Meiksin, A., Rees, M. 2000, ApJ, 528, 597-606
Uson, J., Bagri, D., Cornwell, T. 1991, Phys. Rev. Let, 67, 3328-3331
Walter, F., Bertoldi, F., Carilli, C., et al. 2003, Nature, 424, 406-408
Wang, X., Tegmark, M., Santos, M., Knox, L. 2005, ApJ, in press (astroph-0501081)
White, R., Becker, R., Fan, X., Strauss, M. 2003, AJ, 126, 1-14
Wouthuysen, S. 1952, AJ, 57, 31-33
Wyithe, J.S., Loeb, A., Carilli, C. 2005, ApJ, 628, 575-582
Wyithe, J.S., Loeb, A. 2004, Nature, 432, 194-196
Wyithe, J.S., Loeb, A. 2004, ApJ, 610, 117
Wyithe, J.S., Loeb, A., Barnes, D. 2005, ApJ, submitted, (astroph-0506045)
Yu, Q.& Tremaine, S. 2002, MNRAS, 335, 965-976
Zaldarigga, M., Furlanetto, S., Henquist, L. 2004, ApJ, 608, 622-635
Zaldarigga, M., Furlanetto, S., Loeb, A. 2004, ??
Zaroubi, S. & Silk, J. 2005, MNRAS, 360, L64 -67
Zygelman, B. 2005, ApJ, 622, 1356-1362
http://www.atnf.csiro.au/SKA/intmit/
web.phys.cmu.edu/past/
www.lofar.org/
Corbin, M. et al. 2005, exploratory proposal to NASA
Maccone, C. 2004, 35th COSPAR Assembly, p. 1415-1419
web.haystack.mit.edu/arrays/MWA/LFD/index.html
cfa-www.harvard.edu/dawn/
www.skatelescope.org/
lwa.unm.edu/index.shtml
www.ncra.tifr.res.in/ncra\_hpage/gmrt/gmrt.html
Subrahmanyan, R., Chippendale, A., Ekers, R. 2005, ATNF newsletter, 56, 18-19
[^1]: or $u = \frac{l}{2\pi} = 500$, or $\theta \sim \frac{180^o}{l} = 3.6'$, or baselines, $b = \lambda u = 1.2$ km, or wavenumber $k \sim
\frac{l}{10^4} {\rm Mpc}^{-1} = 0.3 {\rm Mpc}^{-1}$ (comoving).
[^2]: Minihalos have masses $< 10^7$ M$_\odot$, and virial temperatures $< 10^4$ K. These halos cannot cool via atomic hydrogen lines, and therefore ’saturate’ at over-densities of $\delta \sim 100$. They will not form stars unless an alternate cooling mechanism can be found. One candidate is molecular hydrogen, which is an effective coolant down to virial temperatures of 100 K [@BL00].
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'It is addressed the issue of black holes with nonlinear electromagnetic field, focussing mainly in the Born-Infeld case. The main features of these systems are described, for instance, geodesics, energy conditions, thermodynamics. Also it is revised some black hole solutions of alternative nonlinear electrodynamics and its inconveniences.'
author:
- |
N. Bretón and R. García-Salcedo.\
Departamento de Física, Cinvestav-IPN, Mexico City.
title: Nonlinear electrodynamics and black holes
---
Introduction
============
Since the begining of the past century, proposals were made of nonlinear electrodynamics with the aim to cure the singularity of Maxwell’s solution to the field of a point charge at the charge’s position. Among the most succesful is the formulation of Born and Infeld (BI). The proposal by Born and Infeld [@BI] in 1934 has several desirable properties in a physical theory, for instance finiteness of the electric field and energy at the charge’s position, the freedom of duality rotations, propagation of discontinuities of the electromagnetic field on single characteristic surfaces, among others.
Born and Infeld inspired in a finiteness principle chose a particular “nonlinear” action with a maximum field strength $b$; they solved the resulting field equations for the static spherically symmetric solution corresponding to a point charge. The resulting field was very different from a Coulombian field in the neighborhood of the point charge where the fields are most intense; they found, in fact, that the field was finite everywhere, including the location of the point charge.
Solutions to the Einstein equations coupled with Born-Infeld nonlinear electromagnetic theory were found as soon as 1937 by Hoffman and Infeld [@hoffmann]. Nonlinear electrodynamics coupled to general relativity was also explored by Peres [@Peres], finding some static spherically symmetric solutions as well as wavelike solutions.
Pellicer and Torrence [@PT] faced BI theory coupled with gravitational field and obtained a class of nonsingular static spherically symmetric (SSS) solutions, corresponding to a point charge source, that asymptotically behaves as a Reissner-Nordstrom (RN) solution. RN is the SSS solution of Einstein-Maxwell equations.
During the eighties, Plebañski [@Pleban1] and colaborators made an extensive study of nonlinear electrodynamics (NLED) in general relativity, in particular they addressed Einstein equations coupled to Born-Infeld field in metrics of type D in Petrov classification [@Pleban2], [@GSP]; amog these solutions there is the nonlinear generalization of the Reissner-Nordstrom solution. They also investigated the causal propagation of signals and gave a classification of the characteristic surfaces (along which discontinuities of the field propagate) for NLED [@Pleban3].
Demianski [@Demian] found in 1986 a SSS solution of Einstein-Born-Infeld equations that is regular at the origin, the so called [*EBIon*]{}. This solution has a particle like structure with a regular center.
Critics are the arbitrariness in choosing the NLED Lagrangian, since there is a lot of Lagrangians that fulfill the physical requirements like the linear weak field limit. In this sense, BI Lagrangian is exceptional since it can be considered as an effective Lagrangian for quantum electrodynamics (QED) in the one-loop approximation, as was shown by Heisenberg and Euler [@HE] and later on by Schwinger [@Schwinger] (see also the review by Delphenich [@Delphenich]). If another justification was needed it came from string theory, where solutions of the BI equations represent states of D-branes [@Gibbons].
In this contribution we review aspects of NLED black holes according to the following plan: In Sec. 2 NLED formalism is introduced; Sec. 3 deals with Einstein-NLED solutions that can be interpreted as black holes; we focus in the static spherically symmetric (SSS) solutions. Sec. 4 addresses the NLED black hole thermodynamics. In Sec. 5 the relationship between the ADM and horizon masses of NLED solitons and black holes is presented. Sec. 6 contains stability aspects of NLED SSS solutions and comments on recent related research are at the end.
NLED formalism
==============
NLED Lagrangian depends in nonlinear way of the elecvtromagnetic invariants. We assume that the nonlinear electromagnetic field can be described by a vector potential $A_{\mu}$,
$$F_{\mu \nu}=2A_{[\mu,\nu]},$$
$F_{\mu \nu}$ possesses just one independent invariant and one independent pseudo-invariant:
$$F=\frac{1}{4}F_{\alpha \beta}F^{\alpha \beta}, \quad
\tilde{G} =\frac{1}{4}F_{\alpha \beta} \tilde{F}^{\alpha \beta},$$
where $\tilde{F}^{\alpha \beta}$ denotes the dual of ${F}^{\alpha \beta}$ defined by $\tilde{F}^{\alpha \beta}=(i/2 \sqrt{-g})\epsilon^{\alpha \beta
\gamma \delta}F_{\gamma \delta}$. If NLED Lagrangian is to be invariant under Lorentz group including reflections, then it must depend on $F$ and $\tilde{G}^2$. The condition that for weak fields the nonlinear theory should approximate the linear one must be imposed as well.
($F, \tilde{G}$) and ($P, \tilde{Q}$) frameworks
------------------------------------------------
It is convenient to introduce the canonical formalism for the system, related to Lagrangian by a Legendre transformation. Defining
$$P^{\alpha \beta}= 2 \frac{\partial L}{\partial {F}_{\alpha \beta}}=
\frac{\partial L}{\partial F} {F}^{\alpha \beta}+\frac{\partial
L}{\partial \tilde{G}} \tilde{F}^{\alpha \beta},
\label{mateqFG}$$
$$H= \frac{1}{2}P^{\alpha \beta}F_{\alpha \beta}-L(F,G^2),$$
we can work with $H(F,\tilde{G}^2)$ or with $H=H(P,\tilde{Q}^2)$ the later depending on the invariants associated to $P^{\alpha \beta}$,
$$P=\frac{1}{4}P_{\alpha \beta}P^{\alpha \beta}, \quad
\tilde{Q} =\frac{1}{4}P_{\alpha \beta} \tilde{P}^{\alpha \beta}.$$
The material or constitutive equations (\[mateqFG\]) express $P_{\alpha
\beta}$ through ${F}^{\alpha \beta}$, $F$ and $\tilde{G}$ ; we shall assume that they can be inverted to express $F_{\alpha \beta}$ through ${P}^{\alpha \beta}$ and $P, \tilde{Q}$. These equations (also called Hamilton’s equations) are
$$F^{\alpha \beta}= 2 \frac{\partial H}{\partial {P}_{\alpha \beta}}=
\frac{\partial H}{\partial P} {P}^{\alpha \beta}+\frac{\partial
H}{\partial Q} \tilde{P}^{\alpha \beta}.$$
The $(P, \tilde{Q})$ framework is an alternative form of NLED obtained from the original one, the $(F, \tilde{G})$ framework, by a Legendre transformation. Physically reasonable conditions must be imposed on $H(P,
\tilde{Q})$: $H$ is real; for weak fields nonlinear effects are negligible and the theory must have the limit of linear electrodynamics, i.e. $H(P,
\tilde{Q})=P+O(P^2, \tilde{Q}^2)$; if parity is to be conserved, then under transformation of $\tilde{Q}$ into $- \tilde{Q}$, $H$ must stay invariant, therefore, $H$ depends on $P$ and $\tilde{Q}^2$. Often it is also required that strong or dominant energy conditions hold.
In this formalism, ${F}^{\alpha \beta}$ is the physically significant electromagnetic field tensor, while ${P}^{\alpha \beta}$ is a tensor with no direct physical meaning. While the various nonlinear field theories succeed in making ${F}^{\alpha \beta}$ well behaved, the tensor ${P}^{\alpha \beta}$ is, in general, divergent at the location of point sources. The reason is that $\delta$-function sources continue to play a role in these theories and ${P}^{\alpha \beta}$ [*absorbs*]{} the singularities of the sources allowing the field tensor ${F}^{\alpha
\beta}$ be well behaved.
Note that by passing from $(P, \tilde{Q})$ framework to $(F, \tilde{G})$, we are changing from one NLED theory, characterized by some lagrangian $L(P, \tilde{Q})$, to another, in general different, corresponding to the lagrangian $L(F, \tilde{G})$; in the case of Maxwell electrodynamics, both theories coincide, $L=F=H=P$.
The coupled gravitational and NLED equations are derived from the action
$$S = \int{d^4x \sqrt{-g} \{ R (16\pi)^{-1}-L \} },
\label{NLEDaction}$$
where $R$ denotes the scalar curvature, $g:= {\rm det} \vert g_{\mu \nu}
\vert$ and $L$, the electromagnetic part, is assumed to depend in nonlinear way on the invariants of $P_{\mu \nu}$, in the $(P, \tilde{Q})$ framework, or $L$ depending on the invariants of $F_{\mu \nu}$ in $(F,
\tilde{G})$ scheme. Let us refer here to the former one:
$$L=\frac{1}{2} P^{\mu \nu} F_{\mu \nu} -H(P, \tilde{Q}),
\label{Lagr}$$
The energy-momentum tensor and the scalar of curvature are given, respectively, by
$$\begin{aligned}
4 \pi T_{\mu \nu}&=& H_{,P} P_{\mu \alpha} P^{\alpha}_{ \nu}-g_{\mu
\nu}(2PH_{,P} +\tilde{Q}H_{, \tilde{Q}}-H), \nonumber\\
\label{Tmunu}
R&=& 8(PH_{,P} +\tilde{Q}H_{, \tilde{Q}}-H),\end{aligned}$$
where $\partial H/ \partial P= H_{,P}$. Note that the curvature scalar, $R$, and consequently the trace of $T_{\mu \nu}$, may differ from zero.
The Born-Infeld nonlinear electrodynamics is given by the structural function $H(P, \tilde{Q})$,
$$H=b^2 \left(1-\sqrt{1-{2P}/{b^2}+ \tilde{Q}^2/{b^4}} \right),
\label{BIH}$$
where $b$ is the maximum field strength and the relevant parameter of the BI theory.
NLED energy conditions
----------------------
Using a timelike vector, $V^{\alpha}$, $V_{\alpha}V^{\alpha} <1$, imposing local energy density being non-negative amounts to $T_{\mu
\nu}V^{\mu}V^{\nu} \ge 0$; while that the local energy flow vector be nonspacelike requires that $T_{\alpha \beta}T^{\alpha}_{
\gamma}V^{\beta}V^{\gamma} \le 0$; these are, respectively, the weak energy condition (WEC) and the dominant energy condition (DEC); both conditions hold provided
$$H_{,P} >0, \quad (PH_{,P} +\tilde{Q}H_{, \tilde{Q}}-H ) \ge 0.$$
The strong energy condition (SEC) $R_{\mu \nu}V^{\mu}V^{\nu} \ge 0$, using the Einstein equation can be settled as,
$$R_{\mu \nu}V^{\mu}V^{\nu}=8 \pi(T_{\mu \nu}V^{\mu}V^{\nu} +\frac{T}{2})
\ge 0.$$
Note that NLED matter can violate SEC if the trace of the energy-momentum tensor is negative enough; for instance, BI energy-momentum tensor can violate SEC. In the case of Maxwell ED ($T=0$) the fulfilment of WEC implies SEC.
NLED Black holes
=================
Hoffmann and Infeld (HI) [@hoffmann] solved the Einstein-Born Infeld coupled equations (EBI) for the spherically symmetric case, imposing the condition of regularity on the electromagnetic tensor $F_{kl}$ and its first derivative and the same condition to the metric tensor $g_{kl}$. The requirement that there be no infinities in $g_{kl}$ forces the identification of gravitational with electromagnetic mass. For the SSS line element
$$ds^2= - \psi dt^2+\psi^{-1}dr^2+r^2(d \theta^2 + \sin^2{\theta}d
\phi^2),
\label{sss-metric}$$
HI find the solution
$$\psi = 1- \frac{8 \pi}{r} \int_{0}^{r}{(\sqrt{r^4+1}-r^2)dr}.$$
However, this form leads to a conical singularity. In the same paper HI found the regular solution given by the fields $D_{,r}= 1/r$, $E_{,r}=
r^2/(r^4+1)$ and the metric function
$$\psi_{HI} = 1- \frac{k}{r}+\frac{8 \pi \gamma}{r}
\int_{0}^{r}{\left(r^2 \ln \left[\frac{r^4}{1+r^4}\right] \right)dr},$$
where $k$ is a constant of integration corresponding to $(-2m)$ in Schwarzschild solution. To have regularity at $r=0$ it must be taken $k=0$. Therefore we must consider as the gravitational mass the quantity $4 \pi \int_{0}^{r}r^4T^{t}_{t}dr$, that is the total electromagnetic mass within a sphere having its center at $r=0$. Thus the regularity condition shows that electromagnetic and gravitational mass are the same.
Pellicer and Torrence (PT) [@PT] following the lines of HI [@hoffmann] searched for a EBI SSS solution well behaved at the origin; they imposed the continuity of $F_{\mu \nu}$ and the Lagrangian in the neighborhood of the charge; continuity on $F_{\mu \nu}$ and $L(F)$ leads to conditions on $H(F)$. They considered $\tilde{G}=0$, but were not restricted to BI.
For the SSS line element (\[sss-metric\]) PT obtained
$$\psi_{PT}= 1+\frac{d}{r} + \frac{8 \pi}{r} \int_{0}^{r}{H(x)x^2dx},
\label{metrPT}$$
the comparison with Schwarzschild solution gives $d=2m$, the constant $d$ is related to the mass parameter $m$. If one considers the mass as arising from electromagnetic properties, one can put $m=d=0$. The electromagnetic field for a charge $e$ is given by
$$F_{\mu \nu}= -\frac{e}{r^2} \frac{\partial H(P,0)}{\partial P}2
\delta^{[0}_{\mu} \delta^{r]}_{\nu},
\label{emfieldPT}$$
also $P=-e^2/4r^4$. Regularity takes place if the following conditions hold: the integral in (\[metrPT\]) exists and is finite; the field (\[emfieldPT\]) must be finite as $r \to 0$; besides, asymptotically, for large $r$, $H(P,0) \approx P$. These conditions guarantee the solution be well behaved. Note that there is still some freedom in chosing the function $H(P)$ so it can be selected [*ad hoc*]{} and numerous examples may be built.
Related to this point we shall address some cases of regular black hole solutions obtained with ad hoc NLEDs.
The existence of two alternative ways, $(P, \tilde{Q})$ and $(F,
\tilde{G})$, of setting NLED, can generate some confusion. Let us refer to those regular electric black hole solutions that have been obtained recently [@Ayon]. In deriving the solutions, in the $(P,
\tilde{Q}=0)$ framework, an ad hoc NLED $H(P)$ was chosen to obtain a regular metric that describes black holes with regular center and RN asymptotics. The [*appropriate*]{} $H(P)$ was found imposing regularity conditions on the metric, in a similar way to the sketched by PT. The so constructed solutions [@Ayon] were using:
$$\begin{aligned}
H(P)&=&P \frac{1-3 \Pi}{(1+ \Pi)^3}+\frac{6}{q^2s}
\left( \frac{\Pi}{1+\Pi} \right)^{5/2}, \\
H(P)&=&P / \cosh^2(s \sqrt{\Pi}), \\
H(P)&=&P\frac{\exp(-s \sqrt{\Pi})}{(1+ \Pi)^{5/2}} \left( 1+
\frac{3\Pi}{s}+ \Pi \right),\end{aligned}$$
where ${\Pi}=\sqrt {-q^2P/2}$ and $s= \mid q \mid /(2m)$; $q=q_e$ and $m$ are, respectively, the parameters identified as charge and mass. The above presented functions $H(P)$ behave like $P$ at small $P$ and tend to finite limits as $P \to - \infty$. The solutions are called regular due to the finiteness of three invariants: $R, R_{\alpha \beta}R^{\alpha \beta}$ and $R_{\alpha \beta \gamma \delta}R^{\alpha \beta \gamma \delta}$.
However, there is a clear contradiction with no go theorems that forbid the existence of SSS solutions with a regular center for whatever $L(F)$ chosen if it is required that for weak fields, $L \sim F$, i.e. if for weak fields it is demanded a Maxwellian behavior [@Bronnikov2], [@Bronnikov].
The explanation of this apparent contradiction is the following: one can obtain regular solutions choosing ad hoc $H(P, \tilde{Q})$ as was explained above. These solutions correspond to certain lagrangian $L(F,
\tilde{G})$ that is the one from which one derives the dynamical equations. That this Lagrangian be well behaved is not guarateed by an [*adecuate*]{} selection of the corresponding $H(P, \tilde{Q})$. This is the case, for instance in [@Ayon]: those solutions correspond to Lagrangians $L(F)$ that suffer branching. So, in spite that solutions are well behaved in the $P$ framework ($\tilde{G}=0$), they correspond to different Lagrangians in different regions of space. The branching in $L(F)$ is due to extrema in the electromagnetic field; associated to extrema there are singularities that are seen only by photons. These singularities can either be hidden behind a horizon or naked depending on the value of one parameter. Additional features of those “regular" solutions include, for instance, that the energy density of electromagnetic field may be negative for some interval of the radial coordinate; this and more were investigated by Novello et al [@Novello], [@Novello2].
Note, however, that regular solutions with only magnetic charge may exist [@Bronnikov]. Another way to avoid the prohibition of SSS electrically charged regular structures is to resign of having a regular center, like the solution by Dymnikova in [@Dymnikova] that has a de Sitter center. It is not excluded neither the possibility of regular solutions that correspond to Lagrangians depending on both invariants of the electromagnetic field, $L(F, \tilde{G})$.
Type-D solutions with EBI
-------------------------
BI theory has the property that its equations have an exact $SO(2)$ electric-magnetic duality invariance. In [@Pleban2] Type-D solutions were constructed for BI coupled with Einstein equations allowing for the freedom of duality rotations. Under the assumption that the natural tetrads of type-D metrics coincide with the eigenvectors of the algebraically general nonlinear electromagnetic field and assuming that the two principal null directions are geodesic and shear-free, all D-type solutions that are compatible with the scheme of NLED endowed with duality rotations were found. It turned out that the presence of the acceleration or rotation parameters prohibits the existence of NLED type-D solutions with duality rotations. Moreover, by adding axion and dilaton fields, this invariance may be extended to $SL(2, {\bf R})$ S-duality, relevant to string theory, which implies a strong-weak coupling duality of such theories; for dualities in the context of more general nonlinear electrodynamics see [@Gibbons-Rasheed], [@PP], [@Gibbons-Hashimoto].
All type-D solutions of the coupled Einstein and Born-Infeld equations were determined in [@GSP]. There are two classes of them: static and stationary. The static solutions are exhausted by the BI generalizations of (i) the Bertotti-Robinson solution, (ii) Reissner-Nordstrom (RN) and (iii) anti-Reissner-Nordstrom. The stationary solutions belong to two subfamilies: the BI generalization of the NUT $\tilde{B}(+)$ metric which contains as a limit the BI generalization of RN, and the BI generalization of the anti-NUT $\tilde{B}(-)$ solution that contains as a limit the static case (iii). The NUT solution includes the NUT parameter, $m$, while anti-NUT changes from $m$ to $(-n)$. These EBI solutions were derived including a cosmological constant $\Lambda$ in their energy-momentum tensor. Since BI generalization of RN solution is of most interest, we analyze it in detail in what follows.
Born-Infeld black hole and EBIon
--------------------------------
The EBI solution for a SSS spacetime (\[sss-metric\]) is given by the metric function $\psi_{BI}(r)$
$$\begin{aligned}
\psi_{BI}(r)&=& 1-\frac{2m}{r} + \frac{2}{3}b^2 (r^2 - \sqrt{r^4+a^4})+
\frac{4g^2}{3r}G(r), \\
G'(r)&=&- (r^4+a^4)^{- \frac{1}{2}},
\label{BImetrfunc}\end{aligned}$$
where $G'(r)$ denotes the derivative of $G(r)$ with respect to the radial variable, $m$ is the mass parameter, $g$ is the magnetic (or electric) charge (both in lenght units), $a^4=g^2/b^2$ and $b$ is the Born-Infeld parameter given in units of $[\rm{lenght}]^{-1}$. The nonvanishing components of the electromagnetic field are
$$F_{rt}= g (r^4+ a^4)^{- \frac{1}{2}},
\quad P_{rt}= \frac{g}{r^2}.
\label{FrtBI}$$
The black hole solution given by García-Salazar-Plebañski [@GSP] corresponds to
$$G(r)= \int^{\infty}_{r}{\frac{ds}{\sqrt{s^4+a^4}}}=\frac{1}{2a}
{\mathbb{F}} \left[ \arccos{ \left( \frac{r^2- a^2}{r^2+a^2} \right)},
\frac{1}{\sqrt{2}} \right],
\label{gPleb}$$
where ${\mathbb{F}}$ is the elliptic integral of the first kind. On the other side, the particle-like solution given by Demianski [@Demian] is
$$G(r)= \int^{r}_{0}{\frac{-ds}{\sqrt{s^4+a^4}}}=- \frac{1}{2a}
{\mathbb{F}} \left[ \arccos{ \left(
\frac{a^2-r^2}{a^2+r^2}\right)},\frac{1}{\sqrt{2}} \right].
\label{gDem}$$
The selection of $G(r)$ as in Eq. (\[gPleb\]) or Eq. (\[gDem\]) has as a consequence a different behavior of the solution at $r=0$. The metric function $\psi_{BI}(r)$ with $G(r)$ given by Eq. (\[gPleb\]) diverges at $r \to 0$ (even when $m=0$), corresponding to the black hole solution. The other one, meaning $\psi_{BI}(r)$ with $G(r)$ given by Eq. (\[gDem\]), is the so called EBIon solution that is finite at the origin (for $m=0$). The integrals of Eqs. (\[gPleb\]) and (\[gDem\]) are related by
$$\int^{\infty}_{r}{\frac{ds}{\sqrt{s^4+a^4}}}+\int^{r}_{0}
{\frac{ds}{\sqrt{s^4+a^4}}}=\frac{1}{a}{\rm K} \left[\frac{1}{2} \right],$$
where ${\rm K}[\frac{1}{2}]$ is the complete elliptic integral of the first kind. In the limit of large distances, $r \to \infty$, asymptotically the solution approaches Reissner-Nordstrom (RN) solution, the SSS solution to Einstein-Maxwell equations. Also when the BI parameter goes to infinity, $b \to \infty$, we recover the linear electromagnetic (Einstein-Maxwell) RN solution. In the uncharged limit, $b=0$ (or $g=0$), it is recovered the Schwarzschild black hole. Note that due to the duality rotation both charges, electric $e$ and magnetic $g$, can be included in the solution by substituting $g \to \sqrt{e^2+g^2}$.
Since this solution is of type D, there is only one nonvanishing Weyl scalar, $\Psi_2$. For the black hole solution it is
$$\Psi_2= \frac{m}{r^3}-\frac{g^2r^3}{6} \partial_{,rr} \left( \frac{1}{r^2}
\int_{r}^{\infty} \frac{ds}{s^2+ \sqrt{s^4+a^4}} \right),$$
The invariants depend on $\Psi_2^2$, then at $r=0$ there is a singularity of order $1/r^6$, coming from the gravitational mass term, alike the Schwarzschild and RN singularities; furthermore, the second term also diverges at $r=0$. There are also zeros in the metric function $\psi_{BI}(r)$ that are coordinate singularities which can be removed using analytical extensions [@Graves]. The zeros of $\psi_{BI}(r)$ can be localized numerically and might be one, two or none, depending on the relative values of the parameters $g, m, b$; it is illustrated in Fig.\[psi\]. These parameters determine the position and size of the horizon as well [@Breton1]. Note that distinct $b$ corresponds to different NLED theories.
Trajectories of test particles in BI black hole
-----------------------------------------------
Since the SSS spacetime possesses two Killing vectors, $\partial_t$ and $\partial_{\phi}$, the test particle conserves two motion quantities: its energy $E$ and its angular momentum $l$. Moreover, if we restrict ourselves to the equatorial plane ($\theta= \pi/2$), the timelike and null geodesics for the EBI spacetimes can be reduced to the problem of ordinary one-dimensional motion in an effective potential $U_{eff}$, alike in the Schwarzschild and RN cases.
### Massive particles
Trajectories of massive particles are determined by the Lorentz equation. For a test particle of charge $\epsilon$ and mass $\mu$ it is
$$\frac{d^2x^{\nu}}{d \tau^2}+ \Gamma_{\alpha \beta}^{\nu}
\frac{dx^{\alpha}}{d \tau}\frac{dx^{\beta}}{d \tau}= -
\frac{\epsilon}{\mu}F_{\sigma}^{\nu}
dx^{\sigma}d \tau,$$
where $\tau$ is the affine parameter along the trajectorie. Using the two conserved motion quantities: energy $E$ and angular momentum $l$, the geodesic for $t$ can be integrated once, obtaining the first derivative with respect to the proper time $\tau$,
$$\dot{t} \psi_{BI}=E+ \frac{\epsilon g}{\mu} \sqrt{\frac{b}{4g}}
{\mathbb{F}} \left[
\arccos(\frac{r^2-g/b}{r^2+g/b}), \frac{1}{\sqrt{2}}\right].$$
From the line element for timelike geodesics we have,
$$1= \psi \dot{t}^2- \psi^{-1} \dot{r}^2-r^2 \dot{\phi}^2$$
substituting $l=g_{\phi \phi}\dot{\phi}$ it is obtained
$$\dot{r}^2+ \psi \left( \frac{l^2}{r^2} +1 \right)-
\left[E+ \frac{\epsilon g}{\mu} \sqrt{\frac{b}{4g}}{\mathbb{F}}
\left[
\arccos(\frac{r^2-g/b}{r^2+g/b}), \frac{1}{\sqrt{2}}\right]
\right]^2=0.$$
Comparison with $\frac{1}{2}\dot{r}^2+U_{eff}(E, l, r)=0$, gives the effective potential that a charged test particle feels. The shape of the effective potential shows attractive regions with stable equillibrium positions. The equilibrium positions are of lower energy for greater angular momentum of the particles. If test particles reach the singularity or not depends on the value of the maximum fiel strength $b$.
### Light trajectories in NLED
Discontinuities of the fields propagate obeying the equation of the characteristic surfaces, that in ordinary optics is the so-called [*eikonal equation*]{}. Locally these surfaces are normal to the light rays trajectories. In General Relativity when a linear Maxwellian field is present, its characteristics do coincide with the Einsteinian cone; but NLED photons do not propagate along null geodesics of the background geometry. Instead, they propagate along null geodesics of an effective geometry which depends on the nonlinear electromagnetic field. In a curved spacetime the characteristic surfaces equation is [@Pleban3], [@Boillat]
$$g^{\mu \nu}S_{,\mu}S_{\nu}=0.
\label{null-geod}$$
However, if in the spacetime it is present a nonlinear electromagnetic field, the corresponding equation is
$$(g^{\mu \nu}+ \frac{4 \pi}{b^2}T^{\mu \nu}_{NLED})S_{,\mu}S_{\nu}=
\gamma^{\mu \nu}S_{,\mu}S_{\nu}=0,
\label{phot-traj}$$
where $T^{\mu \nu}_{NLED}$ is the electromagnetic energy momentum density; remind that the linear limit is obtained when $b \to \infty$, hence in linear limit light trajectory, that is normal to characteristic surfaces, occurs on null geodesics. Hence, as far as $b$ is finite, there is a distinction between the propagation of gravitational discontinuities (gravitons, abusing of languaje) along Eq. (\[null-geod\]) and the propagation of electromagnetic discontinuities (photons) ruled by Eq. (\[phot-traj\]). For BI black holes, actually both trajectories converge at the horizon.
Using the two constants of motion, $E=- \psi \dot t$, $l=\dot \phi r^2$ and the line element for null geodesics, we obtain the derivative of $r$ respect to an affine parameter,
$$\dot r = \sqrt{E^2- \frac{\psi_{BI} l^2}{r^2}}.$$
The trajectories for light rays, however, are given by an effective geometry, considering the same constants of motion $E$ and $l$:
$$\dot r_{ph} = \sqrt{E^2- \frac{\psi_{BI} l^2}{r^2} \left( 1+
\frac{a^4}{r^4}
\right)^{-1}}.$$
There is a correction factor in the last formula, $(1+a^4/r^4)^{-1}$, due to the effective geometry; it vanishes if $a=0$ ($b \to \infty$) or for photons with $l=0$. Remind that usually it is considered $\hbar E$ and $\hbar l$ as the total energy and angular momentum of photon, respectively. The corresponding effective potentials are shown in Fig. \[U\_eff\].
From the above expressions it is easy to check that $(dr/dt)_{photon}<
(dr/dt)_{grav}$, i. e. light travels slower that gravitational waves due to nonlinear effects, such as photon-photon interaction, for instance. This effect can also be described as if photons ruled by Maxwell ED were propagating inside a dielectric medium, with dielectric “constant” depending in nonlinear way of the fields. The possibility is not excluded of superluminal signals for other NLED [@Novello].
NLED black hole thermodynamics
==============================
In a general context the zeroth and first laws of black hole mechanics (BHM) refer to equilibrium situations and small departures therefrom. First law of BHM is an identity relating the changes in mass, angular momentum and horizon area of a stationary black hole when it is perturbed. The variation applies for perturbations from one stationary axisymmetric solution of Einstein equations to another; moreover, it has been shown that the validity of this law depends only on very general properties of the field equations [@Wald]. For the horizon mass $M_{\Delta}$ the first law, when static spherically symmetric solutions are considered, is
$$\delta M_{\Delta}= \frac{\kappa}{8 \pi} \delta{a_{\Delta}}
+ \Phi_{\Delta} \delta Q_{\Delta},
\label{first}$$
where $\kappa$ is the surface gravity at the horizon, $a$ is the area of the horizon, $Q$ is the electric charge and $\Phi$ is the electric potential; the subindex $\Delta$ indicates that the quantity is evaluated at the horizon of the black hole.
On the other side, the total mass is given by the Smarr’s formula
$$M_{\Delta}= \frac{\kappa a_{\Delta}}{4 \pi} +
\Phi_{\Delta} Q_{\Delta} .
\label{smarr}$$
In the case of Einstein-Maxwell theory, it is possible to deduce one, Eq. (\[first\]), directly from the other, Eq. (\[smarr\]), using the homogeneity of the mass as a function of $\sqrt{a}$ and $Q$. In the work by Ashtekar, Corichi and Sudarsky [@ACS] the first law of BHM, for quantities defined only at the horizon, arises naturally as part of the requirements for a consistent Hamiltonian formulation.
Work on NLED black hole thermodynamics includes the derivation of the first law of black hole physics for some nonlinear matter models [@Heusler]. Rasheed [@Rasheed] studied the zeroth and first laws of black hole mechanics in the context of non-linear electrodynamics coupled to gravity. In this case, the zeroth law, which states that the surface gravity of a stationary black hole is constant over the event horizon, is shown to hold even if the Dominant Energy Condition is violated. In NLED one no longer has homogeneity of the mass function and a priori one has no reason to expect that neither the first law or Smarr formula hold. Rasheed found that the usual first law (the general mass variation formula) holds true for the case of non-linear electrodynamics but the formula for the total mass (Smarr’s formula) does not.
However, we can propose the form that must have a Smarr-type formula for the horizon mass in order to be consistent with the variations expressed by the first law of BHM that indeed holds,
$$M_{\Delta}= \frac{\kappa a_{\Delta}}{4 \pi} +
\Phi_{\Delta} Q_{\Delta} + V(a_{\Delta}, Q_{\Delta}, P_{\Delta}),
\label{hormass0}$$
where $V$ is a so far undetermined potential that depends on the horizon parameters, $a_{\Delta}, Q_{\Delta}, P_{\Delta}$ and also on the coupling constants of the theory. In the variational principle this term plays no role, however in the Hamiltonian description it becomes essential.
Note that in the first law, Eq. (\[first\]), only variations of the electric charge are involved, and not variations of the magnetic charge, $P_{\Delta}$. On the other hand, the horizon mass, Eq. (\[hormass0\]) might depend on $P_{\Delta}$ through $V$.
The equations to determine the potential $V(a_{\Delta},
Q_{\Delta}, P_{\Delta})$ arise from the condition that the first law holds and demanding consistency between Eq.(\[first\]) and the variations of Eq.(\[hormass0\]), these are [@ulises],
$$\begin{aligned}
a_{\Delta} \frac{\partial \psi}{\partial a_{\Delta}}+
8 \pi r_{\Delta} Q_{\Delta} \frac{\partial \Phi}{\partial
a_{\Delta}} + 8 \pi r_{\Delta} \frac{\partial V}{\partial
a_{\Delta}}&&=0, \nonumber\\
\frac{r_{\Delta}}{2} \frac{\partial \psi}{\partial Q_{\Delta}}+
Q_{\Delta} \frac{\partial \Phi}{\partial Q_{\Delta}}
+ \frac{\partial V}{\partial Q_{\Delta}}&&=0,
\label{Veqs}\end{aligned}$$
where $\psi$ is the metric function in a SSS line element (\[sss-metric\]), $\psi= 1-2m'(r)$, $a_{\Delta}=4 \pi r_{\Delta}^2$; $r_{\Delta}$ is the radius of the horizon.
The condition of consistency determines the set of parameters that can vary independently; in this case, the magnetic charge becomes a function of the area and electric charge, $P_{\Delta}= P_{\Delta}(r_{\Delta},
Q_{\Delta})$. To illustrate the point, in what follows we shall determine the horizon mass from a Smarr type formula in agreement with the first law of BHM for the Bardeen black hole.
Smarr’s formula for Bardeen black hole
---------------------------------------
The Bardeen model was proposed some years ago as a regular black hole, however, only recently it has been shown [@ABGBardeen] that it is an exact solution of the Einstein equations coupled to a kind of nonlinear electrodynamics characterized by the Lagrangian
$${\cal L}(F)= \frac{2}{2sg^2}(\frac{2g^2F}{1+ \sqrt{2g^2F}})^{5/2},$$
where $g$ and $F$ are the magnetic charge and electromagnetic invariant, respectively and $s=g/m$. The corresponding energy momentum tensor fulfills the weak energy condition and is regular everywhere. For a SSS space (\[sss-metric\]), the corresponding metric function is given by
$$\psi_{B}= 1- \frac{2m(r)}{r}= 1- \frac{2mr^2}{(r^2+g^2)^{3/2}},
\label{psi-bardeen}$$
This solution is a self-gravitating magnetic monopole with charge $g$. The solution is regular everywhere, although the invariants of the electromagnetic field exhibit the usual singular behaviour of magnetic monopoles, $F= g^2/2r^4$. In the asymptotic behaviour of the solution the constant $g$ vanishes as $1/r^3$, and not as a Coulombian term ($1/r^2$), that allows to interpret the constant $g$ as a magnetic charge. The Bardeen solution does not involve electric charge, then the horizon mass depends only on the area of the horizon,
$$M_{\Delta}= \frac{1}{8 \pi} \int{\kappa da}= \int{(1-m')dr},
\label{hormass2}$$
the condition that the horizon mass be positive, from Eq. (\[hormass2\]), gives that $m(r) \le r$; it also guarantees that $\psi_{B} \ge 0$. Using the expression for $\psi_{B}$ it amounts to $(r^2+g^2)^3 \ge 4m^2r^4$. In this case when $g^2=\frac{16}{27}m^2$ the two horizons that could be present shrink into a single one, being this value of $g$ the corresponding to the extreme black hole; for $g^2<
\frac{16}{27}m^2$ there exist both inner and event horizon. The potential $V$ for the Smarr-type formula, Eq. (\[hormass0\]), for the Bardeen black hole turns out to be, undetermined until an integration constant which we have put zero,
$$V= mr^3 \frac{2g^2-r^2}{(g^2+r^2)^{\frac{3}{2}}},$$
Substituting $V$ in the Smarr-type formula one obtains the horizon mass $$M_{\Delta}= \frac{r}{2}-\frac{mr^3}{(r^2+g^2)^{\frac{3}{2}}}
\label{hormass3}$$
This value for the horizon mass coincides with the one determined by integrating the first law, Eq. (\[hormass2\]). Note that the horizon mass of the Bardeen black hole involved dependence only on the horizon area, since the magnetic charge is not considered as a varying parameter of the horizon. In this case there is a full agreement between the horizon mass calculated with the first law of BHM and when it is determined by adding the apropriate potential to a Smarr-type formula; it is shown in Fig. \[Bbhextreme\].
The horizon mass, calculated with a Smarr type formula that is consistent with the fist law of BHM, applies only to the magnetic sector of NLED solutions. If the variation of electric charge is taken into account in the potential $V$ of the Smarr formula, the mentioned consistency does not longer hold. The potential $V$ determined in agreement with the first law of BHM can not give the appropriate dependence for the terms corresponding to electric charge; Eqs. (\[Veqs\]) do not describe confidently the potential $V$ in general situations where nonlinear electromagnetic fields are present. Therefore, when the dependence of $V$ on the electromagnetic field is of nonlinear nature Eqs.(\[Veqs\]) are useless to determine $V$.
Isolated horizon framework and mass relation
============================================
Remarkable properties of nonlinear electrodynamics black holes arise in the context of the isolated horizon formalism, recently put forward by Ashtekar [*et al*]{} [@Ashtekar]. In this approach it is pointed out the unsatisfactory (uncomplete) description of a black hole given by concepts such as ADM mass and event horizon, for instance, specially if one is dealing with hairy black holes. To remedy this uncompleteness, Ashtekar [*et al*]{} have proposed alternatively the isolated horizon formalism, that furnishes a more complete description of what happens in the neighborhood of the horizon of a hairy black hole.
In the isolated horizon formalism one considers spacetimes with an interior boundary, which satisfy quasi-local boundary conditions that insure that the horizon remains isolated. The boundary conditions imply that quasi-local charges can be defined at the horizon, which remain constant in time. In particular one can define a horizon mass, a horizon electric charge and a horizon magnetic charge.
Moreover, Ashtekar-Corichi-Sudarsky (ACS) conjecture about the relationship between the colored black holes and their solitonic analogs [@ACS]: the Arnowitt-Deser-Misner (ADM) mass contains two contributions, one attributed to the black hole horizon and the other to the outside hair, captured by the solitonic residue. In this model, the hairy black hole can be regarded as a bound state of an ordinary black hole and a soliton. The proposed formula relating the horizon mass and the ADM mass of the colored black hole solution with the ADM mass of the soliton solution of the corresponding theory is
$$M^{(n)}_{sol}=M^{(n)}_{ADM} -M^{(n)}_{\Delta},
\label{massrel}$$
where the superscript $n$ indicates the colored version of the hole; in the papers of Ashtekar [*et al*]{} this $n$ refers to the Yang-Mills hair, labeled by this parameter, corresponding to $n=0$ the Schwarszchild limit (absence of YM charge). This relation has been proved numerically to work for the Einstein-Yang-Mills (EYM) black hole.
For the EBI black hole the location and size of the horizon depends on the parameter $bq$, so $b$ and $q$ are not independent parameters; however, at infinity it is undistinguisable from a RN black hole characterized only by its charge $q$ and mass $m$. Provided that in the EBI theory there exist both exact solutions: the black hole and the soliton like solution, in spite that the EBI black hole is not a coloured one, we shall probe it with ACS model, considering $b$ as a free parameter, for a fixed charge. Then for the case studied here the $n$ version shall correspond to the distinct black holes labeled by distinct (continuous) BI parameter, $b$.
It turns out that the EBI black hole and the corresponding EBIon solution fulfill the relation between the masses as well as most of the properties of the model as for the colored black hole [@Breton2]. For the EBI solution the horizon and ADM masses as functions of the horizon radius $r_{\Delta}$ are given, respectively, by
$$M^{(b)}_{\Delta}(r_{\Delta})= \frac{r_{\Delta}}{2}+\frac{b^2
r_{\Delta}}{3}(r_{\Delta}^2-\sqrt{r_{\Delta}^4+a^4})-\frac{2g^2}{3}
\int^{r_{\Delta}}_{0}{\frac{ds}{\sqrt{a^4+s^4}}},
\label{Mhor}$$
$$M^{(b)}_{ADM}(r_{\Delta})= \frac{r_{\Delta}}{2}+\frac{b^2
r_{\Delta}}{3}(r_{{\Delta}}^2-\sqrt{r_{{\Delta}}^4+a^4})+\frac{2g^2}{3}
\int^{\infty}_{r_{\Delta}}{\frac{ds}{\sqrt{a^4+s^4}}}.
\label{Madm}$$
In Figs. \[M\_ADM\] and \[M\_hor\] are displayed $M^{(BI)}_{ADM}$ and $M^{BI}_{\Delta}$ in comparison with the Schwarzschild and Reissner-Norsdtrom cases. The mass of the soliton can be obtained by letting $r_{\Delta} \to 0$ in the ADM mass, Eq. (\[Madm\]), obtaining $M_{sol}^{(b)}=2g \sqrt{gb}{\rm K}[\frac{1}{2}]/3$. From these expressions one can trivially check that they satisfy Eq. (\[massrel\]).
Other predictions of ACS model that are fulfilled by BI black holes and its soliton counterpart are: BI horizon mass is less than Schwarzschild horizon mass (see Fig. \[M\_hor\]); horizon masses satisfy the inequality $M^{RN}_{\Delta} >M^{Schw}_{\Delta}>M^{BI}_{\Delta}$; for all $b$ and all $r_{\Delta}$, the surface gravity of the BI black hole is less than the one for Schwarzschild; $M^{BI}_{\Delta}$ and $\kappa_{BI}$ as functions of $r_{\Delta}$ are monotonically decreasing functions of $b$. However, EBI solutions do not fulfil that $M^{BI}_{\Delta}$ as function of $r_{\Delta}$ increases monotonically for all values of $b$ (see Fig. \[M\_hor\] for $b=1$)
Since most of the ACS features are fulfilled, we can say that the static sector of the EBI theory is described by the heuristic model for the colored black holes proposed by Ashtekar [*et al*]{} when the BI parameter $b$ varies keeping the charge fixed.
Stability of NLED black holes
=============================
Stability properties in self-gravitating nonlinear electrodynamics were studied by Moreno and Sarbach [@M-S]. They derived sufficient conditions for linear stability with respect to arbitrary linear fluctuations in the metric and in the gauge potential, $\delta g_{\mu
\nu}$ and $\delta A_{\mu \nu}$, respectively; the conditions were obtained in the form of inequalities to be fulfilled by the nonlinear electromagnetic Lagrangian $L(F)$ and its derivatives. The application of this criterion is restricted to static, spherically symmetric solutions of NLED coupled to gravity, that are purely electric or purely magnetic ($
\tilde{G}=0$). For these systems a gauge invariant perturbation formalism was used obtaining that linear fluctuations around a SSS purely electric (or purely magnetic) solution are governed by a wavelike equation with symmetric potential. The stability conditions are translated into some requirements on the Lagrangian and its derivatives, $L(F),L_{F}, L_{FF}$; in terms of the variable $y= \sqrt{2g^2F}$ these conditions are
$$L(y) >0, \quad L(y)_{,y} >0, \quad L(y)_{,yy} >0 .
\label{stabcond1}$$
Besides, there are more inequalities to be fulfilled, that arise from the pulsation equations in the even-parity sector
$$f(y) \equiv yL_{,yy}/L_{,y}>0, \quad
f(y)N(y)< 3.
\label{stabcond2}$$
where $N(y)$ is the metric function in the SSS line element (\[sss-metric\]).
We shall apply this criterion to test the stability of the purely magnetic or purely electric EBI particle-like and black hole solutions [@Breton3]. In the former case the boundary point is the origin, $r=0$, while for the black hole case the conditions must be held in the domain of outer communication (DOC), i.e. positions outside the horizon, $r> r_{\Delta}$, $r_{\Delta}$ being the radius of the horizon of the black hole.
The BI Lagrangian fulfills the stability conditions; in terms of the variable $y$, the BI Lagrangian, with $\tilde{G}=0$, is given by
$$L(y)=b^2[\sqrt{1+\frac{y^2}{b^2g^2}}-1]>0,$$
and the rest of the inequalities (\[stabcond2\]) read as:
$$\begin{aligned}
L_{,y}&=&\frac{y}{g^2}(1+\frac{y^2}{b^2g^2})^{-\frac{1}{2}}>0,
\nonumber\\
L_{,yy}&=&\frac{1}{g^2}(1+\frac{y^2}{b^2g^2})^{-\frac{3}{2}}>0,\nonumber\\
f(y)&=& y\frac{L_{,yy}}{L_{,y}} =(1+\frac{y^2}{b^2g^2})^{-1}>0,
\label{BIineq} \end{aligned}$$
Conditions (\[BIineq\]) are fulfiled in all the range of $y$. Moreover, $f(y)$ is monotonically decreasing with $f(y=0)=1$, $0<f(y) \le 1$; then the last stability condition $f(y)N(y) < 3$ reduces to prove that $N(y)=
\psi_{BI}(y)< 3$; for the black hole it must be fulfilled in DOC ($r>r_{\Delta}$), while for the particle-like solution the domain to be considered is $0 \le r < \infty$.
In the black hole case, the metric function $\psi_{BI}(r)$ has a minimum in the extreme case ($g=m$) for $bm=0.5224$ at $r_{\Delta}=0.346 m$; DOC is considered for distances larger than the radius of the horizon, $r>
r_{\Delta}=0.346 m$. In terms of $y$, considering that $F=g^2/2r^4$ then $y=g^2/r^2$, the metric function $N(y)=\psi_{BI}(y)$ is
$$\psi_{BI}(y)=1-\frac{2m \sqrt{y}}{g}+
\frac{2b^2g^2}{3y}[1-\sqrt{1+\frac{y^2}{b^2g^2}}]+
\frac{2 \sqrt{gby}}{3}{\mathbb{F}} \left[
\arccos({\frac{gb-y}{gb+y}}),
\frac{1}{\sqrt{2}} \right],$$
In the range $0<y<y_{\Delta} =8.35$ it turns out that $0< \psi_{BI}(y) \le
1$ with $\psi_{BI}(0)=1$ therefore, $0<N(y)=\psi_{BI}(y)<1<3$, fulfilling the last inequality required as sufficient conditions for linear stability of the EBI black hole.
For the particle-like solution of the EBI equations, the metric function $\psi_{BI}(r)$ in terms of $y$ is
$$\psi_{BI}(y)=1-\frac{2m \sqrt{y}}{g}+
\frac{2b^2g^2}{3y}[1-\sqrt{1+\frac{y^2}{b^2g^2}}]-
\frac{2 \sqrt{gby}}{3}{\mathbb{F}} \left[
\arccos({\frac{y-gb}{gb+y}}),
\frac{1}{\sqrt{2}} \right],$$
The last stability condition $N=\psi < 3$ in fact occurs since $N(y=0)=1$ and the function is monotonically decreasing, having $N(y) \le 1$, for $bg
\ne 0$; the finiteness in the origin of $N(r)$ is valid when $m=0$. Therefore, as far as this analysis proves, the EBI solutions, both black hole and particle-like one, are stable.
On the other side, stability of self-gravitating structures can also be approached from the isolated horizon framework: the ACS heuristic model explains the instability of the coloured black holes in terms of the instability of the solitons. The bound state of a bare black hole and a soliton is going to be unstable if the total energy in the initial bound state $E_{initial}= M_{\Delta}^{(n)}(r_{\Delta}^{initial})+M_{sol}^{(n)}$ exceeds the energy in the final black hole, $E_{final}=M_{\Delta}^{(0)}(r_{\Delta}^{final})$.
The available energy can be expressed (for a fixed $r_{\Delta}^{initial}$) as
$$E^{(n)}_{avail}=M_{ADM}^{(n)}-M_{ADM}^{(0)},$$
where the superscript $n$ indicates the colored version of the hole and $M_{ADM}^{(0)}$ is the ADM mass of the bare (Schwarzschild) black hole. The positivity of the difference between the ADM mass and the horizon mass, $M_{ADM}-M_{\Delta}=E >0$, indicates that there exists an energy $E$ available to be radiated. For static black holes this result can be interpreted as a potential unstability, i.e. a slightly perturbation in the initial data will lead the solution to decay to a Schwarzschild black hole.
The difference between the ADM mass of the BI black hole and the bare black hole using (\[Madm\]), turns out to be
$$M^{(b)}_{ADM}(r_{\Delta}) - \frac{r_{\Delta}}{2}=
\frac{b^2}{3}
\left(r_{{\Delta}}^3(1-\sqrt{1+\frac{a^4}{r_{\Delta}^4}})+
a^3{\mathbb{F}} \left[ \arccos{ \left( \frac{r_{\Delta}^2-
a^2}{r_{\Delta}^2+a^2} \right)},
\frac{1}{\sqrt{2}} \right] \right),$$
the difference is greater than zero (except if $a=0$ that reduces to Schwarzschild black hole). On this basis, one might conjecture that NLED black holes are unstable. However, above was shown that SSS BI solutions are stable under linear perturbations. Hence for BI black hole there is no relation between the positivity of the masses difference and stability under perturbations of the metric and the electromagnetic potential.
ACS stability conjecture has been tested for other SSS NLED solutions (in the magnetic sector), proving to be true; we conclude then that ACS unstability conjecture does not apply generically for NLED solutions.
In relation to Born-Infeld black hole stability recently was presented in [@Fernando] the analysis of quasinormal modes for the gravitational perturbations, deriving a one dimensional Schrodinger type wave equation for the axial perturbations. From the behavior of the potentials it was concluded in [@Fernando] that the EBI black holes are classically stable, this in agreement with the analysis based on Lagrangian inequalities.
Current trends in Born-Infeld black holes include the coupling of EBI fields with dilaton and axion field as well as non Abelian Born Infeld structures [@Kunz]. Another aspect that has been explored is BI black holes as gravitational lenses: deflections depend on the BI parameter $b$ [@Eiroa], [@Mosquera]. Lately also has been arised interest in BI black hole thermodynamics in extra dimensions [@Riazi].
[99]{}
Born, M., Infeld, L.: [*Foundations of the New Field Theory*]{}, Proc. R. Soc. (London) [**A144**]{}, 425-451 (1934).
Hoffmann, B., Infeld, L.: [*On the Choice of the Action Function in the New Field Theory*]{}, Phys. Rev. [**51**]{} 765-773, (1937).
Peres, A. :[*Nonlinear Electrodynamics in General Relativity*]{} Phys. Rev. [**122**]{} 273-274, (1961).
Pellicer, R., Torrence, R. J.:[*Nonlinear Electrodynamics and General Relativity*]{}, J. Math. Phys. [**10**]{}, 1718-1723 (1969).
Plebañski, J. F.:[*Lectures on Non-linear Electrodynamics*]{}, (Copenhagen, NORDITA, 1970).
Salazar, H., García, A., Plebañski, J.F.: [*Duality rotations and type D solutions to Einstein equations with nonlinear electromagnetic sources*]{}, J. Math. Phys. [**28**]{}, 2171-2181, (1987).
García, A., Salazar, H., Plebañski, J. F.: [*Type-D solutions of the Einstein and Born-Infeld Nonlinear Electrodynamics Equations*]{}, Nuovo Cim. [**84**]{}, 65-90 (1984).
Dudley, A., Alarcón, S., Plebañski, J. F.: [*Signals and discontinuities in general relativistic nonlinear electrodynamics*]{}, J. Math. Phys. [**22**]{}, 2835-2848 (1981).
Demianski, M.: [*Static Electromagnetic Geon*]{}, Found. of Phys. [**16**]{}, 187-190 (1986).
Heisenberg, W., Euler, H.: [*Folgerungen aus der Diracschen Theorie des Positrons*]{}, (1936) Zeit. f. Phys. [**98**]{} 714-732. Weisskopf, W.S. [*On the self-energy and the electromagnetic field of the electron*]{}, Phys. Rev. [**56**]{} 72 (1939).
Schwinger, J.: [*On Gauge Invariance and Vacuum Polarization*]{}, [Phys. Rev.]{}[**82**]{} 664-679 (1951).
Delphenich, D. H.:[*Nonlinear Electrodynamics and QED*]{}, arXiv: hep-th/0309108.
Gibbons, G. W.: [*Born-Infeld particles and Dirichlet p-branes*]{}, Nucl. Phys. [**B 514**]{} 603-639 (1998).
Ayón-Beato, E., García, A.: [*Regular black hole in general relativity coupled to nonlinear electrodynamics*]{}, Phys. Rev. Lett., [**80**]{}, 5056-5059 (1998); Ayón-Beato, E., García, A. :[*Nonsingular charged black hole solution for nonlinear source*]{}, Gen. Rel. Gravit., [**31**]{}, 629-633 (1999); Ayón-Beato, E., García, A.: [*New regular black hole solution from nonlinear electrodynamics*]{}, Phys. Lett. B [**464**]{}, 25-28, (1999).
Bronnikov, K. A.: [*Comment on “Regular black hole in general relativity coupled to nonlinear electrodynamics"*]{}, Phys. Rev. Lett. [**85**]{}, 4641 (2000).
Bronnikov, K. A.: [*Regular magnetic black holes and monopoles from nonlinear electrodynamics*]{}, Phys. Rev. D [**63**]{}, 044005 (2001).
Novello, M. De Lorenci, V. A., Salim, J. M., Klippert, R.: [*Geometrical Aspects of Light Propagation in Nonlinear Electrodynamics*]{} Phys. Rev. D. [**61**]{} (2000) 045001.
Novello, M., Perez Bergliaffa, S.E., Salim, J. M.: [*Singularities in general relativity coupled to nonlinear electrodynamics*]{} Class. Quant. Grav. [**17**]{}, 3821-3831, (2000).
Dymnikova, I : [*Regular electrically charged vacuum structures with de Sitter center in Nonlinear Electrodynamics coupled to General Relativity* ]{}, Class. Quant. Grav. [**21**]{}, 4417-4429, (2004).
Gibbons, G. W., Rasheed, D. A.: [*Electric-magnetic duality rotations in non-linear electrodynamics*]{}, Nucl. Phys. [**B 454**]{} 185 (1995).
Plebañski, J. F., Przanowski, M.: [*Duality Transformations in Electrodynamics*]{}, Int. J. Theor. Phys. [**33**]{} 1535-1551, (1994).
Gibbons, G. W., Hashimoto, K.: [*Non-linear Electrodynamics in Curved Backgrounds*]{}, JHEP [**0009**]{} 013 (2000).
Graves, J. C., Brill D. R.: [*Oscillatory Character of Reissner-Nordström Metric for an Ideal Charged Wormhole* ]{}, Phys. Rev. [**120**]{} 1507-1513 (1960)
Bretón, N.: [*Geodesic structure of the Born-Infeld black hole*]{}, Class. Quant. Grav. [**19**]{}, 601-612, (2002).
Boillat, G.: [*Nonlinear Electrodynamics: Lagrangians and Equations of Motion*]{}, J. Math. Phys. [**11**]{} 941-951 (1970).
Wald, R.: [*The First Law of Black Hole Mechanics*]{}, In ‘College Park 1993, Directions in general relativity, vol. 1’ 358-366. arXiv: gr-qc/9305022.
Ashtekar, A., Corichi, A., Sudarsky, D.: [*Hairy black holes, horizon mass and solitons*]{}, Class.Quant.Grav. [**18**]{}, 919-940, (2001).
Heusler, M., Straumann, N.: [*The Fist law of black hole physics for a class of nonlinear matter models*]{}, Class. Quant. Grav. [**10**]{}, 1299-1322, (1993).
Rasheed, D. A.: [*Non-Linear Electrodynamics: Zeroth and First Laws of Black Hole Mechanics*]{}, arXiv: hep-th/9702087.
Corichi, A., Nucamendi, U., Sudarsky, D.: [*Einstein-Yang-Mills isolated horizons: phase space, mechanics, hair and conjectures*]{}, Phys. Rev. [**D 62**]{}, 044046 (2000).
Ayón-Beato, E., García, A.: [*The Bardeen model as a nolinear magnetic monopole*]{}, Phys. Lett. B [**493**]{}, 149-152, (2000).
Ashtekar, A., Fairhurst, S., Krishnan, B.: [*Isolated horizons: Hamiltonian evolution and the first law*]{}, Phys. Rev. [**D 62**]{}, 104025 (2000).
Bretón, N.: [*Born-Infeld black hole in the isolated horizon framework*]{}, Phys. Rev. [**D 67**]{}, 124004 (2003).
Moreno, C., Sarbach, O. [*Stability properties of black holes in selfgravitating nonlinear electrodynamics*]{}, Phys. Rev. D [**67**]{}, 024028 (2003).
Bretón, N.: [*Stability of nonlinear magnetic black holes* ]{}, Phys. Rev. [**D 72**]{}, 044015 (2005).
Fernando, S.: [*Gravitational perturbations and quasi-normal modes of charged black holes in Einstein-Born-Infeld gravity*]{}, Gen. Rel. Grav. [**37**]{} 585-604 (2005). hep-th/0407062.
Wirschins, M., Sood, A., Kunz, J.: [*Non-Abelian Einstein-Born-Infeld black holes*]{}, Phys. Rev. [**D 63**]{}, 084002 (2001).
Eiroa, E. F.: [*Gravitational lensing by Einstein Born Infeld Black holes*]{}, Phys. Rev. [**D 73**]{}, 043002 (2006).
Mosquera-Cuesta, H. J., de Freitas Pacheco, J. A., Salim, J. A.: [*Einstein’s gravitational lensing and nonlinear electrodynamics*]{}, Int. J. Mod. Phys. [**A 21**]{}, 43-55 (2006).
Sheykhi, A., Riazi, N.: [*thermodynamics of black holes in (n+1) dimensional Einstein-Born-Infeld dilaton gravity*]{} arXiv: hep-th/0610085.
| {
"pile_set_name": "ArXiv"
} |
---
author:
- 'S. Hayes, A. Hundemer, E. Milliken, T. Moulinos\*'
title: 'The Real Dynamics of Bieberbach’s Example'
---
ABSTRACT.
INTRODUCTION
============
Bieberbach considered two domains $\Omega^+_{\mathbb{C}}$ and $\Omega^-_{\mathbb{C}}$ which are basins of attraction of the same automorphism $$f(z,w):=(w, \frac{z}{2}-w^3+\frac{3}{4}w)$$ Both basins are biholomorphic to ${\mathbb{C}}^2$, according to a result originating with Poincaré, but obviously not all of ${\mathbb{C}}^2$, since they are disjoint. These basins are symmetric with respect to the origin, and Bieberbach’s map $f$ is one of the simplest having two basins in such a geometric relationship to each other.
For over two decades now, with the incentive from the visualization possibilities offered by computer graphics, renewed interest in higher dimensional complex dynamics has led to many interesting topological results. For example, the basins $\Omega^+_{\mathbb{C}}$ and $\Omega^-_{\mathbb{C}}$ have the same boundary in ${\mathbb{C}}^2$ and that boundary is never a topological manifold (\[BS2\], Theorem 2). Surprisingly enough, however, computer pictures of the real sections of these boundaries look smooth. The purpose of this paper is to present a proof of that fact. It will be shown that the boundaries of the real basins $\Omega_+:=\Omega^+_{\mathbb{C}}\cap{\mathbb{R}}^2$ and $\Omega_-:=\Omega^-_{\mathbb{C}}\cap{\mathbb{R}}^2$ in ${\mathbb{R}}^2$ coincide and are composed exactly of the real stable manifolds of 3 saddle points. Whereas in the standard literature it is sometimes stated that basin boundaries are smooth on the basis of computer studies or numerical calculations (see \[R\], pg.503), an explicit proof is given here.
\
\
The paper is organized as follows: In the next section, first a closed polygon $R$ in $\bf {R}^2$ will be shown to contain all real points with bounded forward as well as backward orbits. Then the fate of the forward and the backward orbit of every point in $R$ will be described. In the third section, the stable and the unstable manifolds of every saddle point will be located. The real filled Julia sets and the real Julia sets are calculated explicitly in the fourth section in terms of the 5 real periodic points. The real basin boundaries can be completely described in the last section which also contains a revealing computer generated image of those basins.\
\* We thank Jeff Galas, who generated Figures 1-7, and Korrigan Clark for their contributions.
ORBIT BEHAVIOR
==============
Consider $f (x, y) = (y, \frac{x}{2} - y^3 + \frac{3}{4} y)$ as a self-map of $\bf {R}^2$. The points $p_+ = (\frac{1}{2},\frac{1}{2}), p_- = (-\frac{1}{2}, -\frac{1}{2})$ are obviously fixed points. The third fixed point is at the origin and that is a saddle. There is also a period two saddle at $p = (-\frac{\sqrt{5}}{2}, \frac{\sqrt{5}}{2})$, $p' = (\frac{\sqrt{5}}{2}, -\frac{\sqrt{5}}{2}).$ The backward iterate of $(x, y)$ is $f^{-1}(x, y) = (2x^3 -\frac{3}{2} x + 2y, x)$.\
[Hr]{}[0.4]{}
{width="38.00000%"}
The first objective is to locate the set $K_{\mathbb{R}}$ of points in $\bf{R}^2$ with bounded forward and bounded backward orbits, since they are the observables. $K_{\mathbb{R}}$ also generates the set $K^+_{\mathbb{R}}$ of points with bounded forward orbits as well as the set $K^-_{\mathbb{R}}$ of points with bounded backward orbits (see section 4). We will use a partitioning of the real plane similar to that in \[K, p.132\]. As a first estimate it will be shown that $K_{\mathbb{R}}$ is contained in a closed polygon $R$ with corners given by the 8 points: $p, (\frac{1}{2}, \frac{\sqrt{5}}{2}), p_+, (\frac{\sqrt{5}}{2}, \frac{1}{2}), p', (-\frac{1}{2}, -\frac{\sqrt{5}}{2}), p_-, (-\frac{\sqrt{5}}{2}, -\frac{1}{2})$.\
Moreover, except for the periodic points $ p, p', p_+, p_-$, the set $K_{\mathbb{R}}$ is in the interior of $R$. The proof will use the backward iterates
**Lemma 2.1** $K_{\mathbb{R}}\subset R $ and\
$K_{\mathbb{R}}\setminus \{p, p', p_+, p_-\} \subset int \, R$
The proof will show that outside of the interior of $R$ every point except $p, p', p_+, p_-$ escapes to infinity either under forward or under backward iteration of $f$.\
The complement of the interior of $R$ will be partitioned into the following $4$ closed quadrants $Q_k, 1 \leq k \leq 4$ and their reflections $Q_k ' = \sigma(Q_k)$ at the origin where $\sigma(x,y)= (-x, -y)$:
$$Q_1 : = \{(x,y) \in {\mathbb{R}}^2 : x \leq -1/2, \hspace{1mm} y \leq -1/2 \},\qquad Q_2 := \{(x,y) \in {\mathbb{R}}^2 : x \leq -\sqrt{5}/2, \hspace{1mm} y \leq \sqrt{5}/2 \}$$ $$Q_3 := \{(x,y) \in {\mathbb{R}}^2 : x \leq -\sqrt{5}/2, \hspace{1mm} y \geq \sqrt{5}/2 \} , \qquad Q_4 := \{(x,y) \in {\mathbb{R}}^2 : x \geq -\sqrt{5}/2, \hspace{1mm} y \geq \sqrt{5}/2 \}$$
[0.28]{} 
[0.28]{} 
[0.28]{} 
It will be shown that $f^n(x,y) \to \infty$ if $(x,y)$ is in $Q_3\setminus\{ p\}$ and $f^{-n}(x,y) \to \infty$ when $(x,y)$ is in $Q_1, Q_2,$ or $Q_4$ but is not $p$ or $ p_-$. If we interchange $Q_k$ and $Q'_k$, the corresponding statements hold, since $f \circ \sigma= \sigma \circ f$ and $f^{-1} \circ \sigma = \sigma \circ f^{-1}$. The quadrants are mapped as follows: $$f^{-1}(Q_1) \subset Q_1, f^{-1}(Q_2) \subset Q_4 ', f(Q_3) \subset Q_3', f^{-1}(Q_4) \subset Q_2'$$
[0.28]{} 
[0.28]{} 
[0.28]{} 
We will first consider $Q_3$. If $(x,y)$ lies in $Q_3$ but $y \not= \frac{\sqrt{5}}{2}$, then its image under $f$ lies further away from the origin with respect to the pseudonorm $|(x,y)|= |y - \frac{x}{2}|$ which is obviously the $y-$ intercept of the line through $(x, y)$ with slope $\frac{1}{2}$. To see this, notice that $|(x, y)| = y - \frac{x}{2}$ if $(x, y) \in Q_3$ and when $(x, y) \in Q'_3$, then $|(x, y)| = \frac{x}{2} - y$. Since $f(Q_3) \subset Q'_3$, if $(x, y) \in Q_3$ with $y \not= \frac{\sqrt{5}}{2}$, then
$$|f(x,y)| - |(x,y)| = -\frac{x}{2} + y^3 -\frac{y}{4} - y + \frac{x}{2} = y(y^2 - \frac{5}{4}) >0.$$
Using the fact that the pseudonorm is preserved by the reflection $\sigma$ at the origin, it immediately follows that when $(x, y) \in Q'_3\setminus\{p'\}$, then $|f(x, y)| - |(x, y)| > 0$ if $y \neq -\frac{\sqrt{5}}{2}$.
If $ (x, y) \in Q_3\setminus \{p\}$ with $y \neq \frac{\sqrt{5}}{2}$, then $y_1 \neq -\frac{\sqrt{5}}{2}$ for $f(x, y) = (x_1, y_1) \in Q'_3\setminus \{p'\}$ and $$|f^2(x, y)| > |f(x, y)| > |(x, y)|.$$ The same inequalities hold for $(x, y) \in Q'_3 \setminus \{p'\}$ with $y \neq -\frac{\sqrt{5}}{2}$ due to the properties of $\sigma$. By induction, for a point $(x,y)$ in $( Q_3 \cup Q'_3) \setminus \{p, p'\}$ with $|y| \neq \frac{\sqrt{5}}{2}$ the sequence $(|f^n (x,y)|)_{n \in \mathbb{N}}$ is strictly monotonically increasing.\
Whenever $| y| = \frac{\sqrt {5}}{2}$, after one iterate $|y_1| \neq \frac{\sqrt{5}}{2}$ for $f(x, y) = (x_1, y_1)$. Consequently, $(|f^n (x,y)|)_{n \geq 1}$ is strictly monotonically increasing for all points $(x,y)$ in $( Q_3 \cup Q'_3) \setminus \{p, p'\}$. That sequence is also unbounded; otherwise, it would converge to some value r where $|a|=r$ for every accumulation point $a$ of $(f^n(q))$. Because $Q_3$ is closed, $a \in Q_3$. This is a contradiction, since then $|f(a)| > |a|$ would follow, contradicting the fact that $f(a)$ is also an accumulation point in $Q_3$ and therefore $|f(a)| =|a|$. It follows that $f^n(x,y) \to \infty$ as $n \to \infty$ for $(x,y) \in Q_3 $.
Now consider $Q_1$. If $(x,y) \in Q_{1}$, then $f^{-1}(x,y) \in Q_{1}$. Using the pseudonorm $\lvert (x,y) \rvert = \lvert y+\frac{x}{2} \rvert$ in $Q_1$, we see that for $(x,y) \in Q_{1}$,
$$\lvert (x,y) \rvert = -y - \frac{x}{2} \quad and \quad \lvert f^{-1}(x,y) \rvert =\lvert (x_{-1},y_{-1}) \rvert = -y_{-1} - \frac{ x_{-1}}{2} = -x - x^{3} + \frac{3x}{4} - y = - \frac{x}{4} - x^3 - y.$$
Therefore, $$\lvert f^{-1}(x,y) \rvert - \lvert (x,y) \rvert = -x - x^{3} + \frac{3x}{4} - y +y + \frac{x}{2} = -x^{3} + \frac{x}{4} = -x (x^2 - \frac{1}{4}) .$$
This difference is positive for $ x < -\frac{1}{2}$. If $x = -\frac{1}{2}$ and $(x, y)\in Q_1$, then $f^{-1}(x,y) = (x_{-1}, y_{-1})$ satisfies $x_{-1} < - 1/2$ if and only if $y < -\frac{1}{2}$, implying by induction that for $(x, y) \in Q_1 \setminus p_-$ the sequence $(|f^{-n}(x,y)|)_{n \geq 1}$ is strictly monotonically increasing. As above, it is unbounded and $f^{-n}(x,y)\rightarrow \infty$ as $n \rightarrow \infty$ for $(x,y) \in Q_{1}\setminus p_-$ follows.
Finally, we will consider $Q_2$ and $Q_4$. Note that $Q_2$ and $Q_1$ overlap, and once any backward iterate of a point $(x, y)$ in $Q_2$ lands in $Q_1$, then its fate is sealed and $f^{-n}(x,y)\rightarrow \infty$ as $n \rightarrow \infty$. Thus, we only need to consider points $(x,y) \in Q_{2} $ such that $f^{-n}(x, y) \notin Q_1$ for every $n$.
Using the maximum norm $\lvert (x,y) \rvert = max\{ \lvert x \rvert, \lvert y \rvert \}$, we arrive at the following: If $ (x,y) \in Q_{2} \setminus Q_{1}$, then $x\leq -\frac{\sqrt{5}}{2}, -\frac{1}{2} \leq y \leq \frac{\sqrt{5}}{2}$ and $ \lvert (x,y) \rvert =\lvert x \rvert = - x.$ If $(x, y) \in Q'_4 \setminus Q_1$, then $ -\frac{1}{2} \leq x \leq \frac{\sqrt{5}}{2} , y \leq -\frac{\sqrt{5}}{2}$ and $ \lvert (x,y) \rvert =\lvert y \rvert = - y.$ Therefore, when $(x, y) \in Q_2 \setminus Q_1$ and $f^{-1}(x, y) \in Q'_4 \setminus Q_1$, the difference\
$$\lvert f^{-1}(x,y) \rvert - \lvert (x,y) \rvert =-x +x = 0.$$
However, when $(x, y) \in Q'_4 \setminus Q_1$, then it is no restriction to assume that $f^{-1}(x, y) \in Q_2 \setminus Q_1$, and the difference\
$$\lvert f^{-1}(x,y) \rvert - \lvert (x,y) \rvert = -2x^3 +\frac{3}{2}x - y$$ is positive if $ y < -\frac{\sqrt{5}}{2}$.
Consequently, $ \lvert f^{-2(n+1)}(x,y) \rvert > \lvert f^{-2n}(x,y) \rvert$ and $|f^{-(2n+1)}(x,y)| > | f^{-(2n-1)}(x,y)|$ for $n \in \mathbb{N}$ and $(x, y) \in Q_2 \setminus Q_1$ with $ y < -\frac{\sqrt{5}}{2}$. That means the sequence of the absolute values of the backward images is itself not strictly monotonically increasing, but the subsequence of the absolute values of the odd inverse images as well as the subsequence of the absolute values of the even inverse images are both strictly monotonically increasing and therefore unbounded, implying that $f^{-n} (x, y) \to \infty $.
**Corollary 2.2** The backward orbit of every point $q$ on the boundary of $R$ which is not $p, p', p_+$ or $p_-$ escapes, i.e. $f^{-n}(q) \to \infty$.
Since $q$ lies in one of $Q_1, Q_2, Q_4 $ or their reflections, Lemma 2.1 gives the result.
The orbit behavior inside $R$ will be studied in two steps. All points outside the open unit square $S = \{(x, y)\in {\mathbb{R}}^2 : | (x, y) | < \frac{1}{2}\} $ will be denoted by $T$ and will be considered first; $| (x, y)|$ denotes the maximum norm for points $(x, y)$ in $T$.
We subdivide $T$ as follows. Let $$T_{1}= \{(x,y) \in T : -y \leq x \leq \frac{1}{2}, \frac{1}{2} \leq y \leq \frac{\sqrt{5}}{2}\}, \qquad T_{2}= \{(x,y) \in T : -\frac{\sqrt{5}}{2} \leq x \leq -y, \frac{1}{2} \leq y < -x \}$$ $$T_{3}= \{(x,y) \in T :\frac{-\sqrt{5}}{2} \leq x \leq \frac{-1}{2}, \frac{-1}{2} \leq y \leq \frac{1}{2} \}$$\
As before, $\sigma (x, y) = (-x, -y)$ denotes the reflection at the origin, and for each $i \in \{1, 2, 3\}$, $\sigma(T_i)= T_i '$. It is clear that $T_{+} \cup T_{-} = T$ where $T_{+} = \bigcup_{i=1} ^{n=3}T_{i}$ and $T_{-} = \bigcup_{i=1} ^{n=3}T_{i}'$.\
[0.28]{} 
[0.28]{} 
We show now that the following mapping properties hold:\
$$f(T_1\setminus \{p\}) \subset T_2' \cup T_3', \quad f(T_{2}) \subset T_2' \cup T_3', \quad f(T_3 ) \subset S \cup T_1' ,\quad f(S) \subset S , \quad f(\bar S) \subset \bar S, \quad f^2(\bar S ) \subset S\cup \{p_+, p_-\}$$\
from which it follows that $f $ is forward invariant, i.e. $f(R) \subset R$. The mapping properties result from the simple fact that $-\frac{1}{4} \leq g(y) \leq \frac{1}{4}$ for the function $g(y) = -y^3 + \frac{3}{4}y$.\
[0.28]{}
[0.28]{}
**Proposition 1** $f(S)\subset S$ and $ f(\bar S) \subset \bar S$.
Since $S = (-\frac{1}{2},\,\frac{1}{2}) \times (-\frac{1}{2},\,\frac{1}{2})$ and $ f(x, y) = (x_1, y_1) = ( y, \frac{x}{2} + g(y))$, to show $f(S)\subset S$ we need only note that $| y_1|<\frac{1}{2} $. This implies $\overline{f(S)} \subset \bar S$. Futhermore, $S=f^{-1}(f(S) \subset f^{-1}(\overline{f(S)})$, which is a closed set since the pre-image of a closed set is closed under a continuous map. Thus, $\bar S \subset f^{-1}(\overline{f(S)})$ which yields $f(\bar S) \subset \overline{f(S)} \subset \bar S$.
**Proposition 2** $f^2(\bar S) \subset S \cup \{p_+,\,p_-\}$
$f( S) \subset S$ implies that $f^2(S) \subset S $. Therefore, we need only show that $f^2$ maps $\partial S$ into $ S \cup \{p_+,\,p_-\}$. The boundary of $S$ is composed of four line segments: $$\ell_1 = \{ (x, y): |x| \leq \frac{1}{2}, y = \frac{1}{2}\}, \qquad \ell_2 = \{(x,y): |x| \leq \frac{1}{2}, y= - \frac{1}{2}\}$$ $$\ell_3 = \{(x,y):|y| \leq \frac{1}{2}, x= - \frac{1}{2}\}, \qquad \ell_4 = \{(x,y): |y| \leq \frac{1}{2}, x = \frac{1}{2}\}$$ Denote $f(x, y) = (x_1, y_1) = (y, \frac{x}{2} + g(y))$. If $(x,y) \in \ell_1$ , then $x_1= \frac{1}{2}$ and $0\leq y_1\leq\frac{1}{2}$. Let $\ell'_4 = \{ (\frac{1}{2}, y): 0\leq y \leq\frac{1}{2}\}$. Thus, $f(\ell_1) \subset \ell'_4 \subset \ell_4$. When $(\frac{1}{2}, y) \in \ell'_4\setminus \{p_+\}$, then $0\leq y < \frac{1}{2}$ and $\frac{1}{4} \leq y_1 < \frac{1}{2}$, which means that $f(\ell'_4\setminus \{p_+\}) \subset S$ and therefore $f^2(\ell_1\setminus \{p_+\} )\subset S$. Since $\sigma (\ell_1) = \ell_2$, it follows that $ f^2(\ell_2\setminus \{p_-\}) \subset S$.\
The images of $\ell_3$ and $\ell_4$ behave differently under $f$, namely except for the corners $(-\frac{1}{2}, \frac{1}{2})$ and $p_-$, they land in $S$ after one iteration. To see this, consider $\ell''_4 = \{ (\frac {1}{2}, y) : -\frac{1}{2} < y <0 \}$. If $(\frac{1}{2}, y) \in \ell'_4 $, then $g(y) <0$ and $ -\frac{1}{2} < y_1 < \frac{1}{4} $, which implies $f(\ell''_4\setminus \{(\frac{1}{2}, -\frac{1}{2})\}) \subset S$. Since $f^2(\frac {1}{2}, -\frac{1}{2}) \in S$, it follows that $f^2(\ell_4\setminus \{p_+\}) \subset S$. Because $\sigma (\ell_4) = \ell_3$, we have $ f^2(\ell_3 \setminus \{p_-\}) \subset S$ and Proposition 2 follows.
{width="90.00000%"}
**Proposition 3** $f(T_1 \setminus \{p\}) \subset T_2' \cup T_3 ' $, $f(T_2) \subset T_2' \cup T_3 '$, $f(T_3) \subset S \cup T_1'$
Let $(x_1, y_1)= f(x,y)$. If $(x,y) \in T_1 \setminus \{p\}$, obviously $x_1 =y \in [\frac{1}{2}, \frac{\sqrt{5}}{2}]$. It will be enough to show that $ -x_1 < y_1= \frac{x}{2} + g(y) \leq \frac{1}{2}$. Now $y_1\leq \frac{1}{2}$, because $x \leq \frac{1}{2} $ and $g(y) \leq \frac{1}{4}$. For the lower bound, since $-y \leq x$, we have $ y_1 \geq - \frac{y}{2}+ g(y) \geq -\frac{y}{2} + \frac{1}{4} > - y$ if and only if $y> \frac{1}{2}$. However, when $y = \frac{1}{2}, y_1 \geq -\frac{1}{4} + \frac{1}{4} = 0 > -y = -\frac{1}{2}$. $> -y$ when $-y(y^2 - \frac{5}{4}) > 0$. That inequality is true for $y \in [\frac{1}{2},\frac{\sqrt{5}}{2}) $ but not for $ y = \frac {\sqrt{5}}{2}$. However, for $ y = \frac {\sqrt{5}}{2}$, $y_1 > -\frac{\sqrt{5}}{2}$ if and only if $x > -\frac{\sqrt{5}}{2}$, implying that $( x, y) $ cannot be $p$. Thus, $f(T_1 \setminus \{ p\}) \subset T_2' \cup T_3 '$.\
Let $(x,y) \in T_2$. Then, $ -\frac{\sqrt{5}}{2}\leq x \leq -y$ and $ \frac{1}{2} \leq y < -x$. Obviously, $\frac{1}{2} \leq x_1 \leq \frac{\sqrt{5}}{2}$. Next, we verify that $ -x_1 < y_1 \leq \frac{1}{2}$. Note that $ y_1 \geq -\frac{\sqrt{5}}{4} - y(y^2 - \frac{3}{4})\geq -\frac{\sqrt{5}}{4} - \frac{y}{2} \geq -y $, because $|y^2 - \frac{3}{4}| \leq \frac{1}{2}$ and $y \geq \frac{\sqrt{5}}{2}$. Therefore $(x_1, y_1) \in T_2' \cup T_3 '$.\
Let $(x,y) \in T_3 $. Then $-\frac{\sqrt{5}}{2} \leq x \leq -\frac{1}{2}$ and $ |y| \leq \frac{1}{2}$. It suffices to show that $-\frac{\sqrt{5}}{2} \leq y_1 \leq 0$. But $y_1 \leq -\frac{1}{4} + g(y) \leq 0$ and $y_1 \geq -\frac{\sqrt{5}}{4} + g(y) \geq -\frac{\sqrt{5}}{4} - \frac{1}{4} \geq -\frac{\sqrt{5}}{2}.$
The next Lemma treats all forward and all backward orbits of points in $T$.
**Lemma 2.3** The forward orbit of a point $q$ in $T$ which is not $p$ or $p'$ either eventually lands in $S$ or it stays in $T$ and converges to $p_+$ or $p_-$, i.e. $q \in \Omega_+ \cup \Omega_-$. The backward orbit of a point q in T which is not $p_+$ or $p_-$ either eventually lands outside $R$ and escapes, i.e. $q\in W^u(\infty)$, or it remains in T and converges to $\{p,\,p'\}$, i.e. $q\in W^u(p,\,p')$.
Note that the maximum norm $|(x, y)|$ is $|y| = y $ for $(x, y) \in T_1$ and $|(x, y)| = |x| = -x $ for $(x, y) \in T_2 \cup T_3.$ Similar statements hold when $T_i$ is replaced by $T'_i$. Furthermore, when $(x, y) \in T_1$, then $|f(x, y)| = |(x, y)| = |y|$. When $(x, y) \in T_2, $ then $|(x, y)| = |x| > |f(x, y)| = |y|$, because $|x| = -x , |y| = y$ and $ -x > y$ by definition of $T_2$.
However, if $f(x, y) \in S$, the next Lemma will treat that forward orbit. Consequently, after Proposition 3, we need only consider $(x, y) \in T_3$ with $f(x, y) \in T'_1$. In that case, $|(x, y)| = |x| = -x$ and $|f(x, y)| = -\frac{x}{2} + y^3 -\frac{3}{4}y $. Thus, $|(x, y)| > |f(x, y)| $ if and only if $g(y) = y(y^2 - \frac{3}{4}) < -\frac{x}{2} $. However, $ \frac{1}{4} \leq -\frac{x}{2} $ and $g(y) < \frac{1}{4}$ if $y > -\frac{1}{2} $. When $ y = -\frac{1}{2}$, then $|(x, y)| > |f(x, y)|$ if and only if $x < -\frac{1}{2}.$ Therefore, if $(x, y) \in T_3 \setminus \{p_-\} $, then $|(x, y)| > |f(x, y)|$. Similarly, if $(x, y) \in T'_3 \setminus \{ p_+\} $ with $ f(x, y) \in T_1$, then $|(x, y)| > |f(x, y)|$ . To summarize, for all points $(x, y) \in T $ whose forward orbit remains in $T$, the sequence $(| f^n(x, y)| )_n$ is monotonically decreasing. Furthermore, the sequence of the norms of all even forward iterates $(|f^{2n}(x, y)|)_n$ is strictly monotonically decreasing for every point $(x, y) \in T \setminus \{p_+, p_-\}$ It remains to show that $f^n(x,y) \to p_+ $ or $f^n(x,y) \to p_-$. Since $|f^n(x, y)| \geq \frac{1}{2}$ for every $n\in \mathbb{N}$, this sequence converges. Let $r= \lim_{n \to \infty} |f^n(x, y)|$. Then $r\geq \frac{1}{2}$.
We will see now that $r=\frac{1}{2}$. Because $T$ is compact, the forward iterates $f^n(x, y)$ have at least one accumulation point $a$. Then $|a|$ is an accumulation point of the convergent sequence $(|f^n(x, y)|)_n$ and $|a| = r$. Since $f(a)$ and $f^2(a)$ are also accumulation points of $f^n(x, y)$, we have $|a| = |f(a)| = |f^2(a)| = r$. If $r> \frac{1}{2}$, then every accumulation point $a$ would lie in $T$ and the contradiction $|f^2(a)| < |a| = r = \frac{1}{2} $ would follow.\
Thus, every accumulation point $a$ must lie on the boundary of $S $. The only possible accumulation points however are $p_+$ and $p_-$, since otherwise $f^2(a) \in S $ as noted above in Proposition 2. The backward invariance of the two basins $\Omega_+, \Omega_-$ implies then that $(x,y)$ must lie in one of them and consequently $f^n(x,y) \to p_+ $ or $f^n(x,y) \to p_-$.
Consider a point q in T which is not $p_+$ or $p_-$. We will investigate the behavior of the backward orbit $O^-_f(q)=\{f^{-n}(q):\;n\in \mathbb{N}\}$ of $q$. For $q\in T$, $O^-_f(q)\cap S=\emptyset$ because otherwise the forward invariance of S (Proposition 1) results in the contradiction $q \in S$. Thus, either $O^-_f(q)\subset T$ or there is an $n$ with $f^{-n}(q)$ not in $T$ and therefore not in $R$. If $f^{-n}(q)$ is not in $R$, it cannot be in $Q_3\cup Q'_3$, due to forward invariance, and therefore $f^{-n}(q) \to \infty$ by Lemma 2.1.
Consider the case $O^-_f(q)\subset T$. The sequence $(|f^{-n}(q)|)_n$ will be shown to be monotonically increasing. If $q= (x, y) \in T_2 \cup T_3$ and $f^{-1}(q) = ( x_{-1}, x)$ then obviously $|f^{-1} (q)| \geq |x| = |q|$ by the definition of the maximum norm. If $q =(x,y)\in T_1$, then $ |q|=|y| = y$ and $f^{-1}(q)\subset T'_3 \cup (\mathbf{R^2}\setminus R)$. When $f^{-1}(q) \in T'_3$, then $|x| \leq \frac{1}{2} $ and $|f^{-1}(q)| = |x_{-1}| = x_{-1} = x( 2x^2 - \frac{3}{2}) + 2y. $ Consequently, $|f^{-1}(q)| = x_{-1} \geq |q| = y $ if and only if $x(2x^2 - \frac{3}{2}) \geq -y $. But $-y \leq -\frac{1}{2}$, and thus $-\frac{1}{2} \leq x(2x^2 - \frac{3}{2})$, since $-x \geq -\frac{1}{2}$ and $ - (2x^2 - \frac{3}{2}) \geq 1$, proving the monotonicity. If $a$ is any accumulation point of $(f^{-n}(q))_n$, let $r= |a|$. Obviously, $r \leq \frac{\sqrt{5}}{2}$. Because $f^{-m}(a)$ is also an accumulation point, it follows that $ |f^{-m}(a)| = |a| = r$ for every $m \in \mathbb{N}$. Hence, $a $ is on the boundary of $R$ and $r = \frac{\sqrt{5}}{2}$, implying that $a$ is not $p_+$ or $p_-$. Due to 2.2, $a$ must be either $p$ or $p'$ proving the claim.\
To treat the orbit behavior inside $S$, subdivide $S\setminus (0,0)$ into 4 open squares
$$S_{1}=\{(x,y)\in \mathbb{R}^2 : -\frac{1}{2} < x < 0,\space - \frac{1}{2} < y < 0 \}, \qquad S_{2}=\{(x,y)\in \mathbb{R}^{2} : -\frac{1}{2} < x < 0, 0 < y < \frac{1}{2} \},$$ $$S'_1 = \sigma (S_1), \qquad S'_2 = \sigma (S_2)$$.
The preimages of $S_i$ and $T_i$ under $f$ are depicted below:
{width="90.00000%"}
{width="90.00000%"}
**Proposition 4** $f(S_1) \subset S_1$, $f(S_2) \subset S'_1 \cup S'_2$\
$f^{-1}(S_1) \subset S_1 \cup S_2' \cup T_3$, $f^{-1}(S_2) \subset S'_2 \cup T'_3 \cup (\mathbb{R}\setminus R)$
Let $(x, y) = S_1$. Clearly, $\frac{x}{2} +y(\frac{3}{4} - y^2) < 0$, since $\frac{3}{4} - y^2 > \frac{1}{2}$. From $- \frac{1}{2} - \frac{x}{2} < -\frac{1}{4} < \frac{y}{2} < y (\frac{3}{4} - y^2) $, it follows that $\frac{x}{2} +y(\frac{3}{4} - y^2) > -\frac{1}{2}$. and $f(x, y) \in S_1$. If $(x, y) \in S_2$, then $f(x, y) $ is in the right half plane. Due to Proposition 1, $f(S) \subset S$, implying $f(S_2) \subset S'_1 \cup S'_2$.
To see that $f^{-1}(S_1) \subset S_1 \cup S_2' \cup T_3$, let $(x,y)$ be a point in $S_1$ and let $f^{-1}(x,y)=(x_{-1},y_{-1})$. It is clear that $-\frac{1}{2} < y_{-1} = x < 0$. We note that $-1 < x_{-1} = 2y +2x^3 -\frac{3}{2}x < \frac{1}{2}$. This is because $-\frac{1}{2} < x < 0$ and $2x^2 - \frac{3}{2} < -1$ imply $x_{-1} < \frac {1}{2}$ whereas $-1 < 2y < 0 < 2x^3 - \frac{3}{2} x $ implies $ x_{-1} > -1$. To prove $f^{-1}(S_2) \subset S'_{2} \cup T_3 ' \cup ({\mathbb{R}}\setminus R)$, let $(x,y) \in S_2$. It will be enough to show that $x_{-1} > 0$ which follows from $x_{-1} \geq x(2x^2 -\frac{3}{2})$ due to $x<0$ and $x^2 <\frac{1}{4}$.
Obviously, the corresponding statements are true if $S'_i$ is interchanged with $S_i$ and $T'_3$ with $T_3$.
The axes inside $S$ are mapped into $S_1 \cup S'_1$ after at most two forward iterations:\
**Proposition 5** $f(\{(0, y): 0<y <\frac{1}{2}\}) \subset S'_1, \qquad f(\{(0, y): -\frac{1}{2} < y < 0\}) \subset S_1$,\
$ f^2(\{(x, 0): 0<x< \frac{1}{2}\}) \subset S'_1,\qquad f^2(\{(x, 0): -\frac{1}{2} < x < 0\}) \subset S_1.$
The positive $y-$axis in $S$ is mapped into $S'_1$, because $ 0<y <\frac{1}{2}$, implies $0 < y(\frac{3}{4} - y^2) < \frac{1}{2} (\frac{3}{4} - y^2) < \frac{1}{2}$. The negative $y$-axis in $S$ is mapped into $S_1$ due to $\sigma \circ f = f \circ \sigma$. The negative $x-$ axis in $S$ is mapped into the negative $y-$ axis in $S$ by $f$ and thus after another iteration it is mapped into $S_1$. Similarly, the positive $x-$ axis in $S$ is mapped into $S'_1$ after two iterations.
All forward and all backward orbits of points in $S$ are treated next:
**Lemma 2.4** The forward orbit of a point q in S stays in S and converges to either $(0, 0), p_+$ or $p_-$, i.e. $q \in W^{s}(0) \cup \Omega_-\cup \Omega_+$. The backward orbit of $q$ either eventually leaves $S$ for $T$ or it remains in $S$ and converges to the origin i.e. $q\in W^u(0)$. Furthermore, $\Omega_+$ contains $S'_1$ and $S_1$ is in $\Omega_-$.
First consider points $q$ in $S_1 \cup S'_1$. We will show that $f^n(q) \to p_+$ for $q \in S'_1$, from which $f^n(q) \to p_-$ for $q \in S_1$ follows, due to $ \sigma \circ f = f \circ \sigma$. Let $q = (x, y) \in S'_1$. Using the pseudonorm $|(x, y)| = y + \frac{x}{2}$, $$|f(x,y)| - |(x,y)| = \frac{x}{2} - y^3 + \frac{3}{4}y + \frac{y}{2} - y - \frac{x}{2} = -y(y^2 - \frac{1}{4} ) > 0$$. By induction, the sequence $(|f^{n}(q)|)_n$ is strictly increasing. It is obviously bounded, therefore it must converge. Let $r $ denote the limit. Let $a$ denote an accumulation point of the forward orbit $(f^n(q))_n$. Then $|a| = r$. Since $f^m(a)$ is also an accumulation point for every $m$, it follows that $|a| = |f^m(a)| = r$ for every $m$, implying that $a$ must be on the boundary of $S'_1$, which means that $a$ is on the boundary of $T$ or on the positive $x-$ or $y-$ axis in $S$. The latter case cannot happen, because then $a$ would be mapped into $S'_1$ after two iterations by Proposition 5. Hence, $a$ must be in $T$, and Lemma 2.3 then shows that $f^n(q) \to p_{+}$.
Consider now points $q$ in $S_2 \cup S'_2$. Since forward orbits landing in $S_1 \cup S'_1$ have already been treated, because of Proposition 4 it suffices to look at points $q \in S_2$ with $f^{2n}(q) \in S_2 $ for all $n\in \mathbb{N}$ and show that $f^{2n}(q) \to (0, 0)$.
Using the pseudonorm $ |(x, y)| = |y - \frac{x}{2}|$, we have $|f(q)| - |q| = y(y^2 - \frac{5}{4}) $ for $q = (x, y)$ which is negative for $q = (x, y) \in S_2 $ with $f(q) \in S'_2$ and for $q\in S'_2 $ with $f(q) \in S_2$. Therefore, $(|f^{n}(x,y)|)_n$ is strictly monotonically decreasing, as is $(|f^{2n}(q)|)_n$. Let $r = \lim_{n \to \infty}|f^{2n}(q)|$. Then $r \geq 0$. Since $|a| = |f^2(a)| = r $ for every accumulation point $a$ of the forward orbit $(f^{2n}(q))_n$, $a$ cannot be in $S_2 $ and must be on $ \partial S_2 $. Due to Proposition 5, $a$ must be the origin or on $\partial S_2 \cap \partial S$. But if $a = (x, \frac{1}{2}), -\frac{1}{2} \leq x \leq 0$, then $|a| \geq \frac{1}{2},$ implying that $a$ cannot be an accumulation point. Similarly, if $a= (-\frac{1}{2}, y), 0 \leq y \leq \frac{1}{2}$, then $|a| \geq \frac{1}{4}$ which means that such an $a$ also cannot be an accumulation point. Consequently, $a $ is the origin, $r = 0 = \lim _{n \to \infty}| f^{2n}(q)|$. Then $f^{2n + 1}(q) \to (0, 0)$, since the origin is a fixed point for $f$, and hence $f^n(q) \to (0, 0)$ follows.
We turn to the backward orbits of points $q$ in $S$. It suffices to consider the two cases: $O^-_f(q)\subset S_1$ and $O^-_f(q)\subset S_2 \cup S'_2$ . Suppose now that $ q = (x,y) \in S_1$ and $f^{-n}(q) \in S_1$ for all $n\in \mathbb{N}$. Letting $|(x,y)|=|y + \frac{x}{2}|$, we show that the sequence $(|f^{-n}(q)|)_{n \in \mathbb{N}}$ is strictly monotonically decreasing and converges to $0$. To see this, we note that for all such $(x,y) \in S_1$, the following inequality holds:
$$\begin{aligned}
|f^{-1}(x,y)| - |(x,y)| & = |\frac{1}{4}x + x^3 + y| -|y+\frac{x}{2}|\\
&= (-\frac{1}{4}x - x^3 -y) - (-y -\frac{x}{2}) \\
&=\frac{1}{4}x - x^3<0 \\\end{aligned}$$
which is true for $-\frac{1}{2}< x < 0$. As the sequence $(|f^{-n}(q|)_{n \in \mathbb{N}}$ is strictly monotonically decreasing and bounded from below by zero, we conclude that $f^{-n}(q) \to (0,0)$. Otherwise $(|f^{-n}(q)|)_{n \in \mathbb{N}}$ would converge to some constant $r>0$. If the point $a$ is an arbitrary accumulation point for the backward orbit $(f^{-n}(q))_{n \in \mathbb{N}}$, then $|a|=r$. Since $f^{-1}(a)$ is also an accumulation point, it follows that $|f^{-1}(a)|=|a|=r$ and $a$ cannot be in $S_1$, implying that $a$ is on $\partial S_1$. By Proposition 5, $a$ cannot be on the negative $x-$ or $y-$ axes, and thus $a$ is either the origin or on $\partial S$. The latter situation would mean $f^{-n}(a) \in \partial S$ for all $n$, contradicting Lemma 2.3 which states that $a \in W^u(p, p')$. Therefore, $f^{-n}(x,y) \to (0,0)$.
Using the pseudonorm defined by $|y - \frac{x}{2}|$, we now prove that $|f^{-1}(x,y)| - |(x,y)|> 0$ for every point $(x,y)\in S_2 \cup S_2'$. Without loss of generality, suppose $(x,y)$ is in $S_2$. By definition $|f^{-1}(x,y)|= |\frac{7}{4}x - x^3 -y|$. Since $\frac{7}{4}x - x^3 -y$ is negative and $y-\frac{x}{2}$ is positive for all relevant values of $x$ and $y$, the result will follow if $- (\frac{7}{4}x - x^3 -y) > y-\frac{x}{2}$, i.e. $-\frac{7}{4}x +x^3 > -\frac{x}{2}$ which is true for all $x$ in the interval $(-\frac{1}{2}, 0)$. Thus, $(|f^{-n}(x,y)|)_{n \in \mathbb{N}}$ is strictly monotonically increasing for all $(x,y) \in S_2 \cup S_2'$. From this we may deduce that $(S_2 \cup S_2') \cap W^{u}(0) = \emptyset$.\
**Remark** $S \subset W^s(0)\cup\Omega_+\cup\Omega_- $, $S \subset W^u(\infty) \cup W^u(p, p') \cup W^u(0)$,\
$W^s(0) \cap S \subset S_2 \cup S'_2 $, $ W^u(0) \cap( S_2 \cup S'_2) = \emptyset$\
Combining Lemmas 2.3 and 2.4, we know the fate of every forward and every backward orbit of points in $R$:\
**Lemma 2.5** The forward orbit of a point $q$ in $R$, which is not $p$ or $p'$, converges to $0$, $p_+$, or $p_-$, i.e. $q\in W^s(0)\cup\Omega_+\cup\Omega_-$. The backward orbit of a point $q$ in $R$ different from $p_+$ or $p_-$ converges to $0$, $\{p,p'\}$, or escapes, i.e. $q\in W^u(0)\cup W^u(p,p') \cup W^u(\infty)$.\
In particular, we now know that $R$ is contained in the set $K_{\mathbb{R}}^+$ of real points with bounded forward orbits. Moreover, Lemma 2.1 implies the following result contained in \[K, Proposition 7.10\]:\
**Remark** There are only 5 real periodic points for $f$, namely the three fixed points $(0,0), p_+, p_-,$ and the period 2 cycle $ p, p'$.\
STABLE AND UNSTABLE MANIFOLDS
=============================
There is an unstable eigenvector $v$ of $D\,f^2(p)$ pointing into the interior of $Q_3$ and a parameterization $\Gamma:\;{\mathbb{R}}\to{\mathbb{R}}^2$ of the unstable manifold $W^u_{f^2}(p)$ of $p$ with respect to $f^2$ having $v$ as its tangent at $p$. A similar statement holds for a parameterization $\Gamma'$ of $W^u_{f^2}(p')$ replacing $p$ by $p'$ and $Q_3$ by $Q'_3$. Denote $\Gamma({\mathbb{R}}_-)\cup\Gamma'({\mathbb{R}}_-)$ by $W^u_-(p,p')$ for the negative real numbers ${\mathbb{R}}_-$and $W^u_+(p,p'):=\Gamma({\mathbb{R}}_+)\cup\Gamma'({\mathbb{R}}_+)$. Note that $W^u(p,p')=W^u_-(p,p')\cup W^u_+(p,p')\cup\{p,p'\}$.
**Theorem 3.1**
\(i) $ W^u(0)\subset S_1\cup S'_1\cup \{0\} $\
(ii) $\overline{W^u(0)}=W^u(0)\cup \{p_+,p_-\} $\
(iii) $W_-^u(p,p')\subset int R\;and\;W_+^u(p,p')\subset Q_3\cup Q'_3$\
(iv) $ W^u_-(p,p')\not\subset W^s(0)$\
(v) $ W^s(p,p')\cap R=\{p,p'\}$
$(i)$ If $q \not \in R$ and $q \not \in Q_3 \cup Q'_3$, then $f^{-n}(q) \to \infty$ follows from Lemma 2.1 and therefore $ q \not \in W^u(0)$ showing that $W^u(0) \subset R \cup (Q_3 \cup Q'_3)$. Because $W^u(0)$ is connected, if there was a point in $W^u(0) \cap( Q_3 \cup Q'_3)$, then $p$ or $p'$ would have to be on $W^u(0)$, a contradiction. Thus, $W^u(0) \subset R$. However, if $q\in W^u(0) \cap T$, Lemma 2.3 implies that $q \in W^u(p, p')$ would result, a contradiction, and consequently, $W^u(0) \subset S$. By the above remark, $ W^u(0) \cap (S_2 \cup S'_2) = \emptyset$ and (i) follows.
$(ii)$ If $p_-$ is a limit point of the unstable manifold $W^u(0)$, then by symmetry ( using the reflection map ), $p_+$ is also a limit point. To show that $p_-$ is a limit point, take any point $q \in W^u(0)$ which is not the origin. By $ (i)$ it is no restriction to assume that $q \in S_1$. After Lemma 2.4, $f^n(q) \to p_-$. The invariance of the unstable manifold implies that $f^n(q) \in W^u(0)$ and hence $p_- \in \overline{W^u(0}$. To see that $p_-$ and $p_+$ are the only limit points, let $q \in L = \overline{W^u(0)} \setminus W^u(0)$. By (i), $q \in \overline{S_1} \cup \overline {S'_1}$. Without restriction, let $q \in \overline {S_1}$. Then 2.5 implies that $q \in \partial S_1$, because otherwise $q \in W^u(\infty) \cup W^u(p, p')$. Since the set $L$ of limit points is closed and invariant, $f^{-n}(q) \in L$ for all $n$, implying that $q \notin W^u(\infty)$ and thus $q \in W^u(p, p')$ must follow. This in turn would mean that $\{p, p'\} \subset L \subset \overline{S}$ which is a contradiction. If $q \in \partial S_1$ but $q$ is not $p_-$, then by 2.3 it would follow that $q \in W^u(p, p')$ if $q$ was not on an axis in $S$ and once again the same contradiction. If $q$ was on an axis in $S$, then $f^2(q) \in S_1$ by Proposition 5 and the contradiction $q \in W^u(p, p')$ again follows from 2.5.
$(iii)$ Let $J_{f^2}(p)$ and $J_{f^2}(p')$ represent the Jacobian matrices of $f^2$ at $p$ and $p'$ respectively. Then, by the chain rule, $J_{f^2}(p)=J_f(p)\cdot J_f(p')=J_{f^2}(p') = $
$$\begin{bmatrix} \frac{1}{2} & -3 \\ -\frac{3}{2} & \frac{19}{2}
\end{bmatrix}$$ which yields the unstable eigenvector $e_u=\langle -.32, 1 \rangle$. Hence, by the unstable manifold theorem, $W^u_{f^2}(p)$ (resp. $W^u_{f^2}(p')$) is tangent to the parameterization $\Gamma:\mathbb{R}\to\mathbb{R}^2$ defined by $t \mapsto p+t\cdot e_u$ (resp. $\Gamma':\mathbb{R}\to\mathbb{R}^2$ defined by $t \mapsto p'-t\cdot e_u$). Thus, it is clear that $W_-^u(p,p')=\Gamma(\mathbb{R}_-)\cup\Gamma'(\mathbb{R}_-)\subset R$ by the forward invariance of $R$. Furthermore, since for $q\in Q_3\cup Q'_3, \;f^n(q)\to\infty$ as $n\to\infty$, it follows that $W_+^u(p,p')=\Gamma(\mathbb{R}_+)\cup\Gamma'(\mathbb{R}_+)\subset Q_3\cup Q'_3$.
$(iv)$ If $ W^u_-(p,p')\subset W^s(0)$ were to hold, assume that $W^u_-(p, p') = W^s(0) \setminus \{0\} $ would follow (a fact that will be proved later). Now it will be shown that $W^u_-(p, p') = W^s(0) \setminus \{0\} $ implies a contradiction. This assumption means that $W^s(0) \subset int R$ by $(iii)$. However, a point $q \in \partial R \cap W^s(0) $ can be constructed giving the contradiction. Let $a = (\frac{\sqrt{5}}{2}, -\frac{1}{3})$. Obviously, $a \in \partial R$. Furthermore, $f^2(a) \in S_1'$ which means by Lemma 2.4 that $f^2(a) \in \Omega_+$. Consequently, $a\in \Omega_+$ due to the invariance of basins of attraction. Let $b = (\frac{\sqrt{5}}{2}, -\frac {33}{32})$. Obviously, $b\in \partial R$. A calculation shows that $f^5(b) \in S_1$ and therefore $b \in \Omega_-$, again because of 2.4 and invariance.\
The line segment connecting $a$ and $b$ on $\partial R$ intersects $\Omega_+$ as well as its complement and therefore must contain a point $q \in \partial
\Omega_+$. Since $q $ is in $R$ but neither in $\Omega _+$ nor in $\Omega_-$, by 2.5 $q\in W^s(0)$ must follow.\
The final step is to prove that $W^{u}_{-}(p,p') \subset W^s(0)$ implies $W^{u}_{-}(p,p')= W^{s}(0) \setminus \{0\}$. We denote the two path components of $W^s(0) \setminus \{0\}$ by $C_1$ and $C_2$. Let $\Gamma: {\mathbb{R}}\to {\mathbb{R}}^2$ and $\Gamma ': {\mathbb{R}}\to {\mathbb{R}}^2$ denote the parametrizations of $W^u_{f^2}(p)$ resp. $W^u_{f^2}(p')$. By definition, $W^{u}_{-}(p,p') = \Gamma({\mathbb{R}}_-) \cup \Gamma ' ({\mathbb{R}}_-)$. By our assumption, $\Gamma({\mathbb{R}}_-),$ and $\Gamma ' ({\mathbb{R}}_-)$ are path connected subsets of $W^s(0) \setminus \{0\}$. Since $\Gamma({\mathbb{R}}_-)$ and $\Gamma ' ({\mathbb{R}}_-)$ are both forward and backward invariant under $f^2$, we may conclude without restriction that $\Gamma({\mathbb{R}}_-)= C_1$ and $\Gamma'({\mathbb{R}}_-)= C_2$.
$(v)$ By Lemma 2.5, the forward orbit of a point $q$ in $R$ which is not $p$ or $p'$ converges either to $0$, $p_+$ or $p_-$. In other words, such a point $q$ does not belong to $W^s(p, p')$. Hence, $W^s(p, p') \cap R = \{p, p'\}$.
Note that Lemma 2.5 and Theorem 3.1(i) imply the\
**Remark** $W^u(0)\setminus\{0\} \subset \Omega_+ \cup \Omega_-$\
JULIA SETS
==========
The real filled Julia sets $K_{\mathbb{R}}:= K\cap {\mathbb{R}}^2,\;K^+_{\mathbb{R}}:=K^+\cap {\mathbb{R}}^2,\;K^-_{\mathbb{R}}:=K^-\cap{\mathbb{R}}^2$ and the real Julia sets $J^+_{\mathbb{R}}:=J^+\cap{\mathbb{R}}^2,\;J^-_{\mathbb{R}}:=J^-\cap{\mathbb{R}}^2$ and $J_{\mathbb{R}}:=J\cap{\mathbb{R}}^2$ for $f$ can now be calculated in terms of the stable and unstable manifolds of the 5 periodic points $\{0,p_+,p_-,p,p'\}$. Note that $K^+ $ and $K^-$ are closed and $K$ is compact in ${\mathbb{C}}^2$ by \[FM\]. By definition, $J^+= \partial K^+, J^- = \partial K^-$ and $J= J^+\cap J^-$. Note also that $K^+_{{\mathbb{R}}} = W^s(K_{{\mathbb{R}}}) = \{ q \in {\mathbb{R}}^2 : d(f^n(q), K_{{\mathbb{R}}}) \to 0\}$ and $K^-_{{\mathbb{R}}} = W^u(K_{{\mathbb{R}}}) = \{ q \in {\mathbb{R}}^2 : d(f^{-n}(q), K_{{\mathbb{R}}}) \to 0\}$ where $d$ denotes the Euclidean metric in ${\mathbb{R}}^2$, see \[BS\]. In \[BS\] it is shown that $J^+ $ is the closure in ${\mathbb{C}}^2$ of the stable manifold of any saddle point in ${\mathbb{C}}^2$ and $J^+ $ is also the boundary in ${\mathbb{C}}^2$ of every complex basin of attraction.
**Theorem 4.1** $$\begin{aligned}
(i &) \; K_{\mathbb{R}}= \overline{W^u(0)} \cup W^u_-(p,p') \cup \{p,p'\}\\
(ii &) \; K_{\mathbb{R}}^+=\Omega_+\cup\Omega_-\cup W^s(0)\cup W^s(p,p') \\
(iii &) \; K^-_{\mathbb{R}}=\overline{W^u(0)}\cup W^u(p,p')\\
(iv &) \;J^+_{\mathbb{R}}=W^s(0) \cup W^s( p,p') \\
(v &) \;J^-_{\mathbb{R}}=K^-_{\mathbb{R}}\\
(vi &) \;J_{\mathbb{R}}=\{0,p,p'\}\cup(W^u_-(p,p')\cap W^s(0))\end{aligned}$$
\(i) Clearly, $K_{{\mathbb{R}}} \supset \overline{W^u(0)} \cup W^u_-(p,p') \cup \{p, p' \}$, since $W^{u}(0) \subset R$, implying $\overline{W^u(0)} \subset R,$ and $W^u_-(p,p') \subset R$ by Theorem 3.1 and because $R \subset K_{{\mathbb{R}}}^+$ by Lemma 2.5.\
For the reverse inclusion, let $q$ be a non-periodic point in $K_{{\mathbb{R}}}$. As $K_{{\mathbb{R}}}$ is invariant, the backward orbit of $q$ stays in $K_{\mathbb{R}}$ hence in $R$ due to Lemma 2.1. Then, $ q \in W^u(0) \cup W^u(p, p')$ by Lemma 2.5.\
(ii) The inclusion of the right hand side of the equation in the left hand side is immediate. Let $q \in K_{\mathbb{R}}^+$. If the forward orbit $O^+_f(q)$ eventually lands in $R$, then $q\in \Omega_+\cup \Omega_- \cup W^s(0)$ by Lemma 2.5. Now let $O^+_f(q) \cap R = \emptyset$. Then $ d(f^n(q), K_{{\mathbb{R}}}) \to 0$ and $q \notin \{p, p', p_+, p_-\}$. But $K_{{\mathbb{R}}} \setminus \{p, p', p_+, p_-\}$ is in the interior of $R$ by Lemma 2.1, implying that $d(f^n(q), \{p, p', p_+, p_-\}) \to 0$. If $p_+ $ and $p_-$ are limit points of $O^+_{f}(q)$, then $q \in \Omega_+\cup \Omega_- $. If $q \notin \Omega_+\cup \Omega_- $, then $d(f^n(q), \{p, p'\}) \to 0$ and $q \in W^s(p, p')$.\
(iii) The inclusion $\supset$ is obvious. To show the opposite inclusion, let $q $ be a point in $ K_{\mathbb{R}}^-$ which is different from $p, p', p_+$ and $ p_-$. If $q \in R$, then $q \in K_{\mathbb{R}}$, since $R$ is forward invariant, and the claim follows from (i). If $q \notin R$, then the entire backward orbit $O^-_f(q) $ is not in $R$. However, $K_{\mathbb{R}}^- = W^u(K_{\mathbb{R}})$ implies $d(f^{-n}(q), K_{\mathbb{R}}) \to 0$, and again from Lemma 2.1 it follows that $d(f^{-n}(q), \{p, p'\})\to 0$, i.e. $q\in W^u(p, p')$.\
(iv) To prove the inclusion $\subset$, let $q\in J^+_{\mathbb{R}}$. Then $q\in K^+_{\mathbb{R}}$, since $ J^+_{\mathbb{R}}= \partial K^+ \cap {\mathbb{R}}^2 \subset K^+ _{\mathbb{R}}$ because $K^+$ is closed. On the other hand, $q\in J^+ = \partial \Omega^C_+ = \partial\Omega^C_-$ by Theorem 2 of \[BS\], where $ \Omega^C_+$ is the basin of attraction of $p_+$ in $C^2 $ and $ \Omega^C_-$ that of $p_-$ . It follows that $q\notin \Omega_+$ and that $q\notin \Omega_-$. Part (ii) shows that $q \in W^s(0)\cup W^s(p,p') $. The opposite inclusion follows from Theorem 1 in \[BS\] which proves that the closure of the stable manifold in $C^2$ of any saddle point is $J^+$.\
(v) The Jacobian determinant $det \,Df$ of $f$ is $-\frac{1}{2}$ which implies by \[FM, Lemma 3.7\] that the 4-dimensional Lebesque measure of $K^-$ is zero. Consequently, $K^-$ has no interior points and $J^- = \partial K^- = K^-$ gives the claim $J^-_{\mathbb{R}}= K^-_{\mathbb{R}}$.\
(vi) Using (iv), (v), and (iii),
$$J_{\mathbb{R}}= (W^s(0) \cup W^s(p, p')) \cap (\overline{W^u(0)} \cup W^u(p, p'))$$ $$=(W^s(0) \cap \overline{W^u(0)}) \cup (W^s(p, p')\cap \overline{W^u(0)}) \cup (W^s(0)\cap W^u(p, p')) \cup (W^s(p, p')\cap W^u(p, p')).$$\
We will treat the last intersection first and show that $W^s(p, p')\cap W^u(p, p') = \{p, p'\}$. Now $W^s(p, p') \cap (Q_3\cup Q'_3)
= \{p,p'\}$, because by Lemma 2.1 points in $Q_3 \cup Q'_3$ different from $p$ or $p'$ have orbits which escape to infinity under forward iteration. Then $W^u_+(p, p') \subset Q_3 \cup Q'_3$ implies $W^s(p, p')\cap W^u_+(p, p') = \emptyset $. Because $W^u_-(p, p')$ is in the interior of $R$, Theorem 3.1(v) shows that $W^s(p, p')\cap W^u_-(p, p') = \emptyset $. The second intersection is also the empty set by Theorem 3.1(v), since by 3.1(i) and (ii), $\overline{W^u(0)}$ is in $R$. The first intersection is equal to $\{0\}$, because $W^u(0)\subset S_1 \cup S'_1 \cup \{0\}$ by 3.1(i) and $S_1 \cup S'_1 \subset \Omega_+ \cup \Omega _-$ after 2.4. Finally, the third intersection is equal to $W^s(0) \cap W^u_-(p, p')$, due to $W^s(0) \cap \{p, p'\} = \emptyset$ and $W^s(0) \cap W^u_+(p, p') = \emptyset$, using the fact that $W^u_+(p, p') \subset Q_3 \cup Q'_3$.
BASIN BOUNDARIES
================
$\partial\Omega_+=\partial\Omega_-=W^s(0)\cup W^s(p,p') =J^+_{\mathbb{R}}$
It is clear that $\Omega_+$ and $\Omega_-$ both lie in $K^+_{\mathbb{R}}$. Then $\partial \Omega_+ \cup \partial \Omega_- \subset K^+_{\mathbb{R}}$ follows, since $K^+_{\mathbb{R}}$ is closed, implying by Theorem 4.1(ii) that $\partial \Omega_+ \cup \partial \Omega_- \subset W^s(0)\cup W^s(p,p')$.\
To prove the opposite inclusion, it will first be shown that $W^s(0) \subset \partial \Omega_+$. Due to the invariance of the basin boundary, it suffices to show that the local stable manifold of the origin $W^s_{\epsilon}(0)$ of size $\epsilon$ is contained in $\partial \Omega_+$. Without restriction, $W^s_{\epsilon}(0) \subset S_2 \cup S'_2$. Let $q \in W^s_{\epsilon}(0) \cap S_2$, and let $U$ be a polydisk around $q$. Every point $q'$ in $U \setminus W^s(0) $ is in $ \Omega _+ \cup \Omega_-$ by 2.4. If $q'\in \Omega_+$, then $q \in \partial \Omega_+$ and $\sigma(q) \in \partial\Omega_-$.\
It remains to show $W^s(p, p') \subset \partial \Omega_+ $, since $W^s(p, p') \subset \partial \Omega_-$ follows due to $\sigma(W^s(p, p')) = W^s(p, p'))$ and $\sigma(\partial\Omega_+) = \partial \Omega_-$. By 3.1(iv), there is a point $ q \in W^{u}_{-}(p,p^{'})$ such that $q \notin W^{s}(0)$. Then $q $ is in the interior of $R$ and by 2.5, $q \in \Omega_{+} \cup \Omega_{-}$. Without restriction, let $q \in \Omega_+$. An application of the Lambda Lemma (see\[R\]) will be used. Take a curve $C$ through $q$ transversal to $W^u_-(p, p')$ which is contained in $\Omega_+$. Then parts of the backward iterates $f^{-n}(C)$ converge to the local stable manifold $W^s_{\epsilon}(p, p')$ of $\{p, p'\}$ in the $C^1$ topology proving $W^s(p, p') \subset \partial \Omega_+$.
**REFERENCES**\
\[BS\] E. Bedford and J. Smillie, *Fatou-Bieberbach domains arising from polynomial automorphisms*, Indiana Univ. Math. J. **40** (1991) 789-792.
\[BS2\] E. Bedford and J. Smillie, *Polynomial Automorphisms of ${\mathbb{C}}^2$. II: Stable Manifolds and Recurrence*, J. Amer. Math.Soc. **4** (1991) 657-679.
\[FM\] S. Friedland and J. Milnor, *Dynamical properties of plane polynomial automorphisms*, Ergod. Th. and Dynam. Sys. **9** (1989), no.1, 67-99.
\[FS\] J.E. Fornaess and N. Sibony, *Complex Hénon mappings in ${\mathbb{C}}^2 $ and Fatou-Bieberbach domains*, Duke Math. J. **65** (1992), no.2, 345-380.
\[H\] S. Hayes, Fatou-Bieberbach Gebiete in ${\mathbb{C}}^2$, Deutsche Mathematiker Vereinigung Mitteilungen, (1995), 14-18.
\[K\] T. Kimura, *On Fatou-Bieberbach domains in ${\mathbb{C}}^2$*, J. Fac. Sci. Univ. Tokyo, Sect. IA Math. **35** (1988), no.1, 103-148.
\[R\] C. Robinson, *Dynamical Systems (Stability, Symbolic Dynamics, and Chaos)*, CRC Press, 1999.\
Sandra Hayes\
Department of Mathematics, City College of CUNY, New York, NY 10031: [email protected]
Axel Hundemer\
Department of Mathematics and Statistics, McGill University Montreal, Quebec, CA: [email protected]
Evan Milliken\
Department of Mathematics, University of Florida, Gainesville, FL : [email protected]
Tasos Moulinos\
Department of Mathematics, Statistics and Computer Science The University of Illinois at Chicago: [email protected]
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The 3D block matching (BM3D) filter belongs to the state-of-the-art techniques for eliminating additive white Gaussian noise from single-frame images. There exist four multi-frame extensions of BM3D as of today. In this work, we combine these extensions with a variance stabilising transformation (VST) for eliminating Poisson noise. Our evaluation reveals that the extension which retains the original noise model of the noisy images and additionally has a comprehensive connectivity of 2D and temporal image information at both pixel and patch levels, gives the best results. Additionally, we find a surprising change in performance of one the four extensions due to the specific application of the VST. Finally, we also introduce a simple low-pass filtering as a preprocessing step for the best performing extension. This can give rise to a significant additional improvement of 0.94 dB in the output according to the peak signal to noise ratio.'
bibliography:
- 'myrefs.bib'
title: 'Poisson Noise Removal Using Multi-Frame 3D Block Matching'
---
Multi-frame denoising, non-local patch methods, Poisson noise, 3D block matching
Introduction {#sec:intro}
============
Image acquisition through CCD/CMOS sensors is dominated by Poisson noise distribution [@rooms2003PoissonDenoising]. Astronomical imaging [@borkowski2010Astronomy], medical imaging [@rodrigues2008DenoisingPoisson], and electron microscopy [@jinschek2008EMPoisson] are some of the specific applications where we encounter Poisson noise. There are two categories of imaging techniques that are designed to eliminate Poisson noise: In one of the categories, additive white Gaussian noise (AWGN) elimination algorithms [@DFKE07; @boulanger2008Denoising; @zhang2008WaveletsPoisson; @fryzlewicz2012HaarPoisson; @portilla2003Denoising] are combined with variance stabilising transformations (VST) [@anscombe1948PoissonTrans; @makitalo2011optimal; @makitalo2011ClosedForm]. The other category comprises of algorithms that are designed for eliminating Poisson noise directly [@lefkimmiatis2009BayesianPoisson; @luisier2010WaveletPoisson; @willett2003PlateletsPoisson]. Among all these techniques, 3D block matching (BM3D) [@DFKE07] when combined with a VST, gives the best denoising results in most cases [@makitalo2011optimal].
All the above mentioned techniques concentrate on denoising single-frame images. There has not been much investigation in the field of multi-frame image denoising. In applications such as electron microscopy, CT imaging and multi-spectral imaging, we encounter situations where multiple images of the same scene can be acquired. Further processing is required for fusing the information from all the images to one single image. Several approaches [@blu2007sure; @hasan2018denoising; @boulanger2010patch; @dong2015low; @Buades2009; @tico2008multi; @buades2010MultiFrame; @fang2012sparsity; @luisier2009MultiFrame; @zhang2009MultiFrame] that have been designed to remove both AWGN and Poisson noise from multi-frame image datasets, are inspired from algorithms that are originally designed for single-frame image denoising.
Non-local patch-based methods [@DFKE07; @LBM2013; @Lebrun2012] are among the best performing methods for single-frame image denoising. Efforts in the direction of optimally extending non-local patch based methods to multi-frame elimination of signal dependent noise [@Buades2009; @buades2010MultiFrame], made use a of hybrid filtering scheme. In particular, depending on the temporal standard deviation after registration, they use a combination of values computed through simple averaging of the frames and primitive spatio-temporal non-local patch-based methods. We have revived the research in this direction in [@bodduna2019MultiFrame], by studying all possible spatio-temporal extensions of 3D block matching in the multi-frame scenario. In particular, we have evaluated our proposed extensions and existing ones for eliminating AWGN. We have learnt from this evaluation that using multi-frame extension ideas which retain the noise model and simultaneously enhance the connectivity between frames, give superior results. The optimal performance of such ideas in the AWGN layout, motivates us to also analyse their capability in the multi-frame Poisson noise scenario.
[**Our Contribution.**]{} There exist four extensions of BM3D for multi-frame AWGN removal [@bodduna2019MultiFrame]. In the present work, we perform a comparative evaluation of these extensions for eliminating Poisson noise from multi-frame image datasets. For this purpose, we combine all four extensions with the Anscombe transform and the closed-form approximation of the exact unbiased inverse Anscombe transform. The difference in the qualitative performance according to peak-signal-to-noise (PSNR) ratio, between the closed-form approximation and the exact unbiased inverse is small, but the former is faster [@makitalo2011ClosedForm]. Hence, we use the closed-form approximation for computing the inverse variance stabilising transformation. Additionally, we introduce a preprocessing step for the best performing extension after the evaluation, in the form of simple low-pass filtering. This extra step significantly improves the denoising output.
[**Paper Structure.**]{} In Section \[sec:modelling\], we introduce our novel framework that uses various extensions of BM3D for eliminating Poisson noise from multi-frame image datasets. In Section \[sec:expAndDisc\], we present our experimental results. We also provide explanations behind the respective positions of different extensions in the ranking. Finally, in Section \[sec:concAndOutlook\], we present a summary of our work and give an outlook to future work.
Modelling of denoising algorithm {#sec:modelling}
================================
There are two trivial and two non-trivial extensions among the four existing extensions of BM3D for multi-frame AWGN removal [@bodduna2019MultiFrame]. In Sections \[trivial\] and \[non-trivial\] we review these four extensions. In Sections \[variance\_stab\] and \[low-pass\] we explain in detail the exact procedure that makes optimal use of these four extensions for multi-frame Poisson noise elimination.
Trivial Extensions of BM3D {#trivial}
--------------------------
[**BM3D-1:**]{} In the first trivial extension [@bodduna2019MultiFrame], we first average all the frames after registering them. In the second step, we use the original BM3D method [@DFKE07] to denoise the averaged image.\
\
[**BM3D-2:**]{} In the second trivial extension [@bodduna2019MultiFrame], we first denoise every single frame using the original BM3D method. In the second step, we average all the denoised frames after registering them.
Image BM-1 BM-2 BM-3 BM-M $\sigma$/ BM-M$_\sigma$
------------- ------- ------- ------- ------- ------------------------- --
House (1) 18.29 22.73 22.94 24.28 95/ **24.74**
House (2) 21.19 26.02 25.53 27.11 100/ **27.28**
House (3) 23.15 27.29 26.49 28.13 110/ **28.32**
House (4) 24.69 28.14 27.23 29.06 105/ **29.32**
House (5) 25.97 28.87 27.91 29.74 120/ **29.92**
Lena (1) 19.00 24.54 23.90 25.33 195/ **25.88**
Lena (2) 21.64 26.53 25.82 27.31 200/ **27.50**
Lena (3) 23.43 27.67 26.77 28.44 220/ **28.62**
Lena (4) 24.84 28.33 27.41 29.13 215/ **29.24**
Lena (5) 25.86 28.94 27.95 29.67 210/ **29.82**
Bridge (1) 18.02 20.79 20.40 21.16 130/ **21.85**
Bridge (2) 19.85 21.93 21.50 22.36 145/ **22.88**
Bridge (3) 20.95 22.59 22.08 23.08 140/ **23.55**
Bridge (4) 21.73 23.04 22.48 23.55 145/ **24.00**
Bridge (5) 22.26 23.38 22.80 23.89 145/ **24.33**
Peppers (1) 20.48 24.75 23.97 25.53 170/ **26.10**
Peppers (2) 22.78 26.68 25.83 27.46 195/ **27.63**
Peppers (3) 24.49 27.69 26.77 28.41 205/ **28.54**
Peppers (4) 25.64 28.42 27.48 29.14 205/ **29.24**
Peppers (5) 26.50 28.89 27.94 29.57 205/ **29.67**
BM-1 BM-2 BM-3 BM-M $\sigma$/ BM-M$_\sigma$
------- ------- ------- ------- -------------------------
17.55 22.98 23.34 25.20 80/ **26.11**
20.51 26.41 25.79 27.93 90/ **28.26**
22.67 27.77 26.73 29.26 105/ **29.58**
24.25 28.60 27.55 30.09 100/ **30.48**
25.60 29.28 28.23 30.73 100/ **31.06**
18.16 24.78 24.12 25.95 165/ **26.91**
20.98 26.92 26.07 28.24 180/ **28.71**
22.92 28.05 26.99 29.41 195/ **29.75**
24.42 28.75 27.64 30.13 195/ **30.42**
25.55 29.28 28.08 30.64 195/ **30.94**
17.41 20.90 20.53 21.46 115/ **22.53**
19.49 22.10 21.64 22.86 130/ **23.66**
20.77 22.72 22.18 23.61 135/ **24.32**
21.62 23.20 22.61 24.19 145/ **24.81**
22.23 23.55 22.90 24.57 145/ **25.15**
19.65 24.99 24.20 26.22 160/ **27.07**
22.16 26.99 26.09 28.30 175/ **28.71**
23.99 28.02 26.95 29.31 180/ **29.60**
25.23 28.78 27.65 30.03 175/ **30.25**
26.14 29.24 28.12 30.47 185/ **30.69**
\[table1\]
Non-Trivial Extensions of BM3D {#non-trivial}
------------------------------
[**BM3D-3:**]{} The original BM3D method is a two-step process. Initially, for every reference patch considered, a corresponding 3D group of patches is formed by searching for the most similar patches in the image with respect to the $L_2$ distance. In the first non-trivial extension BM3D-3 [@buades2005denoising; @tico2008multi; @Buades2009; @buades2010MultiFrame], we consider a reference frame initially. For every reference patch in this frame we increase the search area for finding the most similar patches from this particular frame to all the frames. The further procedure is similar to the original BM3D method.\
\
[**BM3D-4:**]{} The second non-trivial extension [@bodduna2019MultiFrame] is also a two-step method with three sub-steps in each step like the original BM3D method:\
[*Step 1.1 - Grouping.*]{} Unlike BM3D-3, here we consider reference patches from all the frames but not just one frame. A corresponding 3D group for every reference patch is then created by finding the most similar patches in all the frames using $L_2$ distance. We remove the threshold parameters for $L_2$ distance. This is done because there is a risk of losing similar patches for high amplitudes of noise. However, we retain the parameters that specify the maximum number of patches in a 3D group.\
[*Step 1.2 - Filtering.*]{} We apply the following techniques on every obtained 3D group, in the same order: 2D bi-orthogonal spline wavelet transform, 1D Walsh-Hadamard transform, hard thresholding, 1D Walsh-Hardamard back transform, and a 2D bi-orthogonal spline wavelet back transform.\
[*Step 1.3 - Aggregation.*]{} Now, every pixel in every frame is denoised at least once after the second sub-step. The denoised versions of every pixel is present in 3D groups belonging to the frame the pixel belongs to, as well as in 3D groups from other frames. Thus, we first compute a weighted aggregation of all the denoised versions of every pixel within the frame in which it is present. This gives us as many initial denoised frames as there are input frames. To compute a final denoised frame after the first main step, we use the following equation for weighted aggregation across 3D groups in all frames: $$\label{equation}
\bm{u}^{\textrm{\textbf{basic}}}({\bm{x}}) =
\frac{\sum\limits_\ell \sum\limits_{P_\ell}
w_{P_\ell}^{\textrm{\textbf{hard}}}
\sum\limits_{Q \in P(\textrm{P}_\ell)} \chi_Q(\bm{x})
\bm{u}^{\textrm{\textbf{hard}}}_{Q,P_\ell}(\bm{x})}
{\sum\limits_l \sum\limits_{P_\ell}
w_{P_\ell}^{\textrm{\textbf{hard}}}
\sum\limits_{Q \in P(\textrm{P}_\ell)} \chi_Q(\bm{x}) }.$$ Here, $\bm{x}$ is a position in the 2D image domain $Q$ and $\bm{u}^{\textrm{\textbf{basic}}}$ is the initial denoised image. The set of most similar patches to the reference patch $P_\ell$ belonging to frame $\ell$, are denoted using $\mathcal P (P_\ell)$. We have $\chi_Q(\bm{x}) = 1 $ if $\bm{x} \in Q$ and 0 otherwise, for every patch $Q$ in the set $\mathcal P (P_\ell)$. The symbol $\bm{u}^{\textrm{\textbf{hard}}}_{Q,P_{\ell}}(\bm{x})$ denotes the estimation of the value at pixel position $\bm{x}$, belonging to the patch $Q$, derived after the hard thresholding (with coefficients $w_{P_\ell}^{\textrm{\textbf{hard}}}$) of the reference patch $P_\ell$.\
[*Step 2.1 - Grouping.*]{} The same grouping strategy as in *1.1* is employed by us, using the reference patches from the initial denoised frames computed after the first main step.\
[*Step 2.2 - Filtering.*]{} We execute the following techniques on every obtained 3D group and the corresponding initial noisy 3D group, in the same order: 2D discrete cosine transform, 1D Walsh-Hadamard transform, Wiener filtering on a combination of both the corresponding 3D groups, 1D Walsh-Hadamard back transform, and a 2D discrete cosine back transform.\
[*Step 2.3 - Aggregation.*]{} The same corresponding strategy as in *1.3* is exploited by us to estimate the final denoised image. This aggregation strategy can be represented as: $$\label{equation_2}
\bm{u}^{\textrm{\textbf{final}}}({\bm{x}}) =
\frac{\sum\limits_\ell \sum\limits_{P_\ell}
w_{P_\ell}^{\textrm{\textbf{wien}}}
\sum\limits_{Q \in P(\textrm{P}_\ell)} \chi_Q(\bm{x})
\bm{u}^{\textrm{\textbf{wien}}}_{Q,P_\ell}(\bm{x})}
{\sum\limits_l \sum\limits_{P_\ell}
w_{P_\ell}^{\textrm{\textbf{wien}}}
\sum\limits_{Q \in P(\textrm{P}_\ell)} \chi_Q(\bm{x}) }.$$ Here $\textrm{\textbf{wien}}$ represents Wiener filtering and $\bm{u}^{\textrm{\textbf{final}}}$ is the final denoised image. The rest of the symbols have the same meaning as in (\[equation\]). Moreover, we can also represent the initial denoised image aggregated using BM3D-3 by . Choosing $\ell = 1$ in gives us the initial estimate of the single-frame BM3D algorithm. The final denoised image using BM3D-3 and single-frame BM3D can both be computed by selecting $\ell = 1$ in .
We apply the particular filtering steps 1.2 and 2.2 in the Fourier domain because of better differentiation between signal and noise when compared to the Cartesian domain. The original work [@DFKE07] provides more specific details regarding the single-frame BM3D algorithm.
In the first extension, we have a change in the noise distribution because of averaging the noisy frames. In BM3D-2, we denoise every frame first, which is a sub-optimal solution since we have limited amount of signal in each of the frames. BM3D-3 avoids the above two modelling disadvantages by searching for similar patches in all the frames. However, we consider just one reference frame and this does not allow us to make use of the complete information. In the final extension BM3D-4, we consider every frame as the reference frame. This allows us to compute similar patches from all the frames in both the main steps, unlike BM3D-3. Such a modelling also allows us to perform a better aggregation for obtaining the final denoised image. This is due to the presence of more denoised versions of every pixel in the 2D image domain, when compared to the third extension.
{width="0.2\linewidth"} {width="0.2\linewidth"} {width="0.2\linewidth"} {width="0.2\linewidth"} {width="0.2\linewidth"} {width="0.2\linewidth"} {width="0.2\linewidth"} {width="0.2\linewidth"} {width="0.2\linewidth"} {width="0.2\linewidth"} {width="0.2\linewidth"} {width="0.2\linewidth"} {width="0.2\linewidth"} {width="0.2\linewidth"} {width="0.2\linewidth"} {width="0.2\linewidth"}
Denoising with Variance Stabilisation {#variance_stab}
-------------------------------------
We carry out the following procedure for eliminating Poisson noise by using the various BM3D extensions mentioned above: All the noisy datasets undergo the Anscombe transformation for variance stabilisation, affine rescaling to \[0,1\] greyscale range, filtering using one of the four extensions, affine rescaling back to the original range, and finally the closed-form approximation of the exact unbiased inverse Anscombe transformation.
In BM3D-1, we have applied the transformation on the averaged image while in all other extensions we have applied the VST directly on the initial noisy images. One can find more details regarding the above general procedure and parameter selection in [@makitalo2011optimal; @makitalo2011ClosedForm; @hou2010comments].
Low-Pass Filtering {#low-pass}
------------------
As already mentioned in Section \[sec:intro\], we use a simple low pass filter [@Frangakis2001] as a preprocessing step for the best performing extension. The following low-pass filter is applied on the noisy images directly (without variance stabilisation or affine rescaling), in order to make the process of finding similar patches that form a 3D group more robust: $$H(\bm{\hat{x}}) = \begin{cases}
1 &\text{$ \textrm{for} \ \ |\bm{\hat{x}}| < \frac{\sigma}{2}$},\\
\textrm{exp} \left(\frac{-(\hat{x} - (\frac{\sigma}{2}))^2
-(\hat{y} - (\frac{\sigma}{2}))^2}
{2(\frac{\sigma}{2})^2} \right)
&\textrm{for} \ \ |\bm{\hat{x}}| \ge \frac{\sigma}{2}.
\end{cases}$$ Here, $\bm{\hat{x}} := (\hat{x} , \hat{y})^\top$ represents the 2D frequency vector in the Fourier domain and $\sigma$ specifies the shape of the filter. The Gaussian-type decay of the filter coefficients is intended to reduce the transform-domain filtering artifacts. We use the low-pass filtered image only for forming the 3D groups in the first main step and nowhere else in the entire algorithm.
Experiments and Discussion {#sec:expAndDisc}
==========================
In multi-frame denoising applications, the frames are first registered using motion registration algorithms before denoising them. In this work we want to specifically test the denoising capability of different extensions of BM3D. To this end, we assume that the frames are perfectly registered similar to [@bodduna2019MultiFrame]. This assumption makes even more sense because it has already been verified [@Buades2009; @buades2010MultiFrame] that denoising methods (even without temporal support) are superior to averaging at regions where there is high temporal standard deviation after registration. Thus, verifying the performance of different denoising methods assuming perfect registration also covers the remaining case of less temporal standard deviation. Thereby, to create this experimental setting, we have added Poisson noise to the Lena, House, Peppers and Bridge[^1] images with noise peaks varying from 1 to 5. Due to the signal dependent nature of Poisson noise, we have high noise amplitude when the intensity of the signal decreases. The noise peak value controls this trade-off. Lesser noise peak value indicates higher amount of noise. For every image and every noise peak, we have created two datasets each with 5 and 10 realisations of noise.
Dissolving the threshold parameters for $L_2$ distance while forming the 3D groups is advantageous to all the extensions. Hence, we have excluded this parameter in both the main steps of BM3D for all the four extensions.
For BM3D-3, in the results which we showcase shortly, we present the best PSNR after every frame is considered as the reference frame.
From Table \[table1\] and Figure \[fig:res\], we can conclude that both visually and in terms of PSNR, the non-trivial extension BM3D-4 outperforms all other extensions significantly. The comprehensive inter-frame connectivity strategy used in BM3D-4 is the reason behind its significantly better performance. Also, using a low-pass filter for preprocessing (BM3D-4$_\sigma$) gives BM3D-4 a noteworthy improvement (0.94 dB for the Lena 10-image dataset).
There can be imprecision in variance stabilisation when the VST is applied on a single noisy image [@azzari2016IterativePoisson]. The imprecision is even higher because of the initial averaging step in BM3D-1, which modifies the noise distribution. This is clearly evident from Table \[table1\] because denoising 5 images is surprisingly better than denoising 10 images using BM3D-1. We can avoid such a problem using BM3D-4 as it does not average the initial noisy images but retains the original noise model.
BM3D-4 is slightly more than $2L$ times (two main steps and the extensive patch-matching) slower than single-frame BM3D, for a dataset with $L$ frames. Our CUDA implementation of BM3D-4 on an NVIDIA GeForce GTX 1070 device takes 1.76 seconds for a dataset sized $256 \times
256 \times 5$.
Conclusions and Outlook {#sec:concAndOutlook}
=======================
In this work we have evaluated the capability of different existing extensions of BM3D to remove multi-frame Poisson noise, by combining them with a variance stabilising tranformation. Our evaluation revealed that the extension (BM3D-4) which retains the original noise model and additionally makes use of the complete available information through a comprehensive connectivity of 2D and temporal information, is the best method for multi-frame Poisson noise elimination. Our new idea to preprocess the data using low-pass filtering gives significant additional quality improvement to the best performing extension BM3D-4.
In the future, we will explore the application of BM3D-4 for eliminating Poisson and Poisson-Gaussian mixture noise of varying amplitudes in the same dataset. We also plan to use BM3D-4 for denoising datasets acquired with electron microscopes, by combining it with state-of-the-art motion registration algorithms.
J.W. has received funding from the European Research Council (ERC) under the European Union’s Horizon 2020 research and innovation programme (grant agreement no. 741215, ERC Advanced Grant INCOVID).
[^1]: http://sipi.usc.edu/database/
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In this note we study the partition of $\mathbb{R}_{\geq0}^{n}$ given by the regions where the mixed test ideals $\tau(\mathfrak{a}_{1}^{t_{1}}...\mathfrak{a}_{n}^{t_{n}})$ are constant. We show that each region can be described as the preimage of a natural number under a $p$-fractal function $\varphi:\mathbb{R}_{\geq0}^{n}\rightarrow\mathbb{N}$. In addition, we give some examples illustrating that these regions do not need to be composed of finitely many rational polytopes.'
address: 'Department of Mathematics, University of Michigan, 530 Church Street, Ann Arbor, MI 48109, USA'
author:
- Felipe Pérez
title: On the constancy regions for mixed test ideals
---
Introduction
============
In this note, we study the dependence of mixed test ideals on parameters, and show that the emerging picture is quite different from that in the case of mixed multiplier ideals in characteristic zero.
Multiplier ideals have been intensively studied over the last two decades, as they play an important role in birational geometry, see for example [@Laz]. Given a smooth complex variety $X$ and a nonzero ideal sheaf $\mathfrak{a}$, one can define for any parameter $c>0$ an ideal $\mathcal{J}(\mathfrak{a}^{c})$, called multiplier ideal. This ideal is described via a log resolution $\pi:X'\rightarrow X$ of the pair $(X,\mathfrak{a})$, i.e. a proper birational map, with $X'$ smooth, and such that $\mathfrak{a}\mathcal{O}_{X'}=\mathcal{O}_{X'}(-E)$, where $E$ is a simple normal crossing divisor. Then, $$\mathcal{J}(\mathfrak{a}^{c}):=\pi_{*}\mathcal{O}(K_{X'/X}-\lfloor cE\rfloor),$$ where $K_{X'/X}$ is the relative canonical divisor.
$\textit{Mixed multiplier ideals}$ extend the previous definition to the case of several ideals: for nonzero ideals $\mathfrak{a}_{1},\ldots,\mathfrak{a_{n}}$ and positive numbers $c_{1},\ldots,c_{n}$ we take a log resolution for the pair $(X,\mathfrak{a}_{1}\cdots\mathfrak{a}_{n})$ and set the mixed multiplier ideal to be
$$\mathcal{J}(\mathfrak{a}_{1}^{c_{1}}\cdots\mathfrak{a}_{n}^{c_{n}}):=\pi_{*}\mathcal{O}(K_{X'/X}-\lfloor c_{1}E_{1}+\ldots+c_{n}E_{n}\rfloor),$$ where $\mathcal{O}_{X'}(-E_{i})=\mathfrak{a}_{i}\mathcal{O}_{X'}$.
Test ideals were introduced by Hara and Yoshida in [@HY] as an analogue of multiplier ideals in positive characteristic. One question that was studied since [@HY] is which properties of multiplier ideals have analogues for test ideals. For example, for multiplier ideals the $\textit{jumping}$ $\textit{numbers}$ of $\mf[a]$ are defined as the positive real numbers $c$ such that $\mathcal{J}(\mf[a]^{c})\not=\mathcal{J}(\mf[a]^{c-\epsilon})$ for every $\epsilon>0$ (cf. [@ELSV]). It is easy to see from the definition (1.1) that for each $\mf[a]$ these numbers are discrete and rational. Thus it was expected that this was the case also in positive characteristic. Blickle, Mustaţă and Smith proved discreteness and rationality of the analogous positive characteristic invariants in [@BMS], but the proof was more involved.
In the mixed multiplier ideal setting, it follows from the above description in terms of a log resolution that for every $b_{1},\ldots,b_{n}$ the region $$\{(c_{1},\ldots,c_{n})\in\mathbb{R}_{\geq0}^{n}|\, c_{i}\leq b_{i}\mbox{ for all }i\}$$ can be decomposed in a finite set of rational polytopes with nonoverlapping interiors, such that on the interior of each face of each polytope the mixed multiplier ideal $\mathcal{J}(\mathfrak{a}_{1}^{c_{1}}\cdots\mathfrak{a}_{n}^{c_{n}})$ is constant. It was expected that in the positive characteristic setting we would have a similar picture.
In the present note we prove that this is not the case, but we can still get a nice decomposition. This decomposition depends on a $p$-fractal function, that is, a function $\varphi:\mathbb{R}_{\geq0}^{n}\rightarrow\mathbb{N}$ satisfying the following property. If we restrict $\varphi$ to a bounded domain $D$, then the vector space generated by the functions $\phi(t_{1},\ldots,t_{n})=\varphi((t_{1}+b_{1})/p^{e},\ldots,(t_{n}+b_{n})/p^{e})$ with $b_{i}$ integers and $((t_{1}+b_{1})/p^{e},\ldots,(t_{n}+b_{n})/p^{e})\in D$ , is finite dimensional (Definition \[Def Fractal function\]). Explicitly, we show:
$[$Theorem \[thm:Main Theorem\]$]$ For an $F$-finite, regular ring $R$ essentially of finite type over a finite field of positive characteristic and non zero ideals $\mathfrak{a}_{1},\ldots,\mathfrak{a}_{n}$ of $R$, there is a $p$-fractal function $\varphi:\mathbb{R}_{\geq0}^{n}\rightarrow\mathbb{N}$ such that $$\tau(\mathfrak{a}_{1}^{c_{1}}...\mathfrak{a}_{n}^{c_{n}})=\tau(\mathfrak{a}_{1}^{d_{1}}...\mathfrak{a}_{n}^{d_{n}})\Longleftrightarrow\varphi(c_{1},\ldots,c_{n})=\varphi(d_{1},\ldots,d_{n}),$$ and therefore the constancy regions are of the form $\varphi^{-1}(i)$ for $i\in\mathbb{N}$.
Roughly speaking, this shows that each constancy region has a $p$-fractal structure that, as we see in the examples in Section 5, can be intricate.
This note is structured as follows. In section 2 we recall the definition of test ideals and mixed test ideals following [@BMS] and state some of the theorems that were proved there. In section 3 we give our main definitions and deduce some basic consequences of these definitions. We prove our main theorem in section 4. In the last section, we give an example of a constancy region that is not a finite union of polyhedral regions.
Acknowledgments {#acknowledgments .unnumbered}
===============
I would like to thank Mircea Mustaţă for introducing me to this problem and for guiding my work. I would also like to thank Angélica Benito, Luis Nuñez and Axel Stäbler for comments and suggestions on earlier drafts. Finally, I would like to thank Rui Huang for doing the graphs in section 5.
Preliminaries
=============
Recall that a ring $R$ of positive characteristic is $\emph{F-finite}$ if the Frobenius morphism $F:R\rightarrow R$ is finite. Throughout this note we let $R$ be a regular ring essentially of finite type over an $F$-finite field $k$ of positive characteristic $p$. In particular, $R$ is $F$-finite as well.We now recall the basic definitions and properties related to test ideals and refer to [@BMS] for proofs and details.
Given an ideal $\mf[b]$ in $R$, we denote by $\mathfrak{b}^{[1/p^{e}]}$ the smallest ideal $\mathfrak{J}$ such that $\mathfrak{b}\subseteq\mathfrak{J}^{[p^{e}]}:=(f^{p^{e}}|f\in\mathfrak{J})$. The existence of a smallest such ideal is a consequence of the flatness of the Frobenius map in the regular case. The following proposition gives an explicit description of $\mf[b]^{[1/p^{e}]}$ when $R$ is free over $R^{p^{e}}$.
\[thm: root ideals\][@BMS Proposition 2.5] Suppose that $R$ is free over $R^{q}$, for $q=p^{e}$, and let $e_{1},\ldots,e_{N}$ be a basis of $R$ over $R^{q}$. If $h_{1},\ldots,h_{n}$ are generators of an ideal $\mathfrak{b}$ of $R$, and if for every $i=1,\ldots,n$ we write $$h_{i}=\sum_{j=1}^{N}a_{i,j}^{q}e_{j}$$ with $a_{i,j}\in R$, then $$\mathfrak{b}^{[1/p^{e}]}=(a_{i,j}|i\leq n\mbox{ and }j\leq N).$$
Test ideals were introduced by Hochster and Huneke [@HH] as a tool in their tight closure theory , and were later generalized by Hara and Yoshida [@HY] in the context of pairs $(R,\mathfrak{a}^{c})$, where $\mathfrak{a}$ is an ideal in $R$ and $c$ is a real parameter. Blickle, Mustaţă, and Smith [@BMS] gave an elementary description of these ideals in the case of a regular $F$-finite ring $R$ . It is this description which we take as our definition.
Given a non-negative number $c$ and a nonzero ideal $\mathfrak{a}$, we define the $\textit{generalized}$ $\textit{test ideal}$ $\textit{of \ensuremath{\mf[a]\:}with exponent}$ $c$ to be $$\tau(\mathfrak{a}^{c})=\bigcup_{e>0}(\mathfrak{a}^{\lceil cp^{e}\rceil})^{[1/p^{e}]},$$ where $\lceil c\rceil$ stands for the smallest integer $\geq c.$
The ideals in the above union form an increasing chain of ideals; therefore as $R$ is Noetherian, they stabilize. Hence for $e$ large enough $\tau(\mathfrak{a}^{c})=(\mathfrak{a}^{\lceil cp^{e}\rceil})^{[1/p^{e}]}.$ In the principal ideal case we can say more.
\[Computing test ideals for principal ideals\][@BMS2 Lemma 2.1] If $\lambda=\frac{m}{p^{e}}$ for some positive integer $m$, then $\tau(f^{\lambda})=(f^{m})^{[1/p^{e}]}.$
It can be shown that as the parameter $c$ varies over the reals, only countably many different test ideals appear; moreover, we have:
[@BMS Proposition 2.14] For every nonzero ideal $\mf[a]$ and every non-negative number $c$, there exists $\epsilon>0$ such that $\tau(\mf[a]^{c})=\tau(\mf[a]^{c'})$ for $c<c'<c+\epsilon$.
A positive real number $c$ is an $\textit{F-jumping exponent}$ of $\mf[a]$ if $\tau(\mf[a]^{c})\neq\tau(\mf[a]^{c-\epsilon})$ for all $\epsilon>0$
The $F$-jumping exponents of an ideal $\mf[a]$ form a discrete set of rational numbers, that is, there are no accumulation points of this set. In fact, they form a sequence with limit infinity (see [@BMS Theorem 3.1]).
As in the case of one ideal, one can define the mixed test ideal of several ideals as follows.
\[Def mixed test ideals\]Given nonzero ideals $\mf[a]_{1},...,\mf[a]_{n}$ of $R$ and non-negative real numbers $c_{1},...,c_{n}$, we define the $\textit{mixed generalized test ideal with exponents }c_{1},\ldots,c_{n}$ as: $$\tau(\mathfrak{a}_{1}^{c_{1}}\cdots a_{n}^{c_{n}})=\bigcup_{e>0}(\mathfrak{a}_{1}^{\lceil c_{1}p^{e}\rceil}\cdots\mathfrak{a}_{n}^{\lceil c_{n}p^{e}\rceil})^{[1/p^{e}]}.$$
As in the case of $\tau(\mathfrak{a}^{c})$, we have $\tau(\mathfrak{a}_{1}^{c_{1}}\cdots\mf[a]_{n}^{c_{n}})=(\mathfrak{a}_{1}^{\lceil c_{1}p^{e}\rceil}\cdots\mathfrak{a}_{n}^{\lceil c_{n}p^{e}\rceil})^{[1/p^{e}]}$ for all $e$ large enough.
\[thm:bound on the degree generators mixed case\] Let $\mf[a]_{1},\ldots,\mf[a]_{n}$ be nonzero ideals in the polynomial ring $R=k[x_{1},\ldots,x_{r}]$, and let $c_{1}=r_{1}/p^{s},\ldots,c_{n}=r_{n}/p^{s}$ be such that $r_{1},\ldots,r_{n}$ are natural numbers. If each $\mf[a]_{i}$ can be generated by polynomials of degree at most $d,$ then the ideal
We argue as in [@BMS Proposition 3.2], where the result was proven for the case of one ideal. We know that $R$ is free over $R^{p^{e}}$ with basis $$\{\beta_{j}x_{1}^{\alpha_{1}}\cdots x_{r}^{\alpha_{r}}|0\leq\alpha_{i}<p^{e}\mbox{ and \ensuremath{\beta_{j}}part of a basis for \ensuremath{k}over \ensuremath{k^{p^{e}}}}\}.$$ The ideal $\mf[a]_{1}^{\lceil p^{e}c_{1}\rceil}\cdots\mf[a]_{n}^{\lceil p^{e}c_{n}\rceil}$ can be generated by polynomials of degree at most $d\lceil p^{e}c_{1}\rceil+\ldots+d\lceil p^{e}c_{n}\rceil$. Hence taking $e>s$ large enough by Proposition \[thm: root ideals\] the ideal $$\tau(\mathfrak{a}_{1}^{c_{1}}\cdots\mf[a]_{n}^{c_{n}})=(\mathfrak{a}_{1}^{\lceil p^{e}c_{1}\rceil}\cdots\mathfrak{a}_{n}^{\lceil p^{e}c_{n}\rceil})^{[1/p^{e}]}$$ is generated by polynomials of degree at most $(d\lceil p^{e}c_{1}\rceil+\ldots+d\lceil p^{e}c_{n}\rceil)/p^{e}=(dp^{e-s}r_{1}+\ldots+dp^{e-s}r_{n})/p^{e}=d(r_{1}+...+r_{n})$.
Some Sets Associated to mixed test ideals
=========================================
In this section we introduce the definitions needed for our study of mixed test ideals, and derive some basic properties. Recall that $R$ denotes a regular ring essentially of finite type over an $F$-finite field $k$ of positive characteristic.
In order to simplify notation we denote $\mf[a]_{1}^{c_{1}}...\mf[a]_{n}^{c_{n}}$ by $\mf[a]^{\bs[c]}$, where $\mf[a]=(\mf[a]_{1},...,\mf[a]_{n})$, $\bs[c]=(c_{1},...,c_{n})\in\mathbb{R}_{\geq0}^{n}$. We similarly denote the vector $(\lceil r_{1}\rceil,...,\lceil r_{n}\rceil)$ by $\lceil\bs[r]\rceil$, where $\bs[r]=(r_{1},\ldots,r_{n})\in\mathbb{R}_{\geq0}^{n}$.
\[Def: Approximation sets\]Given nonzero ideals $\mf[a]_{1},\ldots,\mf[a]_{n},$ and $I$ in $R$, we define $$V^{I}(\mf[a],p^{e})=\left\{ \frac{1}{p^{e}}\bs[c]=\left(\frac{c_{1}}{p^{e}},\ldots,\frac{c_{n}}{p^{e}}\right)\in\frac{1}{p^{e}}\mathbb{Z}_{\geq0}^{n}|\;\mf[a]^{\bs[c]}\not\subseteq I^{[p^{e}]}\right\}$$
and $$B^{I}(\mf[a],p^{e})=\bigcup[0,l_{1}]\times\ldots.\times[0,l_{n}]\subset\mathbb{R}^{n},$$ where the union runs over all $(l_{1},\ldots,l_{n})\in V^{I}(\mf[a],p^{e})$.
From this definition it follows that if $e'\geq e$ then $V^{I}(\mf[a],p^{e})\subseteq V^{I}(\mf[a],p^{e'})$ and $B^{I}(\mf[a],p^{e})\subseteq B^{I}(\mf[a],p^{e'}).$ Indeed, if $\mf[a]^{\bs[c]}\not\subseteq I^{[p^{e}]}$, then there is an element $f\in\mf[a]^{\bs[c]}$ with $f\notin I^{[p^{e}]}$, and by the flatness of the Frobenius morphism we get $f^{p^{e'-e}}\in\mathfrak{a}^{p^{e'-e}\bs[c]}$ but $f^{p^{e'-e}}\notin I^{[p^{e'}]}$. Therefore $\mf[a]^{p^{e'-e}\bs[c]}\not\subseteq I^{[p^{e'}]}$, hence we get the first inclusion. The second one is then straightforward.
\[Def characteristic function\]Let $B^{I}(\mf[a])=\bigcup_{e>0}B^{I}(\mf[a],p^{e})$ and define $\chi_{\mf[a]}^{I}:\mathbb{R}^{n}\rightarrow\mathbb{N}$ to be the characteristic function of the set $B^{I}(\mf[a])$. That is, $\chi_{\mf[a]}^{I}(\boldsymbol{c})$ is $1$ if $\boldsymbol{c}$ is in $B^{I}(\mf[a])$ and it is $0$ otherwise.
In order to study the sets $B^{I}(\mf[a])$ it is crucial to understand how they intersect any increasing path. This motivates the following definition.
Let $\mf[a]_{1},\ldots,\mf[a]_{n},$ and $I\not=R$ be nonzero ideals as before and let $\bs[r]=(r_{1},\ldots,r_{n})\in\mathbb{Z}_{\geq0}^{n}$ be such that that $\mf[a]^{\bs[r]}\subseteq\mbox{rad}(I)$. We denote $$V_{\bs[r]}^{I}(\mf[a],p^{e})=\mbox{max}\{m\in\mathbb{Z}_{\geq0}|\mbox{ }\mf[a]^{m\bs[r]}\not\subseteq I^{[p^{e}]}\}.$$
While in the definition of $V^{I}(\mf[a],p^{e})$ one does not require any relation between $\mf[a]$ and $I$, observe that we require that $\mf[a]^{\bs[r]}\subseteq\mbox{rad}(I)$ when we consider $V_{\bs[r]}^{I}(\mf[a],p^{e})$.
Note that if $\mf[a]^{m\bs[r]}\not\subseteq I^{[p^{e}]}$ then $a^{pm\bs[r]}\not\subseteq I^{[p^{e+1}]}$. Therefore $pV_{\bs[r]}^{I}(\mf[a],p^{e})\le V_{\bs[r]}^{I}(\mf[a],p^{e+1})$, hence $$\left(\frac{V_{\bs[r]}^{I}(\mf[a],p^{e})}{p^{e}}\right)_{e\geq1}\label{Decreasing sequence}$$ is a non-decreasing sequence.
The sequence \[Decreasing sequence\] is bounded, hence it has a limit.
If $\mf[a]^{\bs[r]}$ is generated by $s$ elements, then $\mf[a]^{(s(p^{e}-1)+1)\bs[r]}\subseteq(\mf[a]^{\bs[r]})^{[p^{e}]}.$ For $l$ large enough such that $\mf[a]^{l\bs[r]}\subseteq I$, we have $V_{\bs[r]}^{I}(\mf[a],p^{e})\leq l(s(p^{e}-1)+1)-1$ for all $e$. Therefore $V_{\bs[r]}^{I}(\mf[a],p^{e})/p^{e}\leq ls$, thus the sequence is bounded.
We call this limit the *$F$-threshold of $\mathfrak{a}$ associated to $I$ in direction $\boldsymbol{r}=(r_{1},\ldots,r_{n})$*, and we denote it by $C_{\bs[r]}^{I}(\mf[a])$.
In the case $n=1$ we recover the usual definition of $F$-threshold [@MTW], [@BMS Section 2.5].
\[Contaiment in the positive direction\]Let $\frac{1}{p^{e}}\bs[b]=(\frac{b_{1}}{p^{e}},\ldots,\frac{b_{1}}{p^{e}})$ and $\frac{1}{p^{e'}}\bs[c]=(\frac{c_{1}}{p^{e'}},\ldots,\frac{c_{n}}{p^{e'}})$ be two elements in
It follows as in [@BMS Lemma 2.8]. The condition $b_{i}\leq c_{i}p^{e-e'}$ implies that $\mf[a]_{i}^{b_{i}}\supseteq\mf[a]_{i}^{c_{i}p^{e-e'}}$for every $i$. Therefore $$(\mf[a]^{\bs[b]})^{[1/p^{e}]}\supseteq(\mf[a]^{p^{e-e'}\bs[c]})^{[1/p^{e}]}\supseteq(\mf[a]^{\bs[c]})^{[1/p^{e'}]}.$$
\[Constancy in positive directions\]Given any $\bs[c]=(c_{1},\ldots,c_{n})\in\mathbb{R}_{\geq0}^{n}$, there is $\bs[\epsilon]=(\epsilon_{1},\ldots,\epsilon_{n})\in\mathbb{R}_{>0}^{n}$ such that for every $\bs[r]=(r_{1},\ldots,r_{n})$ with $0<r_{i}<\epsilon_{i}$, we have $\tau(\mf[a]^{\bs[c]})=\tau(\mf[a]^{\bs[c]+\bs[r]})$.
We argue as in the proof of [@BMS Proposition 2.14]. We first show that there is a vector $\bs[\epsilon]=(\epsilon_{1},\ldots,\epsilon_{n})$, with $\epsilon_{i}>0$ for all $i$, such that for all vectors $\bs[r]=(r_{1},\ldots,r_{n})\in\mathbb{Z}^{n}$ with $c_{i}<\frac{1}{p^{e}}r_{i}<c_{i}+\epsilon_{i}$ we have that $(\mf[a]^{\bs[r]})^{[1/p^{e}]}$ is constant. Indeed, otherwise there are sequences $\bs[r]_{m}=(r_{m,1},\ldots,r_{m,n})\in\mathbb{Z}_{\geq0}^{n}$ and $e_{m}\in\mathbb{Z}_{\geq0}$ such that $\frac{1}{p^{e_{m}}}\bs[r]_{m}$ converges to $\bs[c]$, $\left(\frac{1}{p^{e_{m}}}r_{m,i}\right)_{m}$ is a decreasing sequence for every $i$, $e_{m}\leq e_{m+1}$, and $(\mf[a]^{\bs[r]_{m}})^{[1/p^{e_{m}}]}\not=(\mf[a]^{\bs[r]_{m+1}})^{[1/p^{e_{m+1}}]}$. It follows from Lemma \[Contaiment in the positive direction\] that $(\mf[a]^{\bs[r]_{m}})^{[1/p^{e_{m}}]}\subsetneq(\mf[a]^{\bs[r]_{m+1}})^{[1/p^{e_{m+1}}]}$ for all $m$, but this contradicts the fact that $R$ is Noetherian.
Assume now that $\bs[\epsilon]=(\epsilon_{1},\ldots,\epsilon_{n})$ is as above and let $I=(\mf[a]^{\bs[r]})^{[1/p^{e}]}$ for all $\bs[r]=(r_{1},...,r_{n})\in\mathbb{Z}^{n}$ with $c_{i}<\frac{1}{p^{e}}r_{i}<c_{i}+\epsilon_{i}$. We show that $I=\tau(\mf[a]^{\bs[c]})$. Take $e$ large enough such that $\tau(\mf[a]^{\bs[c]})=(\mf[a]^{\lceil p^{e}\bs[c]\rceil})^{[1/p^{e}]}$ and $\frac{\lceil p^{e}c_{i}\rceil}{p^{e}}<c_{i}+\epsilon_{i}$ for every $i$. If all $p^{e}c_{i}$ are non-integers then $\frac{\lceil p^{e}c_{i}\rceil}{p^{e}}>c_{i}$ and $\tau(\mf[a]^{\bs[c]})=I$. Let us suppose that $p^{e}c_{i}$ is an integer precisely when $i=i_{1},...,i_{l}$. Let $\bs[d]=(d_{1},\ldots,d_{n})$ be the vector whose $i_{j}$ coordinates are $1$ and all the other are $0$. As $e$ is arbitrarily large we may also assume that $c_{i}<c_{i}+\frac{1}{p^{e}}d_{i}<c_{i}+\epsilon_{i}$ for all $i\in\{i_{1},\ldots,i_{l}\}$, hence $I=(\mf[a]^{\lceil p^{e}\bs[c]\rceil+\boldsymbol{d}})^{[1/p^{e}]}\subseteq(\mf[a]^{\lceil p^{e}\bs[c]\rceil})^{[1/p^{e}]}=\tau(\mf[a]^{\bs[c]})$.
The reverse inclusion follows by showing $\mf[a]^{\lceil p^{e}\bs[c]\rceil}\subseteq I^{[p^{e}]}$. Let $u\in\mf[a]^{\lceil p^{e}\bs[c]\rceil}$. If $e'>e$ and $e'$ is large enough, then $c_{i}<c_{i}+\frac{1}{p^{e'}}<c_{i}+\epsilon_{i}$, hence $\mf[a]^{\lceil p^{e'}\bs[c]\rceil+\boldsymbol{1}}\subseteq I^{[p^{e'}]}$. Here $\boldsymbol{1}$ denotes the vector whose coordinates are all $1$. Thus, for $v$ a nonzero element in $\mf[a]_{1}\cdots\mf[a]_{n}$ we have $$vu^{p^{e'-e}}\in\mbox{\ensuremath{\mf[a]}}^{p^{e'-e}\lceil p^{e}\bs[c]\rceil+\boldsymbol{1}}\subseteq\mf[a]^{\lceil p^{e'}\bs[c]\rceil+\boldsymbol{1}}\subseteq(I^{[p^{e}]})^{[p^{e'-e}]}.$$ This implies that $u$ is in the tight closure of $I^{[p^{e}]}$, but as $R$ is a regular ring, the tight closure of $I^{[p^{e}]}$ is equal to $I^{[p^{e}]}$(see [@HH]). This gives $\mathfrak{a}^{\lceil p^{e}\boldsymbol{c}\rceil}\subseteq I^{[p^{e}]}$ hence, by definition, $\tau(\mathfrak{a}^{\boldsymbol{c}})=(\mathfrak{a}^{\lceil p^{e}\boldsymbol{c}\rceil})^{[1/p^{e}]}\subseteq I$.
A positive real number $c$ is called an *$F$-jumping number of $\mf[a]$ in the direction $\bs[r]\not=\bs[0]\in\mathbb{Z}_{\geq0}^{n}$*, if $c$ is such that $\tau(\mf[a]^{c\bs[r]})\not=\tau(\mf[a]^{(c-\epsilon)\bs[r]})$ for every real number $\epsilon>0$.
\[Mixed test ideals =00003D normal test ideals\]If $r\in\mathbb{Z}_{\geq0}^{n}$ and $\lambda\in\mathbb{R}_{\geq0}$, then $$\tau(\mf[a]^{\lambda r_{1}}\cdots\mf[a]^{\lambda r_{n}})=\tau(\mf[J]^{\lambda}),$$
where $\mf[J]=\mf[a]_{1}^{r_{1}}\cdots\mf[a]_{n}^{r_{n}}$.
By Propostion \[Constancy in positive directions\], we may assume $\lambda=\frac{s}{p^{e'}}$ with $s\in\mathbb{Z}_{\geq0}$. For $e$ sufficiently large, we have $$\tau(\mf[a]^{\lambda r_{1}}\cdots\mf[a]^{\lambda r_{n}})=(\mf[a]^{\lceil\lambda r_{1}p^{e}\rceil}\cdots\mf[a]^{\lceil\lambda r_{n}p^{e}\rceil})^{[1/p^{e}]}=(\mf[a]^{sr_{1}p^{e-e'}}\cdots\mf[a]^{sr_{n}p^{e-e'}})^{[1/p^{e}]}$$ $$=((\mf[a]^{r_{1}}\cdots\mf[a]^{r_{n}})^{sp^{e-e'}})^{[1/p^{e}]}=((\mf[a]^{r_{1}}\cdots\mf[a]^{r_{n}})^{\lambda p^{e}})^{[1/p^{e}]}=\tau(\mf[J]^{\lambda}).$$
The $F$-threshold of $\mathfrak{a}$ associated to $I$ in the direction $\boldsymbol{r}=(r_{1},\ldots,r_{n})$ is equal to the $F$-threshold of $\mf[a]_{1}^{r_{1}}\cdots\mf[a]_{n}^{r_{n}}$ associated to $I$.
The set of $F$ -jumping numbers of $\mf[a]$ in direction ${\bf r}$ is equal to the set of $F$-jumping numbers of $\mf[a]^{{\bf r}}$.
Therefore [@BMS Corollary 2.30] implies the following.
The set of $F$-jumping numbers of $\mf[a]$ in the direction $\bs[r]$ is equal to the set of $F$-thresholds of $\mf[a]$, associated to various ideals $I$, in the direction $\bs[r]$.
Given $l_{1},\ldots,l_{n}$ positive real numbers we denote by $[\boldsymbol{0},\boldsymbol{l}]$ the set $[0,l_{1}]\times\ldots\times[0,l_{n}].$
\[Finite many test ideals\] Given nonzero ideals $\mf[a]_{1},\ldots,\mf[a]_{n}$ of $R$, where $R$ is a regular, $F$-finite ring essentially of finite type over a finite field, the set $\{\tau(\mf[a]^{\bs[c]})|\:\bs[c]\in[\bs[0],\bs[l]]\}$ is finite.
Since $R$ is assumed to be essentially of finite type over $k$, arguing as in the proof of [@BMS Theorem 3.1], one can see that the assertion for all such $R$ follows if we know it for $R=k[x_{1},\ldots x_{r}]$, with $r\geq1.$ We will therefore assume that we are in this case.
By Lemma \[Constancy in positive directions\], we may assume that $\bs[c]=(\frac{\alpha_{1}}{p^{e}},...,\frac{\alpha_{n}}{p^{e}})$ with $\alpha_{i}\in\mathbb{N}$ and $e\geq1$. Let $d$ be an upper bound for the degrees of the generators of $\mf[a_{i}]$, for all $i$. By Theorem \[thm:bound on the degree generators mixed case\] we have that $\tau(\mf[a]^{\bs[c]})$ is generated by polynomials of degree $\leq ndL$, where $L=\mbox{max}\{l_{i}\}$. Since $k$ is finite, there are only finitely many sets consisting of polynomials of bounded degree and therefore only finitely many ideals $\tau(\mf[a]^{\bs[c]})$ where $\bs[c]\in[\bs[0],\bs[l]]$.
The $\textit{constancy region}$ for a test ideal $\tau(\mf[a]^{\boldsymbol{c}})$ is defined as the set of points $\boldsymbol{c'}\in\mathbb{R}_{\geq0}^{n}$ such that $\tau(\mf[a]^{\boldsymbol{c}})=\tau(\mf[a]^{\boldsymbol{c}'})$.
Assume first that ${\bf c}=(\frac{\alpha_{1}}{p^{e}},\ldots,\frac{\alpha_{n}}{p^{e}})$ with $\alpha_{i}\in\mathbb{N}$. Choose a representation of ${\bf c}$ with $e$ large enough such that $\tau(\mbox{\ensuremath{\mf[a]}}^{{\bf c}})=(\mbox{\ensuremath{\mf[a]}}^{{\bf {\bf \alpha}}})^{[1/p^{e}]}.$ In this case we have
$${\bf c}\in B^{J}(\mf[a])\Longleftrightarrow\mbox{\ensuremath{\mf[a]}}^{{\bf \alpha}}\not\subseteq J^{[p^{e}]}\Longleftrightarrow(\mbox{\ensuremath{\mf[a]}}^{{\bf \alpha}})^{[1/p^{e}]}\not\subseteq J\Longleftrightarrow\tau(\mbox{\ensuremath{\mf[a]}}^{{\bf c}})\not\subseteq J.$$
For the general case, let ${\bf c}\in B^{J}(\mf[a])$, this implies that ${\bf c}\in B^{J}(\mf[a],p^{e})$ for some $e$. Therefore we can find ${\bf r}=(\frac{\alpha_{1}}{p^{e}},\ldots,\frac{\alpha_{n}}{p^{e}})\in B^{J}(\mf[a],p^{e})\subseteq B^{J}(\mf[a])$, with $\alpha_{i}\in\mathbb{N}$, $\frac{\alpha_{i}}{p^{e}}\geq c_{i}$. By the first part this implies $\tau(\mbox{\ensuremath{\mf[a]}}^{{\bf r}})\not\subseteq J$, but as $\frac{\alpha_{i}}{p^{e}}\geq c_{i}$ for all $i$, we have that $\tau(\mbox{\ensuremath{\mf[a]}}^{{\bf r}})\subseteq\tau(\mbox{\ensuremath{\mf[a]}}^{{\bf c}})$ hence $\tau(\mbox{\ensuremath{\mf[a]}}^{{\bf c}})\not\subseteq J$.
For the reverse inclusion, let ${\bf c}\in\mathbb{R}_{\geq0}^{n}$ be such that $\tau(\mf[a]^{{\bf c}})\not\subseteq J.$ By Proposition \[Constancy in positive directions\] there is a point ${\bf r}=(\frac{\alpha_{1}}{p^{e}},\ldots,\frac{\alpha_{n}}{p^{e}})$ with $\alpha_{i}\in\mathbb{N}$, $\frac{\alpha_{i}}{p^{e}}\geq c_{i}$ and $\tau(\mf[a]^{{\bf c}})=\tau(\mf[a]^{{\bf r}})$, therefore $\tau(\mf[a]^{{\bf r}})\not\subseteq J$. We use the first part again and conclude ${\bf r}\in B^{J}(\mf[a])$, but as $\frac{\alpha_{i}}{p^{e}}\geq c_{i}$ for all $i$, we deduce that ${\bf c}\in B^{J}(\mf[a])$.
If $\mf[a]_{1},...,\mbox{\ensuremath{\mf[a]}}_{n}$ are all contained in a maximal ideal $\mf[m]$ and the base field $k$ is finite, then for each $\boldsymbol{c}\in\mathbb{R}_{\geq0}^{n}$, there exist ideals $I_{1},...,I_{d}$ and $J$ such that the constancy region for the test ideal $\tau(\mf[a]^{\bs[c]})$ is given by $\underset{i=1,...,d}{\bigcap}B^{I_{i}}(\mf[a])\backslash B^{J}(\mf[a])$.
We first show that this constancy region is bounded. As $\mf[a]_{i}\subseteq\mf[m]$ for all $i$ we have that for any $\boldsymbol{c'}\in\mathbb{R}_{\geq0}^{n}$and $e$ sufficiently large $$\tau(\mf[a]^{{\bf c'}})=(\mf[a]_{1}^{\lceil c'_{1}p^{e}\rceil}\cdots\mf[a]_{n}^{\lceil c'_{n}p^{e}\rceil})^{[1/p^{e}]}\subseteq(\mf[m]^{\lceil c'_{1}p^{e}\rceil+\ldots+\lceil c'_{n}p^{e}\rceil})^{[1/p^{e}]}$$ $$\subseteq\mbox{\ensuremath{\mf[(]}}\mf[m]^{\lceil c'_{1}p^{e}+\ldots+c'_{n}p^{e}\rceil-n})^{[1/p^{e}]}\subseteq\mf[m]^{\lceil c'_{1}+\ldots+c'_{n}\rceil-n+1}.$$
Since $\cap_{s}\mf[m]^{s}=0$, there is $L$ such that $\tau(\mf[a]^{{\bf c}})\not\subseteq\mbox{\ensuremath{\mf[m]}}^{L}$, we deduce that for any ${\bf c}'$ in the constancy region $\tau(\mf[a]^{{\bf c}'})=\tau(\mf[a]^{{\bf c}})\not\subseteq\mbox{\ensuremath{\mf[m]}}^{L}$, hence $c'_{1}+\ldots+c'_{n}\leq L$. This implies that the constancy region for $\tau(\mf[a]^{\bs[c]})$ is bounded.
To deduce our description consider a sufficiently large hypercube $[\bs[0],\bs[l]]$ containing the constancy region for $\tau(\mf[a]^{{\bf c}})$. By Proposition \[Finite many test ideals\], we know that the set $\mathcal{A}=\{\tau(\mf[a]^{\bs[c]})|\:\bs[c]\in[\bs[0],\bs[l]]\}$ is finite. Let $I_{1},\ldots,I_{d}$ be the ideals in $\mathcal{A}$ that are strictly contained in $\tau(\mf[a]^{{\bf c}})$ and let $J=\tau(\mf[a]^{{\bf c}})$. We claim that the constancy region for $\tau(\mf[a]^{{\bf c}})$ is equal to $\underset{i=1,...,d}{\bigcap}B^{I_{i}}(\mf[a])\backslash B^{J}(\mf[a])$. Lemma \[lem:Description of the approximation\] implies that the set $\underset{i=1,...,d}{\bigcap}B^{I_{i}}(\mf[a])\backslash B^{J}(\mf[a])$ is equal to the set of all ${\bf r}$ such that $\tau(\mf[a]^{{\bf r}})\not\subseteq I_{i}$ for all $i$ and $\tau(\mf[a]^{{\bf c}})\subseteq\tau(\mf[a]^{{\bf r}})$, or equivalently, $\tau(\mf[a]^{{\bf r}})=\tau(\mf[a]^{{\bf c}})$ by our choice of $I_{i}$.
\[Regions\]We can remove the condition that all ideals $\mf[a]_{i}$ are contained in a maximal ideal and still get a similar description. Explicitly, in each hypercube $[\bs[0],\bs[l]]$ the constancy region is given by $\underset{i=1,...,d}{\bigcap}\big(B^{I_{i}}(\mf[a])\backslash B^{J}(\mf[a])\big)\cap[\bs[0],\bs[l]],$ for suitable $I_{1},\ldots,I_{d}$ and $J$.
We now give a version of Skoda’s theorem for mixed test ideals (see [@BMS Proposition 2.25] for the case of one ideal). This theorem allows us to describe the constancy regions in the first octant by describing only the constancy regions in a sufficiently large hypercube $[\bs[0],\bs[l]]=[0,l_{1}]\times\ldots\times[0,l_{n}]$.
\[thm:(Skoda’s-Theorem)\](Skoda’s Theorem) Let $\bs[e_{1}],...,\bs[e_{n}]$ be the standard basis for $\mathbb{R}^{n}$, and assume $1\leq i\leq n$. If $\mf[a]_{i}$ is generated by $m_{i}$ elements, then for every $\bs[s]=(s_{1},\ldots,s_{n})$ with $s_{i}\geq m_{i}$, we have $$\tau(\mf[a]^{\bs[s]})=\mf[a]^{\bs[e_{i}]}\tau(\mf[a]^{\bs[s]-\bs[e_{i}]}).$$
We only need to prove $(\mf[a]^{\lceil p^{e}\bs[s]\rceil})^{[1/p^{e}]}=\mf[a]^{\bs[e_{i}]}(\mf[a]^{\lceil p^{e}(\bs[s]-\bs[e_{i}])\rceil})^{[1/p^{e}]}$ for $e$ large enough.
Let $\bs[d]=(d_{1},..,d_{n})$ be a vector with integer coordinates and $d_{i}\geq p^{e}s_{i}$. We want to show that $$(\mf[a]^{\bs[d]})^{[1/p^{e}]}=\mf[a]^{\bs[e_{i}]}(\mf[a]^{\bs[d]-p^{e}\bs[e_{i}]})^{[1/p^{e}]},$$ from which the result follows.
Since $\mf[a]^{\bs[d]-p^{e}\bs[e_{i}]}\cdot\mf[a]_{i}^{[p^{e}]}\subseteq\mbox{\ensuremath{\mf[a]}}^{\bs[d]}\subseteq((\mf[a]^{\bs[d]})^{[1/p^{e}]})^{[p^{e}]}$, we have $$\mf[a]^{\boldsymbol{d}-p^{e}\bs[e_{i}]}\subseteq\big(((\mf[a]^{\bs[d]})^{[1/p^{e}]})^{[p^{e}]}:\mf[a]_{i}^{[p^{e}]}\big)=\big((\mf[a]^{\bs[d]})^{[1/p^{e}]}:\mf[a]_{i}\big)^{[p^{e}]},$$ where the equality is consequence of the flatness of Frobenius. Therefore $$(\mf[a]^{\bs[d]-p^{e}\bs[e_{i}]})^{[1/p^{e}]}\subseteq((\mf[a]^{\bs[d]})^{[1/p^{e}]}:\mf[a]_{i}),$$ that is, $$\mf[a]^{\bs[e_{i}]}(\mf[a]^{\bs[d]-p^{e}\bs[e_{i}]})^{[1/p^{e}]}\subseteq(\mf[a]^{\bs[d]})^{[1/p^{e}]}.$$ For the reverse inclusion, note that since $d_{i}\geq m_{i}(p^{e}-1)+1$, in the product of $d_{i}$ of the generators of $\mf[a]_{i}$ at least one should appear with multiplicity $\geq p^{e}$. Therefore $\mf[a]^{\bs[d]}=\mf[a]_{i}^{[p^{e}]}\cdot\mf[a]^{\bs[d]-p^{e}\bs[e_{i}]}$, hence $$\mf[a]^{\bs[d]}\subseteq\mf[a]_{i}^{[p^{e}]}\cdot\mf[a]^{\bs[d]-p^{e}\bs[e_{i}]}\subseteq\mf[a]_{i}^{[p^{e}]}\cdot\big((\mf[a]^{\bs[d]-p^{e}\bs[e_{i}]})^{[1/p^{e}]}\big)^{[p^{e}]}=\big(\mf[a]^{\bs[e_{i}]}\cdot(\mf[a]^{\bs[d]-p^{e}\bs[e_{i}]})^{[1/p^{e}]}\big)^{[p^{e}]},$$ which clearly implies $(\mf[a]^{\bs[d]})^{[1/p^{e}]}\subseteq\mf[a]^{\bs[e_{i}]}(\mf[a]^{\bs[d]-p^{e}\bs[e_{i}]})^{[1/p^{e}]}.$
If $c$ is an $F$-jumping number in the direction $\bs[r]=(r_{1},\ldots,r_{n})$ then also $cp$ is an $F$-jumping number in the direction $\bs[r]$.
Note that $V_{\bs[r]}^{I}(\mf[a],p^{e+1})=V_{\bs[r]}^{I^{[p]}}(\mf[a],p^{e})$, hence $pC_{\bs[r]}^{I}(\mf[a])=C_{\bs[r]}^{I^{[p]}}(\mf[a])$.
The constancy regions
=====================
In this section we prove our main result, Theorem \[thm:Main Theorem\] below. We begin by recalling our definition of $p$-fractals.
Let $\mathcal{F}$ be the algebra of functions $\phi:\mathbb{R}_{\geq0}^{n}\rightarrow\mathbb{Q}$. For each $q=p^{e}$ and every $\bs[b]=(b_{1},\ldots,b_{n})\in\mathbb{Z}^{n}$ with $0\leq b_{i}<q$ we define a family of operators $T_{q|\bs[b]}:\mathcal{F}\rightarrow\mathcal{F}$ by $$T_{q|\bs[b]}\phi(t_{1},\ldots,t_{n})=\phi((t_{1}+b_{1})/q,\ldots,(t_{n}+b_{n})/q).$$
\[Def Fractal function\]Let $\phi:[0,l]^{n}\rightarrow\mathbb{Q}$ be a map and let denote also by $\phi$ its extension by zero to $\mathbb{R}_{\geq0}^{n}$. We say that $\phi$ is a $p\textit{-fractal}$ if all the $T_{q|\bs[b]}\phi$ span a finite dimensional $\mathbb{Q}$-subspace $V$ of $\mathcal{F}$. Furthermore, we say that an arbitrary $\phi\in\mathcal{F}$ is a $p\textit{-fractal}$ if its restriction to each hypercube $[\bs[0],\bs[l]]$ is a $p$-fractal.
This definition is similar to the one in [@MT Definition 2.1]. The only difference is that in [@MT Definition 2.1] the domain of the functions is the hypercube $[0,1]\times\ldots\times[0,1].$
In this section we assume that $R$ is a regular, $F$-finite ring essentially of finite type over a finite field of characteristic $p>0$, and $\mf[a]_{i}\subseteq R$ are nonzero ideals.
\[lemma for the main theorem\] Let $\bs[c]=(c_{1},\ldots,c_{n})\in\mathbb{R}_{\geq0}^{n}$, and $\mf[a]_{1},\ldots\mf[a]_{n}$ be nonzero ideals of $R$ then $\tau(\mf[a]^{\bs[c]})^{[1/p^{e}]}=\tau(\mf[a]^{\frac{1}{p^{e}}\bs[c]})$.
Taking $k$ large enough $$\tau(\mf[a]^{{\bf c}})^{[1/p^{e}]}=\left((\mf[a]_{1}^{\lceil c{}_{1}p^{k}\rceil}\cdots\mf[a]_{n}^{\lceil c{}_{n}p^{k}\rceil})^{[1/p^{k}]}\right)^{[1/p^{e}]}$$ and by [@BMS Lemma 2.4] the later contains $$(\mf[a]_{1}^{\lceil c{}_{1}p^{k}\rceil}\cdots\mf[a]_{n}^{\lceil c{}_{n}p^{k}\rceil})^{[1/p^{k+e}]}=(\mf[a]_{1}^{\lceil\frac{c{}_{1}}{p^{e}}p^{k+e}\rceil}\cdots\mf[a]_{n}^{\lceil\frac{c{}_{n}}{p^{e}}p^{k+e}\rceil})^{[1/p^{k+e}]}$$ $$=\tau(\mf[a]^{\frac{1}{p^{e}}\bs[c]}).$$
Therefore $\tau(\mf[a]^{{\bf c}})^{[1/p^{e}]}\supseteq\tau(\mf[a]^{\frac{1}{p^{e}}\bs[c]}).$
For the other inclusion note that $$\tau(\mf[a]^{{\bf c}})=(\mf[a]_{1}^{\lceil c{}_{1}p^{k}\rceil}\cdots\mf[a]_{n}^{\lceil c{}_{n}p^{k}\rceil})^{[1/p^{k}]}$$ $$=(\mf[a]_{1}^{\lceil\frac{c{}_{1}}{p^{e}}p^{k+e}\rceil}\cdots\mf[a]_{n}^{\lceil\frac{c{}_{n}}{p^{e}}p^{k+e}\rceil})^{[p^{e}/p^{k+e}]}$$ that by [@BMS Lemma 2.4] is contained in $$\left((\mf[a]_{1}^{\lceil\frac{c{}_{1}}{p^{e}}p^{k+e}\rceil}\cdots\mf[a]_{n}^{\lceil\frac{c{}_{n}}{p^{e}}p^{k+e}\rceil})^{[1/p^{k+e}]}\right)^{[p^{e}]}=\tau(\mf[a]^{\frac{1}{p^{e}}\bs[c]})^{[p^{e}]}$$ but this is equivalent to say $$\tau(\mf[a]^{{\bf c}})^{[1/p^{e}]}\subseteq\tau(\mf[a]^{\frac{1}{p^{e}}\bs[c]}).$$
\[characterisct functions fractal behavior\] Let $\boldsymbol{l}=(l_{1},\ldots l_{n})\in\mathbb{Z}^{n}$ be such that $l_{i}$ is the minimum number of generators of the ideal $\mf[a_{i}]$. Let $\bs[b]\in\mathbb{Z}^{n}$ such that $l_{i}-1\leq b_{i}$ . For all $e$, we have where $\chi_{\mf[a]}^{I}$ denotes the characteristic function introduced in Definition \[Def characteristic function\].
We have that $$T_{p^{e}|\bs[b]}\chi_{\mf[a]}^{I}(\bs[t])=\chi_{\mf[a]}^{I}\left(\frac{1}{p^{e}}(\bs[t]+\bs[b])\right)$$
is equal to $1$ if and only if, by Lemma \[lem:Description of the approximation\], to $$\tau(\mf[a]^{\frac{1}{p^{e}}(\bs[t]+\bs[b])})\not\subseteq I,$$
and by Lemma \[lemma for the main theorem\] this is $$\tau(\mf[a]^{\bs[t]+\bs[b]})^{[1/p^{e}]}\not\subseteq I,$$
but the later is equivalent to $$\tau(\mf[a]^{\bs[t]+\bs[b]})\not\subseteq I^{[p^{e}]}.$$
As $b_{i}\geq l_{i}-1$ by Skoda’s Theorem the previous expresion becomes $$\mf[a]^{\bs[b-l+1]}\cdot\tau(\mf[a]^{\bs[t]+\bs[l]-\bs[1]})\not\subseteq I^{[p^{e}]}$$
Wich in turn is equivalent to $$\tau(\mf[a]^{\bs[t]+\bs[l]-\bs[1]})\not\subseteq(I^{[p^{e}]}:\mf[a]^{\bs[b-l+1]})$$ but this is the case if and only if $$T_{p^{0}|(\bs[l]-\bs[1])}\chi_{\mf[a]}^{(I^{[p^{e}]}:\mf[a]^{\bs[b-l+1]})}(\bs[t])=\chi_{\mf[a]}^{(I^{[p^{e}]}:\mf[a]^{\bs[b-l+1]})}(\bs[t]+\bs[l]-\bs[1])$$
is equal to $1$
Note that a point of the form $\frac{1}{p^{k}}\bs[r]+(\boldsymbol{l}-\boldsymbol{1})$ with $\bs[r]\in\mathbb{Z}^{n}$ is in $B^{(I^{[p^{e}]}:\mf[a]^{b-l+1})}(\mf[a])$ if and only if $\mf[a]^{\bs[r]+p^{k}(\bs[l]-\bs[1])}\not\not\subseteq(I^{[p^{e}]}:\mf[a]^{\bs[b]-\bs[l]+\bs[1]})^{[p^{k}]}$ if and only if $\mf[a]^{\bs[r]}\cdot\mf[a]^{p^{k}(\bs[l]-\bs[1])}\cdot(\mf[a]^{\bs[b]-\bs[l]+\bs[1]})^{[p^{k}]}\not\subseteq I^{[p^{e+k}]}$ this by Lemma \[lemma for the main theorem\] occurs if and only if $\mf[a]^{\bs[r]+p^{k}\bs[b]}\not\subseteq I^{[p^{e+k}]}$, or equivalently $\frac{1}{p^{e+k}}\bs[r]+\frac{1}{p^{e}}\bs[b]\in B^{I}(\mf[a])$. From this the result follows easily.
This lemma is especially useful when the ideals are principal, as we will see in the examples of Section 5.
\[lem:In-each-hypercube\] In each hypercube $[\bs[0],\bs[l]]$ there are only finitely many functions $\chi_{\mf[a]}^{I}$. That is, the set $\{\chi_{\mf[a]}^{I}|_{[\bs[0,l]]};I\subseteq R\}$ is finite.
By Lemma \[lem:Description of the approximation\], $B^{I}(\mf[a])$ is the set of all points $\bs[c]=(c_{1},\ldots,c_{n})\in\mathbb{R}_{\geq0}^{n}$ such that $\tau(\mf[a]^{\bs[c]})\not\subseteq I$, hence $B^{I}(\mf[a])$ is a union of constancy regions. By Lemma \[Finite many test ideals\], we know that there are only finitely many constancy regions for bounded exponents, therefore there are only finitely many functions $\chi_{\mf[a]}^{I}|_{[\bs[0],\bs[l]]}$.
and therefore the constancy regions are of the form $\varphi^{-1}(i)$ for some number $i.$
We first show that the functions $\chi_{\mf[a]}^{I}$ are $p$-fractal. We want to prove that all the $T_{p^{e}|\bs[b]}\chi_{\mf[a]}^{I}$ span a finite dimensional space. Lemma \[characterisct functions fractal behavior\] states that all but finitely many of these functions have the form $T_{p^{0}|(\bs[l]-\bs[1])}\chi_{\mf[a]}^{J}$ for different ideals $J$. Lemma \[lem:In-each-hypercube\] ensures that there are only finitely many of those in each hypercube $[\bs[0],\bs[l]]$. From this it follows that $\chi_{\mf[a]}^{I}$ is a $p$-fractal.
For $\bs[c]\in\mathbb{R}_{\geq0}^{n}$, let $\eta_{\bs[c]}$ be the characteristic function associated to the constancy region $\tau(\mf[a]^{\bs[c]})$. Remark \[Regions\] implies that that in each hypercube $[\bs[0],\bs[l]]$, $\eta_{\bs[c]}|_{[\bs[0],\bs[l]]}=\left(\chi_{\mf[a]}^{I_{1}}\cdots\chi_{\mf[a]}^{I_{d}}-\chi_{\mf[a]}^{J}\right)\big|{}_{[\bs[0],\bs[l]]}$ for some ideals $I_{1},\ldots,I_{d}$ and $J$, therefore $\eta_{\bs[c]}$ is $p$-fractal.
Clearly there are countably many constancy regions, so we can numerate them. For every $i$, let $\bs[c]_{i}=(c_{i1},\ldots,c_{in})$ a point in the $i$-th constancy region, and we define $$\varphi=\sum_{i\in\mathbb{N}}i\cdot\eta_{\bs[c]_{i}}.$$ This function satisfies the desired conditions.
\[cor-caracteristic functions of constancy regions are p-fractal\]Let $\eta_{\bs[c]}$ be the characteristic function associated to the constancy region of $\tau(\mf[a]^{\bs[c]})$, then $\eta_{\bs[c]}$ is a $p$- fractal.
An Example
==========
In section 4 we showed that the characteristic functions of the constancy regions are $p$-fractal functions, Corollary \[cor-caracteristic functions of constancy regions are p-fractal\]. We use this fact and Proposition \[thm: root ideals\] to compute an explicit example. Throughout this section we use a subscript $*_{p}$ to denote that the number is written in base $p$. One of the main tools for computing examples is the following theorem:
\[thm:(Lucas’-Theorem)\](Lucas’ Theorem [@E]) Fix non-negative integers $m\geq n\in\mathbb{N}$ and a prime number $p$. Write $m$ and $n$ in their base $p$ expansions: $m=\sum_{j=0}^{r}m_{j}p^{j}$ and $n=\sum_{j=0}^{r}n_{j}p^{j}.$Then modulo $p,$ $$\binom{m}{n}=\binom{m_{0}}{n_{0}}\cdot\binom{m_{1}}{n_{1}}\cdots\binom{m_{r}}{n_{r}},$$ where we interpret $\binom{a}{b}$ as zero if $a<b$. In particular, $\binom{m}{n}$ is non-zero mod $p$ if and only if $m_{j}\geq n_{j}$ for all $j=1,\ldots r$.
\[nonzero coefficients\]In particular if $m=p^{k}-1$ all the coefficients in the expansion of $(x+y)^{m}$ are nonzero.
(The Devil’s Staircase) Let $R=\mathbb{F}_{3}[x,y]$, $f_{1}=x+y$, and $f_{2}=xy$. We want to describe the constancy regions for the test ideals $\tau(f^{\boldsymbol{c}})$.
We first show that there are five different test ideals in the region $[0,1]\times[0,1]$. More precisely, we show that $$\tau(f^{\boldsymbol{c}})=\begin{cases}
R\mbox{ or }(x,y), & \mbox{\ensuremath{\boldsymbol{c}\in[0,1)\times[0,1)}}\\
(x+y), & \boldsymbol{c}\in\{1\}\times[0,1)\\
(xy), & \boldsymbol{c}\in[0,1)\times\{1\}\\
(xy(x+y)), & \boldsymbol{c}=(1,1).
\end{cases}$$
We want to compute the test ideal at $(\frac{1}{3},\frac{2}{3})$ . By Proposition \[Mixed test ideals =00003D normal test ideals\] $$\tau(f^{(0.1_{3},0.2_{3})})=\tau((f_{1}\cdot f_{2}^{2})^{\frac{1}{3}}).$$ By Proposition \[Computing test ideals for principal ideals\], $$\tau((f_{1}\cdot f_{2}^{2})^{\frac{1}{3}})=((x+y)(xy)^{2})^{[\frac{1}{3}]}=(x^{3}y^{2}+x^{2}y^{3})^{[\frac{1}{3}]}.$$ Finally, Proposition \[thm: root ideals\] gives $$(x^{3}y^{2}+x^{2}y^{3})^{[\frac{1}{3}]}=(x,y),$$ and therefore $$\tau(f_{1}^{c_{1}}\cdot f_{2}^{c_{2}})\subseteq(x,y)\mbox{ if \ensuremath{c_{1}\geq1/3}and \ensuremath{c_{2}\geq2/3}.}$$ In particular, the test ideal associated to the points $(1-\frac{1}{3^{k}},1-\frac{1}{3^{k}})$ is contained in $(x,y)$. Now $$\tau(f^{(1-\frac{1}{3^{k}},1-\frac{1}{3^{k}})})=((x+y)^{3^{k}-1}(xy)^{3^{k}-1})^{[\frac{1}{3^{k}}]}$$ $$=((x^{2}y+xy^{2})^{3^{k}-1})^{[\frac{1}{3^{k}}]}.$$
Since the terms $x^{2(3^{k}-1)}y^{3^{k}-1}$and $x^{3^{k}-1}y^{2(3^{k}-1)}$appear in the expansion of $(x^{2}y+xy^{2})^{3^{k}-1}$ with nonzero coefficient, Remark \[nonzero coefficients\]. We conclude that $\tau(f^{(1-\frac{1}{3^{k}},1-\frac{1}{3^{k}})})\supseteq(x,y).$ Therefore $$\tau(f^{(1-\frac{1}{3^{k}},1-\frac{1}{3^{k}})})=(x,y).$$
Thus there are only two test ideals in the region $[0,1)\times[0,1)$, these are $R$ and $(x,y)$.
Clearly $\tau(f^{(1,0)})=(x+y)$, and by Skoda’s Theorem $$\tau(f^{(1,1-\frac{1}{3^{k}})})=f_{1}\cdot\tau(f^{(0,1-\frac{1}{3^{k}})})$$ $$=(x+y)\cdot((xy)^{3^{k}-1})^{[\frac{1}{3^{k}}]}$$ $$=(x+y),$$
hence the only test ideal in the region $[0,1)\times\{1\}$ is $(x+y)$.
In a similar way, $\tau(f^{(0,1)})=(xy)$ and $$\tau(f^{(1-\frac{1}{3^{k}},1)})=f_{2}\cdot\tau(f^{(1-\frac{1}{3^{k}},0)})$$ $$=(xy)\cdot((x+y)^{3^{k}-1})^{[\frac{1}{3^{k}}]}$$ $$=(xy).$$
Thus $(xy)$ is the only test ideal that appears in the region$\{1\}\times[0,1)$.
Lastly, note that the test ideal at $(1,1)$ is $$\tau(f^{(1,1)})=((x+y)xy).$$
We now show that $(\frac{1}{3},\frac{2}{3})$ is a point in the boundary of $B^{(x,y)}(f)$ and then use the $p$-fractal structure to sketch the constancy regions.
For every $k$ $$\tau(f^{(\frac{1}{3}-\frac{1}{3^{k}},\frac{2}{3}-\frac{1}{3^{k}})})$$ $$=((x+y)^{3^{k-1}-1}(xy)^{2\cdot3^{k-1}-1})^{[\frac{1}{3^{k}}]}.$$ But in the expansion of $(x+y)^{3^{k-1}-1}$ every term appears with nonzero coefficient, Remark \[nonzero coefficients\]. In particular the term $(xy)^{\frac{3^{k-1}-1}{2}}(xy)^{2\cdot3^{k-1}-1}$ appears with non-zero coefficient when expanding the product $(x+y)^{3^{k-1}-1}(xy)^{2\cdot3^{k-1}-1}$. Since the degrees in $x$ and $y$ of this monomial are smaller than $3^{k}$, by Proposition \[thm: root ideals\] we conclude that $\tau(f^{(\frac{1}{3}-\frac{1}{3^{k}},\frac{2}{3}-\frac{1}{3^{k}})})=R$ . Thus $$\chi_{f}^{(x,y)}(\frac{1}{3},\frac{2}{3})=0$$
and $$\chi_{f}^{(x,y)}([0,\frac{1}{3})\times[0,\frac{2}{3}))=1.$$
The later shows that the point $(\frac{1}{3},\frac{2}{3})$ is in the boundary of constancy regions for $R$ and $(x,y)$. We can use the $p$-fractal structure to find more points in this boundary. The idea is to break the region $[0,1]\times[0,1]$ into squares of length $1/3$ and find which of these must contain a boundary point. Then we apply the $p$-fractal structure to these squares to find the points.
For the points $(0,\frac{2}{3})$, $(\frac{2}{3},\frac{1}{3})$, $(\frac{1}{3},1)$, and $(1,\frac{2}{3})$ we have: $$\tau(f^{(0,\frac{2}{3})})=((xy)^{2})^{[\frac{1}{3}]}=R,$$ $$\tau(f^{(\frac{2}{3},\frac{1}{3})})=((x+y)^{2}xy)^{[\frac{1}{3}]}=(x^{3}y-x^{2}y^{2}+xy^{3})^{[\frac{1}{3}]}=R$$
and $$\tau(f^{(\frac{1}{3},1)})=((x+y)(xy)^{3})^{[\frac{1}{3}]}=(xy)\subset(x,y),$$ $$\tau(f^{(1,\frac{2}{3})})=((x+y)^{3}x^{2}y^{2})^{[\frac{1}{3}]}=(x+y)\subset(x,y).$$
Therefore there should be boundary points in the squares $[0,1/3)\times[2/3,1)$ and $[2/3,1)\times[0,1/3)$. Is easy to check that there are not boundary points in all the other squares. From this and lemma \[characterisct functions fractal behavior\] we know that $T_{3|(0,2)}\chi_{f}^{(x,y)}=T_{3|(2,1)}\chi_{f}^{(x,y)}=\chi_{f}^{(x,y)}$, since $\chi_{f}^{(x,y)}$ is the only characteristic function that is non constant in $[0,1)\times[0,1)$. Moreover, $$\chi_{f}^{(x,y)}(0.01_{3},0.22_{3})=\chi_{f}^{(x,y)}(0_{3}+0.01_{3},0.2_{3}+0.02_{3})$$ $$=\chi_{f}^{(x,y)}(\frac{0_{3}+0.1_{3}}{3},\frac{2_{3}+0.2_{3}}{3})=T_{3|(0,2)}\chi_{f}^{(x,y)}(0.1_{3},0.2_{3})$$ $$=\chi_{f}^{(x,y)}(0.1_{3},0.2_{3})=\chi_{f}^{(x,y)}(\frac{1}{3},\frac{2}{3})=0$$ in a similar way $$\chi_{f}^{(x,y)}(0.21_{3},0.12_{3})=0$$ and $$\chi_{f}^{(x,y)}([0,0.01_{3})\times[0,0.22_{3}))=\chi_{f}^{(x,y)}([0,0.21_{3})\times[0,0.12_{3}))=1.$$ This is the points $(0.01_{3},0.22_{3})$ and $(0.21_{3},0.12_{3})$ are also in the boundary. We can repeat the proccess by subdividing the squares $[0,1/3)\times[2/3,1)$ and $[2/3,1)\times[0,1/3)$ into smaller squares of length $1/9$ and obtain more points of the boundary. This process can be sumarized as follows. Let $A$ is the set of points obtained from $(0.1_{3},0.2_{3})$ by successively applying the operations $$(0.a_{1}\ldots a_{n}1_{3},0.b_{1}\ldots b_{n}2_{3})\mapsto\begin{cases}
(0.a_{1}\ldots a_{n}01_{3},0.b_{1}\ldots b_{n}22_{3})\\
(0.a_{1}\ldots a_{n}21,0.b_{1}\ldots b_{n}12_{3})
\end{cases}$$ then $$\chi_{f}^{(x,y)}(\boldsymbol{p})=0$$ and $$\chi_{f}^{(x,y)}([\boldsymbol{0},\boldsymbol{p}))=1$$ for all $\boldsymbol{p}\in A$. This is, the points of $A$ are points in the boundary. We can now sketch the regions of constancy in $[0,1]\times[0,1]$:
Using Skoda’s theorem, we can describe the whole diagram of test ideals:
We choose the name Devil’s Staircase for this example, because the resemblance to the Devil’s Staircases or Cantor functions that appear in the basic courses of analysis.
In a similar way, it can be shown that for any characteristic $p$ the same polynomials give a staircase that has infinitely many steps. Indeed, $$\tau(f^{(\frac{1}{p^{k}},1-\frac{1}{p^{k}})})=((x+y)(xy)^{p^{k}-1})^{[\frac{1}{p^{k}}]}=(x,y)$$
but $$\tau(f^{(\frac{2}{p^{k}},1-\frac{2}{p^{k}})})=((x+y)^{2}(xy)^{p^{k}-2})^{[\frac{1}{p^{k}}]}=R$$ and so we have many different points in the line $x+2y=2$ with test ideal equal to $(x,y)$ and infinitely many with test ideal equal to $R$. Therefore we can not expect that there are characteristics for which the region given by the test ideals will be the same as the one given by the multiplier ideals
[BMS2]{} M. Blickle, M. Mustaţă, K. Smith, Discreteness and rationality of $F-\mbox{thresholds}$, $\textit{Michigan Math. J.}$ $\boldsymbol{57}$ (2008), 43–61.
Blickle, Manuel; Mustaţă, Mircea; Smith, Karen E. F-thresholds of hypersurfaces. *Trans. Amer. Math. Soc*. 361 (2009), no. 12, 6549–6565.
E. Lucas, Theorie des Fonctions Numeriques Simplement Periodiques, *Amer. J. Math* **1** (1878), no. 3, 197–240.
L. Ein, R. Lazarfeld, K.E. Smith, and D. Varolin, Jumping coefficients of multiplier ideals, *Duke Math. J.* **123** (2004), 469–506.
C. Huneke and M. Hochster, Tight closure, invariant theory and the Briançon- Skoda theorem, *J. Amer. Math. Soc.* **3** (1990), 31–116.
N. Hara and K. -i. Yoshida, A generalization of tight closure and multiplier ideals, $\textit{Trans. Amer. Math. Soc. }$$\boldsymbol{355}$ (2003), 3143–3174. 1,2,6,8,9,10,11,13,15.
R. Lazarfeld: $\textit{Positivity in Algebraic Geometry}$ II. Ergebnise der Mathematik und ihrer Grenzegebiete. 3. Folge, Vol. 49, Springer-Verlag, Berlin, 2004.
P. Monsky and P. Teixeira, $p$-Fractals and power series-I Some 2 variable results, *Journal of Algebra* **280** (2004), 505–536.
M. Mustaţă, S. Takagi and K.-i. Watanabe, $F-$thresholds and Bernstein-Sato polynomials, European Congress of Mathematics, 341–364, Eur. Math. Soc., Zürich, 2005. 1, 2, 3, 13.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Using hydrodynamical simulations, we explore the use of the mean and percentiles of the curvature distribution function to recover the equation of state of the high-$z$ ($2 < z < 4$) intergalactic medium (IGM). We find that the mean and percentiles of the absolute curvature distribution exhibit tight correlation with the temperatures measured at respective characteristic overdensities $\bar{\Delta}_i$’s at each redshift. Hence, they provide [complementary]{} probes of the same underlying temperature-density distribution, and can in principle be used to simultaneously recover both parameters $T_0$ and $\gamma$ of the IGM effective equation of state. We quantify the associated errors in the recovered parameters $T_0$ and $\gamma$ from the intrinsic scatter in the characteristic overdensities and the uncertainties in the curvature measurement.'
author:
- |
Hamsa Padmanabhan$^{1}$[^1], R. Srianand$^1$[^2], T. Roy Choudhury$^2$[^3]\
$^{1}$ Inter-University Centre for Astronomy and Astrophysics, Pune 411007, India\
$^{2}$ National Centre for Radio Astrophysics, Tata Institute of Fundamental Research, Pune 411007, India
bibliography:
- 'mybib.bib'
date:
title: 'Measuring the equation of state of the high$-z$ intergalactic medium using curvature statistics'
---
dark ages, reionization, first stars - intergalactic medium - quasars : absorption lines
Introduction
============
The Lyman-$\alpha$ absorption lines seen in the spectra of high-redshift quasars arise predominantly from the low to moderate overdensity intergalactic medium [IGM; @cen; @zhang; @petitjean1995; @hernquist; @miralda]. The balance between photoionization heating and adiabatic cooling in the IGM leads to a mean temperature-density relation that is well-approximated [@huignedin] by a power-law, $T = T_0 \Delta^{\gamma - 1}$ for overdensities $\Delta \leq 10$. The parameters of this “equation of state” depend upon the reionization history of the universe and their predicted values at different epochs vary in different reionization scenarios, so long as either hydrogen or helium reionization happened relatively recently. Otherwise, the temperature-density relation is predominantly set by the shape of the ionizing ultraviolet background and the adiabatic expansion or collapse of large-scale structure [@huignedin]. Non-radiative mechanisms of energy injection like, e.g. blazar heating [@chang2012; @puchwein2012] may also distort the $T-\Delta$ relation. Hence, it is important to measure the temperature-density relation at various redshifts in order to constrain the epoch and evolution of reionization and the properties of the ionizing and heating sources.
Detailed studies of different properties of the IGM have contributed to understanding the thermal history of the IGM, by using observed data together with the results of numerical simulations. The methods include (a) using the Doppler $b$-parameter-column density ($b-N$) or $b$-distribution of the Lyman-$\alpha$ absorption lines originating from the IGM [@haehnelt1998; @schaye1999; @schaye2000; @mcdonald2001; @tirthanandtp1; @ricotti2000a; @bryan2000; @bolton10; @bolton12], (b) using the small-scale power spectrum of the Lyman-$\alpha$ forest [@theuns2000a; @zaldarriaga2001], and (c) using the wavelet decomposition of the Lyman-$\alpha$ lines [@theuns2000; @zaldarriaga2002; @theuns2002; @theuns2002a; @lidz2010]. The latter two methods do not require Voigt profile decomposition of the spectral lines. Some evidence for a peak in the intergalactic medium temperature around $z \sim 3$, together with a nearly isothermal profile (i.e. $\gamma = 1$) signifying the end of reionization, has been reported [@schaye2000; @theuns2002a; @lidz2010], however, e.g. @mcdonald2001 do not find evidence supporting this claim. Interestingly, the parameters ($T_0$ and $\gamma$) derived using the above methods typically have large uncertainties ($\gtrsim$ 30%).
Recently, the temperature of the IGM over redshifts 4.5 to $\sim 2.8$ has been measured to a high precision ($\lesssim 10$%) by @becker11 [and extended upto $z \sim 1.5$ by @boera2014] using the curvature statistic, which also avoids the fitting of individual spectral lines and can be used to detect the additional heating effect in high-redshift quasar near-zones [@hp2014 hereafter Paper I]. The curvature is normally sensitive to both $T_0$ and $\gamma$, however, it is found that at a characteristic overdensity $\bar{\Delta}$, it becomes more sensitive to $T(\bar{\Delta})$ and fairly independent of $\gamma$. The value of $\bar{\Delta}$ increases with decreasing redshifts, going towards non-linear densities at lower redshifts. The actual value of $\bar{\Delta}$ at each redshift also depends on the effective optical depth in the simulations used to calibrate the observations [@boera2014]. This procedure leads to significantly smaller uncertainties in the determination of $T(\bar{\Delta})$ compared to the typical errors in $T_0$ derived using the other methods discussed above. The results indicate a steady rise in $T (\bar{\Delta})$ across $z \sim 4$ to $z \sim 2$, consistent with additional heat input either from the prolonged reionization of or non-radiative energy injection from blazar heating, etc. A significant contribution to the rise in $T (\bar{\Delta})$ comes from the increase of $\bar{\Delta}$ itself with decreasing redshifts. However, in order to obtain $T_0$ from $T(\bar{\Delta})$, one has to assume the value of $\gamma$ [assumed to be in the range 1.3 - 1.5 in @becker11]. Thus, any uncertainty in $\gamma$ will propagate into the uncertainty in the derived value of $T_0$.
In this letter, we use hydrodynamical simulations to understand the specific regions of the Lyman-$\alpha$ forest that contribute to $\bar{\Delta}$ and their significance for measuring the temperature of the IGM using the curvature statistic. As the curvature distribution is non-Gaussian, we explore the results on using the percentiles of the distribution together with the mean. We find that this leads to simultaneous constraints on $T_0$ and $\gamma$, thus breaking the degeneracy introduced by $T(\bar{\Delta})$. We propagate the uncertainties in the curvature measurement to associate a resultant error with the recovered values of $T_0$ and $\gamma$. Throughout this letter, we use the cosmological parameters $\Omega_m =
0.26$, $\Omega_{\Lambda} = 0.74$, $\Omega_b h^2 = 0.024$, $h = 0.72$, $\sigma_8
= 0.85$, and $n_s = 0.95$, which are consistent with the third-year Wilkinson Microwave Anisotropy Probe (WMAP) and Lyman-$\alpha$ forest data [@seljak2006; @viel2006]. The helium fraction by mass is assumed to be 0.24 [@oliveskillman].
IGM simulations and curvature statistics {#sec:simul}
========================================
We perform smoothed-particle hydrodynamical (SPH) simulations using [@gadget2] with $512^3$ each of dark matter and baryonic particles in a periodic box of length $10 h^{-1}$ comoving Mpc, which corresponds to redshift intervals of $\Delta z \sim 0.0092, 0.014$ and 0.02 at redshifts 2, 3 and 4 respectively. The ionization correction is applied using the equation of state to assign temperatures to overdensities at each pixel[^4] and the updated values of background photoionization rates measured by @becker2013 ($ \Gamma_{\rm HI} \times 10^{12} = $ 1.035, 0.789, 0.847 $s^{-1}$ at redshifts 2, 3 and 4 respectively). Simulated spectra (with a spectral sampling of 2.65 km/s per pixel) are generated at each of these redshifts for 100 such randomly drawn lines-of-sight, for 30 different equations of state with $T_0$ ranging from 5000 to 25000 K in steps of 5000 K, and $\gamma$ from 1.1 to 1.6 in steps of 0.1. In each case, the mean and percentiles of the absolute curvature distribution for all the pixels having normalized transmitted fluxes in the range $0.1 \leq$ flux $\leq 0.9$ are computed. [^5]
The distribution of absolute curvature $|\kappa|$ for the equation of state with $T_0$ = 10000 K and $\gamma = 1.3$ at redshift 3 is plotted in Fig. \[fig:curvdist\]. It is clear from the top panel that the curvature distribution is skewed, the mean value is always higher than the median. The mean curvature was shown to follow a tight relationship with the gas temperature at a characteristic overdensity [@becker11; @boera2014]. In the present work, we explore the possibility of using the inferred characteristic overdensities at different percentiles of the curvature distribution to simultaneously constrain $T_0$ and $\gamma$.
At any given $z$, for each of the input models, we find the mean $|\kappa|$ of the simulated spectra. Then, for a given value of $\Delta$, we find $T(\Delta)$ for each model using the assumed $T_0$ and $\gamma$. We plot the values of $T(\Delta)$ versus $\log\langle|\kappa|\rangle$ for all the input models, and fit the relation using a power law fit with the free parameters $A$ and $\alpha$: $$\log \langle|\kappa|\rangle = - \left(\frac{T(\Delta)}{A}\right)^{1/\alpha}
\label{powerfit}$$ We then vary the value of $\Delta$ in , and find the value of $\Delta$ at which the fit (varying $A$ and $\alpha$) leads to the minimum $\chi^2$. The value of $\Delta$ thus obtained is denoted by $\bar{\Delta}_1$, and defined as the “characteristic overdensity” associated with the mean curvature. A similar procedure is applied to find the characteristic overdensities associated with the percentiles $D_1, Q_1, Q_2, Q_3, D_9$ ([the lower (first) decile, lower (first) quartile, median (second quartile), upper (third) quartile and upper (ninth) decile respectively]{}) of the curvature distribution. Thus, at each redshift, the six characteristic overdensities, $\bar{\Delta}_i$, $i = 1$ to 6 for the mean and the five percentiles ($D_1, Q_1, Q_2, Q_3, D_9$) respectively are obtained, with corresponding best-fit parameters $A_i$ and $\alpha_i$. Table \[table:fit\] indicates the values of $A_i$ and $\alpha_i$ for each redshift along with the values of $\bar{\Delta}_i$. The best-fit curves for the mean and the percentiles $D_1, Q_1$ and $D_9$ of the curvature are plotted at redshifts 2 and 3 in Fig. \[fig:red3mm\], which shows that the mean and percentiles of the absolute curvature follow tight relations with $T(\bar{\Delta}_i$) at each redshift.[^6] We notice that the relationship shown in Fig. \[fig:red3mm\] is independent of $\gamma$, even when we use the other percentiles $Q_2, Q_3$ of the curvature.
-0.2in
The scatter in the value of $\Delta$ may be quantified by using the $\chi^2$ of the power law fit. Since we do not have estimates on the systematic errors in $|\kappa|$, we scale the statistical errors in $|\kappa|$ such that the reduced $\chi^2$ at the minimum has the value unity. With this rescaling, we estimate the $1\sigma$ range in the value of the characteristic overdensity, which is found to be $\sigma(\log(\Delta)) \simeq 0.10$. This error is of the same order as that (0.15) estimated by @schaye2000 using the minimization of the scatter in the $\log b_{\Delta} - \log T(\Delta)$ distribution (where $b_{\Delta}$ is the Doppler $b$-parameter corresponding to the density contrast $\Delta$) at the optimal $\Delta$.
$z$ $A_i$ $\alpha_i$ $\bar{\Delta}_i$
----- ------------------------ --------------------- ---------------------
2 \[46.44, 1.10, 6.54, \[5.72, 7.39, 6.60, \[9.30, 6.05, 5.80,
29.07, 83.83, 164.99\] 5.88, 5.41, 5.10\] 6.45, 8.55, 10.45\]
3 \[70.57, 2.97, 14.39, \[5.17, 6.44, 5.78, \[2.65, 1.55, 1.70,
48.51, 90.33, 224.32\] 5.29, 5.17, 4.63\] 2.05, 2.50, 3.15\]
4 \[73.90, 2.37, 12.71, \[5.08,6.67,5.92, \[1.55, 1.25, 1.30,
47.65, 121.77,253.05\] 5.28, 4.79,4.43\] 1.35, 1.40, 1.60\]
: The best fit parameters $A_i$ and $\alpha_i$, $i = 1$ to 6 for the mean and the five percentiles ($D_1, Q_1, Q_2, Q_3, D_9$) respectively, of at three different redshifts. Corresponding values of $\bar{\Delta}_i$ are quoted in the last column. The relative errors on $\bar{\Delta}_i$ are of the order of 0.10.[]{data-label="table:fit"}
Recovery of temperature {#sec:recovery}
=======================
In the present section, we explore the use of the mean and the percentiles of the curvature distribution to recover the parameters of the IGM effective equation of state. The temperature-density relation may be reasonably well-approximated by a power law. We note that the mean and the percentiles of the curvature distribution are typically sensitive to different characteristic overdensities $\bar{\Delta}_i$. In addition, the mean and percentiles are related to specific temperatures $T_i \equiv T(\bar{\Delta}_i)$, through calibration curves analogous to Fig. \[fig:red3mm\]. Hence, they provide [complementary]{} probes of the same underlying temperature-density relation. We may, therefore, conjecture using them together to recover both the parameters $T_0$ and $\gamma$ of the equation of state.
The characteristic overdensities $\bar{\Delta}_i$ [for the mean and the $D_1, Q_1, Q_2, Q_3, D_9$ percentiles]{} of the curvature are provided in Table \[table:fit\]. From the curvature distribution, we can recover the temperatures $T_i$ associated with the mean and these percentiles. A power law fit to the recovered temperatures $T_i$ and $\bar{\Delta}_i$ then provides the best-fitting values of $T_0$ and $\gamma$. These can then be compared to the original model.
Applying this procedure at redshift 3 to the spectra drawn from the simulation with $T_0 = 10000 $ K, $\gamma = 1.3$, we recover the best-fit parameters $T_0 = 10399$ K and $\gamma = 1.325$ from a power-law fit to all the percentiles. Hence, the model parameters are recovered to an accuracy of 2-4%. This is however an ideal case, since the effects of noise and scatter in the temperature-density relation have to be folded into the analysis in the case of realistic spectra, which we describe in the following section.
Errors in the recovered parameters {#sec:errors}
==================================
We have seen that the parameters $T_0$ and $\gamma$ of the IGM effective equation of state can be simultaneously recovered by using the mean and percentiles of the curvature distribution. We now explore the effects of addition of noise and incorporating the deviation from the power-law equation of state.
We begin by estimating the errors in the recovery of $T_0$ due to the uncertainty in the $\bar{\Delta}$ measurement. We propagate the errors on $\bar{\Delta}$ to the temperature measurement by using the relation [@schaye2000]: $$\sigma^2(\ln T_0) = \sigma^2(\ln T(\Delta)) + \sigma^2(\gamma)[\ln(\Delta)]^2 + (\gamma - 1)^2\sigma^2[\ln(\Delta)]$$ with $$\sigma^2(\ln T(\Delta)) = \sigma^2(|\kappa|) + \sigma^2_{\rm fit}$$ where $\sigma^2_{\rm fit}$ is the scatter of the data points around the fit. We find $\sigma (|\kappa|) \sim 0.025$,[^7] and with the conservative estimate of $\sigma_{\rm fit} \sim 0.15$, we obtain $\sigma(\ln T(\Delta)) \sim 0.16$. Since $\gamma$ is not known a priori in the realistic case, the error in $\gamma$ is estimated from the recovered value, and this uncertainty in $\gamma$, in turn, propagates to the uncertainty in the recovered value of $T_0$.
It is known [Paper I, @becker11; @boera2014] that the addition of noise significantly influences the value of the curvature statistic. In @becker11, the noisy spectra are smoothed by using a $b$-spline fitted to the data. In this work, we add noise to have a continuum signal-to-noise ratio of 50 (a typical SNR of e.g. the UVES/HIRES spectra used in IGM studies) and smooth the noisy spectra with a Gaussian filter of width 12 km/s, using a procedure analogous to that described in Paper I.
The temperature of the gas closely follows a power-law equation of state for overdensities $\Delta \leq 10$ [@huignedin; @puchwein2014] and flattens above $\Delta \sim 10$ [@mcdonald2001]. To take this into account, we simulate noise-added spectra with 10% Gaussian scatter in the $T-\Delta$ relation, where the equation of state is assumed upto $\Delta = 10$, with a subsequent flattening for $\Delta \gtrsim 10$.[^8]
Figure \[fig:scatter\] shows the joint distribution (at redshift 3) of the relative errors in recovery of $T_0$ and $\gamma$ for the noise added spectra incorporating the deviation from the power-law equation of state. The relative error in the recovery is computed using a best-fitting power law from the mean and all the percentiles of the curvature distribution. The different symbols and sizes indicate the different values of the input parameters $\gamma$ and $T_0$. The blue histograms on the top and right panels indicate the relative errors in recovery for the entire sample. We note that in most realistic scenarios of reionization, the $(T_0,\gamma)$ pairs of (5000 K, 1.1), (5000 K, 1.2), (20000 K, 1.6) and (25000 K, 1.6) are disfavoured, since they correspond to very low or high values of both $T_0$ and $\gamma$. Hence, we plot the recovery for the remaining subset of models as the transparent red histograms, which denote realistic uncertainties for the more physically motivated range in the IGM equations of state.
Discussion
==========
We have studied the mean and the percentiles of the curvature distribution of the Lyman-$\alpha$ forest, and their associated characteristic overdensities as probes of the thermal state of the intergalactic medium.
The characteristic overdensities $\bar{\Delta}_i$ we derive here may be related to the optimal overdensities in the calibration of the $b-T$ relation in @schaye1999 [@schaye2000; @bolton2014]. It was found that at the optimal overdensity, the $b_{\Delta} - T(\Delta)$ relation becomes independent of $\gamma$, even though e.g. the $b_{\Delta} - T_{0}$ relation is dependent on $\gamma$. In practice, the value of $\Delta$ is varied and the value corresponding to the minimal scatter in the $b-T$ relation is chosen as the optimal overdensity. Unlike the $b-T$ relation, the $\kappa-T$ relation is nonlinear. However the same trends are found to be valid, i.e. the $\kappa-T$ relation is independent of $\gamma$ even though the $\kappa-T_0$ relation is $\gamma$-dependent [@becker11].
In @schaye2001, it is argued that the characteristic size of the absorbers in the Lyman-$\alpha$ forest is of the order of the Jeans’ scale. This implies the relation between the neutral hydrogen column density, $N_{\rm HI}$ and the characteristic overdensity $\Delta$ of the absorbers at various redshifts: $N_{\rm HI} \propto \Delta^\alpha \Gamma^{-1} (1+z)^{4.5}$ where $\Gamma$ is the hydrogen photoionization rate. The value of $\alpha$ has been found to be of order 1.37 - 1.43 from the results of numerical simulations [@dave1999; @schaye2001]. For a fixed column density, the characteristic overdensity is hence expected to scale with redshift as $\Delta \propto (1+z)^{-4.5/\alpha} \Gamma^{1/\alpha}$. We find that the characteristic overdensities roughly follow this scaling (plotted for $\bar{\Delta}_1$ and $\bar{\Delta}_4$ in Fig. \[fig:chardens\]), assuming the evolution of $\Gamma$ according to @becker2013.[^9] Hence, we conclude that these characteristic overdensities trace particular column densities in the Lyman-$\alpha$ forest, which are $N_{\rm HI} \sim 1.6 \times 10^{14}$ cm$^{-2}$ for $\bar{\Delta}_1$ and $9.1 \times 10^{13}$ cm$^{-2}$ for $\bar{\Delta}_4$ (using the relation between the characteristic overdensity and column density). These column densities (as well as those associated with the other percentiles) are in the range $10^{13.8} - 10^{14.3}$ cm$^{-2}$. This is within the range $10^{12.5} - 10^{14.5}$ cm$^{-2}$ where it is found [@schaye1999] that the $b-T$ relation shows minimal scatter.
We have shown that the mean and percentiles of the curvature act as [complementary]{} probes of the same underlying $T-\Delta$ relation since they are sensitive to different regions of the Lyman-$\alpha$ forest. Together, they may be used to recover the parameters $T_0$ and $\gamma$ of the IGM effective equation of state. The errors on the recovery of $T_0$ arising from the combination of the scatter in $\bar{\Delta}$, the addition of noise to the spectra and the modifications to the power-law equation of state for realistic models are of the order of $\lesssim 30$%.
The above procedure may be applied to the observations of quasar spectra to possibly place interesting constraints on the parameters $T_0$ and $\gamma$ at various redshifts. With the noise as well as the flattening of the equation of state, the final constraints we arrive at are comparable to previous works (e.g. @ricotti2000a), but one advantage of using the curvature is that Voigt profile line decomposition is not required. The present sample size is consistent with the samples in observations of quasar spectra, e.g. 100-300 sightlines in each redshift interval available from 61 QSO spectra in @becker11. It is to be noted that increasing the sample size would also considerably reduce the errors. In the realistic case, calibration errors, continuum fitting errors, the impact of Jeans smoothing and the possibility of larger scatter in the equation of state at low densities have to be folded into the analysis, which we plan to carry out in future work.
Acknowledgements {#acknowledgements .unnumbered}
================
The research of HP is supported by the SPM research grant of the Council for Scientific and Industrial Research (CSIR), India. The hydrodynamical simulations were performed using the Perseus cluster of the IUCAA High Performance Computing Centre. We thank Patrick Petitjean for useful discussions, and the anonymous referee for helpful comments that improved the content and presentation.
===
[^1]: Electronic address: [email protected]
[^2]: Electronic address: [email protected]
[^3]: Electronic address: [email protected]
[^4]: This is a simplification at high overdensities, however, we return to the issue of deviation from the power law equation of state at high densities in Sec. \[sec:errors\].
[^5]: To start with, we use simulated spectra without noise and without applying extra smoothing to the spectrum for exploration of the different aspects, however we include the effects of noise in the spectra later in Sec. \[sec:errors\].
[^6]: A slight flattening of the power law fit at very low $T(\Delta) (< 10000 $ K) arises due to a larger fraction of lines being near the saturated regime of the curve of growth. This decreases the sensitivity of the measured curvature to changes in $\gamma$.
[^7]: This is the magnitude of the error in log $|\kappa|$ for which the minimum reduced $\chi^2$ has value unity.
[^8]: The flattening of the temperature-density relation for $\Delta \geq 10$ has also been used, for example, in the semi-analytical modelling of the Lyman-$\alpha$ forest power spectrum in @greig2014.
[^9]: At redshift 4, there is a deviation in case of the median, however, note that this range in $\alpha$ is strictly valid only across redshifts 0-3 [@schaye2001].
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We investigate the problem of growing clusters, which is modeled by two dimensional disks and three dimensional droplets. In this model we place a number of seeds on random locations on a lattice with an initial occupation probability, $p$. The seeds simultaneously grow with a constant velocity to form clusters. When two or more clusters eventually touch each other they immediately stop their growth. The probability that such a system will result in a percolating cluster depends on the density of the initially distributed seeds and the dimensionality of the system. For very low initial values of $p$ we find a power law behavior for several properties that we investigate, namely for the size of the largest and second largest cluster, for the probability for a site to belong to the finally formed spanning cluster, and for the mean radius of the finally formed droplets. We report the values of the corresponding scaling exponents. Finally, we show that for very low initial concentration of seeds the final coverage takes a constant value which depends on the system dimensionality.'
author:
- 'N. Tsakiris'
- 'M. Maragakis'
- 'K. Kosmidis'
- 'P. Argyrakis'
title: Percolation of randomly distributed growing clusters
---
Introduction
============
Percolation theory has drawn a continuous interest from the scientific community for several years [@Shante]-[@Newman]. It has been studied in a wide variety of systems ranging from lattices [@Odor] to complex networks [@Newman]. The fields that percolation applies are as diverse as electromagnetism [@Bergqvist]-[@Hu04], chemistry [@Groot], materials science [@Markworth], geology [@McKenzie], social systems [@Palla], wireless networks [@Franceschetti] and many more. In chemistry and materials science it is of major importance for the movement of liquids or gases in porous media. Problems in this area relate to the leakage in seals [@Bottiglione] and the gas permeability in cement paste [@Galle].
Various algorithms have been used to simulate the phase transformation kinetics. In many pattern formation models several small spherical seeds are nucleated at a constant rate (homogeneous nucleation). Seeds can also initiate on defects in the case of heterogeneous nucleation. From the simulation point of view the defects are considered as points in the lattice representing the seeds. The seeds once formed are in a metastable phase and grow at a constant velocity as long as there is adequate available material for adsorption.
Additionally, several models exist that do not allow the adsorption of a new particle in contact with or overlapping with an already adsorbed one. An example is the random sequential adsorption (RSA) model [@Evans]. This model has been extensively used for colloid and globular protein adsorption in heterogeneous surfaces. In such systems discrete lattice sites can act as adsorption sites with attractive short range interactions [@Adamczykrev]. The jamming coverage and the structure of the particle monolayer as a function of the site coverage and the particle/site size ratio have been studied.
Models studying pattern formation ranging in between these two cases have not been used extensively. Andrienko proposed [@Andrienko] the idea of disks and droplets growing at a constant rate on random initial sites over the lattice and stopping once they come in contact. In the so called “Touch and Stop” model the droplets grow at a constant rate in all directions (circular in 2D, spherical in 3D).
The main characteristic of this model is located in the notion that the droplets stop growing after two or more of them come in contact. This can be due to several reasons. In material science it is possible to have a strong surface tension that inhibits the nuclei from taking any shape other than that of a circular or spherical one. Additionally, a significant interacting force between the substrate and the forming droplet can prevent two or more discs from coalescing in the time scale needed for the growth of other islands.
This problem also relates to the well studied Apollonian packing problem [@Kasner] for circles and spheres. In fact it can be considered as a random version of packing with various discrete sizes, where growth velocity is constant but not infinitely large. The Touch and Stop model has also been studied in some variations (random insertion of seeds in time) as a packing limited growth problem [@Dodds].
Model description
=================
The system used can be described as follows. Initially, lattices of $10^6$ sites ($1000\times 1000$ for 2D, and $100\times 100\times 100$ for 3D) are randomly populated with seeds of singular size in a non overlapping way. The initial occupation probability of these sites is $p$. At every time step all seeds are investigated once as to the possibility of growing in size instantaneously in all neighboring sites. Investigation sequence is random in order. Each seed is allowed to grow its periphery by one layer (increase the radius by one) provided that there is no overlapping with other growing seeds. Thus each seed becomes an evolving cluster. As soon as two, or more, clusters touch each other, the growth of all of the adjoined clusters stop. The touching disks or droplets can be considered as belonging to the same stable cluster, a cluster that no longer grows over time unless other evolving clusters happen to touch it. Periodic boundary conditions are applied in both the 2D and 3D systems. In order to ensure a smaller statistical error a large number of runs was used (1000 individual configurations).
The system continues to evolve until no other cluster can grow in time, so that all of them have at least one adjoining cluster. At this point the final occupation probability of each site is generally larger than $p$ and the system is investigated as to whether it has a spanning cluster using the classic Hoshen-Kopelman algorithm [@HoshKop]. The final shape of each disk or droplet of this system is not circular or spherical, in 2D and 3D. It is square and cubic, respectively. Therefore, it is possible that two evolving clusters have more than just one adjoining sites. In order to have only one adjoining site they must touch at their tips. When two such clusters touch in a part of one of their edges (2D) or facets (3D), the resulting stable cluster formed has two clusters which are connected in more than just one site. Snapshots of a typical finite 2D system evolution can be seen in Fig. \[fig\_snaps\], and a larger snapshot of the final state of a system with very low concentration of initial seeds is given in Fig. \[fig\_snaps\_larger\]. It is obvious in Fig. \[fig\_snaps\_larger\] that several clusters are connected in two or more sites.
![Snapshots of the evolution of a 2D system. Each snapshot corresponds to an advance of 1 time unit. The first snapshot is the initial system and the last is the final system configuration, where no percolating cluster is formed. Different colors signify different evolving clusters. At snapshot 4 the evolution of the blue and red droplets stops since they touch and the blue and red droplet now form one stable cluster. All other clusters stop growing at snapshot 5. Periodic boundary conditions are used.[]{data-label="fig_snaps"}](Fig_1.eps){width="1.\textwidth"}
![Snapshot of the final state of a 2D $500\times500$ system with very low concentration of seeds (p=0.001). Each stable cluster has a different color and the largest is shown in black.[]{data-label="fig_snaps_larger"}](Fig_2.eps){width="1.\textwidth"}
Discussion
==========
Starting from an empty lattice, we vary $p$ in the entire domain $0<p<1$, allow the system to evolve, and we monitor the size of the largest, ${S_1}$, and second largest, ${S_2}$, clusters formed in the final stable configuration (Fig. \[fig\_largest2D\] and Fig. \[fig\_largest3D\]). In contrast to the classical percolation model where small concentrations of seeds translates to either isolated sites or very small formed clusters, our model exhibits at first high values for the normalized sizes of the two largest clusters. Initially, the values of ${S_1}$ and ${S_2}$ are quite large, although quickly a sudden drop occurs. This can be explained by the very small number of initial seeds in our system which means that they are randomly placed far apart from one another. This enables them to grow without touching each other for many time steps. Therefore, the largest clusters end up with a high final size value.
![Logarithmic representation of the normalized mean largest and second largest cluster sizes, ${S_1}$ (squares) and ${S_2}$ (circles) respectively, over the initial concentration of seeds, $p$, for the 2D case. Inset shows a linear representation of the same where the sharp drop for low initial concentrations and a second transition near $p_c$ are obvious.[]{data-label="fig_largest2D"}](Fig_3.eps){width="1.\textwidth"}
In fact, if we simply have one initial seed, then that one also forms the spanning cluster. This is obviously a finite size effect and is due to the periodic boundary conditions applied. This will cause the growing droplet to eventually touch itself on two opposite vertices of the evolving cluster.
![Logarithmic representation of the normalized mean largest and second largest cluster sizes, ${S_1}$ (squares) and ${S_2}$ (circles) respectively, over the initial concentration of seeds, $p$, (same as above) for the 3D case. Inset shows a linear representation of the same where the sharp drop for low initial concentrations and a second transition near $p_c$ are obvious.[]{data-label="fig_largest3D"}](Fig_4.eps){width="1.\textwidth"}
As the initial concentration increases, and until it reaches some system specific value (i.e. $p=0.095\pm 0.003$ for the 2D system), ${S_1}$ and ${S_2}$ reduce. The decrease of the final cluster sizes follows a power law over the initial concentration for a major part of this region ($10^{-5}\leq p\leq5\times10^{-2}$ for the 2D, and $10^{-5}\leq p\leq5\times10^{-3}$ for the 3D) with the exponents having a value of $0.77 \pm 0.02$ and $0.65 \pm 0.02$, respectively. This decrease is caused by the fact that for very small concentrations it is possible to have only several, but large in size, clusters being formed. As the initial concentration is still low but increases, the droplets will now touch each other much faster and the resulting largest clusters will be smaller in size. It is worth mentioning that the application of periodic boundary conditions plays an important role in the calculation of the actual values of the exponents and critical thresholds, especially in the very low and low concentrations of initial seeds. In larger system sizes this effect is reduced.
By further increasing the number of initial seeds and when approaching the half coverage region, many droplets will touch each other at the first few steps, or more commonly be randomly generated at neighboring sites, and therefore not grow at all. The size of the two largest clusters will also start to increase again. As in normal percolation, once the initial concentration reaches a critical value (percolation threshold), $p_c$, all smaller clusters quickly join the one giant cluster formed. At this point the second largest cluster of the system will start to decrease rapidly in size (Fig. \[fig\_largest2D\] and Fig. \[fig\_largest3D\]). The initial concentration where this decrease starts is $p=0.497 \pm 0.002$ for the 2D, and $p=0.040 \pm 0.002$ for the 3D case. All major clusters from this point on will touch and form one spanning cluster (the giant component) consisting of many adjoined ones.
In agreement to the size of the largest clusters, the probability for a spanning cluster to occur, ${P_{span}}$, over the initial concentration exhibits two phase transitions (Fig. \[fig\_prob\_spanning\]). The first ones occur at very low initial concentration, while the second ones occur at a value of ${p\approx0.497}$ and ${p\approx 0.040}$ for the 2D and 3D respectively. Similarly to the largest clusters behavior, ${P_{span}}$ decreases with the increase of the density and reaches a minimum value (for the 2D when $p\approx10^{-3}$ it is $P_{span}\approx10^{-3}$). ${P_{span}}$ also shows a power law behavior for a significant region of low concentration ($10^{-5}\leq p\leq10^{-2}$ for the 2D) with the calculated exponents being approximately $0.77 \pm 0.02$ for the 2D and $0.64 \pm 0.02$ for the 3D.
![Probability of a single occupied site to belong to the spanning cluster over the density of initial seeds shown in a logarithmic representation with the power law area in low values of initial concentration. Squares and circles are for the 2D and 3D case respectively. Inset shows a linear representation where the two phase transitions are clearly shown for the 2D case and 3D cases.[]{data-label="fig_prob_spanning"}](Fig_5.eps){width="1.\textwidth"}
We performed simulations in various system sizes from $200\times200$ to $1000\times1000$.The shape of the curves observed in Figs. \[fig\_largest2D\] and \[fig\_prob\_spanning\] is found to be independent of the lattice size with a minimum found on the same value of initial seed concentration. The minimum value of the quantities shown in Figs. \[fig\_largest2D\] and \[fig\_prob\_spanning\], exhibits a power law dependence on the size (see Fig. \[fig\_min\_values\]). The exponent of this power is in both cases $1.65 \pm 0.02$.
![Minimum values of ${P_{span}}$ and ${S_1}$ over the size of the system in the 2D case. Squares are values of ${P_{span}}$ and circles are that of ${S_1}$. Lines are best fit.[]{data-label="fig_min_values"}](Fig_6.eps){width="1.\textwidth"}
As mentioned, cluster evolution stops at various sizes depending on the distance from their neighboring clusters. We calculate the mean radii of the finally formed disks or droplets, $\langle R\rangle$, vs. the initial concentration (Fig. \[fig\_aver\_size\]). We calculate only the increase of the clusters radius and do not include its initial site. For the region of low initial concentration the logarithmic plot reveals a power law behavior of the mean radii which decreases with an exponent $0.52 \pm 0.02$ and $0.37 \pm 0.02$ for the 2D and 3D respectively.
The exact values of $\langle R\rangle$ for very low concentrations relate to the actual system size dimensions. In the 2D case, the maximum value is $\langle R\rangle=500$ and is found when we have initially only one seed. The corresponding value for our 3D system is $\langle R\rangle=50$. As expected from the application of periodic boundary conditions, it is equal to half the lattices linear dimension. As the concentration increases the mean radius decreases and reaches a value of 1 for ${p\approx0.053}$ and ${p\approx 0.020}$ for the 2D and 3D respectively. For higher concentrations many of the randomly placed initial seeds touch other neighboring clusters at their first time step. Therefore, the neighboring clusters do not grow after this and the average value of the mean radius becomes lower than 1.
![Mean radii of the evolving clusters once the system stops. Squares and circles are for the 2D and 3D case respectively. Note that for ${p\geq 0.053}$ for the 2D and ${p\geq 0.020}$ for the 3D case, the value of $\langle R\rangle$ is less than one. This is due to the fact that we measure only the increase of the disks radius and the initial seed is not included.[]{data-label="fig_aver_size"}](Fig_7.eps){width="1.\textwidth"}
It is also interesting to calculate the ratio of the final lattice coverage to the initial concentration of seeds ${p_{fin}/p}$ in such a system, after all the clusters have stopped expanding. Fig. \[fig\_real\_occup\_2D\] shows that the ratio versus the initial concentration exhibits a power-law behavior area in a logarithmic plot where the exponent is calculated to be approximately 1. This means that there is a linear relation between the ratio and the inverse of the initial concentration of seeds. Therefore, we expect a constant final coverage value of the lattice over the initial concentration. Indeed, the inset shows that the final coverage remains constant for both the 2D and the 3D case for very low initial concentration (${p_{fin}=0.35 \pm 0.01}$ for the 2D and ${p_{fin}=0.19 \pm 0.01}$ for the 3D case). These values are found to be independent of the system size.
Overall, the 2D and 3D systems differ qualitatively one with another in the minimum coverage of initial seeds needed for the spanning cluster to occur. In classic percolation, the thresholds of the corresponding 2D and 3D systems studied here are 0.593 and 0.312 respectively. In this model, the 2D thresholds differ much more from the 3D ones, since the 3D threshold is only 0.040, whilst the 2D is 0.497. A 3D system reaches the threshold with a low coverage value. Additionally, the final concentration of the 3D case is much smaller than that of the 2D for very low initial coverages, leaving a large portion of the lattice unoccupied and ending up with a very porous structure.
![The ratio of the final to the initial concentration versus the initial concentration for both the 2D and 3D case. Squares and circles are for the 2D and 3D case respectively. The inset shows that the final concentration remains constant for very low initial concentration for both cases.[]{data-label="fig_real_occup_2D"}](Fig_8.eps){width="1.\textwidth"}
Conclusion
==========
We have applied and studied a model of randomly distributed growing clusters from the percolation point of view. The results produced show that there is a region of very low concentration of initially placed seeds where many properties of the system exhibit a power-law behavior. The size of the two largest clusters, the probability to form a spanning cluster and the mean final radii of the evolving clusters all show a power low behavior for approximately 4 orders of magnitude. The minimum values of the size of the largest cluster and the probability to form a spanning cluster also show a power law dependence on the system size. The smaller the system size the higher the probability to have a spanning cluster. Infinite systems should not have spanning clusters formed at low densities. Additionally, we have found that the final coverage in the model for very low initial concentrations is constant and has a value of $0.35$ for the 2D case and $0.19$ for the 3D.
The accurate determination of the critical percolation threshold of this system requires a more extensive study which is beyond the scope of this manuscript. In [@Second_paper], we calculate the precise values of the percolation threshold and the critical exponents of this transition, and investigate the universality class that this model belongs to.
This model of randomly distributed growing clusters would be valid for the explanation of properties in problems in material science and technological applications in the cement industry. Variations of this model can also potentially be applied to social networks. We are currently investigating the application of this model to other complex systems. Civilizations can expand until they come in contact with each other at which point they may stop growing for some time because they allocate all available resources to this interaction, i.e. in a war. Such a study would require the application of this model in complex geometries (fractals) and networks. In specific fields, commercial companies can be considered to grow at nearly constant rate until they decide to merge with each other. At such a point their growth is inhibited due to the dedication of large amounts of resources and manpower in the merging process. In both cases the time needed for the growth process can make this circumstantial halting seem as a growth stop when compared to the systems evolutionary speed.
Acknowledgment
==============
This work was partially supported by the FP6 Project, INTERCONY NMP4-CT-2006-033200.
[99]{}
V. K. S. Shante, and S. Kirkpatrick, Advances in Physics, [**20**]{}, 325 (1971). D. Stauffer, Phys. Rep., [**54**]{}, 1 (1979); D. Stauffer, and A. Aharony, [*Introduction to Percolation Theory*]{}, 2nd Ed. (Taylor & Francis, London, 1994). M. E. J. Newman, SIAM Rev. [**45**]{}, 167 (2003). G. Ódor, Rev. Mod. Phys., [**76**]{}, 663 (2004). L. Bergqvist, O. Eriksson, J. Kudrnovský, V. Drchal, P. Korzhavyi, and I. Turek, Phys. Rev. Lett., [**93**]{}, 137202 (2004). L. Hu, D. S. Hecht, and G. Grüner, Nano Lett., [**4**]{} (12), 2513 (2004). R. D. Groot and T. J. Madden, J. Chem. Phys., [**108**]{}, 8713 (1998). A.J. Markworth, K.S. Ramesh, and W.P. Parks, J. Mater. Sci., [**30**]{}, 2183 (1995). D. McKenzie and R.K. O’Nions, J. Pet., [**32**]{}, 1021 (1991). G. Palla, A.-L. Barabási and T. Vicsek, Nature, [**446**]{}, 664 (2007). M. Franceschetti, O. Dousse, D. Tse, P. Thiran, IEEE Trans. on Information theory, [**53**]{} (3), 1009 (2007). F. Bottiglione, G. Carbone, G. Mantriota, Tribology International, [**42**]{}, 731 (2009). C. Gallé and J.-F. Daian, Mag. Concr. Res., [**52**]{}, 251 (2000). J. W. Evans, Rev. Mod. Phys. [**65**]{}, 1281 (1993). Z. Adamczyk, K. Jaszczólt, A. Michna, B. Siwek, L. Szyk-Warszyńska and M. Zembala, Adv. Colloid Interface Sci., [**118**]{}, 25 (2005). Y. A. Andrienko, N. V. Brilliantov, and P. L. Krapivsky, J. Stat. Phys., [**75**]{} (3-4), 507 (1994); N.V. Brilliantov, P.L. Krapivsky, and Y.A. Andrienkov, J. Phys. A, [27]{}, L381 (1994). E. Kasner and F. Supnick, Proc. Natl. Acad. Sci. USA., [**29**]{} (11), 378 (1943). P.S. Dodds, and J.S. Weitz, Phys. Rev. E, [**65**]{}, 056108 (2002). J. Hoshen and R. Kopelman, Phys. Rev. B, [**14**]{}, 3428 (1976). N. Tsakiris, M. Maragakis, K. Kosmidis and P.Argyrakis, arXiv:1004.5028v1 (2010).
0 cm
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'New two dimensional systems like surface of topological insulator and graphene offer a possibility to experimentally investigate situations considered “exotic” just a decade ago. One of those is the quantum phase transition of the “chiral” type in electronic systems with relativistic spectrum. Phonon mediated (“conventional”) pairing in the Dirac semimetal appearing on the surface of topological insulator leads to transition into a chiral superconducting state, while exciton condensation in these gapless systems has been envisioned long time ago in the physics of the narrow band semiconductors. Starting from the microscopic Dirac Hamiltonian with local attraction or repulsion, the BCS type gaussian approximation is developed in the framework of functional integrals. It is shown that due to an “ultra-relativistic” dispersion relation there is a quantum critical point governing the zero temperature transition to a superconducting or the exciton condensed state. The quantum transitions that have critical exponents very different from the conventional ones. They belong to the chiral universality class. We discuss the application of these results to recent experiments in which surface superconductivity was found in topological insulators and estimate feasibility of the phonon pairing.'
author:
- Dingping Li
- Baruch Rosenstein
- 'B.Ya. Shapiro'
- 'I. Shapiro'
title: 'Chiral universality class of the normal-superconducting and the exciton condensation transition on the surface of topological insulator '
---
Introduction
============
Topological insulator (TI) is a novel state of matter in materials with strong spin - orbit interactions that create topologically protected surface states [@Zhang]. The electrons (holes) in these states have a linear dispersion relation, see Fig. 1, and can be described approximately by a (pseudo) relativistic two dimensional (2D) Hamiltonian. The system realizes an “ultra-relativistic” 2D electron or hole conducting liquid along with much better studied graphene[@Katsnelson], a 2D one layers sheet of carbon atoms that became a paradigms example of the Dirac semi-metal. In the context of graphene certain quantum phase transitions were theoretically contemplated. The superconductivity in graphene has been repeatedly considered[@supergraphene], however despite great experimental efforts was never achieved. The 2D dimensional phonons seem to be unable to overcome strong Coulomb repulsion in order to create a Cooper pair. The same can be said about attempts to achieve exciton condensate in graphene that was proposed[@Khveshchenko; @Gamayun] even before its discovery. Apparently the repulsion is not strong enough either to create stable electron - hole bound states[@graphenechiral]. The surface of topological insulator therefore became a prime candidate to realize the quantum transitions.
It is known for a long time that similar 2D and quasi-2D metallic systems like the surface metal on twin planes [@Shapiro], layered materials (strongly anisotropic high $T_{c}$ cuprates[@Wen] or organic superconductors[@organic]) may develop 2D (surface) superconductivity. This phenomenon became known as “localized superconductivity”[@Buzdin]. Since best studied TIs possess a quite standard phonon spectrum [phononexp]{}, it was predicted recently [@DasSarma; @Li14] that they become superconducting. The predicted critical temperature of order of $1K$ is rather low (despite a fortunate suppression of the Coulomb repulsion due to a large dielectric constant $\varepsilon \sim 50$), the nature of the “normal” state (so-called 2D Weyl semi-metal) might make the superconducting properties of the system unusual. The ultra-relativistic nature manifests itself mostly when the Weyl cone is very close to the Fermi surface. Especially interesting is the case (that actually was originally predicted for the \[111\] surface of $Bi_{2}Te_{3}$ and $Bi_{2}Se_{3}$[@Zhang1]) when the chemical potential coincides with the Dirac point. Although subsequent ARPES experiments[@Zhang] show the location of the cone of surface states order tenths of $eV$ off the Fermi surface; there are experimental means to shift the chemical potential, for example by the bias voltage [@bias].
Unlike the more customary poor 2D metals with several small pockets of electrons/holes on the Fermi surface (in semiconductor systems or even some high $T_{c}$ materials[@Wen]), the electron gas TI has two peculiarities especially important when pairing is contemplated. The first is the bipolar nature of the Dirac spectrum: there is no energy gap between the upper and lower cones. The second is that the spin degree of freedom is a major player in the quasiparticle dynamics. This degree of freedom determines the pairing channel. The pairing channel problem was studied theoretically on the level of the Bogoliubov-deGennes equation [@Herbut]. Both $s$-wave and $p$-wave are possible and compete due to the breaking of the bulk inversion symmetry by the surface. Various pairing interactions were considered to calculate the DOS measured in $Cu_{x}Bi_{2}Se_{3}$ using self-consistent analysis [@Sato]. As mentioned above the most intriguing case is that of the small chemical potential that has not been addressed microscopically. It turns out that it is governed by[ ]{}a quantum critical point (QCP)[@Sachdev].
The concept of QCP at zero temperature and varying doping constitutes a very useful language for describing the microscopic origin of superconductivity in high $T_{c}$ cuprates and other “unconventional” superconductors[@Wen]. Superconducting transitions generally belong to the $U\left( 1\right) $ class of second order phase transitions[@Herbutbook], however it was pointed out a long time ago[@Rosenstein] that, if the normal state dispersion relation is “ultra-relativistic”, the transition at zero temperature as function of parameters like the pairing interaction strength is qualitatively distinct and belongs to chiral universality classes classified in ref. [@Gat]. The term “chiral” appears following the corresponding discussion of the well studied both theoretically and experimentally chiral symmetry breaking transition in Quantum Chromodynamics. Attempts to experimentally identify second order transitions governed by QCP in condensed matter included quantum magnets [@Sachdev], superconductor - insulator transitions[@SCinsulator] and more recently exciton condensate in graphene[@Katsnelson; @CastroNeto] and other Dirac semi - metals including TI. In the last two cases the broken symmetry is also often termed “chiral”.
Exciton condensation is a very old concept in low dimensional narrow gap semiconductor physics. The best-studied exciton condensate is the quantum Hall bilayer at half-filled Landau levels[@QHE]. Here ingenious methods had been developed to separately contact the two layers so that one can directly probe the order parameter via counterflow superfluidity along the layers and tunneling between the layers. The same idea was extended to bilayer graphene and recently to TI[@excitonTI].
In addition Dirac semimetal was realized in cold atom system[@cold] (following the realization in 2D known as the “synthetic graphene”). Interestingly the sign and strength of the interaction can be controlled. The Dirac semimetal in optically trapped cold atom systems [@cold] is well suited to study this fascinating phenomenon. The Dirac semimetal in optically trapped cold atoms[@cold] offers a well controllable system in which this phenomenon occurs both for repulsive interaction (chiral symmetry breaking) and the attractive one (superconductivity).
In this paper the quantum phase transitions in Dirac semimetal due to local interactions both attractive (superconductivity) and repulsive (exciton condensation) are studied with emphasis on their distinct criticality. The critical exponents belong to chiral universality classes that are identified.
In Section II the general framework that allows to study the surface superconductivity and exciton condensation in general Dirac semi-metal (TI, not necessarily time reversal and reflection invariant or some other of the numerous systems being identified recently) with a general local interaction is presented. The local coupling strength $g$, chemical potential $\mu $ and temperature $T$ will be kept general ($g$ is negative for repulsion leading to the exciton condensation or positive leading to superconductivity). Since the symmetry analysis is crucial, we first discuss the space and spin rotations. In Section III we concentrate on the simplest time reversal and reflection invariant Dirac model and identify its spontaneous symmetry breaking patterns. In Section IV the phase diagram for $g>0$ is obtained for arbitrary temperature $T$ and chemical potential much smaller than the Debye energy $T_{D}$. The latter condition is the main difference from the conventional BCS model in which $\mu >>T_{D}$. A quantum critical point at $%
T=\mu =0$ when the coupling strength $g$ reaches a critical value $g_{c}$ dependent on the cutoff parameter $T_{D}$. We concentrate on properties of the superconducting state in a part of the phase diagram that is dominated by the QCP. Various critical exponents are obtained. In particular, the coupling strength dependence of the coherence length is $\xi \propto \left(
g-g_{c}\right) ^{-\nu }$ with $\nu =1$ , the order parameter scales as $%
\Delta \propto \left( g-g_{c}\right) ^{\beta }$, $\beta =1$. For the repulsion similar transition occurs in the exciton channel in Section V. The critical exponents beyond mean field and experimental feasibility of superconductivity are discussed in Section VI.
Generalized mean field approximation for local four - Fermi interactions
========================================================================
Hamiltonian and the partition function
--------------------------------------
We consider the second quantized electron Hamiltonian via four-Fermi local coupling of strength $g$ $$\begin{gathered}
H=\int d^{2}r\psi _{\alpha }^{\dagger }\left( \mathbf{r}\right) K_{\alpha
\beta }\psi _{\beta }\left( \mathbf{r}\right) \mathbf{-}\frac{g}{2}\psi
_{\alpha }^{\dagger }\left( \mathbf{r}\right) \psi _{\beta }^{\dagger
}\left( \mathbf{r}\right) \psi _{\beta }\left( \mathbf{r}\right) \psi
_{\alpha }\left( \mathbf{r}\right) \text{;} \notag \\
K_{\alpha \beta }\left( \mathbf{\nabla }\right) =\mathcal{E}_{\alpha \beta
}\left( \mathbf{\nabla }\right) -\mu \delta _{\alpha \beta }\text{,}
\label{Hamiltonian}\end{gathered}$$where space is two dimensional, $\mathbf{r=}\left\{ x,y\right\} $ and $\mu $ is the chemical potential. The precise definition of the relevant “single” electronic excitations $\mathcal{E}_{\alpha \beta }\left( \mathbf{\nabla }%
\right) $ will be dependent on the specific model considered and is specified below. The index $\alpha $ of the spinors $\psi $ refers to valley/spin degrees of freedom. The partition function is
$$Z=\mathrm{Tr}\text{ }e^{-H/T}=\int D\psi ^{+}D\psi e^{-S\left[ \psi _{\alpha
}^{+}\mathbf{,}\psi _{\alpha }\right] }\text{,} \label{Z}$$
with measure defined by independent Grassmann variables $D\psi ^{+}=\underset%
{\alpha }{\Pi }d\psi _{\alpha }^{+},D\psi =\underset{\alpha }{\Pi }d\psi
_{\alpha }$. The Matsubara action reads:$$\begin{aligned}
S\left[ \psi ^{+}\mathbf{,}\psi \right] &=&\int_{0}^{1/T}d\tau \int_{r}\psi
_{\alpha }^{+}\left( \tau ,\mathbf{r}\right) \left( \partial _{\tau
}+K_{\alpha \beta }\right) \psi _{\beta }\left( \tau ,\mathbf{r}\right)
\notag \\
&&\mathbf{-}\frac{g}{2}\psi _{\alpha }^{+}\left( \tau ,\mathbf{r}\right)
\psi _{\beta }^{+}\left( \tau ,\mathbf{r}\right) \psi _{\beta }\left( \tau ,%
\mathbf{r}\right) \psi _{\alpha }\left( \tau ,\mathbf{r}\right) \text{,}
\label{S}\end{aligned}$$with the anti-periodic conditions,$$\begin{aligned}
\psi _{\alpha }^{+}\left( \tau ,\mathbf{r}\right) &=&-\psi _{\alpha
}^{+}\left( \tau +1/T,\mathbf{r}\right) , \label{antiperiod} \\
\psi _{\alpha }\left( \tau ,\mathbf{r}\right) &=&-\psi _{\alpha }\left( \tau
+1/T,\mathbf{r}\right) \text{.} \notag\end{aligned}$$The local interaction term is not the most general one, but generalization to more “exotic” local cases (inter-valley[@Fu] the exchange spin - spin coupling[@Rosenstein15]) is quite straightforward.
The normal and anomalous Green’s functions (consistent with definition in ref.[@AGD]) are:$$\begin{aligned}
\left\langle T\psi _{\alpha }\left( X\right) \psi _{\beta }^{+}\left(
X^{\prime }\right) \right\rangle &=&-G_{\alpha \beta }\left( X\mathbf{;}%
X^{\prime }\right) \text{;} \notag \\
\left\langle T\psi _{\alpha }\left( X\right) \psi _{\beta }\left( X^{\prime
}\right) \right\rangle &=&F_{\alpha \beta }\left( X\mathbf{;}X^{\prime
}\right) \text{;} \label{GF_Gorkov} \\
\left\langle T\psi _{\alpha }^{+}\left( X\right) \psi _{\beta }^{+}\left(
X^{\prime }\right) \right\rangle &=&F_{\alpha \beta }^{+}\left( X\mathbf{;}%
X^{\prime }\right) \text{,} \notag\end{aligned}$$where $X=\left( \mathbf{r,}\tau \right) $, $X^{\prime }=\left( \mathbf{r}%
^{\prime },\tau ^{\prime }\right) $. For simplicity, we denote $\left\langle
\psi _{\alpha }\left( X\right) \right\rangle $ as $\psi _{\alpha }\left(
X\right) $, and drop the time ordering operation $T$ in correlators. For example, $\left\langle \psi _{\alpha }\left( X\right) \psi _{\beta }\left(
X^{\prime }\right) \right\rangle $ stands for $\left\langle T\psi _{\alpha
}\left( X\right) \psi _{\beta }\left( X^{\prime }\right) \right\rangle $. Their expressions via partition function, Eq.(\[S\]), are given in Appendix A.
Space and spin rotations symmetries
-----------------------------------
Generally a system may be invariant under both the space rotation and the spin rotation separately. Certain valley symmetries are generally present. In Weyl semimetals, the action is typically only invariant under the combined space rotation and spin/valley rotation. The space rotation, $%
\mathbf{r}^{\prime }=\Lambda \mathbf{r}$, acts on a generalized spinor field as:
$$\begin{aligned}
\psi _{\alpha }^{\prime }\left( \mathbf{r}^{\prime },\tau \right) &=&S\left(
\Lambda \right) _{\alpha \alpha ^{\prime }}\psi _{\alpha ^{\prime }}\left(
\mathbf{r},\tau \right) , \label{space_trans} \\
\psi _{\beta }^{\prime +}\left( \mathbf{r}^{\prime },\tau \right) &=&\psi
_{\beta ^{\prime }}^{+}\left( \mathbf{r},\tau \right) S\left( \Lambda
\right) _{\beta ^{\prime }\beta }^{\dagger }\text{.} \notag\end{aligned}$$
The invariance of the action under the transformation, the correlators satisfy: $$\begin{aligned}
G_{\alpha \beta }\left( X_{1},X_{2}\right) &=&S\left( \Lambda \right)
_{\alpha \alpha ^{\prime }}G_{\alpha ^{\prime }\beta ^{\prime }}\left(
X_{1},X_{2}\right) S\left( \Lambda \right) _{\beta ^{\prime }\beta
}^{\dagger }; \notag \\
F_{\alpha \beta }\left( X_{1},X_{2}\right) &=&S\left( \Lambda \right)
_{\alpha \alpha ^{\prime }}F_{\alpha ^{\prime }\beta ^{\prime }}\left(
X_{1},X_{2}\right) S\left( \Lambda \right) _{\beta ^{\prime }\beta }^{t}
\label{conditions}\end{aligned}$$or in the matrix form:
$$\begin{aligned}
G\left( X_{1},X_{2}\right) &=&S\left( \Lambda \right) G\left(
X_{1},X_{2}\right) S\left( \Lambda \right) ^{\dagger }, \label{inv_mat} \\
F\left( X_{1},X_{2}\right) &=&S\left( \Lambda \right) F\left(
X_{1},X_{2}\right) S\left( \Lambda \right) ^{t} \notag\end{aligned}$$
Assuming that the ground state is homogeneous, $G\left( X,X\right) \equiv
G_{c}$ and $F\left( X,X\right) \equiv g^{-1}\Delta $ are constant matrices satisfying $$G_{c}=S\left( \Lambda \right) G_{c}S\left( \Lambda \right) ^{+},\Delta
=S\left( \Lambda \right) \Delta S\left( \Lambda \right) ^{t}\text{.}
\label{localcorr}$$Let $\Sigma $ be generator of $S\left( \Lambda \right) $. Then $$\left[ \Sigma ,G_{c}\right] =0,\Sigma \Delta +\Delta \Sigma ^{t}=0
\label{rotation_alg}$$$G_{c}$ and $\Delta $ also satisfy the following equations:$$G_{c}^{+}=G_{c},\Delta ^{t}=-\Delta \text{.} \label{Pauli}$$In superconductor $G_{\alpha \beta }^{c}=n_{c}\delta _{\alpha \beta }$ and leads within the BCS approximation just to renormalization of the chemical potential. Therefore we finally obtain the Gor’kov equations, $$\begin{gathered}
\left( -\partial _{\tau }\delta _{\alpha \beta }-K_{\alpha \beta }\left(
\mathbf{\nabla }\right) \right) G_{\beta \gamma }\left( X,X^{\prime }\right)
\mathbf{-}g\times \notag \\
F_{\beta \alpha }\left( X,X\right) F_{\beta \gamma }^{+}\left( X,X^{\prime
}\right) =\delta \left( X-X^{\prime }\right) \delta _{\alpha \gamma };
\label{GeqX} \\
\left( \partial _{\tau }\delta _{\alpha \beta }-K_{\beta \alpha }\left( -%
\mathbf{\nabla }\right) \right) F_{\beta \gamma }^{+}\left( X,X^{\prime
}\right) - \notag \\
gF_{\alpha \beta }^{+}\left( X,X\right) G_{\beta \gamma }\left( X,X^{\prime
}\right) =0\text{.} \notag\end{gathered}$$
We will also discuss the non-superconducting state like the exciton condensate with opposite bulk properties, $G_{\alpha \beta }^{c}\left(
X,X\right) \neq n_{c}\delta _{\alpha \beta }$, $F_{\alpha \beta }\left(
X;X^{\prime }\right) =0$. In this case the Dyson - Schwinger form is more convenient. The gap equation can be recasted (see Appendix A) in matrix form as
$$G^{-1}=G_{0}^{-1}+g\delta \left( X-X^{\prime }\right) G_{c}^{\prime }\text{,}
\label{DSeq}$$
where $G_{\alpha \beta }^{\prime }$ is the traceless part of $G_{\alpha
\beta }$, $$\begin{gathered}
G_{\alpha \beta }^{\prime }\left( X;X^{\prime }\right) =-\left\langle \psi
_{\alpha }\left( X\right) \psi _{\beta }^{+}\left( X^{\prime }\right)
\right\rangle _{c}+ \label{traceless} \\
\frac{1}{4}\delta _{\alpha \beta }\sum \left\langle \psi _{\gamma }\left(
X\right) \psi _{\gamma }^{+}\left( X^{\prime }\right) \right\rangle _{c},
\notag\end{gathered}$$with renormalized chemical potential taking case of the trace as in superconductor.
The Matsubara Green’s functions ($\tau $ is the Matsubara time) for uniform superconducting states can be expressed via Fourier transforms, $$\begin{gathered}
G_{\alpha \beta }\left( \mathbf{r},\tau ;\mathbf{r}^{\prime },\tau ^{\prime
}\right) =\frac{T}{\left( 2\pi \right) ^{D}}\int d^{D}\mathbf{k}\times
\notag \\
e^{i\mathbf{k\cdot }\left( \mathbf{r-r}^{\prime }\right)
}\sum_{n}e^{-i\omega _{n}\left( \tau -\tau ^{\prime }\right) }G_{\alpha
\beta }\left( \mathbf{k,}\omega _{n}\right) \text{;} \label{GFdef} \\
F_{\alpha \beta }^{\dagger }\left( \mathbf{r},\tau ;\mathbf{r}^{\prime
},\tau ^{\prime }\right) =\frac{T}{\left( 2\pi \right) ^{D}}\int d^{D}%
\mathbf{k}\times \notag \\
e^{i\mathbf{k\cdot }\left( \mathbf{r-r}^{\prime }\right)
}\sum_{n}e^{-i\omega _{n}\left( \tau -\tau ^{\prime }\right) }F_{\alpha
\beta }^{\dagger }\left( \mathbf{k,}\omega _{n}\right) \text{,} \notag\end{gathered}$$where $\omega =\pi T\left( 2n+1\right) $ is the Matsubara Fermionic frequency. The Matsubara Green’s functions in Fourier forms can simplify the calculation significantly.
General formula[@AGD] obtained in Appendix B for the energy of the superconducting state reads: $$\frac{d\Omega }{d\left( g^{-1}\right) }\mathcal{=}\frac{V}{2}\text{Tr}\left(
\Delta \Delta ^{\dag }\right) \label{energy_super}$$while for chiral symmetry breaking (exciton condensation) states the gaussian energy is[@CJT] $$\Omega =-\text{Tr}\{-\ln G+[G_{0}^{-1}G-1]\}+\frac{g}{2}\text{Tr}%
G_{c}G_{c}^{\dag }\ \text{.} \label{energy_chiral}$$
The Dirac model and its symmetries
==================================
Hamiltonian of the time reversal invariant TI with local interaction.
---------------------------------------------------------------------
Electrons on the surface of a TI perpendicular to $z$ axis are described by a Pauli spinors $\psi \left( \mathbf{r}\right) $, where the upper plane, $%
\mathbf{r}=\left\{ x,y\right\} $. In principle there are multiple valleys. The case of just one valley describing surface of the topological insulator like $Bi_{2}Te_{3}$ were considered in [@Zhang; @Herbut; @Li14]. It breaks time reversal invariance and does not allow chiral symmetry breaking, so here we consider the simplest case of multiple valleys: the Dirac model in which chiralities of the two Weyl modes are opposite described by field operators $\psi _{fs}\left( \mathbf{r}\right) $, where $f=L,R$ are the valley index (pseudospin) for the left/right chirality bands with spin projections taking the values $s=\uparrow ,\downarrow $ with respect to, for example, $z$ axis. To use the Dirac (“pseudo-relativistic”) notations, these are combined into a four component bi-spinor creation operator, $\psi
^{\dagger }=\left( \psi _{L\uparrow }^{\dagger },\psi _{L\downarrow
}^{\dagger },\psi _{R\uparrow }^{\dagger },\psi _{R\downarrow }^{\dagger
}\right) $, whose index $\gamma =\left\{ f,s\right\} $ takes four values. The non-interacting massless Hamiltonian with Fermi velocity $v_{F}$ and linear dispersion relation, see Fig.1, reads[@Wang13],$$\mathcal{E}_{\gamma \delta }=-i\hbar v_{F}\nabla ^{i}\alpha _{\gamma \delta
}^{i}\text{,} \label{epsilon}$$where two $4\times 4$ matrices, $i=x,y$, $$\mathbf{\alpha }=\left(
\begin{array}{cc}
\mathbf{\sigma } & 0 \\
0 & -\mathbf{\sigma }%
\end{array}%
\right) \text{,} \label{alfa_mat}$$are presented in the block form via Pauli matrices $\mathbf{\sigma }$. They are related to the Dirac $\mathbf{\gamma }$ matrices (in the chiral representation, sometimes termed “spinor”) by $\ \mathbf{\alpha }=\beta
\mathbf{\gamma }$ with$$\beta =\left(
\begin{array}{cc}
0 & \mathbf{1} \\
\mathbf{1} & 0%
\end{array}%
\right) \text{.} \label{beta_mat}$$
{width="8cm"}
The noninteracting Hamiltonian in these notations reads:$$K=\int_{\mathbf{r}}\psi ^{\dagger }\left\{ -i\hbar v_{F}\beta \gamma \cdot
\mathbf{\nabla }-\mu \right\} \psi \text{.} \label{K_Dirac}$$The Matsubara action, Eq.(\[S\]), is conveniently written in pesudo-relativistic notations with $\overline{\psi }=\psi ^{+}\gamma _{0}$ with (Euclidean) $\gamma _{0}=i\beta $:$$\begin{aligned}
S &=&\int_{\tau =0}^{1/T}\int_{\mathbf{r}}\left\{ -\overline{\psi }\left(
\gamma _{0}\partial _{\tau }+\hbar v_{F}\gamma \cdot \mathbf{\nabla -}\mu
\gamma _{0}\right) \psi +\right. \label{relativistic action} \\
&&\left. \frac{g}{2}\overline{\psi }\gamma _{0}\psi \overline{\psi }\gamma
_{0}\psi \right\} \text{.} \notag\end{aligned}$$Since our focus is on symmetry and its spontaneous breaking, let us review known discrete and continuous symmetries.
Continuous symmetries
---------------------
Symmetries of the 2D Dirac model with local interactions Eqs.([Hamiltonian]{},\[epsilon\]) were thoroughly discussed in relation to graphene[@Gusynin]. They include parity, time reversal and the discrete chiral (flavor) transformation: $\psi \rightarrow \gamma _{5}\psi ;\overline{%
\psi }\rightarrow -\overline{\psi }\gamma _{5}$, where $\gamma _{5}=\gamma
_{0}\gamma _{1}\gamma _{2}\gamma _{3}=\gamma _{5}^{\dagger }$. Spontaneous breaking of this symmetry has been comprehensively investigated in the context of graphene[@Gusynin] and will not be stressed here. There are also three continuous symmetries that in principle can lead to ordered phase with massless Goldstone bosons (order parameter waves). The first is the usual electric charge $U\left( 1\right) $, $\psi \rightarrow e^{i\chi }\psi $, that is spontaneously broken in a superconducting state studied in next section. In addition there is the “chiral” flavour rotations $SU\left(
2\right) $ that play an important role in exciton condensation that will be addressed in section V.
### Space time symmetries: just Aphelian space rotation combined with spin (pseudospin) rotation
Unlike the non-interacting model the pseudorelativistic 2+1 dimensional Lorentz invariance is explicitly broken by the static interaction Eq.([Hamiltonian]{}), so that only 2D rotations accompanied by the (pseudo) spin rotation already described in Section II with spin $\Sigma $ operator, $$\Sigma \mathbf{=}\left(
\begin{array}{cc}
\sigma _{z} & 0 \\
0 & \sigma _{z}%
\end{array}%
\right) =\gamma _{1}\gamma _{2}\text{,} \label{spin_Dirac}$$remain a symmetry. The conserved quantity is the abelian angular momentum $$J=\int_{\mathbf{r}}\psi ^{\dagger }\left( \mathbf{r}\right) \left\{
i\varepsilon _{ij}r_{i}\nabla _{j}+\frac{1}{2}\mathbf{\Sigma }\right\} \psi
\left( \mathbf{r}\right) \text{.} \label{J}$$The second part is referred to as the spin rotation, $S=\frac{1}{2}\int \psi
^{\dagger }\mathbf{\Sigma }\psi $.
### Electric charge $U\left( 1\right) $
The usual electric charge $U\left( 1\right) $ with conserved electric charge:
$$Q=\int_{\mathbf{r}}\psi ^{\dagger }\left( \mathbf{r}\right) \psi \left(
\mathbf{r}\right) =\int_{\mathbf{r}}\rho \left( \mathbf{r}\right) \text{.}
\label{Q}$$
Action Eq.(\[relativistic action\]) is invariant under the phase (global gauge) transformation, $\psi \rightarrow e^{i\chi }\psi $
### Flavour (chiral or valley) $SU(2)$
It was noticed early on in relation to graphene[@Gusynin] that there is flavour $SU\left( 2\right) $ symmetry. It is shown in Appendix C that quantities $$Q_{i}=\int_{\mathbf{r}}\psi ^{\dagger }\left( \mathbf{r}\right) T_{i}\psi
\left( \mathbf{r}\right) , \label{Q_i}$$commute with Hamiltonian and thus are conserved quantities.The generator matrices, $$T_{1}=\frac{i}{2}\gamma _{3},T_{2}=\frac{1}{2}\gamma _{5},T_{3}=\frac{1}{2}%
\gamma _{3}\gamma _{5}\text{,} \label{Tmat}$$($\gamma _{5}=\gamma _{0}\gamma _{1}\gamma _{2}\gamma _{3}$) constitute a nonrelativistic $SU\left( 2\right) $ algebra:$$\left[ T_{i},T_{j}\right] =i\varepsilon _{ijk}T_{k}\text{.} \label{SU(2)}$$A discrete chiral symmetry is just the chiral rotation by angle $\pi $.The action Eq.(\[relativistic action\]) is invariant under infinitesimal transformation, $\delta \psi =iT_{i}\psi ;$ $\delta \overline{\psi }=i%
\overline{\psi }\gamma _{0}T_{i}\gamma _{0}$. All three continuous symmetries commute??(meaning charge, Flavour, and space time symmetry commute).
Spontaneously broken electric charge $U\left( 1\right) $ symmetry phases: superconducting pairing channels
----------------------------------------------------------------------------------------------------------
Due to locality of the dominant interactions, the superconducting order parameter is local,$$O=\int_{\mathbf{r}}\psi _{\alpha }^{\dagger }\left( \mathbf{r}\right)
M_{\alpha \beta }\psi _{\beta }^{\dagger }\left( \mathbf{r}\right) ,
\label{O}$$where the constant matrix $M$ should be a $4\times 4$ antisymmetric matrix. Due to the rotation symmetry they transform covariantly under infinitesimal rotations generated by the spin $\Sigma $ operator Eq.(\[spin\_Dirac\]).
Out of 16 matrices of the four dimensional Clifford algebra six are antisymmetric. We will not consider rather exotic phases in which in addition to the charge $U\left( 1\right) $ symmetry neither the 2D rotations or the $SU\left( 2\right) $ chiral transformations are spontaneously broken. Therefore superconducting order parameter is invariant under the remaining symmetries: $\left[ O,J\right] =\left[ O,Q_{i}\right] =0$. Namely it is invariant under the $SU\left( 2\right) $ and is either scalar or pseudoscalar under rotations. The requirement of invariance expressed using Eq.(\[spin\_Dirac\]) takes a form:
$$\begin{aligned}
\left[ O,J\right] &=&\int_{\mathbf{r,r}^{\prime }}\left[ \psi _{\alpha
}^{\dagger }\left( \mathbf{r}\right) M_{\alpha \beta }\psi _{\beta
}^{\dagger }\left( \mathbf{r}\right) \right. , \notag \\
&&\left. \psi _{\gamma }^{\dagger }\left( \mathbf{r}^{\prime }\right)
\mathbf{\Sigma }_{\gamma \delta }\psi _{\delta }\left( \mathbf{r}^{\prime
}\right) \right] \label{transformation} \\
&=&-\int_{\mathbf{r}}\psi ^{\dagger }\left( \mathbf{r}\right) \left( \mathbf{%
\Sigma }M+M\mathbf{\Sigma }^{t}\right) \psi ^{\dagger }\left( \mathbf{r}%
\right) =0\text{.} \notag\end{aligned}$$
Similarly $$\left[ O,Q_{i}\right] =-\int_{\mathbf{r}}\psi ^{\dagger }\left( \mathbf{r}%
\right) \left( T_{i}M+MT_{i}^{t}\right) \psi ^{\dagger }\left( \mathbf{r}%
\right) =0\text{.} \label{rotation1}$$One finds that the only scalar is, $M=i\alpha _{y}$. There is also a pseudoscalar that will not be discussed here, namely we assume that the superconducting state preserves all the other symmetries. Which one of the condensates is realized at zero temperature is determined by the parameters of the Hamiltonian along the line of dynamical calculation presented in section IV for the scalar.
It turns out that there when the effective electron attraction is replaced by repulsion and the superconductivity is not realized there is still a possibility of continuous symmetry breaking that also belongs to a chiral universality class: the chiral $SU\left( 2\right) $.
Chiral $SU\left( 2\right) $ broken excitonic phases
---------------------------------------------------
In this subsection we consider an opposite situation when the charge symmetry is unbroken, that is no superconducting condensate appears. Still due to nontrivial multicomponent situation with the $SU\left( 2\right) $ symmetry there are possible transitions into a gapped exciton condensate phases. It is plausible that rotational symmetry is also unbroken. Still there are two possible patterns, one is breaking down to an $U\left(
1\right) $ subgroup with two Goldstone boson modes and another down to trivial subgroup with three Goldstone models.
General order parameter now is$$P=\int_{\mathbf{r}}\psi _{\alpha }^{\dagger }\left( \mathbf{r}\right)
V_{\alpha \beta }\psi _{\beta }\left( \mathbf{r}\right) , \label{Pdef}$$where the constant matrix $V$ should be an $4\times 4$ hermitian matrix. There are four chiral $SU\left( 2\right) $ triplets of order parameters $%
\mathbf{P=}\left\{ P_{1},P_{2},P_{3}\right\} $ (that can be viewed as the $%
O\left( 3\right) $ vectors). Their commutations with chiral rotations generators $Q_{i}$ defined in Eq.(\[Q\_i\]) are: $$\left[ P_{i},Q_{j}\right] =\int_{\mathbf{r}}\psi ^{\dagger }\left( \mathbf{r}%
\right) \left[ V_{i},T_{j}\right] \psi \left( \mathbf{r}\right) =\varepsilon
_{ijk}P_{k}\text{.} \label{commP}$$Four sets of matrices $V_{\alpha \beta }$ in terms of the Dirac and chiral symmetry matrices are$$\begin{aligned}
V^{\left( 1\right) } &=&\{\gamma _{0}T_{2},\gamma _{0}T_{1},\frac{1}{2}%
i\gamma _{0}\}=\frac{1}{2}\left\{ -\gamma _{1}\gamma _{2}\gamma _{3},i\gamma
_{0}\gamma _{3},i\gamma _{0}\right\} \notag \\
V^{\left( 2\right) } &=&\left\{ T_{1},T_{2},T_{3}\right\} =\frac{1}{2}%
\left\{ i\gamma _{3},\gamma _{0}\gamma _{1}\gamma _{2}\gamma _{3},-\gamma
_{0}\gamma _{1}\gamma _{2}\right\} \notag \\
V^{\left( 3\right) } &=&\left\{ \gamma _{1},\gamma _{1}\gamma _{3},\gamma
_{0}\gamma _{2}\gamma _{3}\right\} , \label{Vmat} \\
V^{\left( 4\right) } &=&\left\{ \gamma _{2},\gamma _{2}\gamma _{3},\gamma
_{0}\gamma _{1}\gamma _{3}\right\} \text{.} \notag\end{aligned}$$There are also four chiral scalars, $\left[ P,Q_{i}\right] =0$, $Q=\int \psi
^{\dagger }I\psi $ (charge)$,S=\int \psi ^{\dagger }\gamma _{1}\gamma
_{2}\psi $ (spin), $H_{i}=\int \psi ^{\dagger }\alpha _{i}\psi ,\alpha
_{i}=\gamma _{0}\gamma _{i}$,$i=x,y$, that complete the Clifford algebra consisting of 16 hermitian matrices. These are not order parameters and hence will not be of interest to us. We also limit ourselves to the rotation invariant phases. The requirement of invariance expressed using Eq.(\[spin\_Dirac\]) takes a form:
$$\left[ P,J\right] =\int_{\mathbf{r}}\psi ^{\dagger }\left( \mathbf{r}\right) %
\left[ V\mathbf{,}\Sigma \right] \psi ^{\dagger }\left( \mathbf{r}\right)
\text{.} \label{rotation}$$
Since $V^{\left( 3\right) }$ and $V^{\left( 3\right) }$ in Eq.(\[Vmat\]) are not invariant under rotations only the first two are considered. If only one of the chiral vector order parameters has a nonzero expectation value, say $\left\langle P_{3}^{\left( 1\right) }\right\rangle \not=0$, the symmetry breaking pattern is $SU\left( 2\right) \rightarrow U\left( 1\right)
$, since $\left[ P_{3}^{\left( 1\right) },Q_{3}\right] =0$. According to the Goldstone theorem there are two soft modes in directions $Q_{1}$ and $Q_{2}$. If in addition the second vector order parameter acquires VEV, $%
\left\langle P_{i}^{\left( 2\right) }\right\rangle \not=0$ and $i\not=3$, the pattern will be $SU\left( 2\right) \rightarrow I$ with three Goldstone modes.
Which symmetry breaking mode is actually realized at given parameters of the system (chemical potential, Fermi velocity, interaction sign and strength, temperature...) is a dynamical question. Therefore now we turn to dynamical aspects of the phase diagram of the Dirac model.
Superconducting state
=====================
Within gaussian approximation, the Green’s functions obey the Gor’kov equations derived in[@Li14] and in last section. For $g>0$ anomalous Green’s functions are nonzero, so that the Gor’kov equations for Fourier components of the Greens functions simplify considerably, $$\begin{aligned}
D_{\gamma \beta }^{-1}G_{\beta \kappa }\left( \omega ,p\right) -\widehat{%
\Delta }_{\gamma \beta }F_{\beta \kappa }^{\dagger }\left( \omega ,p\right)
&=&\delta ^{\gamma \kappa }\text{;} \label{Gorkov_uniform} \\
D_{\beta \gamma }^{-1}F_{\beta \kappa }^{\dagger }\left( \omega ,p\right) +%
\widehat{\Delta }_{\gamma \beta }^{\ast }G_{\beta \kappa }\left( \omega
,p\right) &=&0\text{,} \notag\end{aligned}$$where $\ D_{\gamma \beta }^{-1}=\left( i\omega -\mu \right) \delta _{\gamma
\beta }-v_{F}\varepsilon _{ij}p_{i}\alpha _{\alpha \beta }^{j}$ where the chemical potential is renormalized. The matrix gap function can be chosen as ($\Delta $ real) $$\widehat{\Delta }_{\beta \gamma }=gF_{\gamma \beta }\left( 0\right) =\left(
\begin{array}{cc}
0 & \Delta \\
-\Delta & 0%
\end{array}%
\right) \text{.} \label{delta}$$
These equations are conveniently presented in matrix form (superscript $t$ denotes transposed and $I$ - the identity matrix): $$\begin{aligned}
D^{-1}G-\widehat{\Delta }F^{\dagger } &=&I\text{;} \label{matrixeq} \\
D^{t-1}F^{\dagger }+\widehat{\Delta }^{\ast }G &=&0\text{.} \notag\end{aligned}$$Solving these equations one obtains $$\begin{aligned}
G^{-1} &=&D^{-1}+\widehat{\Delta }D^{t}\widehat{\Delta }^{\ast }\text{;}
\label{solution} \\
F^{\dagger } &=&-D^{t}\widehat{\Delta }^{\ast }G\text{,} \notag\end{aligned}$$with the gap function found from the consistency condition $$\widehat{\Delta }^{\ast }=-g\sum\limits_{\omega q}D^{t}\widehat{\Delta }%
^{\ast }G\text{.} \label{gap eq}$$The off-diagonal component of this equation is: $$\begin{gathered}
\frac{1}{g}=\sum\limits_{\omega p}\left( \Delta ^{2}+v_{F}^{2}p^{2}+\mu
^{2}+\hbar ^{2}\omega ^{2}\right) \times \notag \\
\left( \Delta ^{2}+\hbar ^{2}\omega ^{2}+\left( v_{F}p-\mu \right)
^{2}\right) ^{-1}\times \label{gap} \\
\left( \Delta ^{2}+\hbar ^{2}\omega ^{2}+\left( v_{F}p+\mu \right)
^{2}\right) ^{-1}\text{.} \notag\end{gathered}$$
The spectrum of elementary excitations obtained from the poles of the Greens function coincides with that found within the Bogoliubov - de Gennes approach [@Herbut]: $E_{p}=\pm \sqrt{\Delta ^{2}+\left( v_{F}p-\mu
\right) ^{2}}$.
Zero temperature phase diagram for the superconductor - normal transition.
--------------------------------------------------------------------------
At zero temperature the integrations over frequency and momentum limited by the UV cutoff $\Lambda $ result in $$U=\sqrt{\Delta ^{2}+\mu ^{2}}-\frac{\mu }{2}\log \frac{\sqrt{\Delta ^{2}+\mu
^{2}}+\mu }{\sqrt{\Delta ^{2}+\mu ^{2}}-\mu }\text{,} \label{gapeq1}$$where the dependence on the cutoff is incorporated in the renormalized coupling with dimension of energy defined as $$U=v_{F}\Lambda -\frac{4\pi \hbar ^{2}v_{F}^{2}}{g}\text{.} \label{gren}$$This can be interpreted as an effective binding energy of the Cooper pair in the Weyl semi - metal. We consider only $\mu >0$, since the particle - hole symmetry makes the opposite case of the hole doping, $\mu <0$, identical. Of course the superconducting solution exists only for $g>0$. In Fig. 2 the dependence of the gap $\Delta $ as function of the chemical potential $\mu $ is presented for different values of $U$.
![Order parameter at zero temperature as function of chemical potential of the TI surface Weyl semi-metal at various values of coupling parametrized by the renormalized energy $U$, Eq.(\[gren\]). For positive $U$ (blue lines) the superconductivity is strong and does not vanish even for zero chemical potential. There exists the critical coupling, $U=0$ (red line), at which the second order transition occurs at quantum critical point $\protect\mu =0$. For negative $U$ the superconductivity still exists at $\protect\mu >0$, but is exponentially weak. ](fig2.pdf){width="8cm"}
For an attractive coupling $g$ stronger than the critical one, $$g_{c}=\frac{4\pi \hbar ^{2}v_{F}}{\Lambda }\text{,} \label{g_c}$$(when $U>0$), blue lines in Fig. 2, there are two qualitatively different cases.
(i). When $\mu <<U$ the dependence of $\Delta $ on the chemical potential is parabolic, see [@Li14]. In particular, when $\mu =0,$ the gap equals $U$. As can be seen from Fig. 2, the chemical potential makes a very limited impact in the large portion of the phase diagram.
\(ii) For the attraction just stronger than critical, $g>g_{c}$, namely for small positive $U$, the dependence becomes linear, see red line in Fig. 2, $%
\Delta =0.663\,\mu $. So that the already weak condensate becomes sensitive to $\mu $.
The case (i) is more interesting than (ii) since it exhibits stronger superconductivity (larger $T_{c}$, see below). Finally for $g<g_{c}\,$, namely negative $U$ (green lines), the superconductivity is very weak with exponential dependence similar to the BCS one, $\Delta \approx \mu $ exp$%
\left[ -\left( \left\vert U\right\vert /\mu -1\right) \right] $. As was mentioned above, in the more interesting cases of large $\Delta $ the dependence on the chemical potential is very weak. A peculiarity of superconductivity in TI is that electrons (and holes) in Cooper pairs are created themselves by the pairing interaction rather than being present in the sample as free electrons. Therefore it is shown that it is possible to neglect the effect of weak doping and consider directly the $\mu =0$ particle-hole symmetric case. This point in parameter space is the QCP [Sachdev]{} and will be studied in detail in what follows. Of course, at finite temperature at any attraction, $g>0$, there exists a (classical) superconducting critical point at certain temperature $T_{c}$ that is calculated next.
Dependence of the critical temperature $T_{c}$ on strength of pairing interaction.
----------------------------------------------------------------------------------
Summation over Matsubara frequency and integrations over momenta in the gap equation, Eq.(\[gapeq1\]), at finite temperature and arbitrary chemical potential. The critical temperature as a function of $\mu $ and (positive) $%
U $ is obtained numerically and presented in Fig. 3. Again at relatively large $U$ the dependence of $T_{c}$ on the chemical potential is very weak and parabolic. When $0<g<g_{c}$ the critical temperature is exponentially small albeit nonzero, $\frac{\Delta }{U}\approx 1+\left( \frac{\mu }{U}%
\right) ^{2} $.
{width="8cm"}
Zero chemical potential $\protect\mu =0$.
-----------------------------------------
At zero chemical potential the Hamiltonian Eq.(\[Hamiltonian\]) possesses a particle - hole symmetry. Microscopically, Cooper pairs of both electrons and holes are formed, see Fig. 1a. The system is unique in this sense since the electron - hole symmetry is not spontaneously broken in both normal and superconducting phases. Supercurrent in such a system does not carry momentum or mass. Performing the sum and integral over momenta in the gap equation, Eq.(\[gapeq1\]), analytically (see Appendix A), it becomes (using the definition of $U$ given in Eq.(\[gren\])) for $U>0$: $$U=2T\log \left[ 2\cosh \frac{\Delta }{2T}\right] \text{.} \label{gapren_T}$$At zero temperature $\Delta =U$, while $\Delta \rightarrow 0$ as a power of the parameter $U\propto g-g_{c}$ describing the deviation from quantum criticality $$T_{c}=\frac{1}{2\log 2}U^{z\nu };\text{ \ }z\nu =1\text{.} \label{Tc_mic}$$Here $z$ is the dynamical critical exponent[@Sachdev]. Therefore, as expected, the renormalized coupling describing the deviation from the QCP is proportional to the temperature at which the created condensate disappears.
{width="8cm"}
The temperature dependence of the gap reads: $$\Delta \left( T\right) =2T\cosh ^{-1}\left( \frac{1}{2}\exp \frac{U}{2T}%
\right) \text{.} \label{delta_mic}$$This it typical for chiral universality classes [@Sachdev; @Rosenstein].
It is interesting to compare this dependence with the conventional BCS[AGD]{} for transition at finite temperature, namely away from QCP, see Fig 1b. At zero temperature $\Delta \left( 0\right) /T_{c}=2\log 2$ $%
=\allowbreak 1.\,\allowbreak 39$ (within BCS - $1.76$), while near $T_{c}$ one gets $\Delta /T_{c}=2^{3/2}\log ^{1/2}2\sqrt{1-t}=$ $2.\,\allowbreak 35%
\sqrt{1-t}$ (BCS - $3.07\sqrt{1-t}$), where $t=T/T_{c}$. To describe the behavior of the STI in inhomogeneous situations like the external magnetic field, boundaries, impurities or junction with metals or other superconductors, it is necessary to derive the effective theory in terms of the order parameter $\Delta \left( \mathbf{r}\right) $, where $\mathbf{r}$ varies on the mesoscopic scale.
Phase diagram of STI order parameter as function of $U$ at zero temperature is plotted in Fig.4.
Coherence length and the condensation energy
--------------------------------------------
The quadratic term of the Ginzburg-Landau energy $F_{2}=\sum_{\mathbf{p}%
}\Delta _{p}^{\ast }\Gamma \left( p\right) \Delta _{p}$ is obtained exactly from expanding the gap equation to linear terms in $\Delta $ for arbitrary external momentum. The dependence on $\mathbf{p}$ is non-analytic and within our approximation higher powers of $p$ do not appear. The second term is very different from the quadratic term in the GL functional for conventional phase transitions at finite temperature [@Herbutbook] or even quantum phase transitions in models without Weyl fermions [@Sachdev] and has a number of qualitative consequences. Comparing the two terms in Eq.([Gamma1]{}), one obtains the coherence length as a power of parameter $%
U\propto g-g_{c}$ describing the deviation from criticality: $$\xi \left( U\right) =\frac{\pi }{4}v_{F}\hbar U^{-\nu }\text{; \ \ \ \ }\nu
=1\text{.} \label{coherence}$$This is different from the dependence in non-chiral universality classes that is [@Herbutbook] $\xi \left( T\right) \infty \left( T_{c}-T\right)
^{-\nu },$ $\nu =1/2$ in mean field. Of course in the regime of critical fluctuations this exponent is corrected in both non-chiral [@Herbutbook] and chiral[@Gat] universality classes.
Local terms in the GL energy density are also calculable exactly. Expression for the kernel can be written as a trace: $$\begin{gathered}
\Gamma =\frac{1}{2}tr\left\{ \sum\limits_{\omega q}\sigma ^{y}D_{\omega
q}^{t}\sigma ^{y}D_{\omega ,q-p}+\frac{1}{g}I\right\} \label{C1} \\
=\frac{1}{g}-\sum\limits_{\omega q}\frac{\hbar ^{2}\omega
^{2}-v_{F}^{2}p\cdot q+v_{F}^{2}q^{2}}{\left( \hbar ^{2}\omega
^{2}+v_{F}^{2}q^{2}\right) \left( \hbar ^{2}\omega ^{2}+v_{F}^{2}\left\vert
\mathbf{q-p}\right\vert ^{2}\right) }. \notag\end{gathered}$$Integrating over $\omega $ (at zero temperature) one obtains$$\begin{aligned}
\Gamma &=&-\frac{1}{8\pi ^{2}\hbar ^{2}v_{F}}\int_{q,\phi }\frac{pq\cos \phi
-p^{2}}{\left\vert \mathbf{q-p}\right\vert ^{2}+q\left\vert \mathbf{q-p}%
\right\vert } \label{Gamma} \\
&&-\frac{U}{4\pi \hbar ^{2}v_{F}^{2}}, \notag\end{aligned}$$where $\phi $ is an angle between $\mathbf{q}$ and $\mathbf{p}$. The integral is homogeneous in momentum and therefore is linear in $p=\left\vert
\mathbf{p}\right\vert $ and one arrives at: $$\Gamma \left( p\right) =-\frac{U}{4\pi \hbar ^{2}v_{F}^{2}}+\frac{\left\vert
p\right\vert }{16v_{F}\hbar ^{2}}\text{.} \label{Gamma1}$$
Local terms in Ginzburg - Landau equation and energy
-----------------------------------------------------
For $p=0$ the gap equation Eq.(\[gapeq1\]) reads$$\frac{\Delta }{4\pi \hbar ^{2}v_{F}^{2}}\left( -U+\sqrt{\Delta ^{\ast
}\Delta }\right) =0\text{,} \label{C3}$$This is obtained from the energy functional$$F=\frac{1}{4\pi \hbar ^{2}v_{F}^{2}}\int d^{2}\mathbf{r}\left\{ -U\Delta
^{\ast }\Delta +\frac{2}{3}\left( \Delta ^{\ast }\Delta \right)
^{3/2}\right\} \text{.} \label{C4}$$It is quite nonstandard compared to customary quartic term $\left( \Delta
^{\ast }\Delta \right) ^{2}$ in conventional universality classes. The GL equations in the homogeneous case for the condensate gives $\Delta
_{0}=U^{\beta }$ with critical exponent $\beta =1,$ different from the mean field value $\beta =1/2$ for the $U\left( 1\right) $ universality class[Herbutbook]{}. The condensation energy density is $f_{0}=-\frac{1}{12\pi
\hbar ^{2}v_{F}^{2}}U^{2-\alpha }$ with $\alpha =-1$. The free energy critical exponent at QCP therefore is also different from the classical $%
\alpha =0$. The magnetic field couples to the order parameter field by a standard minimal substitution. The Ginzburg-Landau approach the only available practical tool to study properties of inhomogeneous configurations[@Abrikosov] in external magnetic field like the Abrikosov vortex systems.
Exciton condensation
====================
Chiral symmetry breaking
------------------------
Assuming that the electric charge $U\left( 1\right) $ and rotation symmetry is unbroken (no superconductivity) for insufficiently strong attractive interaction (pairing) one still can have transitions due to always existing effective repulsion. This possibility was already considered in graphene[Gusynin]{}. Here we note that the symmetry patten might be very different from often invoked relativistic $2+1$ Gross - Neveu model thoroughly studied as a toy model in relativistic quantum field theory. Possible chiral symmetry breaking states were reviewed in subsection IId and we consider first a ground state with nonzero order parameter $$\left\langle P_{3}^{\left( 1\right) }\right\rangle =v\not=0\text{.}
\label{vdef}$$It is convenient to parametrize the gaussian variational ground state by the trace of propagator:
$$G\left( X;X\right) =m/g\text{ }\gamma _{0}\text{,} \label{mdef}$$
with “mass” $m$ determining the order parameter. According to gap equation derived in Appendix B, we have $$\begin{aligned}
G^{-1} &=&\left( i\omega +\mu \right) I-\alpha \cdot \left( \hbar
v_{F}k\right) -m\gamma _{0}; \label{DSE} \\
G &=&-\frac{\left( \mu +i\omega \right) I+\alpha \cdot \left( \hbar
v_{F}k\right) +m\gamma _{0}}{m^{2}+\left( \hbar v_{F}k\right) ^{2}-\left(
\mu +i\ \omega \right) ^{2}\ }\text{,} \notag\end{aligned}$$leads to $$\begin{gathered}
mI=gG^{\prime }\left( \tau ,\mathbf{r};\tau ,\mathbf{r}\right) \gamma
_{0}^{-1}=-\frac{g}{\left( 2\pi \right) ^{3}}\times \label{gapchiral} \\
\int_{\omega ,\mathbf{k}}\frac{m}{m^{2}+\left( \hbar v_{F}\mathbf{k}\right)
^{2}-\left( \mu +i\ \omega \right) ^{2}\ }\text{.} \notag\end{gathered}$$Performing integrations with momentum cutoff $\Lambda $ one obtains for $%
m\not=0$:$$1=-\frac{g}{4\pi \left( \hbar v_{F}\right) ^{2}}\left\{
\begin{array}{c}
\left( \sqrt{m^{2}+\Lambda ^{2}}-m\right) ,\mu \leq m \\
\left( \sqrt{m^{2}+\Lambda ^{2}}-\mu \right) ,\mu >m%
\end{array}%
\right. \text{.} \label{simplified gapeq}$$For $\mu =m$ one obtains the first order transition point (for $\mu >0$)
$$\text{ }\frac{4\pi \left( \hbar v_{F}\right) ^{2}}{\left\vert
g_{c}\right\vert }=\sqrt{\mu ^{2}+\Lambda ^{2}}-\mu \text{,}
\label{Tc_chiral}$$
For larger repulsion, $\left\vert g\right\vert \geqslant \left\vert
g_{c}\right\vert $, the mass is larger:$$m=\mu +\frac{1}{\left( \hbar v_{F}\right) ^{2}}\left( \left\vert
g\right\vert -\left\vert g_{c}\right\vert \right) \left( \frac{\Lambda ^{2}}{%
8\pi }+\frac{2\pi \left( \hbar v_{F}\right) ^{4}}{\left\vert g\right\vert
\left\vert g_{c}\right\vert }\right) . \label{largerm}$$
The free energy, using the Abrikosov formula derived in Appendix B, is ($%
\Lambda >>m$) $$\begin{gathered}
\left( \Omega \left( m\right) -\Omega _{0}\left( m=0\right) \right)
/V=\int_{0}^{m}2m^{2}\frac{d\frac{1}{\left\vert g\right\vert }}{dm}dm
\label{energy_chiral1} \\
=-\int_{0}^{m}\frac{m^{2}}{2\pi \left( \hbar v_{F}\right) ^{2}}dm=-\frac{%
m^{3}}{6\pi \left( \hbar v_{F}\right) ^{2}}\text{ .} \notag\end{gathered}$$For $\mu >m$, since $\frac{d\left( 1/\left\vert g\right\vert \right) }{dm}=%
\frac{1}{4\pi \left( \hbar v_{F}\right) ^{2}}\frac{m}{\sqrt{m^{2}+\Lambda
^{2}}}>0$, the chiral symmetry breaking state will have higher energy than the normal one. For $\left\vert g\right\vert <\left\vert g_{c}\right\vert ,$ the stable ground state is the normal state with $m=0$.
Finite temperature properties including the phase diagram and the QCP at zero chemical potential can be studied along the lines similar to the superconducting transition.
Discussion and conclusions
==========================
Having considered both the superconducting and the excitonic transitions for sufficiently strong attraction or repulsion within the gaussian approximation, a natural question is what happens when we approach the quantum critical point. At criticality various renormalization group methods should be used[@Herbutbook].
Criticality beyond the gaussian approximation
---------------------------------------------
Critical (quantum) fluctuations are expected to be significant in this relatively low dimensional (relativistic 2+1 dimensional system) system. Generally they are not as strong as in 2D statistical system at finite temperature, but stronger than in 3D one. The approximation we have made describes reasonably well “gaussian” fluctuation beyond the region where stronger critical fluctuations in these systems appear and should be treated nonperturbatively[@Rosenstein] typically using variants of the renormalization group approach[@Herbutbook]. The critical exponents in this region differ from the one called “quantum gaussian (BCS)” in ref. [Sachdev]{} and available results are obtained using either $\varepsilon $ expansion[@Gat; @Herbut09] ($\varepsilon =4-d$, where $d=2+1$ is the space-time dimension), $1/N$, where $N$ is the number of fermionic species on the surface[@Gat] and functional (strong coupling) RG[@Janssen] and Monte Carlo simulations[@old; @MC] (with reservations specified below). The universality class of the supconducting transition according to classification proposed in ref. [@Gat] is the chiral XY (symmetry of order parameter $U\left( 1\right) $) with $N=1$. The large $N$ expansion is not reliable for the one component system considered here (but the number might be larger in similar systems for which our approach trivially generalizes), so let us use the $\varepsilon $ expansion.
Using the formulas for the anomalous dimensions of the order parameter (see second reference in [@Gat]), $$\gamma _{\Delta }=\eta =1/4\varepsilon +0.044\varepsilon ^{2}+O\left(
\varepsilon ^{3}\right) \approx 0.294 \label{gamma}$$and its square, $$\gamma _{\Delta ^{2}}=\left( 1+\sqrt{11}\right) /10\varepsilon
+0.065\varepsilon ^{2}\simeq 0.43\varepsilon +0.065\varepsilon ^{2},
\label{gamma2}$$critical exponents are obtained from the hyperscaling relations: $$\begin{aligned}
\alpha &=&2-d/\left( 2-\gamma _{\Delta ^{2}}\right) =\left( \varepsilon
-2\gamma _{\Delta ^{2}}\right) /\left( 2-\gamma _{\Delta ^{2}}\right)
=-0.353, \notag \\
\beta &=&\left( 1+\gamma _{\Delta }\right) /2\left( 2-\gamma _{\Delta
^{2}}\right) =0.515\text{.} \label{alfabeta}\end{aligned}$$and can be compared with those in Table 1 in ref.[@Li14].$\ $The exponents from the $\varepsilon $ expansion were found to be consistent for larger values of $N$ with the latest Monte Carlo simulations [@MC], while consistent with the functional RG[@Janssen]. The critical exponents of the chiral transition belongs to this class with $N=2$.
The corresponding chiral universality class for the exciton condensation in Dirac semi-metal is the Heisenberg $N=2$ ($SU\left( 2\right) $). The critical exponents for this case were also calculated in ref.[Gat,Herbut09,Janssen]{} Recently they were invoked in a discussion of second order quantum transitions in Hubard model on honeycomb lattice[@Sorella]. It should be noted that the flavour symmetries are often broken “explicitly” by some kind of anisotropy. In this case Goldstone bosons acquire a small mass (like pions in quantum chromodynamics in which the chiral symmetry is slightly broken by the light quark masses), although their major properties remain intact. This can be taken into account as a small perturbation[@Chiraldyn].
Experimental feasibility of observation of quantum phase transition
-------------------------------------------------------------------
The best candidate to observe the superconductivity is a topological insulator of the $Bi_{2}Se_{3}$ family. To estimate the pairing efficiency due to phonons, one should rely on recent studies of surface phonons in TI [@DasSarma]. The coupling constant in the Hamiltonian, Eq.([Hamiltonian]{}), is obtained from the exchange of acoustic (Rayleigh) surface phonons $g=\lambda v_{F}^{2}\hbar ^{2}/2\pi \mu $, where $\lambda $ is the dimensionless effective electron - electron interaction constant of order $%
0.1$ (somewhat lower values are obtained in ref.[@Guinea]). It was shown in ref. [@DasSarma] that at zero temperature the ratio of $\lambda $ and $\mu $ is constant with well defined $\mu \rightarrow 0$ limit with value $%
g=0.23$ $eV\ nm^{2}$ for $v_{F}\approx 7\ \cdot 10^{5}m/s$ (for $%
Bi_{2}Se_{3} $). The critical coupling constant $g_{c}$, Eq.(\[g\_c\]), can be estimated from the Debye cutoff $T_{D}=200K$ determining the momentum cutoff $\Lambda =T_{D}/c_{s}$, where $c_{s}$ is the sound velocity. Taking value to be $c_{s}=2\cdot 10^{3}m/s$ (for $Bi_{2}Se_{3}$), one obtains $%
g_{c}=4\pi v_{F}c_{s}\hbar ^{2}/\,T_{D}=0.20$ $eV$ $nm^{2}$.
Of course the Coulomb repulsion might weaken or even overpower the effect of the attraction due to phonons, so that superconductivity does not occur. In TI like $Bi_{2}Se_{3}$ however, the dielectric constant is very large $%
\varepsilon =50$, so that the Coulomb repulsion is weak. Moreover it was found in graphene (that has identical Coulomb interaction), that although the semi-metal does not screen [@CastroNeto], the effects of the Coulomb coupling are surprisingly small, even in leading order in perturbation theory. The superconductivity was observed in these systems that howeved had to be either doped in the bulk or on the surface[@Koren] (by a $Cu$) or by applying pressure[@pressureBiSe]. It is not yet clear whether the observed superconductivity is a bulk or a surface effect.
* *The Dirac semimetal in optically trapped cold atoms[@cold] offers a well controllable system in which this phenomenon occurs both for repulsive interaction (chiral symmetry breaking) and in particular the attractive one (superconductivity) becuase there is no Coulomb repulsion as the atoms are neutral.
Recently after experimental discovery of 3D Dirac semi-metals[@Potemski] the new class of questions similar to those discussed in present paper arise. Extraordinary electronic properties of these Dirac materials[3Dtheory]{} including superconductivity[@Cava] and chiral condensate are being studied theoretically and experimentally.
Conclusions
-----------
We have studied continuous phase transitions in a Dirac semi-metal realized recently as a surface of topological insulator. The noninteracting system is characterized by (nearly) zero density of states on the 2D Fermi manifold. It degenerates into a point when the chemical potential coincides with the Weyl point of the surface states as in the original proposal for a major class of such materials[@Zhang1]. The pairing attraction (the most plausible candidate being surface phonons) therefore has two tasks in order to create the superconducting condensate. The first is to create a pair of electrons (that in the present circumstances means creating two holes as well) and the second is to pair them. To create the charges does not cost much energy since the spectrum of the Weyl semimetal is gapless (massless relativistic fermions); this is effective as long as the coupling $g$ is larger than the critical $g_{c}$, see Eq.(\[g\_c\]). The situation is more reminiscent of the creation of the chiral condensate in relativistic massless four - fermion theory (a 2D version[@Rosenstein] was recently contemplated for graphene [@CastroNeto; @Katsnelson]) than to the BCS or even BEC in condensed matter systems with parabolic dispersion law. Due to the special “ultra-relativistic” nature of the pairing transition at zero temperature as a function of parameters like the pairing interaction strength is unusual: even the mean field critical exponents are different from the standard ones that generally belong to the $U\left( 1\right) $ class of second order phase transitions.
To summarize, we studied the phase diagram of the superconducting and chiral transition at arbitrary chemical potential, effective local interaction strength and temperature $T$. The quantum ($T=0$) critical point appears at zero chemical potential and belongs the $U_{N}\left( 1\right) $ chiral universality class (the subscript denotes number of massless fermions at QCP according to classification in [@Gat; @Sachdev]) for the attraction (superconductivity) and $SU_{N}\left( 2\right) $ for repulsion (exciton condensation).
*Acknowledgements.* We are indebted to C.W. Luo, J.J. Lin and W.B. Jian for explaining details of experiments, and T. Maniv and M. Lewkowicz for valuable discussions. Work of D.L. and B.R. was supported by NSC of R.O.C. Grants No. 98-2112-M-009-014-MY3 and MOE ATU program. The work of D.L. also is supported by National Natural Science Foundation of China (No. 11274018). B.R. is grateful to School of Physics of Peking University for hospitslity.
Path integral derivation of the Gorkov and Dyson - Schwinger equations
======================================================================
General correlations and sources
--------------------------------
We introduce the grassmanian source terms into Matsubara action Eq.(\[S\]):
$$\begin{gathered}
S\left[ \psi ^{+}\mathbf{,}\psi ,J^{+},J\right] =S\left[ \psi ^{+}\mathbf{,}%
\psi \right] \label{A1} \\
-\int_{0}^{\beta }d\tau \int_{r}\left[ \psi _{\sigma }^{+}\left( \tau ,%
\mathbf{r}\right) J_{\sigma }\left( \tau ,\mathbf{r}\right) +\psi _{\sigma
}\left( \tau ,\mathbf{r}\right) J_{\sigma }^{+}\left( \tau ,\mathbf{r}%
\right) \right] \text{.} \notag\end{gathered}$$
The generating functional for the disconnected correlations is $Z\left(
J^{+},J\right) $
$$\begin{aligned}
\left\langle \psi _{\alpha }\left( X\right) \right\rangle &=&\frac{1}{Z}%
\frac{\delta Z}{\left[ -\delta J_{\alpha }^{+}\left( X\right) \right] }\text{%
;} \label{A2} \\
\left\langle T\psi _{\alpha }\left( X\right) \psi _{\beta }\left( X^{\prime
}\right) ...\right\rangle _{u} &=&\frac{1}{Z}\frac{\delta Z}{\left[ -\delta
J_{\alpha }^{+}\left( X\right) \right] \left[ -\delta J_{\beta }^{+}\left(
X^{\prime }\right) \right] ...}\text{,} \notag\end{aligned}$$
where $X=\left( \mathbf{r,}\tau \right) $, $X^{\prime }=\left( \mathbf{r}%
^{\prime },\tau ^{\prime }\right) $. The correlators is (Matsubara) time ordered. The generating functional for the connected correlations are $%
W=-\ln Z\left( J^{+},J\right) $: $$\begin{aligned}
\left\langle \psi _{\alpha }\left( X\right) \right\rangle &=&\frac{\delta W%
}{\delta J_{\alpha }^{+}\left( X\right) }\text{,} \label{A3} \\
\left\langle T\psi _{\alpha }\left( X\right) \psi _{\beta }\left( X^{\prime
}\right) ...\right\rangle &=&-\frac{\delta W}{\left[ -\delta J_{\alpha
}^{+}\left( X\right) \right] \left[ -\delta J_{\beta }^{+}\left( X^{\prime
}\right) \right] ...}\text{.} \notag\end{aligned}$$In particular, we define the normal and anomalous Green’s functions (see ref.[@AGD]) as,$$\begin{aligned}
\left\langle T\psi _{\alpha }\left( X\right) \psi _{\beta }^{+}\left(
X^{\prime }\right) \right\rangle &=&-\frac{\delta W}{\delta J_{\alpha
}\left( X\right) \delta J_{\beta }^{+}\left( X^{\prime }\right) }=-G_{\alpha
\beta }\left( X\mathbf{;}X^{\prime }\right) \text{;} \notag \\
\left\langle T\psi _{\alpha }\left( X\right) \psi _{\beta }\left( X^{\prime
}\right) \right\rangle &=&-\frac{\delta W}{\delta J_{\alpha }^{+}\left(
X\right) \delta J_{\beta }^{+}\left( X^{\prime }\right) }=F_{\alpha \beta
}\left( X\mathbf{;}X^{\prime }\right) \text{;} \notag \\
\left\langle T\psi _{\alpha }^{+}\left( X\right) \psi _{\beta }^{+}\left(
X^{\prime }\right) \right\rangle &=&-\frac{\delta W}{\delta J_{\alpha
}\left( X\right) \delta J_{\beta }\left( X^{\prime }\right) }=F_{\alpha
\beta }^{+}\left( X\mathbf{;}X^{\prime }\right) \label{A4}\end{aligned}$$where we denote $\left\langle \psi _{\alpha }\left( X\right) \right\rangle $ as $\psi _{\alpha }\left( X\right) $, and drop the time ordering operation $T
$ in correlators. For example, $\left\langle \psi _{\alpha }\left( X\right)
\psi _{\beta }\left( X^{\prime }\right) \right\rangle $ stands for $%
\left\langle T\psi _{\alpha }\left( X\right) \psi _{\beta }\left( X^{\prime
}\right) \right\rangle $.
Derivation of the Gor’kov equation for superconductivity
--------------------------------------------------------
Using identities, $$\begin{aligned}
\int D\psi ^{+}D\psi \frac{\delta }{\delta \psi _{\alpha }^{+}\left(
X\right) }e^{-\frac{1}{\hbar }S\left[ \psi ^{+}\mathbf{,}\psi ,J^{+},J\right]
} &=&0\text{;} \label{A5} \\
\int D\psi ^{+}D\psi \frac{\delta }{\delta \psi _{\alpha }\left( X\right) }%
e^{-\frac{1}{\hbar }S\left[ \psi ^{+}\mathbf{,}\psi ,J^{+},J\right] } &=&0%
\text{,} \notag\end{aligned}$$the equations of motions lead to $$\begin{gathered}
\left( \hbar \partial _{\tau }\delta _{\alpha \beta }+K_{\alpha \beta
}\left( \mathbf{\nabla }\right) \right) \psi _{\beta }\left( X\right)
\mathbf{-}g\times \notag \\
\left\langle \psi _{\beta }^{+}\left( X\right) \psi _{\beta }\left( X\right)
\psi _{\alpha }\left( X\right) \right\rangle =J_{\alpha }\left( \tau ,%
\mathbf{r}\right) \text{,} \label{A6} \\
\left( \hbar \partial _{\tau }\delta _{\alpha \beta }-\widehat{K}_{\beta
\alpha }\left( -\mathbf{\nabla }\right) \right) \psi _{\beta }^{+}\left(
X\right) \mathbf{+}g\times \notag \\
\left\langle \psi _{\alpha }^{+}\left( X\right) \psi _{\beta }^{+}\left(
X\right) \psi _{\beta }\left( X\right) \right\rangle =J_{\alpha }^{+}\left(
\tau ,\mathbf{r}\right) \text{. } \notag\end{gathered}$$Generally full correlations of the fields can be expressed via connected correlators. For example,
$$\begin{gathered}
\left\langle \psi _{\beta }^{+}\left( X\right) \psi _{\beta }\left( X\right)
\psi _{\alpha }\left( X\right) \right\rangle =\psi _{\beta }^{+}\left(
X\right) \psi _{\beta }\left( X\right) \psi _{\alpha }\left( X\right) -
\notag \\
\psi _{\alpha }\left( X\right) \left\langle \psi _{\beta }\left( X\right)
\psi _{\beta }^{+}\left( X\right) \right\rangle _{c}+\psi _{\beta }\left(
X\right) \times \label{A7} \\
\left\langle \psi _{\alpha }\left( X\right) \psi _{\beta }^{+}\left(
X\right) \right\rangle _{c}+\left\langle \psi _{\beta }^{+}\left( X\right)
\right\rangle \left\langle \psi _{\beta }\left( X\right) \psi _{\alpha
}\left( X\right) \right\rangle \notag \\
+\left\langle \psi _{\beta }^{+}\left( X\right) \psi _{\beta }\left(
X\right) \psi _{\alpha }\left( X\right) \right\rangle _{c}\text{.} \notag\end{gathered}$$
For Gaussian mean field approximation, we omit higher order connected correlations, like $\left\langle \psi _{\beta }^{+}\left( X\right) \psi
_{\beta }\left( X\right) \psi _{\alpha }\left( X\right) \right\rangle $,
$$\begin{gathered}
J_{\alpha }\left( X\right) =\left( \hbar \partial _{\tau }\delta _{\alpha
\beta }+K_{\alpha \beta }\right) \psi _{\beta }\left( X\right) \mathbf{-}%
g\times \label{A8} \\
\psi _{\beta }^{+}\left( X\right) \psi _{\beta }\left( X\right) \psi
_{\alpha }\left( X\right) -\psi _{\alpha }\left( X\right) \left\langle \psi
_{\beta }\left( X\right) \psi _{\beta }^{+}\left( X\right) \right\rangle _{c}
\notag \\
+\psi _{\beta }\left( X\right) \left\langle \psi _{\alpha }\left( X\right)
\psi _{\beta }^{+}\left( X\right) \right\rangle _{c}+\psi _{\beta
}^{+}\left( X\right) \left\langle \psi _{\beta }\left( X\right) \psi
_{\alpha }\left( X\right) \right\rangle _{c} \notag\end{gathered}$$
Performing functional derivative$\frac{\delta }{\delta J_{\gamma }\left(
X^{\prime }\right) }$ (by using the identity, $\frac{\delta }{\delta
J_{\gamma }\left( X^{\prime }\right) }\psi _{\alpha }\left( X\right)
=\left\langle \psi _{\alpha }\left( X\right) \psi _{\gamma }^{+}\left(
X^{\prime }\right) \right\rangle $, the correlators of odd number of Grassmannians vanishing) and taking at the end $J_{\alpha }\left( X\right)
=0 $, one obtains:
$$\begin{gathered}
\delta \left( X\mathbf{-}X^{\prime }\right) \delta _{\alpha \gamma }=\left(
\partial _{\tau }\delta _{\alpha \beta }+\widehat{K}_{\alpha \beta }\right)
\left\langle \psi _{\beta }\left( X\right) \psi _{\gamma }^{+}\left(
X^{\prime }\right) \right\rangle \mathbf{-}g \notag \\
\times \left\{ -\left\langle \psi _{\alpha }\left( X\right) \psi _{\gamma
}^{+}\left( X^{\prime }\right) \right\rangle \left\langle \psi _{\beta
}\left( X\right) \psi _{\beta }^{+}\left( X\right) \right\rangle \right.
\label{A9} \\
+\left\langle \psi _{\beta }\left( X\right) \psi _{\gamma }^{+}\left(
X^{\prime }\right) \right\rangle \left\langle \psi _{\alpha }\left( X\right)
\psi _{\beta }^{+}\left( X\right) \right\rangle \notag \\
+\left. \left\langle \psi _{\beta }^{+}\left( X\right) \psi _{\gamma
}^{+}\left( X^{\prime }\right) \right\rangle \left\langle \psi _{\beta
}\left( X\right) \psi _{\alpha }\left( X\right) \right\rangle \right\} \text{%
,} \notag\end{gathered}$$
or
$$\begin{gathered}
\delta \left( X-X^{\prime }\right) \delta _{\alpha \gamma }=\left( -\partial
_{\tau }\delta _{\alpha \beta }-\widehat{K}_{\alpha \beta }\right) G_{\beta
\gamma }\left( X,X^{\prime }\right) \mathbf{-}g \notag \\
\times \left\{ -G_{_{\alpha }\gamma }\left( X,X^{\prime }\right) G_{\beta
\beta }\left( X,X\right) \right. \label{A10} \\
+G_{\beta \gamma }\left( X,X^{\prime }\right) G_{\alpha \beta }\left(
X,X\right) _{c}+ \notag \\
\left. F_{\beta \gamma }^{+}\left( X,X^{\prime }\right) F_{\beta \alpha
}\left( X,X\right) \right\} \text{.} \notag\end{gathered}$$
Similarly the second equation of motion,
$$\begin{gathered}
J_{\alpha }^{+}\left( \tau ,\mathbf{r}\right) =\left( \partial _{\tau
}\delta _{\alpha \beta }-\widehat{K}_{\beta \alpha }\left( -\mathbf{\nabla }%
\right) \right) \left\langle \psi _{\beta }^{+}\left( \tau ,\mathbf{r}%
\right) \right\rangle \mathbf{+}g \notag \\
\left\{ \left\langle \psi _{\alpha }^{+}\left( \tau ,\mathbf{r}\right) \psi
_{\beta }^{+}\left( \tau ,\mathbf{r}\right) \right\rangle \left\langle \psi
_{\beta }\left( \tau ,\mathbf{r}\right) \right\rangle \right. \label{A11} \\
-\left\langle \psi _{\alpha }^{+}\left( \tau ,\mathbf{r}\right)
\right\rangle \left\langle \psi _{\beta }\left( \tau ,\mathbf{r}\right) \psi
_{\beta }^{+}\left( \tau ,\mathbf{r}\right) \right\rangle \notag \\
+\left. \left\langle \psi _{\beta }^{+}\left( \tau ,\mathbf{r}\right)
\right\rangle \left\langle \psi _{\beta }\left( \tau ,\mathbf{r}\right) \psi
_{\alpha }^{+}\left( \tau ,\mathbf{r}\right) \right\rangle \right\} \text{,}
\notag\end{gathered}$$
gives $$\begin{gathered}
\left( \partial _{\tau }\delta _{\alpha \beta }-\widehat{K}_{\beta \alpha
}\left( -\mathbf{\nabla }\right) \right) F_{\beta \gamma }^{+}\left(
X,X^{\prime }\right) +g\left\{ -F_{\alpha \beta }^{+}\left( X,X\right)
\right. \notag \\
G_{\beta \gamma }\left( X,X^{\prime }\right) +F_{\alpha \gamma }^{+}\left(
X,X^{\prime }\right) G_{\beta \beta }\left( X,X\right) + \label{A12} \\
\left. G_{\beta \alpha }\left( X,X\right) F_{\beta \gamma }^{+}\left(
X,X^{\prime }\right) \right\} =0\text{.} \notag\end{gathered}$$
In the superconducting phase, if the spin rotation invariance is not broken and chiral symmetry is preserved, $G_{\alpha \beta }\left( X,X\right)
=G_{\alpha \beta }^{c}=n_{c}\delta _{\alpha \beta }$. For the $N$ component spinors $n_{c}=n/N$, where $n$ is the density of electrons. The quadratic parts of Eqs.(\[A10\],\[A12\]) simplify: $$\begin{gathered}
\mathbf{-}g\left\{ -G_{\alpha \gamma }\left( X,X^{\prime }\right) G_{\beta
\beta }\left( X,X\right) +G_{\beta \gamma }\left( X,X^{\prime }\right)
G_{\alpha \beta }\left( X,X\right) \right\} \notag \\
=\mathbf{-}g\left\{ -G_{\alpha \gamma }\left( X,X^{\prime }\right)
Nn_{c}+n_{c}G_{\alpha \gamma }\left( X,X^{\prime }\right) \right\} \notag \\
=g\left( N-1\right) n_{c}G_{\alpha \gamma }\left( X,X^{\prime }\right) ;
\label{A13} \\
g\left\{ F_{\alpha \gamma }^{+}\left( X,X^{\prime }\right) G_{\beta \beta
}\left( X,X\right) +G_{\beta \alpha }\left( X,X\right) F_{\beta \gamma
}^{+}\left( X,X^{\prime }\right) \right\} \notag \\
=-g\left( N-1\right) n_{c}F_{\alpha \gamma }^{+}\left( X,X^{\prime }\right)
\text{,} \notag\end{gathered}$$and such terms can be absorbed to the chemical potential term with the chemical potential replaced by the renormalized
$$\mu +g\left( N-1\right) n_{c}=\mu _{R}\text{.} \label{A14}$$
Therefore we finally obtain the Gor’kov equations, given in Eqs.(\[GeqX\]).
Derivation of DS equation and renormalized chemical potential for nonsuperconducting state.
-------------------------------------------------------------------------------------------
We need only to discuss the equation for $\frac{\delta J_{\alpha }}{\delta
\psi _{\beta }}$, as$\frac{\delta J_{\alpha }^{+}}{\delta \psi _{\beta }^{+}}
$ is just the complex conjugate of $\frac{\delta J_{\alpha }}{\delta \psi
_{\beta }}$ and $\frac{\delta J_{\alpha }^{+}}{\delta \psi _{\beta }}=\frac{%
\delta J_{\alpha }}{\delta \psi _{\beta }^{+}}=0$. We reorganize the equation $\frac{\delta J_{\alpha }}{\delta \psi _{\beta }}$as $$\begin{gathered}
\frac{\delta J_{\alpha }}{\delta \psi _{\beta }}=\left( \partial _{\tau
}\delta _{\alpha \beta }+K_{\alpha \beta }\right) \delta \left( X-X^{\prime
}\right) -g\delta \left( X-X^{\prime }\right) \times \notag \\
\left\{ \left\langle \psi _{\alpha }\left( X\right) \psi _{\beta }^{+}\left(
X\right) \right\rangle _{c}-\frac{1}{4}\delta _{\alpha \beta }\sum
\left\langle \psi _{\gamma }\left( X\right) \psi _{\gamma }^{+}\left(
X\right) \right\rangle _{c}\right\} \notag \\
+g\delta _{\alpha \beta }\delta \left( X-X^{\prime }\right) \left\{
\left\langle \psi _{\beta }\left( X\right) \psi _{\beta }^{+}\left( X\right)
\right\rangle _{c}-\right. \label{A15} \\
\left. \sum \left\langle \psi _{\gamma }\left( X\right) \psi _{\gamma
}^{+}\left( X\right) \right\rangle _{c}\right\} \text{,} \notag\end{gathered}$$
The last two terms are proportional to $\delta _{\alpha \beta }$, and can be absorbed to the chemical potential, $$\begin{gathered}
\mu +g\delta _{\alpha \beta }\left\{ \sum \left\langle \psi _{\gamma }\left(
X\right) \psi _{\gamma }^{+}\left( X\right) \right\rangle _{c}\right.
\label{A16} \\
-\left. \left\langle \psi _{\beta }\left( X\right) \psi _{\beta }^{+}\left(
X\right) \right\rangle _{c}\right\} =\mu _{R\text{.}} \notag\end{gathered}$$
The gap equation can be recasted as
$$\begin{aligned}
G^{-1} &=&G_{0}^{-1}+g\delta \left( X-X^{\prime }\right) G^{\prime }\left(
X;X^{\prime }\right) \notag \\
G_{0}^{-1} &=&\left( \partial _{\tau }\delta _{\alpha \beta }+K_{\alpha
\beta }\right) \delta \left( X-X^{\prime }\right) \label{A17} \\
G_{\alpha \beta }^{\prime }\left( X;X^{\prime }\right) &=&-\left\{
\left\langle \psi _{\alpha }\left( X\right) \psi _{\beta }^{+}\left(
X^{\prime }\right) \right\rangle _{c}-\right. \notag \\
&&\left. \frac{1}{4}\delta _{\alpha \beta }\sum \left\langle \psi _{\gamma
}\left( X\right) \psi _{\gamma }^{+}\left( X^{\prime }\right) \right\rangle
_{c}\right\} , \notag\end{aligned}$$
where $G_{\alpha \beta }^{\prime }$ is the traceless part of $G_{\alpha
\beta }$, and the chemical potential in $K_{\alpha \beta }$ is $\mu _{R\text{%
.}}$
A formula for energy of the superconducting state
=================================================
Gap equation in the Nambu notation
----------------------------------
A compact representation of the Gorkov equations for superconductors is the Nambu notations$$\begin{aligned}
\mathcal{G}_{\alpha \beta }\left( X,X^{\prime }\right) &=&%
\begin{pmatrix}
\frac{\delta \psi _{\alpha }\left( X\right) }{\delta J_{\beta }\left(
X^{\prime }\right) } & \frac{\delta \psi _{\alpha }\left( X\right) }{\delta
J_{\beta }^{+}\left( X^{\prime }\right) } \\
\frac{\delta \psi _{\alpha }^{+}\left( X\right) }{\delta J_{\beta }\left(
X^{\prime }\right) } & \frac{\delta \psi _{\alpha }^{+}\left( X\right) }{%
\delta J_{\beta }^{+}\left( X^{\prime }\right) }%
\end{pmatrix}
\label{B1} \\
&=&%
\begin{pmatrix}
-G_{\alpha \beta }\left( x,\tau ;x^{\prime },\tau ^{\prime }\right) &
F_{\alpha \beta }\left( x,\tau ;x^{\prime },\tau ^{\prime }\right) \\
F_{\alpha \beta }^{+}\left( x,\tau ;x^{\prime },\tau ^{\prime }\right) &
G_{\beta \alpha }\left( x^{\prime },\tau ^{\prime };x,\tau \right)%
\end{pmatrix}
\notag\end{aligned}$$
so that
$$\begin{pmatrix}
\frac{\delta J_{\alpha }}{\delta \psi _{\beta }} & \frac{\delta J_{\alpha }}{%
\delta \psi _{\beta }^{+}} \\
\frac{\delta J_{\alpha }^{+}}{\delta \psi _{\beta }} & \frac{\delta
J_{\alpha }^{+}}{\delta \psi _{\beta }^{+}}%
\end{pmatrix}%
=\mathcal{G}_{\alpha \beta }^{-1}\text{.} \label{B2}$$
.The functional identity
$$\begin{gathered}
\frac{\delta J_{\alpha }\left( X\right) }{\delta \psi _{\beta }}\frac{\delta
\psi _{\beta }}{\delta J_{\gamma }\left( X^{\prime }\right) }+\frac{\delta
J_{\alpha }\left( X\right) }{\delta \psi _{\beta }^{+}}\times \label{B3} \\
\frac{\delta \psi _{\beta }^{+}}{\delta J_{\gamma }\left( X^{\prime }\right)
}=\delta _{\alpha \gamma }\left( X-X^{\prime }\right) \text{,} \notag\end{gathered}$$
where abbreviations $$\begin{gathered}
\frac{\delta J_{\alpha }\left( X\right) }{\delta \psi _{\beta }}\frac{\delta
\psi _{\beta }}{\delta J_{\gamma }\left( X^{\prime }\right) }\equiv
\sum_{\beta }\int dX^{\prime \prime }\frac{\delta J_{\alpha }\left( X\right)
}{\delta \psi _{\beta }\left( X^{\prime \prime }\right) }\times \\
\frac{\delta \psi _{\beta }\left( X^{\prime \prime }\right) }{\delta
J_{\gamma }\left( X^{\prime }\right) }\text{, \ \ }\delta _{\alpha \gamma
}\left( X-X^{\prime }\right) =\delta _{\alpha \gamma }\delta \left(
X-X^{\prime }\right) , \notag\end{gathered}$$are used. In the Nambu matrix form it reads
$$\begin{pmatrix}
\frac{\delta J_{\alpha }}{\delta \psi _{\beta }} & \frac{\delta J_{\alpha }}{%
\delta \psi _{\beta }^{+}} \\
\frac{\delta J_{\alpha }^{+}}{\delta \psi _{\beta }} & \frac{\delta
J_{\alpha }^{+}}{\delta \psi _{\beta }^{+}}%
\end{pmatrix}%
\begin{pmatrix}
\frac{\delta \psi _{\beta }}{\delta J_{\gamma }} & \frac{\delta \psi _{\beta
}}{\delta J_{\gamma }^{+}} \\
\frac{\delta \psi _{\beta }^{+}}{\delta J_{\gamma }} & \frac{\delta \psi
_{\beta }^{+}}{\delta J_{\gamma }^{+}}%
\end{pmatrix}%
=%
\begin{pmatrix}
\delta _{\alpha \gamma } & 0 \\
0 & \delta _{\alpha \gamma }%
\end{pmatrix}%
\text{.} \label{B4}$$
The derivatives are:$$\begin{aligned}
\frac{\delta J_{\alpha }\left( X\right) }{\delta \psi _{\beta }\left(
X^{\prime }\right) } &=&\left( \partial _{\tau }\delta _{\alpha \beta
}+K_{\alpha \beta }\right) \delta \left( X-X^{\prime }\right) ; \notag \\
\frac{\delta J_{\alpha }\left( X\right) }{\delta \psi _{\beta }^{+}\left(
X^{\prime }\right) } &=&g\delta \left( X-X^{\prime }\right) \left\langle
\psi _{\alpha }\left( \tau ,\mathbf{r}\right) \psi _{\beta }\left( \tau ,%
\mathbf{r}\right) \right\rangle _{c}; \notag \\
\frac{\delta J_{\alpha }^{+}\left( X\right) }{\delta \psi _{\beta }\left(
X^{\prime }\right) } &=&g\delta \left( X-X^{\prime }\right) \left\langle
\psi _{\alpha }^{+}\left( \tau ,\mathbf{r}\right) \psi _{\beta }^{+}\left(
\tau ,\mathbf{r}\right) \right\rangle ; \label{B5} \\
\frac{\delta J_{\alpha }^{+}\left( X\right) }{\delta \psi _{\beta
}^{+}\left( X^{\prime }\right) } &=&\left( \partial _{\tau }\delta _{\alpha
\beta }-K_{\beta \alpha }\left( -\mathbf{\nabla }\right) \right) \delta
\left( X-X^{\prime }\right) . \notag\end{aligned}$$For the non-interacting model
$$\begin{aligned}
\mathcal{G}_{0}^{-1} &=&%
\begin{pmatrix}
\left( \mathcal{G}_{0}^{-1}\right) _{11} & 0 \\
0 & \left( \mathcal{G}_{0}^{-1}\right) _{22}%
\end{pmatrix}%
\text{,} \notag \\
\left( \mathcal{G}_{0}^{-1}\right) _{11} &=&\left( \partial _{\tau }\delta
_{\alpha \beta }+K_{\alpha \beta }\left( \mathbf{\nabla }\right) \right)
\delta \left( x-x^{\prime },\tau -\tau ^{\prime }\right) ; \label{B6} \\
\left( \mathcal{G}_{0}^{-1}\right) _{22} &=&\left( \partial _{\tau }\delta
_{\alpha \beta }-K_{\beta \alpha }\left( -\mathbf{\nabla }\right) \right)
\delta \left( x-x^{\prime },\tau -\tau ^{\prime }\right) , \notag\end{aligned}$$
and the gap equation can be cast in the Dyson form
$$\begin{aligned}
\mathcal{G}^{-1} &=&\mathcal{G}_{0}^{-1}+%
\begin{pmatrix}
0 & \Sigma _{12} \\
\Sigma _{21} & 0%
\end{pmatrix}%
\text{;} \notag \\
\Sigma _{12} &=&g\delta \left( x-x^{\prime },\tau -\tau ^{\prime }\right)
\left\langle \psi _{\alpha }\left( \tau ,\mathbf{r}\right) \psi _{\beta
}\left( \tau ,\mathbf{r}\right) \right\rangle _{c}\text{;} \\
\Sigma _{21} &=&g\delta \left( x-x^{\prime },\tau -\tau ^{\prime }\right)
\left\langle \psi _{\alpha }^{+}\left( \tau ,\mathbf{r}\right) \psi _{\beta
}^{+}\left( \tau ,\mathbf{r}\right) \right\rangle . \notag\end{aligned}$$
Derivation of the expression for energy density
-----------------------------------------------
The free energy is
$$\Omega \lbrack \mathcal{G}]=-\frac{1}{\beta }\left\{ {\frac{1}{2}}\text{Tr}%
\{-\ln \mathcal{G}+[\mathcal{G}_{0}^{-1}\mathcal{G}-1]\}+\Phi \lbrack
\mathcal{G}]\ \right\} \text{,} \label{B8}$$
where in Gaussian approximation,
$$\Phi \lbrack \mathcal{G}]=\frac{g}{2}\int_{\tau ,r}F_{\alpha \beta }\left(
X;X\right) F_{\beta \alpha }^{+}\left( X;X\right) . \label{B9}$$
The gap equation can be also obtained by $\frac{\delta }{\delta \mathcal{G}}%
\Omega \lbrack \mathcal{G}]=0$.
The energy difference between the superconducting state and normal state can be obtained by the differentiating of the grand canonical potential with respect to the coupling constant:$\frac{d}{dg}\Omega \lbrack \mathcal{G}]$. The green function $\mathcal{G}$ is dependent on $g$, but due to $\frac{%
\delta }{\delta \mathcal{G}}\Omega \left[ \mathcal{G},g\right] =0$, $$\begin{aligned}
\frac{d}{dg}\Omega \left[ \mathcal{G},g\right] &=&\frac{\partial }{\partial g%
}\Omega \left[ \mathcal{G},g\right] \label{B10} \\
&=&-\frac{1}{2\beta }\int_{\tau ,r}F_{\alpha \beta }\left( X;X\right)
F_{\beta \alpha }^{+}\left( X;X\right) \text{.} \notag\end{aligned}$$For homogeneous state, $F_{\alpha \beta }\left( X;X\right) ,F_{\beta \alpha
}^{+}\left( X;X\right) $ are constant (not dependent on $x,\tau $), and we introduce
$$\begin{aligned}
F_{\alpha \beta }\left( X;X\right) &=&\frac{1}{g}\Delta _{\alpha \beta
},F_{\alpha \beta }^{+}\left( X;X\right) \label{B11} \\
&=&\frac{1}{g}\Delta _{\alpha \beta }^{+}=\frac{1}{g}\Delta _{\beta \alpha
}^{\ast }\text{,} \notag\end{aligned}$$
we obtain the free energy formula [@Abrikosov]$$\begin{aligned}
\frac{d}{dg}\Omega \lbrack \mathcal{G]}\mathcal{=-} &&\frac{V}{2}\frac{1}{%
g^{2}}\text{Tr}\left( \Delta \Delta ^{+}\right) \rightarrow \label{B12} \\
d\Omega \lbrack \mathcal{G]} &\mathcal{=}&\frac{V}{2}d\left( \frac{1}{g}%
\right) \text{Tr}\left( \Delta \Delta ^{+}\right) . \notag\end{aligned}$$
Gap equation and free energy for chiral symmetry breaking states
================================================================
Derivation of the gap equation
------------------------------
We will discuss the non-superconducting state with $G_{\alpha \beta }\left(
X,X\right) \neq n_{c}\delta _{\alpha \beta }$, $F_{\alpha \beta }\left(
X;X^{\prime }\right) =0$, which happens for example in the case of chiral symmetry breaking (CSB) state. We need only to discuss the equation for $%
\frac{\delta J_{\alpha }}{\delta \psi _{\beta }}$, as$\frac{\delta J_{\alpha
}^{+}}{\delta \psi _{\beta }^{+}}$ is just the complex conjugate of $\frac{%
\delta J_{\alpha }}{\delta \psi _{\beta }}$ and $\frac{\delta J_{\alpha }^{+}%
}{\delta \psi _{\beta }}=\frac{\delta J_{\alpha }}{\delta \psi _{\beta }^{+}}%
=0$. We reorganize the equation $\frac{\delta J_{\alpha }}{\delta \psi
_{\beta }}$as $$\begin{gathered}
\frac{\delta J_{\alpha }}{\delta \psi _{\beta }}=\left( \partial _{\tau
}\delta _{\alpha \beta }+\widehat{K}_{\alpha \beta }\right) \delta \left(
X-X^{\prime }\right) -g\delta \left( X-X^{\prime }\right) \times \notag \\
\left\{ \left\langle \psi _{\alpha }\left( X\right) \psi _{\beta }^{+}\left(
X\right) \right\rangle _{c}-\frac{1}{4}\delta _{\alpha \beta }\right. \times
\\
\left. \sum \left\langle \psi _{\gamma }\left( X\right) \psi _{\gamma
}^{+}\left( X\right) \right\rangle _{c}\right\} +g\delta _{\alpha \beta
}\delta \left( X-X^{\prime }\right) \times \notag \\
\left\{ \left\langle \psi _{\beta }\left( X\right) \psi _{\beta }^{+}\left(
X\right) \right\rangle _{c}-\sum \left\langle \psi _{\gamma }\left( X\right)
\psi _{\gamma }^{+}\left( X\right) \right\rangle _{c}\right\} \text{,}
\notag\end{gathered}$$
The last two terms are proportional to $\delta _{\alpha \beta }$, and can be absorbed to the chemical potential, $$\begin{gathered}
\mu +g\delta _{\alpha \beta }\left\{ \sum \left\langle \psi _{\gamma }\left(
X\right) \psi _{\gamma }^{+}\left( X\right) \right\rangle _{c}-\right. \\
\left. \left\langle \psi _{\beta }\left( X\right) \psi _{\beta }^{+}\left(
X\right) \right\rangle _{c}\right\} =\mu _{R\text{.}} \notag\end{gathered}$$The gap equation can be recasted as
$$\begin{aligned}
G^{-1} &=&G_{0}^{-1}+g\delta \left( X-X^{\prime }\right) G^{\prime }\left(
X;X^{\prime }\right) \notag \\
G_{0}^{-1} &=&\left( \partial _{\tau }\delta _{\alpha \beta }+\widehat{K}%
_{\alpha \beta }\right) \delta \delta \left( X-X^{\prime }\right) \notag \\
G_{\alpha \beta }^{\prime }\left( X;X^{\prime }\right) &=&-\left\{
\left\langle \psi _{\alpha }\left( X\right) \psi _{\beta }^{+}\left(
X^{\prime }\right) \right\rangle _{c}-\right. \label{B14} \\
&&\left. \frac{1}{4}\delta _{\alpha \beta }\sum \left\langle \psi _{\gamma
}\left( X\right) \psi _{\gamma }^{+}\left( X^{\prime }\right) \right\rangle
_{c}\right\} , \notag\end{aligned}$$
where $G_{\alpha \beta }^{\prime }$ is the traceless part of $G_{\alpha
\beta }$, and the chemical potential in $\widehat{K}_{\alpha \beta }$ is $%
\mu _{R\text{.}}$The free energy is now $$\begin{aligned}
\Omega \lbrack \mathcal{G}] &=&-\left( \beta \right) ^{-1}\left\{ \text{Tr}%
\{-\ln G+[G_{0}^{-1}G-1]\}\right. \\
&&\left. +\Phi \lbrack G]\right\} \text{,} \notag\end{aligned}$$where $$\Phi \lbrack \mathcal{G}]=-\frac{g}{2}\int_{\tau ,r}G_{\alpha \beta
}^{\prime }\left( x,\tau ;x,\tau \right) G_{\beta \alpha }^{\prime }\left(
x,\tau ;x,\tau \right) \text{.}$$A similar free energy equation can be obtained,
$$\begin{aligned}
\frac{d}{dg}\Omega \left[ \mathcal{G},g\right] &=&\frac{\partial }{\partial g%
}\Omega \left[ \mathcal{G},g\right] \label{chiralfree} \\
&=&\frac{1}{2\beta }\int_{\tau ,r}G_{\alpha \beta }^{\prime }\left( x,\tau
;x,\tau \right) G_{\beta \alpha }^{\prime }\left( x,\tau ;x,\tau \right) .
\notag\end{aligned}$$
Details of the calculation of the condensate and energy
-------------------------------------------------------
Using the integral
$$\begin{gathered}
\int_{\omega }\frac{1}{m^{2}+\left( \hbar v_{F}k\right) ^{2}-\left( \mu +i\
\omega \right) ^{2}\ }= \\
\frac{\pi }{\sqrt{m^{2}+\left( \hbar v_{F}k\right) ^{2}}}\Theta \left( \sqrt{%
m^{2}+\left( \hbar v_{F}k\right) ^{2}}-\mu \right) , \notag\end{gathered}$$
and $$\begin{gathered}
\frac{1}{\left( 2\pi \right) ^{3}}\int_{\omega ,p}\frac{1}{%
m^{2}+p^{2}-\left( \mu +i\ \omega \right) ^{2}\ } \notag \\
=\frac{1}{4\pi }\int_{p=0}^{\Lambda }\frac{p\Theta \left( \sqrt{m^{2}+p^{2}}%
-\mu \right) }{\sqrt{m^{2}+p^{2}}} \\
=\left\{
\begin{array}{c}
\frac{1}{4\pi }\left( \sqrt{m^{2}+\Lambda ^{2}}-m\right) ,\mu \leq m \\
\frac{1}{4\pi }\left( \sqrt{m^{2}+\Lambda ^{2}}-\mu \right) ,\mu >m%
\end{array}%
\right. \notag\end{gathered}$$
the gap equation becomes: $$m=-\frac{g}{4\pi \left( \hbar v_{F}\right) ^{2}}\left\{
\begin{array}{c}
\left( \sqrt{m^{2}+\Lambda ^{2}}-m\right) ,\mu \leq m \\
\left( \sqrt{m^{2}+\Lambda ^{2}}-\mu \right) ,\mu >m%
\end{array}%
\right. \text{.}$$
Symmetries of the Dirac model
=============================
The chiral nonrelativistic $SU\left( 2\right) .$
------------------------------------------------
Generally the charge algebra is $\left[ Q,Q_{i}\right] =0;\left[ Q_{i},Q_{j}%
\right] =i\varepsilon _{ijk}Q_{k}$. Indeed all three chiral charges commute with density:$$\begin{aligned}
\left[ \rho \left( x\right) ,Q_{i}\right] &=&\int_{y}\left[ \psi ^{\dagger
}\left( x\right) \psi \left( x\right) ,\psi ^{\dagger }\left( y\right)
T_{i}\psi \left( y\right) \right] \label{C1a} \\
&=&\psi ^{\dagger }\left( x\right) \left[ I,T_{i}\right] \psi \left(
x\right) =0, \notag\end{aligned}$$and the kinetic term,
$$\begin{aligned}
\left[ H,Q_{i}\right] &=&\int_{y}\left[ \psi ^{\dagger }\left( x\right)
\left( \alpha _{1}p_{1}+\alpha _{2}p_{2}\right) \psi \left( x\right) ,\psi
^{\dagger }\left( y\right) T_{i}\psi \left( y\right) \right] \notag \\
&=&\psi ^{\dagger }\left( x\right) \left[ \alpha _{1}p_{1}+\alpha
_{2}p_{2},T_{i}\right] \psi \left( x\right) =0. \label{C2}\end{aligned}$$
The three generator matrices, Eq.(\[Tmat\]), constituting the $SU\left(
2\right) $, $\left[ T_{i},T_{j}\right] =i\varepsilon _{ijk}T_{k}$, commute with both $\gamma _{0}\gamma _{1}$ and $\gamma _{0}\gamma _{2}$ that appear in noninteracting Hamiltonian, Eq.(\[Hamiltonian\]). The density - density interactions part of Hamiltonian also commute with the $SU\left( 2\right) $ charges since
$$\begin{aligned}
\left[ \rho \left( r\right) ,Q_{i}\right] &=&\int_{r^{\prime }}\left[ \psi
^{\dagger }\left( r\right) \psi \left( r\right) ,\psi ^{\dagger }\left(
r^{\prime }\right) T_{i}\psi \left( r^{\prime }\right) \right] \label{C3a}
\\
&=&\psi ^{\dagger }\left( r\right) \left[ I,T_{i}\right] \psi \left(
r\right) =0. \notag\end{aligned}$$
Correspondingly the action Eq.(\[relativistic action\]) is invariant under $\delta \psi =iT_{i}\psi ;\ \delta \overline{\psi }=i\overline{\psi }\gamma
_{0}T_{i}\gamma _{0}$. Indeed both the energy term,
$$\begin{aligned}
\delta A_{0} &=&-\int \left[ i\overline{\psi }\gamma _{0}T_{i}\gamma
_{0}\gamma _{0}\partial _{0}\psi +\overline{\psi }^{s}\gamma _{0}\partial
_{0}iT_{i}\psi \right] \\
&=&-\int \left[ -i\overline{\psi }\gamma _{0}T_{i}\partial _{0}\psi ^{s}+%
\overline{\psi }^{s}\gamma _{0}\partial _{0}iT_{i}\psi ^{s}\right] =0,
\notag\end{aligned}$$
and the momentum terms,
$$\begin{aligned}
\delta A_{1} &=&-\int \left[ i\overline{\psi }\gamma _{0}T_{i}\gamma
_{0}\left( \mathbf{\gamma }\cdot \mathbf{\partial }\right) \psi ^{s}+%
\overline{\psi }\left( \gamma \cdot \partial \right) iT_{i}\psi \right]
\notag \\
&=&-i\int \overline{\psi }\left[ \gamma _{0}T_{i}\gamma _{0}\mathbf{\gamma }+%
\mathbf{\gamma }T_{i}\right] \cdot \mathbf{\partial }\psi =0,\end{aligned}$$
are invariant. The nonrelativistic interaction term$$\begin{aligned}
\delta A &=&v\int \left[ i\overline{\psi }\gamma _{0}T_{i}\gamma _{0}\gamma
_{0}\psi +\overline{\psi }\gamma _{0}iT_{i}\psi \right] \left( \overline{%
\psi }^{r}\gamma _{0}\psi ^{r}\right) \\
&=&iv\int \left[ -\overline{\psi }\gamma _{0}T_{i}\psi +\overline{\psi }%
\gamma _{0}T_{i}\psi \right] \left( \overline{\psi }^{r}\gamma _{0}\psi
^{r}\right) =0. \notag\end{aligned}$$
[99]{} S.-Q. Shen, *Topological Insulators*, Springer-Verlag, Heidelberg (2012); X.-L. Qi and S.-C. Zhang, Rev. Mod. Phys. **83**, 1057 (2011); M.Z. Hasan, C.L. Kane, Rev.Mod. Phys. **82**, 3045 (2010).
M. I. Katsnelson, *Graphene: Carbon in Two Dimensions*, Cambridge University Press, Cambridge (2012);A. H. Castro Neto, F. Guinea, N. M. R. Peres, K. S. Novoselov and A. K. Geim, Rev. Mod. Phys. **81**, 109 (2009).
A. M. Black-Schaffer and S. Doniach, Phys. Rev. B **75**, 134512 (2007); S. Pathak, V. B. Shenoy, and G. Baskaran, Phys. Rev. B **81**, 085431 (2010); R. Nandkishore, L. S. Levitov, and A. V. Chubukov, Nat. Phys. **8** (2011); B. Roy and I. F. Herbut, Phys. Rev. B **82**, 035429 (2010).
D. V. Khveshchenko, Phys. Rev. Lett. **87**, 246802 (2001).
O. V. Gamayun, E. V. Gorbar, V. P. Gusynin, Phys. Rev. B **81**, 075429 (2010); B. Rosenstein and B.J. Warr, Phys. Lett. B **218**, 465 (1989); M. V. Ulybyshev, P. V. Buividovich, M. I. Katsnelson, M. I. Polikarpov, Phys. Rev. Lett. **111**, 056801 (2013).
J. Martin, B. E. Feldman, R. T. Weitz, M. T. Allen, and A. Yacoby, Phys. Rev. Lett. **105**, 256806 (2010); R. T. Weitz, M. T. Allen, B. E. Feldman, J. Martin, and A. Yacoby, Science **330**, 812 (2010); F. Freitag, J. Trbovic, M. Weiss, and C. Schonenberger, Phys. Rev. Lett. **108**, 076602 (2012); J. Velasco Jr., L. Jing, W. Bao, Y. Lee, P. Kratz, V. Aji, M. Bockrath, C.N. Lau, C. Varma, R. Stillwell, D. Smirnov, F. Zhang, J. Jung, and A.H. MacDonald, Nature Nanotechnology **7**, 156 (2012).
V. M. Nabutovskii and B. Ya. Shapiro, Zh. Eksp. Teor. Fiz. **84**, 42243 (Sov. Phys. JETP 57 (I)).
P. A. Lee, N. Nagaosa, X.-G. Wen, Rev. Mod. Phys. **78**, 17 (2006); J. Orenstein and A. J. Millis, Science **288**, 468 (2000).
J. Singleton and C. Mielke, Cont. Phys. **43**, 63 (2002).
I.N. Khlyustikov and A.I. Buzdin, Adv. Phys. **36**, 271 (1987).
X. Zhu, L. Santos, R. Sankar, S. Chikara, C. Howard, F.C. Chou, C. Chamon, M. El-Batanouny, Phys. Rev. Lett. **107**, 186102 (2011); C. W. Luo, H. J. Wang, S. A. Ku, H.-J. Chen, T. T. Yeh, J.-Y. Lin, K. H. Wu, J. Y. Juang, B. L. Young, T. Kobayashi, C.-M. Cheng, C.-H. Chen, K.-D. Tsuei, R. Sankar, F. C. Chou, K. A. Kokh, O. E. Tereshchenko, E. V. Chulkov, Yu. M. Andreev, and G. D. Gu, Nano Lett. **13**, 5797 (2013); X. Zhu, L. Santos, C. Howard, R. Sankar, F.C. Chou, C. Chamon, M. El-Batanouny, Phys. Rev. Lett. **108**, 185501 (2012).
S. Das Sarma and Qiuzi Li, Phys. Rev. B **88**, 081404(R) (2013); Z.-H. Pan, AV Fedorov, D. Gardner, Y. S. Lee, S. Chu, and T. Valla, Phys. Rev. Lett. **108**, 187001 (2012); V. Parente, A. Tagliacozzo, F. von Oppen, and F. Guinea, Phys. Rev.* B*, **88** (2013) 075432; M. Cheng, R. M. Lutchyn, and S. Das Sarma, Phys. Rev B **85**, 165124 (2012).
D. Li, B. Rosenstein, B. Ya. Shapiro, and I. Shapiro, Phys. Rev. B **90,** 054517 (2014).
H. Zhang, C.-X. Liu, X.-L. Qi, X. Dai, Z. Fang, and S.-C. Zhang, Nat. Phys.** 5**, 438 (2009).
J. G. Checkelsky, Y. S. Hor, R. J. Cava, and N. P. Ong, Phys. Rev. Lett. **106**, 196801 (2011); D. Kim, S. Cho, N. P. Butch, P. Syers, K. Kirshenbaum, S. Adam, J. Paglione and M. S. Fuhrer, Nat. Phys. **8**, 459 (2012).
C.-K. Lu and I. F. Herbut, Phys. Rev. B **82**, 144505 (2010).
M. Sato and S. Fujimoto, Phys. Rev B **79**, 094504 (2009).
S. Sachdev, *Quantum Phase Transitions*, Cambridge University Press, Cambridge (2011).
I. Herbut, *A Modern Approach to Critical Phenomena*, Cambridge University Press, Cambridge (2010); D. J. Amit, *Field Theory, The Renormalization Group and Critical Phenomena*, World Scientific, London (2005).
B. Rosenstein, B. J. Warr and S. H. Park, Phys. Rev. Lett. **62**, 1433 (1989); B. Rosenstein, B. J. Warr and S. H. Park, Phys. Reports **205**, 59 (1991).
G. Gat, A. Kovner and B. Rosenstein, Nucl. Phys. \[FS\] B **385**, 76 (1992); B. Rosenstein, H.-L. Yu, and A. Kovner, Phys. Lett. B **314**, 381 (1993).
R. Schneider, A. G. Zaitsev, D. Fuchs, and H. v. Löhneysen, Phys. Rev. Lett. **108**, 257003 (2012).
V. N. Kotov, B. Uchoa, V. M. Pereira, F. Guinea, and A. H. Castro Neto, Rev. Mod. Phys. **84**, 1067 (2012).
H. A. Fertig, Phys. Rev. B **40**, 1087 (1989); S. Q. Murphy, J. P. Eisenstein, G. S. Boebinger, L. N. Pfeiffer, and K. W. West, Phys. Rev. Lett. **72**, 728 (1994); I. B. Spielman, J. P. Eisenstein, L. N. Pfeiffer, and K. W. West, Phys. Rev. Lett. **84**, 5808 (2000); I. B. Spielman, J. P. Eisenstein, L. N. Pfeiffer, and K. W. West, Phys. Rev. Lett. **87**, 036803 (2001); Y. Yoon, L. Tiemann, S. Schmult,W. Dietsche, K. von Klitzing, and W. Wegscheider, Phys. Rev. Lett. **104**, 116802 (2010); A. D. K. Finck, J. P. Eisenstein, L. N. Pfeiffer, and K. W. West, Phys. Rev. Lett. **106**, 236807 (2011); X. Huang, W. Dietsche, M. Hauser, and K. von Klitzing, Phys. Rev. Lett. **109**, 156802 (2012).
B. Seradjeh, J. E. Moore, and M. Franz, Phys. Rev. Lett. **103**, 066402 (2009); Z. Wang, N. Hao, Z. G. Fu, and P. Zhang, New J. Phys. **14**, 063010(2012); D. K. Efimkin, Yu. E. Lozovik, and A. A. Sokolik, Phys. Rev. B **86**, 115436 (2012); S. Rist, A. A. Varlamov, A. H. MacDonald, R. Fazio, and M. Polini, Phys. Rev. B **87**, 075418 (2013).
D.-W Zhang, Z.-D Wang, S.-L. Zhu, Front. Phys. **7**, 31 (2012).
L. Fu and E. Berg , *Phys. Rev. Lett.*, **105,** 097001 (2010).
B. Rosenstein, B. Ya. Shapiro, D. Li and I. Shapiro, J. Cond. Mat. **27** 025701 (2015).
A. A. Abrikosov, L. P. Gor’kov, I. E. Dzyaloshinskii, *Quantum field theoretical methods in statistical physics*, Pergamon Press, New York (1965); E.M. Lifshits, L.P. Pitaeskii, *Course of Theoretical Physics vol.9. Statistical Physics part 2*, Prgamon Press, Oxford (1980).
J.M. Cornwall, R. Jackiw, and E. Tomboulis, Phys. Rev. D **10**, 2428 (1974); R. Haussmann, *Self-consistent Quantum-Field Theory and Bosonization for Strongly Correlated Electron Systems*, Springer, (1999).
Z. J. Wang, Y. Sun, X.-Q. Chen, C. Franchini, G. Xu, H. M. Weng, X. Dai, and Z. Fang, Phys. Rev. B, **85** (2012) 195320; P. Hosur, X. Dai, Z. Fang, X.-L. Qi, Phys. Rev. B **90**, 045130 (2014).
V.P. Gusynin, S.G. Sharapov, and J.P. Carbotte, Int. J. Mod. Phys. B **21**, 4611 (2007).
A. A. Abrikosov, Zh. Eksp. Teor. Fiz. **32**, 1442 (1957) \[Sov. Phys. JETP **5**, 1174 (1957); J. D. Ketterson and S. N. Song, *Superconductivity*, Cambridge University Press, Cambridge (1999); B. Rosenstein and D. Li, Rev. Mod. Phys. **82**, 109 (2010).
I. F. Herbut, V. Juricic, O. Vafek, Phys. Rev. B **80**, 075432 (2009).
L. Janssen, I. F. Herbut, Phys. Rev. B **89**, 205403 (2014).
L. Del Debbio, S. J. Hands, and J. C. Mehegan, Nucl. Phys. B 502, 269 (1997); I. M. Barbour, N. Psycharis, E. Focht, W. Franzki, and J. Jersak, Phys. Rev. D 58, 074507 (1998).
S. Chandrasekharan and A. Li, Phys. Rev. D **85**, 091502 (2012). S. Chandrasekharan, Phys. Rev. D **86**, 021701 (2012); Phys. Rev. D **88**, 021701(R) (2013).
F. F. Assaad and I. F. Herbut, Phys. Rev. B X **3**, 031010 (2013); S. Sorella,Y. Otsuka and S. Yunoki, Scientific Rep. **2**, 992 (2012).
B.W. Lee, *Chiral Dynamics,* Gordon and Breach, New York (1972).
Z.-H. Pan, A. V. Fedorov, D. Gardner, Y. S. Lee, S. Chu, T.Valla, Phys. Rev. Lett. **108**, 187001 (2012); V. Parente, A. Tagliacozzo, F. von Oppen, and F. Guinea, Phys. Rev. B **88**, 075432 (2013).
Y. S. Hor, A. J. Williams, J. G. Checkelsky, P. Roushan, J. Seo, Q. Xu, H. W. Zandbergen, A. Yazdani, N. P. Ong, and R. J. Cava, Phys. Rev. Lett. **104**, 057001 (2010); L. A. Wray, S.-Y., Y. Xia, Y. S. Hor, D. Qian, A. V. Fedorov, H. Lin, A. Bansil, R. J. Cava, and M. Z. Hasan, Nat. Phys. **6**, 855 ; G. Koren, T. Kirzhner, E. Lahoud, K. Chashka, and A. Kanigel, Phys. Rev. B **84**, 224521 (2011); P. H. Le, W.-Y. Tzeng, H.-J. Chen, C. W. Luo, J.-Y. Lin, and J. Leu, APL Mat. **2**, 096105 (2014).
K. Kirshenbaum, P. S. Syers, A. P. Hope, N. P. Butch, J. R. Jeffries, S. T. Weir, J. J. Hamlin, M. B. Maple, Y. K. Vohra, and J. Paglione, Phys. Rev. Lett., **111**, 087001 (2013).
Z. K. Liu, B. Zhou, Y. Zhang, Z. J. Wang, H. M. Weng, D. Prabhakaran, S.-K. Mo, Z. X. Shen, Z. Fang, X. Dai, Z. Hussain, and Y. L. Chen, Science **343,** 864 (2014); S.-Y. Xu, C. Liu, S. K. Kushwaha, T.-R. Chang, J. W. Krizan, R. Sankar, C. M. Polley, J. Adell, T. Balasubramanian, K. Miyamoto, N. Alidoust, G. Bian, M. Neupane, I. Belopolski, H.-T. Jeng, C.-Y. Huang, W.-F. Tsai, H. Lin, F. C. Chou, T. Okuda, A. Bansil, R. J. Cava, M. Z. Hasan, *Observation of a bulk 3D Dirac multiplet, Lifshitz transition, and nestled spin states in $Na_{3}Bi$*, *ArXiv*, 1312.7624 (2013); M. Orlita, D.M. Basko, M.S. Zholudev, F. Teppe, W. Knap, V.I. Gavrilenko, N.N. Mikhailov, S.A. Dvoretskii, P. Neugebauer, C. Faugeras, A.-L. Barra, G. Martinez, and M. Potemski, Nat. Phys. **10**, 233 (2014); G. Xu, H. Weng, Z. Wang, X. Dai, and Z. Fang, Phys. Rev. Lett. **107** (2011) 186806; Z. J. Wang, H. M. Weng, Q. Wu, X. Dai, and Z. Fang, Phys. Rev.* *B, **88**, 125427 (2013); M. Neupane, S. Y. Xu, N. Alidoust, G. Bian, C. Liu, I. Belopolski, T. -R. Chang, H.-T. Jeng, H. Lin, A. Bansil, F. C. Chou, and M. Z. Hasan, Nature Commun. **05**, 3786 (2014).
Y. Fuseya, M. Ogata, and Fukuyama H. , Phys. Rev. Lett*.* **102**, 066601 (2009); P. Hosur, S. A. Parameswaran, and A. Vishwanath, *P*hys. Rev. Lett. 1**08,** 046602 (2012); T. Kariyado and M. Ogata, *J. Phys. Soc. Jpn.*, **80**, 083704 (2011); **81**, 064701 (2012); P. Delplacel, J. Li, and D. Carpentier, Europhys. Lett. **97**, 67004 (2012); M. Lewkowicz and B. Rosenstein, Phys. Rev. B, **88,**045108 (2013).
M. N. Ali, Q. D. Gibson, T. Klimczuk, and R. J. Cava, Phys. Rev.* B*, **89** (2014) 020505(R).
| {
"pile_set_name": "ArXiv"
} |
---
author:
-
title: Voltage Uncertainties in the Presence of Photovoltaic Systems
---
Introduction
============
methods of mitigating voltage violations in the penetration of distributed generation [@c15]-[@c17] may allow renewable energies to enter the grid without causing undue harm to existing systems. When an inverter injects power to the grid, the impedance from the grid and inverter output-circuit, cause a voltage rise at the point of common coupling (PCC). The effect of the voltage rise is considered in the design of PV systems, e.g., the inverter disconnects from the grid if the inverter output voltage exceeds its operating limit. Once the inverter ceases to inject current to the grid, it must monitor the voltage for a while to reconnect to the grid [@c6]. The grid voltage can be adjusted to avoid unnecessary inverter tripping and system losses caused by voltage rise, and may lead to cascading events(cite). There is a need to model the voltage rise caused by PV systems to develop an optimal control algorithm to control the transformer Load Tap Changer (LTC), the switched capacitors, the bidirectional step regulators, and the operating voltage range of the inverters. Traditional analyses presume that loads exhibit linear elements, i.e. can be represented by constant impedances, on the contrary they may exhibit voltage dependent characteristics.
Switching from grid operator distribution, to consumer supplies of power causes difficulties in power flow. For instance, reverse power flow at low voltage distribution grids can cause a voltage rise and, similarly, if generation exceeds the load, the voltage increases. The liklihood of voltage rise increases with PV systems operating at the end of a distribution line with unity power factor [@c1]. As such, voltage fluctuations caused by PV systems, limit their penetration on the grid [@c10].
[@c7] prevent unwanted PV shut-off by applying batteries to store excess energy. [@c8] propose a closed-loop control system to regulate the PCC voltage of a PV system by controlling the reactive power injected into the grid. They do not consider the dynamic behaviour of loads and other inverters connected to the distribution grid on the performance of the designed PCC voltage controller. In order to regulate the PCC voltage of a PV system, the *dynamic behavior* of the grid should be known. Based on the authors’ works, the dynamic effect of loads and PV inverters on the voltage have yet to be studied.
In a grid with high penetration of PV systems, having a flatter voltage is more expensive, as it may require more voltage regulators [@c1]. It has been difficult to quantify the adjustment of voltage regulators and capacitance banks in the presence of high penetration of renewable energies, because of the stochastic nature of the loads and renewable energies. Hence, it is necessary to create a stochastic model for voltage variations through the grid, in order to determine the optimal locations in which to position the voltage regulators, and optimal strategies to operate them. The grid model with a high number of small distributed generators can be described more as a dynamic and non-stationary process. Novel methods for joint analysis of multiple data sets ([@ashkan]-[@ashkan1]) enable us to develop better models of the voltage variations.
Conventional voltage regulators are not designed for the high variability in voltage caused by PV systems. For example high penetration of PV systems can increase the number of tap changes and lead to degradation. A standard time delay is considered to avoid the load tap changer operations, when the voltage is outside the bandwidth for a short duration. To have a reliable grid with high penetration of PV systems, it is necessary to have model-based voltage regulators that respond quickly to voltage variations. Smart inverters can be used to mitigate the voltage variations caused by variability in PV system output and stochastic load [@c5]. This would require developing a national standard for inverters to autonomously regulate the voltage. It is important to determine the impact of PV systems on the voltage, to develop strategies for the voltage regulators. The main contributions of this work are as follows.
Main Contributions
------------------
The regulator tap position is a function of the PV systems’ outputs and loads [@c9]. The effect of PV output and load on the voltage should be modeled in order to determine the tap position at each time and the number of tap changes. In this work, we study the impact of distributed PV systems on the voltage rise of the distribution lines. We quantify the variability of the voltage at different distances from the transformer. We use partial least squares regression to model voltage variation versus injected power in individual PV systems. The accuracy of the model is analyzed by presenting the statistical parameters of voltage variations.\
The rest of the paper is structured as follows. In Section \[SectionII\] the data that is used in the analysis is described and the numerical results for PV profile analysis is presented. Section \[SectionIV\] implements partial least square to predict the voltage rise caused by the inverter output. Finally, concluding remarks are provided in Section \[SectionVII\].
PV Profile {#SectionII}
==========
A customer’s net energy is defined as the difference between the energy supplied by the grid and energy generated by the customer’s PV system. All five of the consumers in this study are equipped with net energy meters that report the net energy and PCC voltage at one minute intervals. All consumers in this study are also equipped with a second meter that reports the PV generation (inverter output) and voltage at one-minute intervals. The reported PV generations are the aggregate of the PV generation over the one-minute intervals. The root mean square (RMS) voltage is measured at each second and the maximum voltage over one-minute intervals is reported. The analysis is based on $160$ days’ data of five PV systems with capacities $1.94$, $3.87$, $9.24$, $11.61$ and $7.31$ kilowatt (KW). The PV systems are positioned at different distances from the distribution transformer on the consumers’ properties. The line impedance from each PV system to the transformer is given in the Table \[table:tb0\]. It is observed that the daily energy for a PV system with capacity of one kilowatt, varies from $3.96$ KWh to $5.17$ KWh. On average, the amount of daily energy for a PV system with capacity of one kilowatt is $4.78$ KWh.
Capacity (KW) 1.94 3.87 7.31 11.61 9.24
------------------------- ------- ------- ------- ------- -------
Impedance (OHM) 0.077 0.060 0.053 0.025 0.011
Average daily energy/SC 4.67 5.13 3.96 5.17 4.98
: Impedance (OHM) and average daily energy per system capacity (SC)[]{data-label="table:tb0"}
The efficiency of a PV system can be impacted if part of a PV panel is shaded, or if the panel is positioned at a non-optimal angle or direction. In figure \[fig:pvirrad\], the PV generated power versus irradiance for all the PV systems under consideration is plotted. Each PV system output power in kilowatts (KW) is divided by its PV System Capacity (SC). The vertical axis represents the generated power per one unit of the PV system capacity, and horizontal axis represents the irradiance (kilowatt per square meter). It is observed that the variations of the PV systems output in irradiance can be approximated as a linear model. The scatter plot shows all the sample data points for all systems and the red line shows the best fit for all the systems using linear regression [@c4]. Again using linear regression, the PV output power is modeled as a linear function of irradiance $\mathit{P}= 0.906 \,\ \mathit{R}$, in which $\mathit{P}$ is the generated power per unit of the PV system, and $\mathit{R}$ is the irradiance power. The following can be observed from PV data:
- The average PV systems’ generation is less than $70$ percent of PV system capacity.
- The efficiency of the PV systems is not related to the capacity of the system.
- Due to differences in PV panel direction, the peak-time of PV profiles varies. This can be used for the flattening of PV generation in the grid.
The effect of irradiance on the PV system output is deterministic on a clear sky day. Because of the irregular presence of the clouds, the irradiance data has been modeled as a random variable [@c2]. Voltage is affected by PV system output and the load. PV system output and load are random variables, and consequently the voltage variations in distribution grids can be modeled as a random process. In the next section, the PCC voltage of the five PV systems, each with different line impedance, is analyzed.
Voltage Analysis {#SectionIV}
================
In figures \[fig:povo1\]-\[fig:povo5\], power and voltage versus time are plotted. The blue dots are the scatter points of the data and the red-lines are the averages of the data at each time. Negative power values imply injection of power into the grid and positive power values represent power flowing in the opposite direction from the grid to the house. The PCC voltage of a PV system is dependent on both active and reactive power injected to the grid [@c11]-[@c14]. In this study, the PV systems only injects active power to the grid, and the amount of reactive load is negligible. Therefore, reactive power is not considered in the voltage analysis. Note that the voltage at each PCC is affected by neighbouring loads, PV generation quantities, and the operation of load tap changers, as well as capacitance banks. The voltage does not rise consistently when injecting more power to the grid at a single PCC, because of the stochastic nature of the load and PV generation at neighboring properties. It is observed that *on average* by injecting more power to the grid, voltage begins to increase. In the next section, the voltage variation is modeled as a function of injected power into the grid, and the statistical parameters about the accuracy of this model are presented.
Partial Least Square (PLS)
--------------------------
In this section, we use partial least squares regression, [@c3], to predict the voltage rise caused by the inverter output at the PCC of an inverter. Voltage decreases over distance and load, and in a conventional grid, the delivered voltage at low load is high, in order to ensure adequate voltage at peak load. Keeping the voltage in an acceptable range during low load and peak load states, is a utility challenge. Utilities regulate the voltage using voltage regulators, switched capacitor banks, and transformer load tap changers. The observed voltage at the PCC is affected by the LTC, switched capacitors, neighbours PV systems, and loads. To model the voltage rise caused by the PV system, the effect of LTC and capacitor banks on the voltage must be filtered out. PLS regression is applied to predict the response of voltage to power injected by the PV system. PV system output is considered as the predictor and voltage is considered as the response. The goal of PLS is to predict the voltage variations from injected power. PLS achieves the maximum correlation between power and the voltage, by maximizing the co-variance between power and voltage. PLS combines information about the variances of both the PV system power (predictors) and the voltage (responses), while also considering the correlations between them. Consider $P$ and $V$ as the vectors of measured power and voltage. We seek a scalar $\beta$ such that $V=\beta P +\epsilon$, where $\epsilon$ stands for the error in the model. In figures \[fig:vvp1\]-\[fig:vvp5\], the voltage versus the power is plotted. The blue dots are the scatter points of the data and the red-lines are the best linear fit to the data using PLS regression. The linear model for the voltage variations versus injected power ($P \leq 0$) for PV system capacities of $1.94$, $3.87$, $7.31$, $11.61$ and $9.24$ KW, are respectively given in (1)-(5).
\[eq:eqno1\] V=-0.0113 P+1.0172,\
\[eq:eqno2\] V=-0.0094 P+1.0220,\
\[eq:eqno3\] V=-0.0036 P+1.0192,\
\[eq:eqno4\] V=-0.0018 P+1.0194,\
\[eq:eqno5\] V=-0.0014 P+1.0237.
From the comparison of the impedance values in Table \[table:tb0\] and equations (1)-(5), it can be concluded that the magnitude of $\beta$ increases with the line impedance. It is observed that the maximum voltage rise for one-kilowatt of injected power is $0.0113$ per voltage unit. Note that the linear model is an approximation to the voltage variation. In the next section more statistics about possible voltage rise are presented.
Statistical Parameters
----------------------
In figures \[fig:beta1\]-\[fig:beta5\], the daily and hourly variations of $\beta$ are plotted. The dashed red-lines are the averages of $\beta$ in the given figure. In the daily plots, each point corresponds to the value of $\beta$ calculated by applying PLS regression to the corresponding day’s data. In the hourly plots, each point corresponds to the value of $\beta$ calculated by applying the PLS regression to the data of the given time over $160$ days. It is observed that the value of $\beta$ has less variation from $12:00$ p.m. to $14:00$ p.m., because of less variation in the loads and stability in the weather conditions.
In Table \[table:tb2\] and \[table:tb3\] the daily and hourly statistics for the $\beta$ values are presented, i.e., minimum, average, standard deviation (STD), and maximum-a-posteriori (MAP) estimate. By comparison of Table \[table:tb2\] and \[table:tb3\], it can be deduced that, by injecting one kilowatt power into the grid at locations with PV sizes of $1.94$, $3.87$, $7.31$, $11.61$, and $9.24$ KW, the highest possible voltage increases are respectively $0.0197$, $0.0141$, $0.0087$, $0.0047$ and $0.0052$ per voltage unit. It is observed from Tables \[table:tb0\], \[table:tb2\] and \[table:tb3\], that the magnitude of minimum, average, STD and MAP for $\beta$ increases with line impedance. Therefore, PV systems with larger impedance exhibit more variability in PCC voltage.
Capacity (KW) 1.94 3.87 7.31 11.61 9.24
--------------- --------- --------- --------- --------- ---------
Minimum -0.0197 -0.0141 -0.0087 -0.0047 -0.0052
Average -0.0112 -0.0095 -0.0040 -0.0019 -0.0017
STD 0.0034 0.0017 0.0014 0.0008 0.0008
MAP -0.0118 -0.0097 -0.0036 -0.0016 -0.0013
: Daily statistics of $\beta$[]{data-label="table:tb2"}
Capacity (KW) 1.94 3.87 7.31 11.61 9.24
--------------- --------- --------- --------- --------- ---------
Minimum -0.0159 -0.0116 -0.0072 -0.0025 -0.0022
Average -0.0120 -0.0093 -0.0034 -0.0016 -0.0012
STD 0.0011 0.0009 0.0009 0.0003 0.0003
MAP -0.0116 -0.0099 -0.0049 -0.0022 -0.0013
: Hourly statistics of $\beta$[]{data-label="table:tb3"}
In figures \[fig:density\] and \[fig:density1\], the probability density function for $\beta$ is plotted. In Table \[table:tb4\], the entropies for each of the PCCs is given. It is observed that a PCC with larger impedance has a larger entropy, this implies more randomness and disorder in the behavior of voltage at a PV system with larger line impedance.
Capacity (KW) 1.94 3.87 7.31 11.61 9.24
---------------- ------ ------ ------ ------- ------
Daily entropy 5.35 4.80 4.80 3.59 3.45
Hourly entropy 6.31 5.77 5.39 4.5 4.5
: Entropy of $\beta$[]{data-label="table:tb4"}
CONCLUSIONS {#SectionVII}
===========
In this work, the effect of injected power into the grid on the voltage fluctuations is quantified. The voltage fluctuations decrease during periods of clear sky, due to less variability in PV generations and loads. Modeling the voltage rise per unit of injected power can be used to develop optimal strategies for operating and placement of voltage regulators. The voltage at each PCC is affected by neighbouring loads and PV systems output.
[99]{} T. Stetz, F. Marten, and M. Braun, Improved low voltage gridintegration of photovoltaic systems in Germany, IEEE Transaction on Sustainability Energy, vol. 4, no. 2, pp. 534–542, 2013. R. Tonkoski, L. A. C. Lopes, and T. H. M. El-Fouly, Coordinated active power curtailment of grid connected PV inverters for overvoltage prevention, IEEE Trans. Sustain. Energy, vol. 2, no. 2, pp. 139–147, 2011. E. Demirok, D. Sera, R. Teodorescu, P. Rodriguez, and U. Borup, Evaluation of the voltage support strategies for the low voltage grid connected pv generators, in Proc. IEEE Energy Conversion Congress and Exposition (ECCE), pp. 710–717, 2010. A. Zeinalzadeh, R. Ghorbani, J. Yee, Stochastic model of voltage variations in the presence of photovoltaic systems, IEEE American Control Conference (ACC), Boston, MA, USA, July 6-8, 2016. T. Stetz, W. Yan, and M. Braun, Voltage control in distribution systems with high level PV-penetration, in 25th European PV Solar Energy Conference, Valencia, Spain, 2010. A review of inverter design and topologies, www.traceengineering.com. A. Zeinalzadeh and Vijay Gupta, Minimizing risk of load shedding and renewable energy curtailment in a microgrid with energy storage, ArXiv e-prints, arXiv:1611.08000. B. Perera, P. Ciufo, and S. Perera., Point of common coupling (PCC) voltage control of a grid-connected solar photovoltaic (PV) system, Annual Conference of the IEEE Industrial Electronics Society. Vienna, Austria. Jan. 2013. A. Zeinalzadeh, T. Wenska and G. Okimoto, Integrated analysis of multiple high-dimensional data sets by joint rank-1 matrix approximations, 2015 IEEE 54th Annual Conference on Decision and Control (CDC), Osaka, Japan, Dec. 15-18, 2015. M. Lee, H. Shen, J. Z. Huang, and J. S. Marron, Biclustering via sparse singular value decomposition, Biometrics, vol. 66, pp. 1087-1095, 2010. V. Castelli, A. Thomasian, Chung-Sheng Li, Clustering and singular value decomposition for approximate similarity searches in high dimensional spaces, IBM Research report, 2000. P. Drineas, A. Frieze, R. Kannan, S. Vempala, V. Vinay, Clustering large graphs via the singular value decomposition, Machine Learning, vol. 56, pp. 9–33, 2004. J. A. Nelder and R. W. M. Wedderburn, Generalized linear models, Journal of the Royal Statistical Society. Series A, Vol. 135, No. 3, pp. 370-384, 1972. A. Zeinalzadeh, An iterated version of the generalized singular value decomposition for the joint analysis of two high-dimensional data sets, University of Hawaii at Manoa, 2013. A. Zeinalzadeh , R. Ghorbani, E. Reihani, Optimal power flow problem with energy storage voltage and reactive power control, The 45th ISCIE International Symposium on Stochastic Systems Theory and Its Applications, Japan, 2013. A. Zeinalzadeh, R. Ghorbani, J. Yee, Voltage control in the presence of photovoltaic systems, ArXiv e-prints, December 2016. J. F. Kenney and E. S. Keeping, Linear regression and correlation. in Mathematics of Statistics, Pt. 1, 3rd ed. Princeton, NJ: Van Nostrand, pp. 252-285, 1962, ch 15. Hollands KGT, Huget RG, A probability density function for the clearness index, with applications, Agricultural and Forest Meteorology, 94, pp. 53-63, 1983. T. Niknam, A. M. Ranjbar, and A. R. Shirani, Impact of distributed generation on volt/var control in distribution networks, in IEEE Power Tech Conference Proceedings,Bologna, Italy, 2003, p. 7 pp. Vol.3. K. Turitsyn, P. Sulc, S. Backhaus, and M. Chertkov, Distributed control of reactive power flow in a radial distribution circuit with high photovoltaic penetration, in IEEE Power and Energy Society General Meeting, pp. 1-6, 2010. T. Stetz, W. Yan, and M. Braun, Voltage control in distribution systems with high level PV-penetration-improving absorption capacity for PV systems by reactive power supply, in Proceeding of 25th European Photovoltaic Solar Energy Conf. and Exhibition, pp. 5000–5006, 2010. Wold, H. . Estimation of principal components and related models by iterative least squares. In P.R. Krishnaiaah (Ed.). Multivariate Analysis. pp. 391-420, New York: Academic Press, 1966.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'This paper introduces the [*Maestro*]{} library. This library for Python focuses on the design of predictive controllers for small to medium-scale energy networks. It allows non-expert users to describe multi-carrier (electricity, heat, gas) energy networks with a range of energy production, conversion, and storage component classes; together with consumption patterns. Based on this description a predictive controller can be synthesized and tested in simulation. This controller manages the dispatch of energy in the network, making sure that the demands are met, while minimizing the total energy cost. Alternative objectives can be specified. The library uses a mixed-integer linear modelling framework to describe the network and can be used in stand-alone based on standardized input files or as part of the larger energy network control platform PENTAGON.'
author:
- 'Tomasz T. Gorecki and William Martin[^1]'
bibliography:
- 'ifacconf.bib'
title: '[*Maestro*]{}: A Python library for multi-carrier energy district optimal control design[^2] [^3]'
---
Introduction
============
The shift from centralized energy generation in few large plants to a more and more decentralized generation infrastructure with a growing penetration of intermittent renewables challenges the management logic of the grid in all its aspects: communication, data management, control [@ilo_linksmart_2016; @reynolds_upscaling_2017]. With the emergence of micro-grids and self-consumption communities, it is expected that local grid control strategies will play an important role in the management of the future power grids. In addition, with the electrification of transport, the increasing penetration of heat pump to serve heating and cooling needs and the emergence of new technology such as power-to-gas systems and fuel cells, power, gas and local heat energy grids are becoming more interconnected. This provides additional opportunities to improve the overall operation and environmental impact of the energy sector, but creates additional challenges in designing efficient and scalable control strategies for the operation of the grid. Multiple control strategies have been proposed for the management of local grids, including particle-swarm based algorithms [@hurtado_smart_2015], evolutionary optimization [@mauser_adaptive_2016]; and hierarchical control design [@tavakoli_two_2018]. Model Predictive Control has been one of the most widely applied control strategies for energy systems [@parisio_cooperative_2017; @patino_economic_2014; @touretzky_integrating_2014] due to its ability to handle constraints, incorporate forecasts and specify complex and dynamic control objectives [@holkar_overview_2010]. A major obstacle in the widespread adoption of advanced control strategies is their high complexity, and the need for multiple expertise to successfully deploy them on real-world systems. To address this issue, the European project PENTAGON [^4] develops a scalable district multi-carrier energy management platform, integrating communication, control, and data management systems in a flexible framework. As part of this platform, we introduce in this paper [*Maestro*]{}, a Python library for the automated design of predictive controllers for energy networks.
Most existing tools are either modelling or decision-support tools [@allegrini_review_2015]. The latter are designed to facilitate design decisions by comparing technological choices, and sizing subsystems. They are often powered by optimization models and solvers. For example, [*Artelys Crystal Energy Planner*]{} and [*DER-CAM*]{} are proprietary tools, while [*ficus*]{} [@atabay_open-source_2017] and [*OMEGAlpes*]{} [@morriet_multi-actor_2019] are open-source tools. Few tools with a focus on control are available. Most notably, the toolbox EHCM [@darivianakis_stochastic_2015] focuses on building control, and is developed in Matlab. To the author’s knowledge, [*Maestro*]{} is the first Python library specifically tackling predictive control design for multi-carrier district energy systems. In comparison with decision modelling tools which typically use a coarse description of the system and simulate “ideal” scenarios with for example perfect knowledge of the future demand, [*Maestro*]{} focuses on control design. It allows a more detailed modelling level of the components and the possibility to simulate controllers in more realistic scenarios, [*e.g.*]{}, observing the effect of imperfect forecasts. In addition, controllers generated with [*Maestro*]{} are deployable as part of a larger energy district control framework which is described briefly below.
[*Maestro*]{} supports a variety of energy consumption, production, conversion and storage systems components. Based on a high-level description of the energy district, [*Maestro*]{} allows to prototype and test controllers for energy districts fully automatically. This makes the library useful for control engineers and energy system designers alike, as it does not require expert knowledge about energy system operation or advanced control strategies. This paper gives an overview of the library’s working principles and present a simple use case to demonstrate its usage. Finally, we also introduce for the first time a model of Power-to-Gas systems suitable for predictive control design. Section \[sec:general\_principles\] introduces the guiding principles used in the library, section \[sec:OCP\] introduces the models used to described individual components of the district, and the design of the MPC problem. Finally, Section \[sec:results\] presents a small case study where [*Maestro*]{} is used to control a small energy grid.\
General principles {#sec:general_principles}
==================
Organisation of the PENTAGON platform
-------------------------------------
Although this paper focuses on predictive control design, we give an overview of the platform in order to contextualize how [*Maestro*]{} fits in a full energy management platform and refer the reader to relevant references for details. The PENTAGON platform integrates fives main software components.
- A tailored graph database [@reynolds_upscaling_2017]
- A forecasting module to predict expected energy consumptions and renewable generation [@ahmad_tree-based_2018]
- A communication bridge from the platform to the physical systems, called Smart Connector [@noauthor_smart_nodate]
- A network simulation engine
- A predictive control module
The general organization and data flow is summarized in Figure \[fig:PENTAGON\_platform\]. In the frame of the PENTAGON platform, all data, including metadata about the network is stored in the database. Data about the system and control specifications is organized according to a standard unified description scheme. It for example specifies unique identifiers for the observations and actuations, and all modules can automatically retrieve and publish data through an API. In the following, we focus on the Predictive Controller module, which directly uses [*Maestro*]{} to build setpoint schedules for the network, based on live system measurement and forecasts for demands.
![Schematics of the PENTAGON platform organization[]{data-label="fig:PENTAGON_platform"}](pentagon_platform.png){width="\columnwidth"}
Modelling and control principles
--------------------------------
In order to flexibly represent different district configurations, [*Maestro*]{} combines and connects elements that represent different part of a multi-carrier network. The two basic type of elements are :
- Components, which represent physical equipment or sets of equipment that are aggregated, [ *e.g.*]{}, inside a consumption profile.
- Nodes that are virtual objects that represent an interface point through which different components transfer/exchange energy
The set of nodes and components represent the full energy network. Each component is modelled according to a set of first principle equations. The combination of all components models, constraints and cost function are compiled in a controller fully automatically. The controller follows the MPC philosophy to calculate the inputs of the different components: it forms an optimization problem that searches for the optimal sequence of inputs to apply over an horizon, denoted $H$. The horizon is divided in time steps of duration $\Delta t$. In closed-loop operation, the decision for the first time step is applied. When new measurements are available, a new optimization step is carried out based on the new information available and the plan is updated and applied in a receding horizon fashion.
In simulation or control operation, the controller takes as input a set of forecasts and a set of measurements and forecasts and outputs a setpoint plan over the horizon for all the components.
Workflow of the library
-----------------------
[*Maestro*]{} uses an automated workflow to generate optimal controllers for multi-carrier networks. This allows users who are not familiar with control to prototype controllers with minimal input. Two possible input configuration description are available:
- A standardized input file. A template spreadsheet file can be used to specify the physical parameters and operational limits of the components, unique identifiers for the observations expected for these components and the actuations they produce (inputs/outputs of the components).
- A programmatic description of the network. A high-level abstraction is used where components of the network can be defined as instances of the library’s classes.
The toolbox implements the following steps:
1. From the input file, a parsing step extracts the components’ parameters and useful metadata.
2. Input data (measurement and forecasts) is gathered from the database API. When [*Maestro*]{} is used standalone, the observations can directly be specified in the input file or programatically.
3. From this data, a directed graph representing the district is built and verifications are performed to ensure consistency of the network.
4. Individual component models are created and descriptive models are formulated using an optimization problem parser. [*Maestro*]{} currently uses the parser PULP [@noauthor_coin-or/pulp_2019], which supports mixed-integer linear (MIL) modelling and is compatible with multiple optimization solvers.
5. Individual models are combined into a network model and the optimization problem is solved.
6. Setpoint schedule are built by processing the solution to the optimization problem, and dispatched to a database if needed.
7. In addition, visualisation routines can create summary plots of the planned operation for the network’s components.
An online demo is under preparation and will be uploaded at (https://www.csem.ch/districtenergy), where a simple network can be configured, and simulations with the automatically generated controller can be run and results displayed.
Optimal control problem construction {#sec:OCP}
====================================
Component models
----------------
Each component is a virtual object that represent a set of equipment in the district. Components are formally described by :
- A set of parameters, describing the physical characteristics and operational limits of the components, [*e.g.*]{}, the battery capacity;
- a set of “observations”, representing dynamic information that is used to calculate the plan of operation for the component. These can be measurements ([*e.g.*]{}, battery state of charge) or forecasts ([*e.g.*]{}, power production forecast for a photo-voltaic (PV) system);
- A set of output nodes, representing points of the network where the component injects power;
- A set of input nodes, representing points of the network where the component draws power.
Components are classified in the following categories:
- *Consumers* represent (possibly aggregated) heat, gas, or electricity consumption from buildings, processes, facilities, etc. Control of consumers is limited to the possibility to shed all or part of their power consumption;
- *Converters* represent systems that convert energy from one carrier to another. This includes heat pumps, gas boilers, electric heating systems;
- *Non-controllable renewable energy sources (RES)* represent renewable energy production means whose output is not fully controllable, including PV systems, Solar Thermal systems, and wind turbines. Control of RES is limited to the possibility to curtail all or part of their generation;
- *Co-generators* represent system that can convert one to several other carriers of energy, such as combined heat and power units (CHP) which produce electricity and heat from gas;
- *Power-to-gas (p2g) systems* are converters synthesizing gas using electric power. Details of the optimization model are described in section \[sec:p2g\];
- *Storage systems* represent energy storage means, such as electric batteries and thermal storage systems, such as cold or hot water tanks;
- *Electric vehicles (EVs)* are specific types of consumers which are controllable. Driving schedules are used to specify the availability and location of EVs, as well as their energy needs. Several charging control types are considered: uni- and bi-directional charging; continuous and on/off charging control;
- *Other generators* represent generation from other types of resources, such as waste incineration, bio-gas plants, …
- *Grid ties* represent the connection of the district to the external electric, heat and/or gas grid. Tariffs and maximum power levels can be attached to the grid ties and considered in the optimization problem.
The platform uses a MIL modelling framework to describe the components’ dynamic behavior and operational constraints. MIL modelling allows to capture a variety of nonlinear behaviors, and in particular the switching nature of many components. For illustration purposes, we describe here two component models available in the library.
### Electric Storage systems
Symbol Description Unit
------------------ ----------------------------- -------
$P^{char}_{max}$ Max. charging Power kW
$P^{disc}_{max}$ Max discharging Power kW
$\eta$ Charging efficiency -
$E_{max}$ Max. state of charge kWh
$E_{min}$ Min. state of charge kWh
$f^{use}$ Cost of cycling the battery €/kWh
$f_{fin}^{soc}$ Value of final SoC €/kWh
: Parameters of the BESS[]{data-label="tb:bess_params"}
A generator convention is used for the battery, [*i.e.*]{}, discharging the battery means that $P_{batt}(\tau) \geq 0$ and charging the battery means that $P_{batt}(\tau)\leq 0$. Constraints on charging and discharging are : $$-P^{char}_{max} \leq P_{batt}(\tau) \leq P^{disc}_{max}, \quad \forall \tau \in [0, H-1]$$ The evolution of the state of charge (SoC) $E_{batt}$ is described by: $$\begin{array}{ll}
E_{batt}(\tau + 1) = E_{batt}(\tau) &- \Delta_t \frac{1}{\eta} \max (P_{batt}(\tau), 0) \\
&- \Delta_t \eta \min (P_{batt}(\tau), 0)
\end{array}$$ with constraints on the SoC : $E_{min} \leq E_{batt}(\tau) \leq E_{max}$. The dynamics can be modeled using an MIL model by introducing binary variables $\delta_{char|disc}$, continuous variables $P^{char}$ and $P^{disc}$ and constraints: $$\begin{array}{c}
0 \leq P^{disc}(\tau) \leq P^{disc}_{max}(1-\delta_{char|disc}) \\
0 \leq P^{char}(\tau) \leq P^{char}_{max}\delta_{char|disc} \\
E_{batt}(\tau + 1) = E_{batt}(\tau) - \frac{\Delta_t}{\eta} P_{disc}(\tau) - \Delta_t \eta P_{char}(\tau)
\end{array}$$
### Power-to-gas {#sec:p2g}
The *PENTAGON* project’s objectives include the construction of an experimental Power-to-Gas facility, and the specific assessment of the impact of Power-to-gas systems in energy districts. In this light, a more detailed study of P2G systems was conducted, including the design of a simplified model which is suitable for predictive control design. While a real-life power to gas plant has various operation modes, a simplified three-stage model consisting of an OFF state (no consumption), an ON state, and an HOT state, which can be seen as a standby state. The model includes:
- A conversion efficiency $\kappa$
- A minimum and maximum electric power consumption in ON mode($P^{min}_{p2g}$ and $P^{max}_{p2g}$)
- A fixed power consumption in HOT state ($P_{hot}$)
- A turn-on time from OFF to HOT state ($t_{OFF-HOT}$)
- A turn-off time from HOT to OFF state ($t_{HOT-OFF}$)
Similar switching times between states ON and HOT can be introduced and models, but they are generally much shorter than the ones between states OFF and HOT and are small compared to the planning time step $\Delta_t$, so they can typically be neglected.
Denoting with $P^{in}_{p2g}$ the electric input power consumption of the plant in kW and $P_{p2g}^{out}$ the gas power produced (which directly translates to gas production rate, assuming a fixed calorific value for the gas produced), we have that, in ON mode, $P^{min}_{p2g} \leq P^{in}_{p2g} \leq P^{max}_{p2g}$ and $P^{out}_{p2g} = P^{in}_{p2g}\kappa$. In HOT mode, $P^{in}_{p2g} = P_{hot}$ and $P^{out}_{p2g} = 0$. Finally, in OFF mode, $P^{out}_{p2g} = P^{in}_{p2g} = 0$. In addition, the power-to-gas system must remain in state HOT for a minimum of $t_{HOT-OFF}$ before switching OFF and must remain in state OFF for a minimum of $t_{OFF-HOT}$ before switching to HOT.
The dynamics of the P2G system can be modeled using an MIL model by introducing binary variables $\delta_{off}$, $\delta_{hot}$, and $\delta_{on}$ to represent the state of the system. Only one state is active at a time, so that: $$\delta_{off}(\tau) + \delta_{hot}(\tau) + \delta_{on}(\tau) = 1, \quad \forall \tau \in [0, H-1]$$ Additional binary variables $\delta_{off2hot}$, $\delta_{hot2off}$, $\delta_{on2hot}$ and $\delta_{hot2on}$ are introduced to express state transitions. Similarly, only one transition is possible at a time: $$\delta_{off2hot}(\tau) + \delta_{hot2off}(\tau) + \delta_{hot2on}(\tau) + \delta_{on2hot}(\tau) \leq 1$$
Constraints ensure that ON to OFF and OFF to ON transitions are not possible: $$\forall \tau \in [0, H],
\begin{array}{ll}
\delta_{off}(\tau) + \delta_{on}(\tau-1) \leq 1 \\
\delta_{on}(\tau) + \delta_{off}(\tau-1) \leq 1
\end{array}$$ Additional constraints are imposed to limit state transitions. In the case where the lag is one time step, we have the logical relation: $$\delta_{off2hot}(\tau - 1) \Longleftrightarrow{} \delta_{off}(\tau-1) \wedge \delta_{hot}(\tau)$$ with $\wedge$ the logical AND operator. In other words, the system can switch from OFF to HOT if and only if the system is in state OFF and comes to state HOT in the next instant. This can be expressed with constraints: $$\label{and_expr}
\begin{aligned}
&\delta_{off2hot}(\tau-1) \leq \delta_{off}(\tau-1) \\
&\delta_{off2hot}(\tau-1) \leq \delta_{hot}(\tau) \\
&\delta_{off2hot}(\tau-1) \geq \delta_{off}(\tau-1) + \delta_{hot}(\tau) -1
\end{aligned}$$ We apply the same system of inequality constraints to the 3 other possible state transitions. A similar derivation can be made for the case where a lag $T$ is present (minimum time in a state before switching). We have then that: $$\begin{aligned}
\delta_{off2hot}(\tau - 1) \Longleftrightarrow{} \left\{
\begin{array}{ll}
\forall i = 1 \ldots T, \delta_{off}(\tau-i)\\\\
\delta_{hot}(\tau)
\end{array}
\right.
\end{aligned}$$ which can be expressed with the following equations: $$\label{and_expr_lag}
\begin{array}{rl}
\delta_{off2hot}(\tau-1) &\leq \delta_{hot}(\tau) \\
\delta_{off2hot}(\tau-1) &\leq \delta_{off}(\tau-i) \quad \forall i = 1 \ldots T\\
\delta_{off2hot}(\tau-1) &\geq \sum_{i=1}^T \delta_{off}(\tau-i) + \delta_{hot}(\tau) -1
\end{array}$$ Equations relax to equations when\
$T = t_{OFF-HOT}/\Delta t= 1$.
Optimization problem
--------------------
### Constraints
To compute the plan to operate the network, the individual models of the components are compiled and a centralized optimization problem is formulated. We denote by $\mathbf x_i$ the set of optimization variables for device $i$ and by $\mathcal S_i$ the feasible set for the optimization of these variables. Energy balance is maintained at each node, so that: $$\label{eq:node}
\sum_{i\in \mathcal{I}^{out}_j} P_{out}^{node, j} = \sum_{i\in \mathcal{I}^{in}_j} P_{out}^{node, j}, \quad \forall \text{ nodes } j$$ where $\mathcal{I}^{in}_j$ and $\mathcal{I}^{out}_j$ are the sets of components that include node $j$ in their input/output nodes, respectively; and $P_{in}^{node, j}$ and $P_{out}^{node, j}$ their input/output power from/to node $j$. The optimization problem reads: $$\begin{aligned}
& \underset{\mathbf x}{\text{min.}}
& & J(\mathbf x) \\
& \text{subject to}
& & \mathbf x_i(x) \in \mathcal S_i, \; i = 1, \ldots, m. \\
&
& & \text{equation} \; \eqref{eq:node}
\end{aligned}$$
### Cost function
The cost function of the problem is an economic cost function, capturing actual operating costs of the system. By default, a global cost function for the entire district is formed as the sum of the individual cost functions of each component. The cost of the net energy exchange with the grid is attached to the grid ties, with a flexible time-varying import and export price profile for the energy carrier considered. Other costs include:
- Cost of curtailing renewable production and shedding loads;
- Cost of primary fuel for generators ([ *e.g.*]{}, biomass);
- Cost of using equipment ( [*e.g.*]{}, a cycling cost can be attached to the battery ) to reflect battery degradation;
- “Regularization” costs to promote secondary objectives such as smooth power output of the components, [*e.g.*]{}, under the form of a power rate cost;
- Terminal costs, as customary in MPC in order to compensate for the finite horizon optimization, [*e.g.*]{}, terminal cost to reward the remaining SoC of the battery at the end of the horizon.
A centralized cost assumes the goal is to minimize the total cost for the district. In order to accommodate more flexibility, components can be assigned to different owners, and the cost function can be formed as the sum of the individual owners’ cost functions, which actually leads to a decentralized controller, where each owner tries to minimize his own cost. Details of the implementation are left out of this paper.
Use cases and results {#sec:results}
=====================
We report the results of a simulation study on a network that illustrates how the library can be used to generate and simulate controllers of energy districts. The system is a small district serving an eight-dwelling apartment building located in Chambery, France. Heat profiles were simulated using a TRNSYS model of the buildings and the electrical profiles were generated from historical data. The district has local generation from wind turbines and a PV plant. The data for wind plant is taken from a nearby wind field historical data; while the PV generation is simulated from weather data. The district is also equipped with a biomass generator, a gas boiler coupled with a thermal storage and a small power to gas system. The components and nodes used to represent the system in [*Maestro*]{} are depicted in Figure \[fig:network\]. The systems’ operational parameters are reported in Table \[tab:parameters\]
![Small district network[]{data-label="fig:network"}](extended_case){width="\columnwidth"}
Parameter value
--------------------------- ------------
Boiler max. power 30 kW
Boiler min. power 5 kW
P2G min. elec. pow. 5 kW
P2G max. elec. power 10 kW
P2G conversion efficiency 0.75
Biomass Boiler max. pow. 10 kW
Thermal tank size 2m$^3$
Electricity buy price 0.2 €/kWh
Electricity sell price 0.04 €/kWh
Gas buy price 0.13 €/kWh
Biomass buy price 0.2 €/kWh
Elec. demand 2383 kWh
Heat demand 15354 kW
PV prod. 637 kWh
Wind power prod 3594 kWh
: Network parameters \[tab:parameters\]
{width="\textwidth"}
w P2G w/o P2G
------------------------ ------- ---------
Gas cost \[€\] 1270 1502
Biomass cost \[€\] 720 720
Electricity cost \[€\] 153 47
Total cost \[€\] 2143 2270
Gas produced \[m$^3$\] 723.3 0
SC ratio\[%\] 95.4 38.3
: KPIs of simulation \[tab:results\]
A one month simulation is performed considering the same network with and without the power to gas system. The controls are applied in simulation are the control computed with the controller generated with [*Maestro*]{}. Figure \[fig:simulation\] gives an snapshot of the control applied to the system during three days of the simulation. Table \[tab:results\] reports the total cost and the breakdown of cost per energy carrier, as well as the amount of gas produced from the power to gas system. We can observe that the controller is able to modify the operation to exploit the excess of renewable energy efficiently, bringing the self-consumption ratio from a low 38% to 95%. To do so, it activates the power to gas system to produce gas, which in turn is used to produced heat stored in the thermal storage. This allows reduced purchasing of gas, and therefore lower operating cost. In this configuration, we observe a decrease of operational costs of 5.5% over a one month period when adding the power to gas system. These savings are modest due to the relatively low price of gas and an excess production of renewable (especially PV) which is limited.
Conclusion
==========
The [*Maestro*]{} library offers the possibility to design and test complex predictive controllers for networks in a matter of minutes. The library remains currently proprietary and readers interested in benefiting from the library are invited to contact the authors. The PENTAGON platform, which incorporates the controllers generated with [*Maestro*]{} is currently being used for the control of a small energy network located in France and will be used to demonstrate thoroughly through simulation the potential of introducing power to gas technology to a urban district of Blaenau-Gwent in Wales.
On the technical side, it has been observed that one of the drawbacks of MIL programming is that due to its NP-hard nature, resolution times even to a modest accuracy cannot be predicted with certainty. Despite excellent performance (worst-case resolution time of about 10 seconds in the study presented), it may become a limitation for large network instances or particularly challenging problems. Therefore, future technical development will focus on the inclusion of a tailored heuristic resolution method which produces good feasible solutions in deterministic time.
[^1]: CSEM PV-center (e-mail: tomasz.gorecki@ csem.ch).
[^2]: This project has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement No 731125
[^3]: This work has been submitted to IFAC for possible publication
[^4]: http://www.pentagon-project.eu/
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
We prove that if $X:M^n\to{{\mathbb H}}^n\times{{\mathbb R}}$, $n\geq 3$, is a an orientable, complete immersion with finite strong total curvature, then $X$ is proper and $M$ is diffeomorphic to a compact manifold $\bar M$ minus a finite number of points $q_1, \dots q_k$. Adding some extra hypothesis, including $H_r=0,$ where $H_r$ is a higher order mean curvature, we obtain more information about the geometry of a neighbourhood of each puncture.
The reader will also find in this paper a classification result for the hypersurfaces of ${{\mathbb H}}^n\times {{\mathbb R}}$ which satisfy $H_r=0$ and are invariant by hyperbolic translations and a maximum principle in a half space for these hypersurfaces.
author:
- Maria Fernanda Elbert and Barbara Nelli
title: ' On the structure of hypersurfaces in ${{\mathbb H}}^n\times {{\mathbb R}}$ with finite strong total curvature '
---
[^1]
Introduction {#introduction .unnumbered}
============
Different notions of total curvature of a manifold M have been used in the literature. Classically a surface has [*finite total curvature*]{} if the norm of the gaussian curvature is integrable on $M$. On the other hand, a hypersurface $M$ of a Riemannian manifold has [*finite extrinsic total curvature*]{} if the norm of the second fundamental form of $M$ belongs to $L^n$. Here, by the norm of the second fundamental form we mean the euclidean norm of the vector formed by the principal curvatures of the hypersurface. Notice that, in the case of minimal surfaces in ${{\mathbb R}}^3,$ the two notions coincide. The geometry of minimal surfaces with finite total curvature have been widely studied (see [@MePe] for a survey). A classical result is due to Huber and Osserman [@Hu; @Oss2]:
[*Let $M$ be a complete oriented, immersed minimal surface in ${{\mathbb R}}^3$ with finite total curvature. Then $M$ is conformally equivalent to a compact Riemann surface $M$ with a finite number of points removed (called the ends of M). Moreover, the Gauss map extends meromorphically to the punctures.*]{}
The extrinsic total curvature was used, for example, by M. Anderson [@An] in order to generalise the previous result to minimal submanifolds of the Euclidean space of higher dimension and by B. White [@Wh], who dealt with surfaces of the Euclidean space satisfying properties different from minimality.
Our aim is to somehow generalize Osserman´s result to hypersurfaces of ${{\mathbb H}}^n\times{{\mathbb R}},$ $n>2,$ where ${{\mathbb H}}^n$ is the hyperbolic space of dimension n. The case $n=2$ has already been addressed in [@HaRo], [@HNST], where the authors prove:
([@HaRo Theorem 3.1 (c)], [@HNST]) [*Let $M$ be a complete minimal immersion in ${{\mathbb H}}^2\times{{\mathbb R}}$ with finite total curvature. Then $M$ is proper, it is conformally equivalent to a compact Riemann surface $M$ with a finite number of points removed (called the ends of M). Moreover the third coordinate of the unit normal vector $n_3$ converges to zero uniformly at each puncture. Finally the asymptotic boundary of each one of its ends can be identified with a special kind of closed polygonal curve in $\partial_\infty ({{\mathbb H}}^2\times{{\mathbb R}}$).*]{}
In the case $n>2$, we first consider a hypersurface with no pointwise assumption on the curvature. Inspired by the ideas of [@DE], we change the hypothesis on finite extrinsic total curvature by that of finite strong total curvature, i.e., we ask that the norm of the second fundamental form of the hypersurface belongs to a special weighted Sobolev space (see Section \[tot-curv-section\] for details). Then, we get the following result (see Theorem \[strong-normal\]).
*Let $X:M\to{{\mathbb H}}^n\times{{\mathbb R}}$, $n\geq 3$, be an orientable complete hypersurface finite strong total curvature. Then:*
- The immersion $X$ is proper.
- $M$ is diffeomorphic to a compact manifold ${\overline}M$ minus a finite number of points $q_1, \dots q_k$.
This result partially generalizes [@DE Theorem 1.1]. Adding some extra hypotheses on our hypersurface of ${{\mathbb H}}^n\times{{\mathbb R}}$, including $H_r=0$, we obtain the geometric behaviour of a neighbourhood of a puncture. Recall that $H_r$, $r\in\{1,\dots,n\},$ is the mean curvature of order $r$ of an $n$-hypersurface (see Section 1 for a precise definition). We prove the following.
For the definition of [*admissible collection*]{} see Definition \[admissible\].
Notice that when working with $n>2$, one looses the technical support of the complex analysis and with $r>1$, one weakens the technical support given by the theory of quasi-linear PDE. Then, it seems somehow reasonable to require a stronger hypothesis on the curvature in our context.
In order to prove (iii) and (iv), we use as barriers a family of hypersurfaces with $H_r=0$ which are invariant by hyperbolic translations, that we are able to construct (Theorem \[classification-theorem\]). As a by product of our construction, we prove Theorem \[iperplano\], which is a maximum principle at infinity for properly immersed hypersurfaces with $H_r=0$. To the best of our knowledge, this is the first maximum principle in a half space for hypersurfaces with $H_r=0.$ For this part of the article, we were inspired by the works [@BS1], [@N-SE-T] and [@ST2].
The paper is organized as follows. After fixing notations in Section \[notation\], in Section \[invariant-hypersurfaces-section\] we describe the family of hypersurfaces that are invariant by hyperbolic translations. In Section \[asymptotic-section\], we analyse the influence of the asymptotic boundary of hypersurfaces with $H_r=0,$ on their shape at finite points. Hypersurfaces with finite strong total curvature are studied in Section \[tot-curv-section\], with no assumption on $H_r.$ Finally in Section \[main-result\] we prove your main results Theorems \[strong-normal\], \[strong-normal-minimal\].
Notations {#notation}
=========
Let $ M^n$ be an orientable Riemannian $n$-manifold and let ${{{\mathbb H}}}^n$ be hyperbolic space (the simply connected Riemannian manifold with constant sectional curvature equal to -1). Let $X: M^n\to {{\mathbb H}}^n\times{{\mathbb R}}$ be an isometric immersion. The image $X(M)$ is a hypersurface of ${{{\mathbb H}}}^n\times{{\mathbb R}}$ and we shall identify $X(M)$ with $M$ throughout the paper.
For each $p\in M$, let $A:T_p M\to T_p M$ be the shape operator of $M$ and $\kappa_1,...,\kappa_ n$ be its eigenvalues corresponding to the eigenvectors $e_1,\ldots,e_n$. The [*higher order mean curvature*]{} of $M$ of order $r$ is defined as
$$H_r(p)=\frac{1}{\binom nr}\sum_{i_1<...<i_r}\kappa_{i_1}...\kappa_{i_r},$$
i.e., the normalized $r^{\rm th}$ symmetric function of $\kappa_1,...,\kappa_ n.$ When $H_{r}=0$, the immersion is called $(r-1)$-minimal. Thus, the classical minimal immersions would be the $0$-minimal ones.
We consider the ball model for the hyperbolic space $${{\mathbb H}}^n=\{x=(x_1,\ldots,x_n)\in{{\mathbb R}}^n \ \vert \ x_1^2+\ldots+x_n^2<1\}$$ endowed with the metric $$g_{{\mathbb H}}:=\frac{dx_1^2+\ldots+dx_{n}^2}{\left ( \frac{1-|x|^2}{2}\right )^2}=\frac{dx_1^2+\ldots+dx_{n}^2}{F^2},$$
where $|x|$ is the euclidean norm of $x$. As in [@ST], we define the [*asymptotic boundary*]{} of ${{\mathbb H}}^n\times{{\mathbb R}}$ as $$\partial_\infty({{\mathbb H}}^n\times{{\mathbb R}})=(\partial_\infty{{\mathbb H}}^n\times{{\mathbb R}})\cup ({{\mathbb H}}^n\times\{-\infty,\infty\})\cup(\partial_\infty{{\mathbb H}}^n\times\{-\infty,\infty\}).$$
Let $\Pi$ be a totally geodesic hyperplane in ${{\mathbb H}}^n.$ The asymptotic boundary of $\Pi$ splits $\partial_{\infty}{{\mathbb H}}^n$ into two connected components. Each component can be identified with a spherical cap of the $(n-1)$-dimensional unit sphere. We set $\partial_\infty {{\mathbb H}}^n=S_-^{n-1}\cup S_+^{n-1}$, where $S_-^{n-1}$ and $S_+^{n-1}$ are the closure of the two spherical caps determined by $\Pi$ .
Let $\Omega\subset {{\mathbb H}}^n\times{{\mathbb R}}$ be a nonempty subset. We say that a point $p_\infty\in \partial_\infty({{\mathbb H}}^n\times{{\mathbb R}})$ is [*an asymptotic point of*]{} $\Omega$ if there is a sequence $\{p_n\}$ of points of $\Omega$ converging to $p_\infty$. The set of asymptotic points of $\Omega$, called the [*asymptotic boundary*]{} of $\Omega$, is denoted by $\partial_\infty\Omega$.
In what follows, we often identify the slice ${{\mathbb H}}^n\times\{0\}$ with ${{\mathbb H}}^n.$ By [*$vertical$ hyperplane*]{} we mean a complete totally geodesic hypersurface $\Pi\times {{\mathbb R}}$, where $\Pi$ is any totally geodesic hyperplane of ${{\mathbb H}}^n.$ We call a [*vertical halfspace*]{} any component of $({{\mathbb H}}^n \times {{\mathbb R}})\setminus P$, where $P$ is a vertical hyperplane.
For a fixed totally geodesic hyperplane $\Pi$ of ${{\mathbb H}}^n\times\{0\},$ let $L^+_{\rho}$ and $L^-_{\rho}$ be the equidistant hypersurfaces to $\Pi$, at distance $\rho$, in the slice ${{\mathbb H}}^n\times \{0\}$. Denote by $Z_{\rho}^+$ the closure of the non mean convex side of the cylinder over the hypersurface $L_{\rho}^+$ in ${{\mathbb H}}^n \times {{\mathbb R}}$. Analogously, we define $Z_{\rho}^-.$ We will call the set $C_{\rho}={{\mathbb H}}^n\times{{\mathbb R}}\setminus Z_{\rho}^+\cup Z_{\rho}^-$ [*$\rho$-cylinder associated to*]{} $\Pi$ (see Figure \[rho-cylinder\]).
 \[rho-cylinder\]
Hypersurfaces with $H_r=0$ invariant by hyperbolic translations {#invariant-hypersurfaces-section}
================================================================
We describe a family of hypersurfaces in ${{\mathbb H}}^n\times{{\mathbb R}}$ with $H_r=0$ which are invariant by a special family of isometries of ${{\mathbb H}}^n\times{{\mathbb R}}.$
Let $\gamma$ be a complete geodesic through the origin $\sigma$ of the hyperbolic space ${{\mathbb H}}^n$, parametrized by the signed distance $\rho$ to $\sigma$. Let $\Pi$ be the hyperbolic hyperplane orthogonal to $\gamma$ at $\sigma$. For each geodesic $\beta$ in $\Pi$, passing through $\sigma$, we consider the hyperbolic translation along $\beta$ in ${{\mathbb H}}^n.$ We notice that the image of any point of $\gamma$ under the hyperbolic translations along all geodesics of $\Pi$ passing through $\sigma$ is an equidistant hypersurface to $\Pi$ in ${{\mathbb H}}^n\times\{0\}.$ We extend the hyperbolic translation along $\beta$ [*slice-wise*]{} to an isometry of ${{\mathbb H}}^n\times{{\mathbb R}}$. By abuse of notation, this isometry of ${{\mathbb H}}^n\times{{\mathbb R}}$ will also be called [*hyperbolic translation along*]{} $\beta.$
We show the existence of a family of hypersurfaces of ${{\mathbb H}}^n\times {{\mathbb R}}$ with $H_r=0$ which are invariant by hyperbolic translations along all geodesics of $\Pi$ passing trough $\sigma$. Moreover we give a complete geometric description of the family. The case of minimal hypersurfaces, that is $r=1$, is treated in [@ST2] and [@BS1].
A generating curve parametrized by $ (\tanh(\rho/2),\lambda (\rho))$ in the vertical 2-plane, $\gamma\times{{\mathbb R}},$ gives rise, under the previous isometry, to a translationally invariant hypersurface $M$ in ${{\mathbb H}}^n\times{{\mathbb R}}$ whose intersection with ${{\mathbb H}}^n \times\{\lambda(\rho)\}$ is the equidistant hypersurface to $\Pi\times\{\lambda(\rho)\}$, at distance $\rho.$
The principal directions of the hypersurface $M$ are the tangent vectors to the generating curve and to the equidistant hypersurface. The corresponding principal curvatures are the following (see [@BS1]):
$$\label{princ-curv}
\kappa_1=\ddot\lambda(\rho)(1+\dot\lambda(\rho)^2)^{-\frac{3}{2}}, \ \
\kappa_2=\dots= \kappa_n=\dot\lambda(\rho)(1+\dot\lambda(\rho)^2)^{-\frac{1}{2}}\tanh(\rho)$$
where $(\ \dot{}\ )$ means the derivative with respect to $\rho.$ It follows that
$$\label{Hr}
nH_r\frac{\cosh^{n-1}(\rho)}{\sinh^{r-1}(\rho)}=\frac{\partial }{\partial\rho}\left[\cosh^{n-r}(\rho)\left(\frac{\dot{\lambda}^2}{1+\dot{\lambda}^2}\right )^{\frac{r}{2}}\right].$$
We prove the following theorem.
\[classification-theorem\] Let $\Pi$ be a totally geodesic hyperplane of ${{\mathbb H}}^n\times\{0\}$ passing through $\sigma$ and let $r\leq n$. Then there exists a one parameter family $\{ {\mathcal M}^r_d,\, d>0\}$ of complete properly embedded hypersurfaces in ${{\mathbb H}}^n \times {{\mathbb R}}$, with $H_r=0$, invariant under hyperbolic translations along all the geodesics of $\Pi$ passing through $(\sigma,0).$ For $r=n$, the parameter $d$ assumes also the value $d=0$. The families are described below.
1. $r=n$:\
- ${\mathcal M}^r_0$ is a slice.
- When $d>0$, ${\mathcal M}^r_d$ is, up to vertical translation, a complete graph, symmetric with respect to $\Pi$, whose asymptotic boundary is composed by $((\Pi\cap\partial_\infty{{\mathbb H}}^n)\times{{\mathbb R}})\cup(\partial S_-^{n-1}\times\{-\infty\})\cup (\partial S_+^{n-1}\times\{+\infty\})$. Here, $\partial S_{\pm}^{n-1}$ are the hemispheres determined by $\Pi$.
2. $r<n$:\
- When $d>1,$ ${\mathcal M}^r_d$ consists of the union of two vertical hypersurfaces of finite height, symmetric with respect to ${{\mathbb H}}^n\times \{0\}$, contained in $Z_{\rho_d}^+$, where $\rho_d=\cosh^{-1}(d^{r/n-r})$.
The asymptotic boundary of ${\mathcal M}^r_d$ is topologically an $(n-1)$-sphere which is homologically trivial in $\partial_{\infty}
{{\mathbb H}}^n \times {{\mathbb R}}$. More precisely, if we set $d=\cosh^{\frac{n-r}{r}} a$, we have that $$h_r(d)=\cosh(a)\int_1^{\infty}\frac{dv}{(v^{2q}-1)^{\frac{1}{2}}(\cosh^2(a)v^2-1)^{\frac{1}{2}}} ,$$
is finite and the asymptotic boundary of ${\mathcal M}^r_d$ consists of the union of two copies of an hemisphere $S_+^{n-1}\times
\{0\}$ of $\partial_\infty {{\mathbb H}}^n \times \{0\}$ in parallel slices $t=\pm h_r(d)$, glued with the finite cylinder $\partial S_+^{n-1}\times [-h_r(d),h_r(d)].$
The vertical height of ${\mathcal M}^r_d$ is then $2h_r(d)$. The height of the family ${\mathcal M}^r_d$ is a decreasing function of $d$ and varies from infinity (when $d\longrightarrow 1$) to $\frac{\pi r }{(n-r)}$ (when $d\longrightarrow \infty$).
- If $d=1$, then, up to reflection over a slice, ${\mathcal M}^r_1$ consists of a complete $($non-entire$)$ vertical graph over a halfspace in ${{\mathbb H}}^n\times \{0\}$, bounded by the totally geodesic hyperplane $\Pi$. It takes infinite boundary value on $\Pi$ and constant value data $c$ on the asymptotic boundary of the halfspace. The asymptotic boundary of ${\mathcal M}^r_1$ is composed by $\partial_{\infty}\Pi\cap\{{{\mathbb H}}^n\times\{\infty\}\}$, by a hemisphere $S_+^{n-1}\times\{c\}$ of $\partial_\infty {{\mathbb H}}^n\times \{c\}$ and by the a half vertical cylinder over $\partial(S_+^{n-1}\times\{c\})$.
- If $d<1$, then ${\mathcal M}^r_d$ is an entire vertical graph with finite vertical height. Its asymptotic boundary consists of a homologically non-trivial $(n-1)$-sphere in $\partial_\infty {{\mathbb H}}^n
\times {{\mathbb R}}.$
In each case, we determine the profile curve. The corresponding hypersurface is given by the orbits of the points of the profile curve by the hyperbolic translations along all the geodesics of $\Pi$ passing through the origin of ${{\mathbb H}}^n\times\{0\}.$ The properties of the hypersurfaces will be clear from our description of the profile curve.
Let us first prove (a). By taking $H_r=0$ and $r=n$ in equation , we easily get that $\dot \lambda(\rho)=d,$ for $d\geq 0$. Then, we have $\lambda (\rho)=d\rho+c$, for a real constant $c.$
Now, we notice that for $d>0$, this straight line gives a profile curve in $\gamma\times{{\mathbb R}}$, parametrized by $(x=\tanh(\rho/2),\lambda(x))$, that is symmetric with respect to $(0,c)$, is increasing from $x=-1$ to $x=1$ and satisfy
$$\lim_{x\to -1}\lambda(x)=-\infty\;\;\;\;\mbox{and}\;\;\;\;\lim_{x\to 1}\lambda(x)=+\infty.$$
This finishes the proof of $(a)$.
Now we prove (b). Taking $H_r=0$ in equation , one easily gets that there exists a constant $d$, with $0<d^r\le \cosh^{n-r}(\rho)$, such that
$$\label{eq1}
\cosh^{n-r}(\rho)\left(\frac{\dot{\lambda}^2}{1+\dot{\lambda}^2}\right )^{\frac{r}{2}}=d^r.$$
We set $q=\frac{n-r}{r}$ and, by a straightforward computation, we obtain
$$\label{eq2}
\dot\lambda^2(\rho)=\frac{d^2}{\cosh^{2q}(\rho)-d^2}.$$
Here, we can choose $\dot\lambda$ to be the positive square root in since, up to a reflection across a slice in ${{\mathbb H}}^n\times {{\mathbb R}}$, the negative root would give rise to the same solution. We divide our study in three cases, depending on $d>1,$ $d=1,$ $0<d<1.$
- $d>1$.\
Let $a>0$ be such that $d=\cosh^{\frac{n-r}{r}}(a).$ Then, after integration, we have
$$\label{eq4}
\lambda(\rho)=\int_a^{\rho}\frac{d}{\sqrt{\cosh^{2q}(\xi)-d^{2}}}d\xi.$$
By the change of variables $v=\frac{\cosh(\xi)}{\cosh(a)}$, we can rewrite as
$$\label{eq5}
\lambda(\rho)=\cosh(a)\int_1^{\frac{\cosh(\rho)}{\cosh(a)}}\frac{1}{(v^{2q}-1)^{\frac{1}{2}}(\cosh^2(a)v^2-1)^{\frac{1}{2}}} dv.$$
It is easy to see that the integral in converges at $v=1$ and when $\rho$ goes to infinity.
Then we can define
$$\label{eq6}
h_r(d):=\cosh(a)\int_1^{\infty}\frac{1}{(v^{2q}-1)^{\frac{1}{2}}(\cosh^2(a)v^2-1)^{\frac{1}{2}}} dv$$
and $2h_r(d)$ will be the height of the hypersurface ${\mathcal M}_d^r.$
A simple computation shows that $$h_r(d)\geq A\int_1^2 [(v-1)(\cosh (a) v-1)]^{-\frac{1}{2}}dt,$$
where $A$ is a positive constant. The latter integral can be computed explicitly and diverges when $a\longrightarrow 0,$ that is when $d\longrightarrow 1.$
Moreover, the limit when $a\longrightarrow \infty$ can be taken under the integral and
$$\label{eq9}
\lim_{a\to\infty}h_r(d)=\int_{1}^{\infty} v^{-1}(v^{2q}-1)^{-1/2}dv=\frac{\pi r}{2(n-r)},$$
where in the last equality we use that $\int v^{-1}(v^{2q}-1)^{-1/2}dv= \frac{1}{q}\arctan(\sqrt{v^{2q}-1})+{\rm const}.$
Finally, since
$$\label{S-derivative}
\frac{dh_r}{da}=-\sinh(a)\int_{1}^{\infty} (v^{2q}-1)^{-1/2}(\cosh^2 (a)v^2-1)^{-3/2}dv<0,$$
we conclude that the function $a\to h_r(a)$ decreases from $\infty$ to $\frac{\pi r}{2(n-r)},$ when $a$ increases from $0$ to $\infty$.
- $d=1.$
By replacing $d=1$ in equation one has that
$$\label{eq10}
\lambda(\rho)=\int_b^{\rho}\frac{1}{\sqrt{\cosh^{2q}(\xi)-1}} d\xi$$
for some constant $b>0.$ It is easy to see that this profile curve tends to $-\infty $, when $\rho\longrightarrow 0$, and tends to a finite value, when $\rho\longrightarrow \infty$.
- $0<d<1.$
In this case one has
$$\label{eq11}
\lambda(\rho)=d\int_0^{\rho}\frac{1}{\sqrt{\cosh^{2q}(\xi)-d^{2}}}d\xi.$$
The curve is defined for every value of $\rho> 0$ and can be extended by symmetry to values $\rho<0.$ Moreover $\lambda$ is bounded. The corresponding hypersurface is an entire vertical graph with finite height.
For future use, we prove a useful property of the hypersurfaces ${\mathcal M}_d^r$.
\[sinal\] For a fixed $r<n$ and for any $d>1$, each hypersurface ${\mathcal M}_d^r$ satisfies:
$$H_j>0, \ \ j<r \ \ {\rm and} \ \ H_j<0, \ \ r<j\leq n.$$
Let us compute $H_j$ for any $0<j\leq n.$ It is straightforward to see that
$$\label{Hj}
nH_j=((n-j)\kappa_2+j\kappa_1)\kappa_2^{j-1},$$
where $\kappa_1$ and $\kappa_2$ are defined as in .
Notice that by deriving one obtains that
$$\label{ddot-lambda}
\ddot \lambda(\rho)=-\frac{(n-r)}{rd^2}\dot\lambda^3(\rho)\tanh(\rho)(\cosh(\rho))^{\frac{2(n-r)}{r}}.$$
By replacing and in , one obtains
$$\label{hj-final}
H_j= \frac{n(r-j)}{rd^2}\kappa_2^{j-1}\tanh(\rho)(\cosh(\rho))^{\frac{2(n-r)}{r}}
\frac{\dot\lambda^3}{(1+\dot\lambda^2)^{\frac{3}{2}}}.$$
This proves the result, since $ \kappa_2$ and $\dot\lambda$ are positive.
Let $\Pi$ and $\bar{\Pi}$ be totally geodesic hyperplanes of ${{\mathbb H}}^n\times\{0\}$, where $\Pi$ passes through the origin. Let $\gamma$ and $\bar{\gamma}$ be the geodesics that are, respectively, orthogonal to $\Pi$ at $\sigma$ and to $\bar{\Pi}$ at a point $p$. Let $\Phi$ be an isometry of the ambient space that takes $\Pi$ into $\bar{\Pi}$, takes $\gamma$ into $\bar{\gamma}$ and that preserves the $t$-coordinate. We notice that, by applying $\Phi$ to each family ${\mathcal M}^r_d$ constructed in Theorem \[classification-theorem\], we obtain a one parameter family of hypersurfaces invariant under hyperbolic translations along the geodesics of $\bar{\Pi}$ passing through $p$. In the next sections, by abuse of notation, we will denote by ${\mathcal M}^r_d$ any hypersurfaces obtained from ${\mathcal M}^r_d$ applying an isometry like $\Phi.$
Maximum Principle and Asymptotic Theorems {#asymptotic-section}
=========================================
In this section, we use the translationally invariant hypersurfaces ${\mathcal M}^r_d$, constructed above, and a maximum principle, in order to investigate how the boundary behaviour of a hypersurface with $H_r=0$ contained in a halfspace, constrains the behaviour of the hypersurface at finite points. Moreover we prove an obstruction result for hypersurfaces with $H_r=0$ and a given boundary.
The suitable version of maximum principle for our purposes is stated below. For further details about such generalized maximum principles, see [@ENS], [@FS] [@HL1],[@HL2].
[@FS Theorem 2.a] [*Let $M$ and $M'$ two oriented hypersurfaces with $H_r=H_r'\equiv 0,$ tangent at a point $p,$ with normal vector pointing in the same direction. Suppose that $M$ remains on one side of $M'$ in a neighborhood of $p.$ Suppose further that $H_j'(p)\geq 0,$ $1\leq j\leq r$ and either $H_{r+1}\not=0$ or $H'_{r+1}\not=0.$ Then $M$ and $M'$ coincide in a neighborhood of $p.$* ]{}
Given $r\in\{1,\dots,n\}$, and a totally geodesic hyperplane $\Pi$ in ${{\mathbb H}}^n\times\{0\},$ we consider the hypersurfaces ${\mathcal M}^r_d$, with $d>1$, described in Section \[invariant-hypersurfaces-section\] (see the last sentence of Section \[invariant-hypersurfaces-section\]). We notice that, by Proposition \[sinal\], all the hypersurfaces ${\mathcal M}^r_d$, $d>1$, satisfy the assumptions of $M'$ in the maximum principle.
Denote by $Q_{\Pi}$ the halfspace determined by $\Pi\times{{\mathbb R}}$ which contains $Z_{\rho_d}^+.$ Clearly, any vertical translation of the hypersurface ${\mathcal M}^r_d$ is contained in $Z_{\rho_d}^+$ and, moreover, any vertical translation of ${\mathcal M}^r_d$ is arbitrarily close to $\partial Q_\Pi=\Pi\times{{\mathbb R}},$ provided $d$ is sufficiently close to one.
The proof of Theorem \[iperplano\] below is inspired by that of [@N-SE-T Theorems 3.2, 4.5].
\[iperplano\] Let $M$ be a hypersurface, with $H_r=0$, properly immersed in ${{\mathbb H}}^n\times{{\mathbb R}}$. Let $P$ be a vertical hyperplane and $P_+$ one of the two halfspaces determined by $P.$ If $\partial M \subset\overline{ P_+}$ ($\partial M$ possibly empty) and $\partial_{\infty}M\cap(\partial_{\infty}{{\mathbb H}}^n\times{{\mathbb R}})\subset\partial_{\infty} P_+,$ then $M\subset \overline{ P_+}$.
Let $\Pi \subset {{\mathbb H}}^ n\times\{0\}$ be a totally geodesic hyperplane and $Q_{\Pi} $ the half space determined by $\Pi$, chosen such that
$$(\mathcal{P}) \ \ \ \ \ \ \ \ \ \ Q_{\Pi} \subset ({{\mathbb H}}^ n\times {{\mathbb R}})\setminus P_+, \ \ {\partial_{\infty}}\Pi\cap {\partial_{\infty}}P=\emptyset$$
We fix a $d>1$ and we consider the family of hypersurfaces ${\mathcal M}^r_d$ contained in $Z_{\rho_d}^+\subset Q_{\Pi}$. The following two properties hold:
1. \[item.asymptotic\]The intersection of $\partial_{\infty}M$ with $\partial_{\infty}({{\mathbb H}}^ n\times{{\mathbb R}})\setminus \partial_{\infty} P_+$ contains no points at finite height.
2. The asymptotic boundary of any vertical translation of ${\mathcal M}^r_d$ is contained in the asymptotic boundary of $ Q_{\Pi}\subset{{\mathbb H}}^n\times{{\mathbb R}}\setminus P_+.$
We will get the result by applying the maximum principle between the hypersurface $M$ and some isometric copy of ${\mathcal M}^r_d$’s.
Let $\gamma$ be the geodesic in ${{\mathbb H}}^n\times \{0\}$, orthogonal to $\Pi$ at a point $p\in\Pi.$ We parametrize $\gamma$ by the signed distance to $p,$ say $s,$ with orientation pointing towards $Q_{\Pi}.$ For any $s,$ we consider the isometry of ${{\mathbb H}}^ n\times{{\mathbb R}}$ that preserves the $t$-coordinate and takes $\Pi$ into the geodesic hyperplane orthogonal to $\gamma$ at a distance $s$ from $p$. By letting $s\longrightarrow \infty$ and by applying the above isometries, we obtain a family of hypersurfaces ${\mathcal M}^r_d(s)$, isometric to ${\mathcal M}^r_d$, that collapses to a vertical segment in $(\partial_{\infty}{{\mathbb H}}^n\times{{\mathbb R}})\cap\partial_\infty Q_{\Pi}$ of height $2h_r(d)$. We claim that, for some $s,$ $M$ and ${\mathcal M}^r_d(s)$ are disjoint. In fact, suppose that, when $s\longrightarrow \infty$, ${\mathcal M}^r_d(s)$ always have a nonempty intersection with $ M$. Then, there would exists a point of the asymptotic boundary of $M$ at finite height in $\partial_{\infty}({{\mathbb H}}^ n\times{{\mathbb R}})\setminus \partial_{\infty}P_+$, giving a contradiction with (I). Then, the claim is proved. Now, starting with a ${\mathcal M}^r_d(s)$ disjoint from $M,$ we apply the (inverse) isometries to come back towards the original position, that is, we let $s\longrightarrow 0$. Then, either we find a first intersection point between $M$ and ${\mathcal M}^r_d(s)$, for some $s,$ contradicting the maximum principle, or we reach the original position without touching $M$. The same process can be done with any vertical translation of ${\mathcal M}^r_d$ and we can conclude that $M$ is contained in ${{\mathbb H}}^ n\times{{\mathbb R}}\setminus Z_{\rho_d}^+$.
Now, we let $d\to 1$ and the maximum principle yields that $M$ is contained in the closed halfspace ${{\mathbb H}}^ n\times{{\mathbb R}}\setminus Q_{\Pi}$. Since this holds for any totally geodesic hyperplane $\Pi$ satisfying the property $(\mathcal P),$ we conclude that $M$ is contained in the closure of $P_+$.
Let us extend Theorem \[iperplano\] to the case of a more general asymptotic boundary.
\[admissible\] Let $\Pi_1,\dots\Pi_k$ be a collection of hyperplanes in ${{\mathbb H}}^n$ such that $\partial_{\infty}\Pi_i=S_i$, where for $i=1,\dots,k,$ $S_i$ is an $(n-2)$-sphere in $\partial_{\infty}{{\mathbb H}}^n$. We say that the hyperplanes $\Pi_1,\dots\Pi_k$ are an [*admissible collection*]{} if it is possible to choose open (n-1)-spheres $B_1,\dots,B_k$ in $\partial_{\infty}{{\mathbb H}}^n$, bounded by $S_1,\dots S_k$, which are mutually disjoint.
\[halfspace-def\] Let $\Pi_1,\dots\Pi_k$ be an admissible collection of hyperplanes in ${{\mathbb H}}^n$ and let $P_j=\Pi_j\times{{\mathbb R}}$, $j=1,\dots,k,$ be the corresponding vertical hyperplanes in ${{\mathbb H}}^n\times{{\mathbb R}}$. Denote by $\tilde P_j$ the half-space such that ${\displaystyle \cup_{i\neq j}}\left (\Pi_i\times{{\mathbb R}}\right )\subset\tilde P_j.$ We define $P(\Pi_1,\dots,\Pi_k):=\cap_{i=1}^k\tilde P_i.$
Notice that $\Pi_i$ and $\Pi_j,$ $i\not=j,$ can meet at most at one point. This yields that $\partial_{\infty}P_i$ and $\partial_{\infty}P_j,$ $i\not=j,$ can meet at most at a vertical line.
\[halfspace-coro\] Let $M$ be a complete hypersurface with $H_r=0,$ possibly with finite boundary, properly immersed in ${{\mathbb H}}^n\times{{\mathbb R}}$ and let $\Gamma=\partial_{\infty}M\cap (\partial_{\infty}{{\mathbb H}}^n\times{{\mathbb R}}).$ Let $\Pi_1,\dots,
\Pi_k$ an admissible collection of hyperplanes. If $\Gamma\subset
\partial_{\infty}P(\Pi_1,\dots, \Pi_k)$ and $\partial M\subset \overline{ P(\Pi_1,\dots, \Pi_k)},$ then $M$ is contained in $\overline{P(\Pi_1,\dots, \Pi_k)}$.
Next result establishes some obstruction to the existence of a hypersurface in ${{\mathbb H}}^n\times{{\mathbb R}}$ with $H_r=0:$ in particular the shape of the asymptotic boundary of a hypersurface may prevents the hypersurface to have $H_r=0.$ The result is a generalization of [@ST2 Corollary 2.2] and [@N-SE-T Theorem 4.6].
\[T.slab.catenoid\] Let $S_\infty \subset \partial_{\infty} {{\mathbb H}}^n \times {{\mathbb R}}$ be a closed set whose vertical projection on $\partial_{\infty} {{\mathbb H}}^n \times \{0\}$ omits an open subset. Assume that $S_\infty $ is contained in an open slab whose height is equal to $\frac{r\pi}{n-r}$. Then, there is no connected hypersurface $M$ with $H_r=0,$ $\partial M=\emptyset,$ properly embedded in ${{\mathbb H}}^n \times {{\mathbb R}},$ with asymptotic boundary $S_\infty$.
Assume, by contradiction, that there exists a hypersurface $M$, satisfying the assumptions and with $\partial_{\infty} M=S_{\infty}$. Then, up to a vertical translation, we can assume that $M$ is contained in the slab ${\mathcal{B}}:=\{(p,y)\in {{\mathbb H}}^n \times{{\mathbb R}};\; t_0\leq t\leq \frac{r\pi}{n-r}-t_0\}$ for some $t_0>0$ (see [@ENS Proposition 3.1]) and $S_\infty\subset \partial_{\infty} \mathcal{B}$. As the vertical projection of $S_{\infty}$ omits an open subset, say $U$, by Theorem \[iperplano\], we find a totally geodesic hyperplane $\Pi\subset {{\mathbb H}}^n\times \{0\} $ such that a component, say $\Pi^+$, of ${{\mathbb H}}^n \times \{0\} \setminus \Pi$ satisfies:
1. $\partial_{\infty}\Pi^+ \subset U$.
2. $M \cap ( \overline{\Pi}^+ \times {{\mathbb R}}) =\emptyset$.
Let ${\mathcal C}\subset {{\mathbb H}}^n \times (0,\frac{r\pi}{n-r})$ be any $n$-catenoid with $H_r=0$, such that a component of its asymptotic boundary stays strictly above $\partial_{\infty}\mathcal{S}$ and the other component stays strictly below $\partial_{\infty}\mathcal{S}$. The existence of such catenoids is proved in [@ENS Theorem 2.1]. There, it is also proved that the the $j$-mean curvatures of the catenoids satisfy $H_j(p)< 0,$ $1\leq j\leq r$ and $H_{r+1}<0$ (see [@ENS Proposition 2.2]).
We define $K={\mathcal B}\cap{\mathcal C}.$ $K$ is compact, connected and its boundary lies in the boundary of the slab ${\mathcal{B}}$.
Let $q\in M$ be a point, let $q_0\in {{\mathbb H}}^n \times \{0\}$ be the vertical projection of $q$ and let $p_\infty$ be a point in $ \partial_{\infty}\Pi^+.$ Denote by $\widetilde \gamma\subset {{\mathbb H}}^n \times \{0\} $ the complete geodesic passing through $q_0$ such that $p_\infty \in \partial_{\infty}\widetilde \gamma$. We can translate $K$ along $\widetilde \gamma$ (with the usual isometry of ${{\mathbb H}}^n\times{{\mathbb R}}$ that preserves the $t$-coordinate), such that the translated $K$ is contained in the halfspace $\Pi^+ \times {{\mathbb R}}$.
Now we come back translating $K$ towards $M$ along $\widetilde \gamma.$ Observe that the boundary of the translated copies of $K$ does not touch $M$. Therefore, doing the translations of $K$ along $\widetilde \gamma$ we find a first interior point of contact between $M$ and a translated copy of $K$. Hence, $M={\mathcal C}$ by the maximum principle, which leads to a contradiction and completes the proof.
Similar techniques may be applied to generalise non existence results analogous to [@N-SE-T Theorem 4.6], [@ST2 Theorem 2.1].
Finite Strong Total Curvature {#tot-curv-section}
=============================
In this section, we deal with a general isometric immersion $X: M\longrightarrow{{\mathbb H}}^n\times{{\mathbb R}}$ without any assumption on $H_r$. The notion of strong total curvature, introduced by the first author and M. Do Carmo [@DE] for hypersurfaces in ${{\mathbb R}}^{n+1}$, is defined as a special norm of the shape operator, $A$, of $M=X(M)$.
Let $p_0$ be a fixed point of $M$ and denote by $\xi(p)$ the intrinsic distance in $M$ from $p$ to $p_0.$
Let $\Omega\subset M$. Given any $q\geq 1$, we define the following two function spaces.
- $L^q _s (\Omega)$ is the [*weighted space* ]{} of weight $s\in{{\mathbb R}}$ of all measurable functions of finite norm $$||u||_{L^q_s (\Omega)}=\left ( \int_\Omega |u|^q \xi^{-qs-n} \;dM\right)^{1/q}\hspace{-15pt}.$$
- $W^{1,q} _s(\Omega)$ is the [*weighted Sobolev space*]{} of weight $s$ of all measurable functions of finite norm
$$||u||_{W^{1,q} _s(\Omega)}=||u||_{L^q _{s} (\Omega)}+||\nabla u||_{L^q _{s-1} (\Omega)},$$
where $\nabla u$ is the gradient of $u$ in $M$.
The latter was used by Bartnik, in a pioneer paper [@B], to define a suitable decay at infinity of the metric of a manifold (asymptotically flat spaces) that guarantees that the ADM-mass is a geometric invariant. We point out that, since then, it was used by a lot of authors and the literature about the subject is wide.
\[defi-finite-strong\] Let $M$ be a hypersurface of ${{\mathbb H}}^n\times{{\mathbb R}}$ and $A$ its shape operator. We define the quantity $||\,|A|\,||_{W^{1,q} _{-1}(M)}$ to be the [*strong total curvature*]{} of the immersion $M$ and we say that the immersion has [*finite strong total curvature*]{} if
$$\label{FSTC-def}|A|\in W^{1,q} _{-1}(M),\;\;{\rm for\; some }\; q>n,$$
where $|A|$ is the norm of the shape operator.
Notice that the definition of strong total curvature does not depend on the choice of the point $p_0$ and that can be written as follows: $$||\,|A|\,||_{W^{1,q} _{-1}(M)}=\left ( \int_M |A|^q \xi^{q-n} \;dM\right)^{1/q}+\left ( \int_M |\nabla |A||^q \xi^{2q-n} \;dM\right)^{1/q}<\infty,\;\;\mbox{for\; some}\; q>n.$$
We point out that the norm $||\,|A|\,||_{W^{1,q} _{-1}(M)}$ is invariant by dilations of the intrinsic metric of $M.$
As in [@DE], we will estimate the rate of the decay at infinity of $|A|$ (see Proposition \[decay\]).
Next lemma is analogous to [@DE Lemma 3.1].
\[compactness-lemma\] Let $B\subset {{\mathbb H}}^n\times{{\mathbb R}}$ be a bounded domain with smooth boundary $\partial B.$ Let $\{W_i\}$ be a sequence of connected n-manifolds and let $X_i:W_i\to {{\mathbb H}}^n\times{{\mathbb R}}$ be hypersurfaces such that $ \partial X_i(W_i)\cap B=\emptyset$ and $ X_i(W_i)\cap B = M_i$ is connected and nonempty. Assume that there exists a constant $C>0$ such that, for every $i,$ ${\displaystyle\sup_{x\in M_i}|A_i(x)|^2<C},$ where $A_i$ is the shape operator of $M_i,$ and that there exists a sequence of points $\{x_i\}$, $x_i\in M_i$, with a limit point $x_0\in B$. Then:
- A subsequence of $(M_i)$ converges $C^{1,\lambda}$ on the compact parts (see the definition below) to a union of hypersurfaces $M_\infty \subset B$, where $\lambda<1$.
- If, in addition, $\left ( \int_{M_i} |A|^q \alpha_i \;dM_i\right)^{1/q}+\left ( \int_{M_i} |\nabla |A||^q \beta_i \;dM_i\right)^{1/q}\to 0,$ for sequences $\{\alpha_i\}$ and $\{\beta_i\}$ of continuous functions on $M_i$ satisfying ${\displaystyle\inf_{x\in
M_i}\{\alpha_i,\beta_i\}\geq \kappa}>0$, then a subsequence of $|A_i|$ converges to zero everywhere.
By $C^{1,\lambda}$ convergence to $M_\infty$ on compact sets we mean that for any $m\in
M_\infty$ and each tangent plane $T_m M_\infty$ there exists a ball $B(m)$ of ${{\mathbb H}}^n\times{{\mathbb R}}$ around $m$ so that, for $i$ large, the image by $X_i$ of some connected component of $X_i^{-1}(B(m)\bigcap M_i)$ is the graph over a part of $T_m M_\infty$ of a function $g_i^m$ and the sequence $g_i^m$ converges $C^{1,\lambda}$ to the function $g_\infty$, that defines $M_\infty$ as a graph over a neighbourhood of $m$ in the chosen plane $T_m M_\infty$.
In i) and ii) we can work in compact subsets $B$ of ${{\mathbb H}}^n\times{{\mathbb R}}$. By using [@ST Proposition 3.1], we can treat $M_i\cap B$ as a sequence of submanifolds of ${{\mathbb R}}^{n+1}$ with uniformly bounded second fundamental form. Then we can use the proof of [@DE Lemma 3.1] in order to conclude our proof.
For the proof of the following proposition we refer the reader to the proof of [@DE Proposition 3.2], with the following precautions.
1. All the rescales of the metric on the hypersurfaces come from a conformal changing on the metric of ${{\mathbb H}}^n\times{{\mathbb R}}$ by a constant conformal factor.
2. The convergences needed in the proof are guaranteed by Lemma \[compactness-lemma\].
\[decay\] Let $M$ be a complete hypersurface in ${{\mathbb H}}^n\times {{\mathbb R}}$ with finite strong total curvature. Then, given $\varepsilon>0$ there exists $R_0>0$ such that, for $R>R_0$,
$$R^2\sup_{x\in M-D_R(p)}|A|^2(x)<\varepsilon.$$
where $D_R(p)$ is the intrinsic open $n$-ball of $M$ centered at a point $p\in M$ of radius $R.$
In Lemma \[identities\] and Lemma \[proper-lemma\] below, we explore the inclusion ${{\mathbb H}}^n\times {{\mathbb R}}\subset{{\mathbb R}}^{n+1}$ and we consider the canonical basis $\{e_1,\dots,e_n,e_{n+1}\}$ of ${{\mathbb R}}^{n+1}$ as a basis at each tangent plane of ${{\mathbb H}}^n\times {{\mathbb R}}$. Let $\Vert\; \Vert^2=\langle\;,\;\rangle$ and ${\overline}\nabla$ denote, respectively, the metric and the covariant derivative of ${{\mathbb H}}^n\times{{\mathbb R}}$. For a vector field $V(q)=\sum v_i(q)e_i+v_{n+1}(q)e_{n+1}$ in ${{\mathbb H}}^n\times{{\mathbb R}}$, where $q=\sum q_ie_i+q_{n+1}e_{n+1}\in {{\mathbb H}}^n\times {{\mathbb R}}$, we have
$$\Vert V(q)\Vert^2=\frac{4\sum_{i=1}^nv_i^2(q)}{\left(1-\sum_{i=1}^nq_i^2\right)^2}+v^2_{n+1}(q),
\ \
\Vert q\Vert^2=\frac{4\sum_{i=1}^nq_i^2}{(1-\sum_{i=1}^nq_i^2)^2}+q^2_{n+1}.$$
Then, it is clear that:
$$\label{limitada}
\Vert q\Vert\to\infty\;\; \mbox{iff} \;\;q\to \partial_{\infty}({{\mathbb H}}^n\times{{\mathbb R}})$$
or, equivalently, that, there exists $K_0>0$ such that if $\Vert q\Vert<K_0,$ there exist $k_0>0,$ $t_0>0$ such that ${\displaystyle\sum_{i=1}^n}q_i^2<k_0<1$ and $q_{n+1}^2<t_0.$
In the next lemma, we consider the vector field $X(p)=\sum x_i(p)e_i+x_{n+1}(p)e_{n+1}$, $p\in M$, given by the immersion and establish some elementary useful identities.
We have
- ${\overline}\nabla_{e_j}X=Le_j$ and $ {\overline}\nabla_{e_{n+1}}X=e_{n+1}$, where $L=(1+\displaystyle{\sum_{i=1}^{n}}x_i^2)(1-\displaystyle{\tiny{\sum_{i=1}^{n}}}x_i^2)^{-1}\geq 1$.
- ${\overline}\nabla_TX=\displaystyle{L\sum_{j=1}^n}t_je_j+t_{n+1}e_{n+1}$, where $T=\displaystyle{\sum_{i=1}^n} t_ie_i+t_{n+1}e_{n+1}$.
\[identities\]
We first recall that the coefficients of the metric in ${{\mathbb H}}^n\times{{\mathbb R}}$ are given by $g_{ij}=\frac{\delta_{ij}}{F^2},$ $g_{n+1,n+1}=1,$ $g_{i,n+1}=0,$ where $i,j=1,\dots,n$ and $F=\frac{1}{2}(1-\sum_{i=1}^nx_i^2)$. Then a straightforward computation gives that the Christoffel symbols satisfy $\Gamma_{ij}^k=0$ if $i,j,k$ are all distinct or if at least one of the indices is $n+1.$ Moreover
$$\Gamma_{ij}^i=\frac{x_j}{F},\ \Gamma_{ii}^j=-\frac{x_j}{F}, \ \Gamma_{ij}^j=\frac{x_i}{F},\ \Gamma_{ii}^i=\frac{x_i}{F}, \mbox{ with } i,j=1,\dots,n.$$
Then, for $j\leq n$ we have ${\overline}\nabla_{e_j}X=e_j+\sum_{i=1}^n x_i{\overline}\nabla_{e_j}e_i$ and
$$\begin{aligned}
\sum_{i=1}^n x_i{\overline}\nabla_{e_j}e_i&= \sum_{i=1}^n x_i\sum_{k=1}^{n+1}\Gamma_{ji}^ke_k\notag\\
&=\sum_{i\not=j}^nx_i\sum_{k=1}^{n}\Gamma_{ji}^ke_k+x_j\sum_{k=1}^{n}\Gamma_{jj}^ke_k\notag\\
&=\sum_{i\not=j}^nx_i[\Gamma_{ji}^je_j+\Gamma_{ji}^ie_i]+x_j[\Gamma_{jj}^je_j+\sum_{i\not=j}\Gamma_{jj}^ie_i]\notag\\
&=\sum_{i\not=j}^n\frac{x_i^2}{F}e_j+ \sum_{i\not=j}^n\frac{x_ix_j}{F}e_i+\frac{x_j^2}{F}e_j+\sum_{i\not=j}^n-\frac{x_ix_j}{F}e_i\notag\\
&=\sum_{i=1}^n\frac{x_i^2}{F}e_j.
\end{aligned}$$
Summing up, we obtain
$${\overline}\nabla_{e_j}X=e_j+\sum_{i=1}^n\frac{x_i^2}{F}e_j=\frac{1+\sum_{i=1}^{n}x_i^2}{1-\sum_{i=1}^{n}x_i^2}e_j=Le_j.$$
The equality ${\overline}\nabla_{e_{n+1}}X=e_{n+1}$ is straightforward and finishes the proof of [*(i)*]{}.
The proof of [*(ii)*]{}, is straightforward
$${\overline}\nabla_TX=\sum_{i=1}^nt_j{\overline}\nabla_{e_j}X+t_{n+1}{\overline}\nabla_{e_{n+1}}X=L\sum_{j=1}^n t_je_j+t_{n+1}e_{n+1}.$$
In Lemma \[proper-lemma\], we generalise [@DE Lemma 3.1] .
\[proper-lemma\] Let $X: M^n\to{{\mathbb H}}^{n}\times{{\mathbb R}}$ be a complete isometric immersion with finite strong total curvature. Then $X$ is proper and the extrinsic distance has no critical points outside a ball of ${{\mathbb H}}^{n}\times{{\mathbb R}}$.
If the immersion is not proper, we can find a ray $\gamma(s)$ issuing from the origin $(\sigma, 0)$, parametrized by the arc length $s$, such that, as $s$ goes to infinity, $Q(s)$ is bounded (see the discussion before Lemma \[identities\]), where $X(s)=X(\gamma(s))$ and $Q(s)=\Vert X(s)\Vert$. Setting $T(s)=\gamma'(s)$, we have
$$Q(s)\geq \langle X(s), T(s)\rangle.$$
In order to estimate $\langle X,T\rangle$ from below, we start by estimating
$$T\langle X,T\rangle=\langle{\overline}{\nabla}_TX,T\rangle+\langle X,{\overline}{\nabla}_T T\rangle.
\label{conta18}$$
Let us estimate the first term in the right hand side of (\[conta18\]). By using Lemma \[identities\] and that $\Vert T(s)\Vert=1$ we have
$$\langle {\overline}{\nabla}_TX, T\rangle=t_{n+1}^2(1-L)+L
\geq 1-L+L=1,
\label{ineq-proper0}$$
where we used that $t_{n+1}^2\leq 1$ and that $L\geq 1.$
Now we estimate the second term in the right hand side of (\[conta18\]). We first notice that, since $\gamma$ is a geodesic in $M$, the tangent component of ${\overline}{\nabla}_TT$ vanishes and we have
$${\overline}{\nabla}_TT=\langle{\overline}{\nabla}_TT,N\rangle
N=-\langle{\overline}{\nabla}_TN,T\rangle N=\langle A(T),T\rangle N.$$
It follows, by Cauchy-Schwarz inequality, that
$$\label{ineq-proper01}
|\langle X,{\overline}{\nabla}_TT\rangle|\le\Vert X\Vert\,\Vert A(T)\Vert\,\Vert T\Vert\le \Vert X\Vert\,|A|.$$
In view of , since we are assuming that $Q(s)$ is bounded, there exist $k_0$ such that ${\displaystyle\sum_{i=1}^nx_i^2}<k_0<1$, then we obtain
$$\label{ineq-proper01.5}
\Vert X\Vert^2< \left(\frac{2}{1-k_0}\right)^2\sum_{i=1}^nx_i^2+x^2_{n+1}< \left(\frac{2}{1-k_0}\right)^2\left( \sum_{i=1}^nx_i^2+x^2_{n+1}\right)=\left(\frac{2}{1-k_0}\right)^2|X|^2_{\tiny {{\mathbb R}}^{n+1}}.$$
hence, by replacing , and in , we get
$$\label{ineq-proper1}
T\langle X,T\rangle\ge1-\frac{2}{1-k_0}| X|_{\tiny {{\mathbb R}}^{n+1}}\,|A|.$$
Now, we notice that $\Vert V\Vert > |V|_{\tiny {{\mathbb R}}^{n+1}}$, for all vector field $V$ tangent to ${{\mathbb H}}^n\times{{\mathbb R}}$. In fact, $$\Vert V(q)\Vert^2=\frac{4\sum_{i=1}^nv_i^2(q)}{\left(1-\sum_{i=1}^nq_i^2\right)^2}+v^2_{n+1}(q)\geq \sum_{i=1}^nv_i^2(q)+v^2_{n+1}(q)=|V|^2_{\tiny {{\mathbb R}}^{n+1}}$$ This implies, since $\gamma$ is a minimizing geodesic, that $$s=\mbox{distance}_M ((X(s),(\sigma,0))>\mbox{distance}_{\;{{\mathbb R}}^{n+1}}(X(s),(\sigma,0))=| X|_{\tiny {{\mathbb R}}^{n+1}}$$
which together with gives
$$T\langle X,T\rangle> 1-\frac{2}{1-k_0}s\,|A|.$$
By using Proposition \[decay\] with ${{\varepsilon}}=\left(\frac{1-k_0}{2m}\right)^2$ we obtain
$$\label{ineq-proper2}
T\langle X,T\rangle(s)> 1-\frac1m,$$
for all $s>R_0$, where $R_0$ is given by Proposition \[decay\]. Integration of from $R_0$ to $s$ gives $$\label{ineq-proper3}
\langle X,T\rangle(s)>\left(1-\frac1m\right)(s-R_0)+\langle
X,T\rangle(R_0).$$
Since $Q(s) = \Vert X(s)\Vert \geq {{\left\langle\right.}}X,T{{\left.\right\rangle}}(s)$, we see from that $Q$ goes to infinity with $s$. This is a contradiction and proves that $M$ is properly immersed.
Now we use that ${{\mathbb H}}^n\times{{\mathbb R}}$ is a Hadamard manifold and we notice that Proposition \[decay\] implies that $X: M^n\to{{\mathbb H}}^{n}\times{{\mathbb R}}$ has tamed second fundamental form (see [@BC Definition 1.1]). Then can use (the proof of) [@BC Theorem 1.2] to conclude there exists a ball of ${{\mathbb H}}^{n}\times{{\mathbb R}}$, centered at the origin, of radius $r_0$ such that the extrinsic distance has no critical points outside this ball.
The technique of the proof of [@BC Theorem 1.2] can also be used to get an alternative proof of the fact that $X$ is properly immersed.
Let $X:M\to{{\mathbb H}}^n\times{{\mathbb R}}$ be a hypersurface with finite strong total curvature. By Lemma \[proper-lemma\] there exists $r_0>0$ such that the distance function in ${{\mathbb H}}^n\times{{\mathbb R}}$ has no critical points in $W=X(M)-(B_{r_0}(p_0)\cap X(M))$, where $B_{r_0}(p_0)$ is an extrinsic ${\rm (n+1)}$-ball of radius $r_0$ in ${{\mathbb H}}^n\times {{\mathbb R}}$. By Morse Theory, $X^{-1}(W)$ is homeomorphic to $X^{-1}[X(M)\cap S_{r_0}(p_0)]\times[0,\infty)$, where $S_{r_0}(p_0)=\partial B_{r_0}(p_0)$.
An [*end*]{} $E$ of $M$ is a connected component of $X^{-1}(W)$. It follows that $M$ has only a finite number of ends. In what follows, we identity $E$ and $X(E)$.
With the same proof of [@DE Lemma (4.2)], we can conclude that, for $r>r_0$, $E\cap B_{r_0}(p_0)$ is connected for each end $E$.
Theorem \[strong-normal\] below is a fundamental result for the characterization of finite strong total curvature hypersurfaces. We notice that it requires no assumption on $H_r$ and that it generalizes part of [@DE Theorem 1.1].
\[strong-normal\] Let $X:M\to{{\mathbb H}}^n\times{{\mathbb R}}$, $n\geq 3$, be an orientable complete hypersurface finite strong total curvature. Then:
- The immersion $X$ is proper.
- $M$ is diffeomorphic to a compact manifold ${\overline}M$ minus a finite number of points $q_1, \dots q_k$.
\(i) has already been proved in Lemma \[proper-lemma\]. To prove (ii), we apply to each end $E_i$ the restriction of the ambient transformation $I\colon ({{{\mathbb H}}}^n\times{{\mathbb R}})-\{(\sigma,0)\}\to({{{\mathbb H}}}^n\times{{\mathbb R}})-\{(\sigma,0)\}$, defined by $I(x)=x/\Vert x\Vert^2$, where the norm is with respect to the metric in ${{{\mathbb H}}}^n\times{{\mathbb R}}.$ Then $I(E_i)\subset B_1((\sigma,0))-\{(\sigma,0)\}$ and as $\Vert x\Vert \to\infty$ in $E_i$, $I(x)$ converges to the origin $(\sigma,0)$. It follows that each $E_i$ can be compactified with a point $q_i$. Doing this for each $E_i$, we obtain a compact manifold $\overline M$ such that ${\overline}M-\{q_1,\dots,q_k\}$ is diffeomorphic to $M$. This prove (ii).
Finite strong total curvature and $H_r=0$ {#main-result}
=========================================
In the next theorem, we deal with an immersion $X:M\to{{\mathbb H}}^n\times{{\mathbb R}}$ with finite strong total curvature and $H_r=0$. The proof is inspired by the proof of [@ST Theorem 2.1], although the assumptions and the result are different in nature.
Let $\Pi_1,\dots\Pi_k$ be an admissible collection of hyperplanes of ${{\mathbb H}}^n$, $P_i$, $i=1,\dots,k,$ the corresponding vertical hyperplanes and let $C^i_{\rho}$ the $\rho$-cylinder associated to $\Pi_i,$ $i=1,\dots,k,$ as defined at the end of Section 1. We say that $M$ is [*asymptotically close*]{} to $(P_1\cup\dots\cup P_k)\times{{\mathbb R}}$ if for any $\rho,$ there is a compact subset $K_{\rho}$ of $M$ such that
$$X(M\setminus K_{\rho})\subset \cup_{i=1}^kC^i_{\rho}.$$
We notice that, there are different notions of closeness at infinity and convergence in [@HNST; @MMR; @ST; @ST1].
\[strong-normal-minimal\] Assume that $X:M\to{{\mathbb H}}^n\times{{\mathbb R}}$ has finite strong total curvature and satisfies $H_r=0$. Let $E$ be an end of $X(M)$ and let $N=(N_1,\dots,N_{n+1})$ be a unit normal vector field on $X(E).$ Let $\Pi_1,\dots\Pi_k$ be an admissible collection of hyperplanes of ${{\mathbb H}}^n$ and $P_i$, $i=1,\dots,k,$ the corresponding vertical hyperplanes, such that $\partial E\subset \overline{P(\Pi_1,\dots,\Pi_k)}$. Suppose that $\partial_{\infty}E\cap (\partial_{\infty}{{\mathbb H}}^n\times{{\mathbb R}})\subset \partial_{\infty }(P_1\cup\dots\cup P_k)$. Then:
- $E$ is asymptotically close to $P_1\cup\dots\cup P_k.$
- For any sequence of points $\{p_m\}\subset E$ converging to a point in $\partial_{\infty}E,$ the sequence $\{N_{n+1}(p_m)\}$ converges uniformly to zero.
We start by proving (i).
Let us first observe the following general facts:
- By Corollary \[halfspace-coro\] one has $$E\subset \overline{P(\Pi_1,\dots,\Pi_k)}.$$
- Consider $\Pi_i$ and $\Pi_j,$ $i\not=j.$ Notice that two cases can happen.
- If $\partial_{\infty}\Pi_i\cap \partial_{\infty}\Pi_j\not=\emptyset,$ then for any $\rho,$ $C^i_{\rho}\cap C^j_{\rho}\not=\emptyset.$
- If $\partial_{\infty}\Pi_i\cap \partial_{\infty}\Pi_j=\emptyset,$ then there exists $\rho$ such that $C^i_{\rho}\cap C^j_{\rho}=\emptyset.$
- Let $\Theta$ be a hyperplane in ${{\mathbb H}}^n$ disjoint from $\Pi_i$ and such that $\partial E$ and $\cup_{i=1}^kP_i$ belong to the same component of $({{\mathbb H}}^n\times{{\mathbb R}})\setminus(\Theta\times{{\mathbb R}}).$ Denote by $(\Theta\times{{\mathbb R}})^-$ the component of $({{\mathbb H}}^n\times{{\mathbb R}})\setminus(\Theta\times{{\mathbb R}})$ that does not contain $\cup_{i=1}^k P_i\cup\partial E.$ Theorem \[iperplano\] yields that $E\cap (\Theta\times{{\mathbb R}})^-=\emptyset.$
- For each $i$, we can choose the corresponding equidistant hypersurface $L^{i +}_\rho$ to be the one which intersects $\overline{P(\Pi_1,\dots,\Pi_k)}$.
Assume, by contradiction, that there exists a positive number $\rho$ such that $E^K:=E\setminus E\cap(\cup_{i=1}^k C^i_{\rho})$ is a non compact set. This means that there is an unbounded sequence of points $p_m=(x_m,t_m)\in E^K.$ Since $\{p_m\}$ is unbounded, we have two possible cases. Either there exists an $i\in\{1,\dots,k\},$ say $i=1,$ such that $\{x_m\}$ has a subsequence converging to a point $\bar{x}$ of $\partial_{\infty}\Pi_1\cap(\partial_{\infty}{{\mathbb H}}^n\times {{\mathbb R}})$ or $\{x_m\}$ is bounded and $\{t_m\}$ is unbounded.
Let us first deal with the case (a subsequence of) $x_m$ converges to $\bar{x}$. Since $E^K\subset E\backslash C^1_{\rho} $ we can assume that the corresponding $\{p_m\}$ is contained in $Z^{1+}_{\rho}$. We can choose a hyperplane $\Theta$ as above in ${{\mathbb H}}^n$ such that $\Theta\cap Z^{1+}_{\rho}\not=\emptyset$ and $\partial_{\infty}\Theta\cap\partial_{\infty}\Pi_1=\bar x.$ This leads to a contradiction with the fact that $E\cap (\Theta\times{{\mathbb R}})^-=\emptyset$ and then we must have that $\{x_m\}$ is bounded.
Now, let $p_m=(x_m,t_m)$ be a sequence in $E^K$ such that $x_m$ is bounded and $t_m$ is unbounded. Without loss of generality we may assume that $t_m\longrightarrow\infty.$ In this case, we get a contradiction using the hypersurfaces ${\mathcal M}^r_d,$ $d>1,$ described in Theorem \[classification-theorem\], constructed with respect to one of the vertical hyperplanes $P_i, $ say $P_1$. We can choose the family ${\mathcal M}^{r}_d$ such that each hypersurface contains the equidistant hypersurface $L^{1+}_{\rho},$ with $\rho=\cosh^{-1}(d),$ and that is contained in the closure of $Z^{1+}_{\rho}.$ Let $V_{d}$ be the closure of the connected component of $({{\mathbb H}}^n\times{{\mathbb R}})\setminus {\mathcal M}_d^{r},$ not containing $P_1.$ By the properties of ${\mathcal M}^r_d,$ the height of $V_{d}$ is bounded. Moreover, as $t_m\longrightarrow\infty,$ there exists $t>0$ such that $\partial E\cap V_d(t)=\emptyset$ and $E\cap V_d(t)\not=\emptyset,$ where $V_d(t)$ is the vertical translation of $V_d$ of height $t.$ Denote by ${\mathcal M}^{r}_d(t)$ the vertical translation of ${\mathcal M}_d^{r}$ of height $t.$ Let $\gamma$ be a geodesic ${{\mathbb H}}^n\times \{0\}$, orthogonal to $P_1$ at a point $p$, whose endpoint is a point $q\in\partial_\infty{{\mathbb H}}^n\times \{0\}$ that is outside all closed balls limited by $\partial\Pi_i$, $i=1,\ldots,n$. Such a point $q$ exists since $n\geq 3.$ Now, let us consider the horizontal translations along $\gamma$ (extended slice-wise to ${{\mathbb H}}^n\times{{\mathbb R}}$) of ${\mathcal M}^r_d(t)$, in the direction of $q$. Since $E$ is properly immersed and $\partial_\infty E\cap (\{q\}\times{{\mathbb R}})=\emptyset$, we can proceed as in the proof of Theorem \[iperplano\] and we find a horizontal translation along $\gamma$ of ${\mathcal M}^r_d(t)$ that has a last contact point with an interior point of $E$. This is a contradiction by the maximum principle. Hence (1) is proved.
Now we prove (2).
Assume, by contradiction, that there exist $\varepsilon>0$ and a sequence of points $p_m=(x_m,t_m)$ converging to a point in $\partial_{\infty} E$ such that $|N_{n+1}(p_m)|>\varepsilon$. Since $\{p_m\}$ is unbounded, we have two possible cases. Either there exists $i=1,\dots,k,$ say $i=1,$ such that $\{x_m\}$ has a subsequence converging to a point of $\partial_{\infty}\Pi_1\cap(\partial_{\infty}{{\mathbb H}}^n\times {{\mathbb R}})$ or $\{x_m\}$ is bounded and $\{t_m\}$ is unbounded.
Let $p_0\in E$ be a fixed point. Since $E$ has finite strong total curvature, Proposition \[decay\] implies that there exist $R_0>0$ and $s>0$ such that
$$\sup_{x\in (E\setminus(E\cap B_{R_0}(p_0)))}|A|^2(x)<s.
\label{constant-s}$$
where $B_{R_0}(p_0)$ is an extrinsic ${\rm (n+1)}$-ball of radius $R_0$ in ${{\mathbb H}}^n\times {{\mathbb R}}$. Notice that, in the previous inequality, we can take the extrinsic ball, because $E$ is properly immersed.
Assume first that (a subsequence of) $x_m$ converges to $\bar{x}\in\partial_{\infty}\Pi_1\cap(\partial_{\infty}{{\mathbb H}}^n\times {{\mathbb R}}).$ For the following constructions, see Figure 2.
 \[for-main-theo\]
Without loss of generality, we can consider a geodesic $\alpha\subset \Pi_1$, passing through the origin, such that $\bar x\in \partial_{\infty}\alpha$ and denote by $\bar y$ the point of $\partial_{\infty}\alpha $ distinct from $\bar x.$ We choose two points $y_1$ and $y_2$ on $\alpha$ such that $y_1$ is between $\bar x$ and $y_2.$ Let $y$ be the point on $\alpha$ equidistant from $y_1$ and $y_2.$ Finally, let $\Lambda$ be the hyperplane in ${{\mathbb H}}^n$ through $y,$ orthogonal to $\alpha$ and $K_i,$ $i=1,2$ be the hyperplanes in ${{\mathbb H}}^n$ passing through $y_i,$ orthogonal to $\alpha.$ For $i=1,2$, denote by $(K_i\times{{\mathbb R}})^+$ the connected component of ${{\mathbb H}}^n\times{{\mathbb R}}\setminus (K_i\times{{\mathbb R}})$ whose asymptotic boundary contains $\bar y$ and by $(K_i\times{{\mathbb R}})^-$ the other connected component. Since $\partial E$ is compact, it is possible to choose $y_1$ and $y_2$ such that $\partial E\subset (K_2\times{{\mathbb R}})^+.$
Let ${\Delta}_1$ and ${\Delta}_2$ hyperplanes in ${{\mathbb H}}^n$, symmetric with respect to $\Pi_1, $ disjoints from $K_1$, such that, for $i=1,2$:
- $ \partial_{\infty }\Delta_i\cap\partial_{\infty }\Pi_1=\bar{x}.$
- $\partial_{\infty}{\Delta}_i\cap\partial_{\infty}K_1=\emptyset$.
- $\partial E$ and $\cup_{i=1}^kP_i$ belong to the same component of $({{\mathbb H}}^n\times{{\mathbb R}})\setminus(({\Delta}_1\times{{\mathbb R}})\cup({\Delta}_2\times{{\mathbb R}})).$
This yields that $({\Delta}_i\times{{\mathbb R}})\subset (K_1\times{{\mathbb R}})^-$, $i=1,2$, and by Corollary \[halfspace-coro\] we conclude that $E\cup\partial E$ is contained in the component of ${{\mathbb H}}^n\times{{\mathbb R}}\setminus (({\Delta}_1\times{{\mathbb R}})\cup({\Delta}_2\times{{\mathbb R}}))$ containing $\Pi_1\times{{\mathbb R}}.$ Without loss of generality, we can choose ${\Delta}_i$ such that $({\Delta}_i\times{{\mathbb R}})\cap E=\emptyset$.
For any $\lambda>0,$ we denote by $T_{\lambda}$ the hyperbolic translation of length $\lambda$ along $\alpha$ oriented from $\bar x$ to $\bar y.$ By abuse of notation, we also denote by $T_{\lambda}$ the extension of $T_{\lambda}$ to ${{\mathbb H}}^n\times{{\mathbb R}}$. For any $\lambda,$ denote by $U_{\lambda}$ the connected component of ${{\mathbb H}}^n\setminus (T_{\lambda}({\Delta}_1)\cup T_{\lambda}({\Delta}_2))$ containing $\Pi_1.$ For any $\delta>0,$ there exists $\lambda(\delta)$ such that the $(n-1)$-planes $T_{\lambda(\delta)}({\Delta}_i),$ $i=1,2,$ are contained in a neighborhood of $\alpha$ of diameter $\delta$ in the Euclidean metric in ${{\mathbb H}}^n$.
Let $D_{\delta}$ be the component of ${{\mathbb H}}^n\setminus(T_{\lambda(\delta)}({\Delta}_1)\cup T_{\lambda(\delta)}({\Delta}_2)\cup K_1\cup K_2)$ containing the point $y=\alpha\cap\Lambda.$ We notice that $D_{\delta}\times {{\mathbb R}}=(U_{\delta}\times{{\mathbb R}})\cap (K_1\times{{\mathbb R}})^+\cap (K_2\times{{\mathbb R}})^-.$ Finally, denote by $\Omega_{\delta}$ the component of ${{\mathbb H}}^n\setminus(T_{\lambda(\delta)}^{-1}(K_2)\cup {\Delta}_1\cup {\Delta}_2)$ such that $({\Omega}_{\delta}\times {{\mathbb R}})\cap (\Pi_1\times{{\mathbb R}})\not=\emptyset$ and $\partial_{\infty}\Omega_{\delta}=\bar x$. By construction, for any $\lambda>\lambda(\delta)$ and any $p\in\Omega_{\delta}\times{{\mathbb R}},$ we have $T_{\lambda}(p)\in U_{\delta}\times{{\mathbb R}}.$ We notice that we can choose $\lambda(\delta)$ such that $ ({\Omega}_{\delta}\times {{\mathbb R}}) \subset ({{\mathbb H}}^n\times{{\mathbb R}})\backslash B_{R_0}(p_0)$.
As $x_m\longrightarrow \bar x,$ we can assume that $p_m\in \Omega_{\delta}\times {{\mathbb R}}$ for $m$ large. Moreover, for any $m$ large, there exists a unique $\lambda_m>0$ such that $T_{\lambda_m}(x_m)\in\Lambda,$ hence $q_m:=T_{\lambda_m}(p_m)\in \Lambda\times{{\mathbb R}}.$ For $m$ large enough, say $m>m_0,$ we have $\lambda_m>\lambda(\delta),$ which implies that $q_m\in(\Lambda\times{{\mathbb R}})\cap (U_{\delta}\times{{\mathbb R}})$.
For any $m>m_0,$ we denote by $E_m(\delta)$ the connected component of $T_{\lambda_m}(E)\cap (D_{\delta}\times{{\mathbb R}})$ containing $q_m.$ By construction, $E_m(\delta)$ is the component of $T_{\lambda_m}(E\cap(\Omega_{\delta}\times{{\mathbb R}}))\cap (D_{\delta}\times{{\mathbb R}})$ containing $q_m$ and for all $m>m_0$, the boundary of $E_m(\delta)$ satisfies
$$\label{boundary-En-1}
\begin{array}{c}
\partial E_m(\delta)\subset \partial(D_\delta\times{{\mathbb R}})\\
\mbox{but}\\
\partial E_m(\delta)\cap( (T_{\lambda(\delta)}({\Delta}_1)\times{{\mathbb R}})\cup (T_{\lambda(\delta)}({\Delta}_2)\times{{\mathbb R}}))=\emptyset.
\end{array}$$
Since $ ({\Omega}_{\delta}\times {{\mathbb R}}) \subset ({{\mathbb H}}^n\times{{\mathbb R}})\backslash B_{R_0}(p_0)$, we can use (\[constant-s\]) in order to conclude that for all $\ p\in E_m(\delta)$, $m>m_0$ and $\delta>0$ it holds
$$\label{bounded-hyp-curv-1}
|A_m(p)|\leq s,$$
where $A_m$ is the shape operator of $E_m(\delta).$ As $D_{\delta}$ is compact, we can look at $D_{\delta}\times{{\mathbb R}}$ as a subset of ${{\mathbb R}}^{n+1}$ where the metric inherited from ${{\mathbb H}}^n\times{{\mathbb R}}$ and the Euclidean metric are $C^1$ close. Then one can prove that the norms of the second fundamental forms of $E_m(\delta)$ induced by the Euclidean and the hyperbolic metric are close (see Proposition 3.1 in the Appendix of [@ST]).
As the norm of the second fundamental forms of $E_m(\delta)$ in the hyperbolic metric is uniformly bounded (see inequality ), the same holds for the norm of second fundamental forms of the family $E_m(\delta)$ measured in the Euclidean metric. By standard arguments, one can prove that this uniform bound implies the existence of a positive number $\eta$, independent on $m$ and $\delta$, such that a part $F_m$ of $E_m(\delta)$ is the Euclidean graph of a function $f_m$ defined on an $n$-ball of radius $\eta$ of the tangent hyperplane of $F_m$ at $q_m$. Moreover, by applying vertical translations, we can assume the points $q_m$ are in a compact set of the Euclidean space and then all the functions $f_m$ have a uniform (Euclidean) $C^1$ bound (see, for instance, the proof of Lemma 2.2 in [@CM]).
Recall that we are assuming, by contradiction, that $|N_{n+1}(p_m)|=|N_{n+1}(q_m)|>\varepsilon,$ for any $m.$ Then if we denote by $\nu$ the Euclidean unit normal vector, we have $\Vert \nu_{n+1}(q_m)\Vert >\varepsilon',$ for some positive $\varepsilon'$ (see the formula in the proof of Proposition 3.2 in [@ST3]). The last inequality implies that the slope of the tangent planes of $E_m(\delta)$ at points $q_m$ is uniformly bounded from below. As the gradient of the functions $f_m$ are uniformly bounded and $\eta$ does not depend on $\delta$ we can choose $\delta$ small enough such that the graph $F_m$ intersect $(T_{\lambda(\delta)}({\Delta}_1)\times{{\mathbb R}})\cup (T_{\lambda(\delta)}({\Delta}_1)\times{{\mathbb R}}),$ that is in contradiction with . This finishes the proof in the case where $\{x_m\}$ converges to a point of $\partial_{\infty}{{\mathbb H}}^n\times{{\mathbb R}}$.
In the case where $t_m$ is unbounded and $x_m$ is bounded the proof is somewhat easier. Without loss of generality, we can assume that $t_m\longrightarrow\infty.$ We proved before that for any $\rho$ there exists $t_{\rho}>0$ such that $E\cap \{|t|>t_{\rho}\}\subset \cup_{i=1}^kC^i_{\rho}.$ Then, there exists $i\in\{1,\ldots,n\}$, say $i=1$, and a subsequence $t_{m_{1}}$, such that $t_{m_{1}}\in C^1_{\rho}$. Since we are assuming that $\{x_m\}$ is bounded, we may assume that $\{p_{m_{1}}\}\subset \omega\times{{\mathbb R}}$, where $\omega\subset {{\mathbb H}}^n\times\{0\}$ is a compact set. Then, we proceed as in the former case, replacing $D_\delta$ by $\omega\cap C^1_{\rho}.$
\[question\] Theorem \[strong-normal-minimal\] can be viewed as a step towards a generalization of the results of [@HaRo Theorem 3.1 (c)] and [@HNST] for minimal surfaces with finite total curvature in ${{\mathbb H}}^2\times{{\mathbb R}}.$ We point out that our technique is completely different from the one in [@HaRo; @HNST] where complex analysis is a key tool.
The authors would like to thank the referee for the careful reading and the valuable suggestions.
[999999]{}
<span style="font-variant:small-caps;">M. Anderson:</span> [*The compactification of a minimal submanifold in Euclidean space by the Gauss map,*]{} Preprint IHES (1985).
<span style="font-variant:small-caps;">R. Bartnik:</span> [*The mass of an asymptotically flat manifold*]{}, Comm. Pure Appl.Math., 39, 661-693 (1986).
<span style="font-variant:small-caps;">G.P. Bessa, M.S. Costa:</span> [*On submanifolds with tamed second fundamental form,*]{} Glasgow Mathematical Journal 51 (3) (2009), 669-680
<span style="font-variant:small-caps;">P. Bérard, R. Sa Earp:</span> [*Minimal hypersurfaces in ${{\mathbb H}}^n\times {{\mathbb R}}$, total curvature and index,*]{} Boll. Unione Mat. Italiana, 9 (3) (2016) 341-362. Calc. Var. Partial Differential Equations 55, 3 (2016).
<span style="font-variant:small-caps;">T. H. Colding, W. P. Minicozzi:</span> [*Minimal surfaces,*]{} Courant Lecture Notes in Math. 4, New York University, Courant Institute of Mathematical Sciences, New York, (1999).
<span style="font-variant:small-caps;">M. Do Carmo, M.F. Elbert:</span> [*Complete hypersurfaces in Euclidean spaces with finite strong total curvature,*]{} To appear in Comm. Anal. Geom..
<span style="font-variant:small-caps;">M.F. Elbert, B. Nelli, W. Santos:</span> [*Hypersurfaces with $H_{r+1} = 0$ in ${{\mathbb H}}^n\times{{\mathbb R}}$,*]{} Manuscripta Mathematica, 149 (2015) 507-521.
<span style="font-variant:small-caps;">F.X. Fontenele, S. L. Silva:</span> [*Maximum principles for hypersurfaces with vanishing curvature functions in an arbitrary Riemannian manifold,*]{} Anais da Ac. Bras. de Ci. 74 (2) (2002) 199-205.
<span style="font-variant:small-caps;">L. Hauswirth, H. Rosenberg:</span> [*Minimal surfaces of finite total curvature in ${{\mathbb H}}\times{{\mathbb R}},$*]{} Workshop on Differential Geometry, Mat. Contemp. 31 (2006), 65-80.
<span style="font-variant:small-caps;">J. Hounie, M. L. Leite:</span> [*Two ended hypersurfaces with zero scalar curvature,*]{} Indiana Univ. Math. Jour. 48 (1999), 817-882.
<span style="font-variant:small-caps;">J. Hounie, M. L. Leite:</span> [*The maximum principle for hypersurfaces with vanishing curvature functions,*]{} J. Differential Geom. 41, 2 (1995), 247-258. <span style="font-variant:small-caps;">L. Hauswirth, B. Nelli, R. Sa Earp, E. Toubiana:</span> [*A Schoen theorem for minimal surfaces in ${{\mathbb H}}^2\times{{\mathbb R}},$*]{} Adv. Math. 274 (2015), 199-240.
<span style="font-variant:small-caps;">A. Huber:</span> [*On subharmonic functions and differential geometry in the large,*]{} Comment. Math. Helvetici, 32, 181-206 (1957).
<span style="font-variant:small-caps;">R. Mazzeo, F. Martin, M. Rodriguez:</span> [*Minimal surfaces with positive genus and finite total curvature in ${{\mathbb H}}^2\times{{\mathbb R}}$,*]{} Geometry and Topology 18 (2014) 141-177.
<span style="font-variant:small-caps;">W. H. Meeks III, J. Perez:</span> [*A survey on classical minimal surface theory,*]{} University Lecture Series (AMS) vol. 60 (2012).
<span style="font-variant:small-caps;">B. Nelli, R. Sa Earp, E. Toubiana:</span> [*Maximum principle and symmetry for minimal hypersurfaces in ${{\mathbb H}}^n\times {{\mathbb R}}$,*]{} Annali della Scuola Normale Superiore di Pisa, Classe di Scienze (5) Vol. XIV (2015) 387-400.
<span style="font-variant:small-caps;">R. Osserman:</span>[*A Survey of Minimal Surfaces,*]{} Dover Publications, New York, 2nd edition, (1986).
<span style="font-variant:small-caps;">R. Sa Earp, E. Toubiana:</span> [*A minimal stable vertical planar end in ${{\mathbb H}}^2\times{{\mathbb R}}$ has finite total curvature,*]{} Jour. of the London Math. Soc. 92, 3 (2015) 712-723. <span style="font-variant:small-caps;">R. Sa Earp, E. Toubiana:</span> [*Concentration of total curvature of minimal surfaces in ${{\mathbb H}}^2\times{{\mathbb R}},$*]{} Math. Annalen, 369, Issue 3-4 (2107) 1599-1621. <span style="font-variant:small-caps;">R. Sa Earp, E. Toubiana:</span> [*An asymptotic theorem for minimal surfaces and existence results for minimal graphs in ${{\mathbb H}}^2\times {{\mathbb R}}$,*]{} Math. Ann. 342 (2008) 309-331. <span style="font-variant:small-caps;">R. Sa Earp, E. Toubiana:</span> [*Minimal graphs in ${{\mathbb H}}^n\times{{\mathbb R}}$ and ${{\mathbb R}}^{n+1},$*]{} Ann. Inst. Fourier 60 (7) (2010) 2373-2402.
<span style="font-variant:small-caps;">B. White:</span> [*Complete surfaces of finite total curvature,*]{} J. Diff. Geometry, 26, 315-326 (1987).
[^1]: Keywords: higher order mean curvature, hyperbolic translation, finite strong total curvature.\
2000 Mathematical Subject Classification: 53C42, 53A10.\
The authors were partially supported by INdAM-GNSAGA, PRIN-2015A35N9B-011.
| {
"pile_set_name": "ArXiv"
} |
Impact-Generated Dust Clouds Surrounding the Galilean Moons
Harald Krüger$^{1}$, Alexander V. Krivov$^{2,3}$, Miodrag Sremčević$^{2}$,\
and Eberhard Grün$^{1,4}$
$^1$ Max-Planck-Institut für Kernphysik,\
Postfach 103980, 69029 Heidelberg, Germany\
E-Mail: [email protected]\
$^2$ Nonlinear Dynamics Group, Institute of Physics, University of Potsdam,\
P.O. Box 601553, 14415 Potsdam, Germany\
E-mail: [email protected]\
$^3$ On leave from: Astronomical Institute, St. Petersburg University,\
Stary Peterhof, 198504 St. Petersburg, Russia\
E-mail: [email protected]\
$^4$ Hawaii Institute of Geophysics and Planetology, University of Hawaii,\
1680 East West Road, Honolulu, HI 96822, USA\
------------------- ----
Manuscript pages: 46
Figures: 16
Tables: 5
------------------- ----
, in press
\
[*Proposed Running Head:*]{}\
DUST CLOUDS OF THE GALILEAN MOONS
[*Corresponding author:*]{}\
[Harald Krüger]{}\
Max-Planck-Institut für Kernphysik\
Postfach 103980\
69029 Heidelberg, Germany\
E-mail: [email protected]
**Abstract**
Tenuous dust clouds of Jupiter’s Galilean moons Io, Europa, Ganymede and Callisto have been detected with the in-situ dust detector on board the Galileo spacecraft. The majority of the dust particles have been sensed at altitudes below five radii of these lunar-sized satellites. We identify the particles in the dust clouds surrounding the moons by their impact direction, impact velocity, and mass distribution. Average particle sizes are between 0.5 and $\rm 1\,\mu m$, just above the detector threshold, indicating a size distribution with decreasing numbers towards bigger particles. Our results imply that the particles have been kicked up by hypervelocity impacts of micrometeoroids onto the satellites’ surfaces. The measured radial dust density profiles are consistent with predictions by dynamical modeling for satellite ejecta produced by interplanetary impactors (Krivov et al., [*Planet. Sp. Sci.*]{}, 2003, 51, 251–269), assuming yield, mass and velocity distributions of the ejecta from laboratory measurements. A comparison of all four Galilean moons (data for Ganymede published earlier; Krüger et al., [*Planet. Sp. Sci.*]{}, 2000, 48, 1457–1471) shows that the dust clouds of the three outer Galilean moons have very similar properties and are in good agreement with the model predictions for solid ice-silicate surfaces. The dust density in the vicinity of Io, however, is more than an order of magnitude lower than expected from theory. This may be due to a softer, fluffier surface of Io (volcanic deposits) as compared to the other moons. The log-log slope of the dust number density in the clouds vs. distance from the satellite center ranges between –1.6 and –2.8. Appreciable variations of number densities obtained from individual flybys with varying geometry, especially at Callisto, are found. These might be indicative of leading-trailing asymmetries of the clouds due to the motion of the moons with respect to the field of impactors.
\
dust\
satellites of Jupiter\
planetary rings
Introduction
============
All celestial bodies without atmospheres are permanently exposed to bombardment by hypervelocity micrometeoroids which knock-off secondary ejecta dust particles from the surfaces of these bodies. Impact ejection of dust particles has been suggested as the main process for maintaining dusty planetary rings like the Jovian rings [@morfill1980b; @horanyi1996; @ockert-bell1999; @burns1999], Saturn’s E ring [@horanyi1992; @hamilton1994] as well as putative dust belts of Mars [@soter1971; @krivov1997; @krivov1999] and Pluto [@thiessenhusen2002]. With the in-situ dust detector on board the Galileo spacecraft [@gruen1992a] a dust cloud formed by impact ejecta particles was for the first time detected surrounding Jupiter’s moon Ganymede [@krueger1999d; @krueger2000a]. The dust cloud was by far too tenuous to be detectable with remote sensing techniques. Particles belonging to the Ganymede dust cloud were identified by their impact direction and impact speed and their sizes were mostly below $\rm 1\,\mu m$. The spatial distribution of the grains as well as their size distribution were in agreement with model predictions based on the impact ejection mechanism.
The Galileo dust measurements can be treated as a natural impact experiment leading to the detection of the ejecta of hypervelocity impacts in space. They can give more insight into the process of hypervelocity dust ejection, for which the laboratory experiments on Earth still do not yield a comprehensive picture. The measurements of the dust cloud at Ganymede stimulated the development of analytical models for impact-generated circumsatellite dust clouds not only for the Galilean moons but also for the Saturnian satellites [@krivov2003; @sremcevic2003]. This is especially important for the dust measurements to be collected at Saturn with the dust instrument [@srama2002] onboard the Cassini spacecraft beginning in 2004.
Since December 1995 Galileo has been on a bound orbit about Jupiter. The spacecraft had a total of 32 targeted close flybys at all four Galilean moons: 7 encounters with Io, 11 with Europa, 6 with Ganymede and 8 with Callisto. During many of the encounters between 1995 and early 1999, the impact rate of dust grains showed a sharp peak within about half an hour centered at closest approach to the moon [@gruen1996c; @gruen1997b; @gruen1998; @krueger1998]. These peaks indicated the existence of dust concentrations not only at Ganymede but also in the close vicinities of Io, Europa and Callisto. During the flybys at the Galilean moons after mid-1999, the spacecraft orientation prevented the detection of dust particles close to the satellites. In November 2002, Galileo had the only opportunity to in-situ measure dust in the close vicinity of a fifth Jovian moon: Amalthea [@krueger2002].
In addition to dust clouds surrounding the Galilean moons, at least three other populations of dust were detected by Galileo in the Jovian system [@gruen1998]. Streams of 10-nanometer dust particles were detected throughout the Jovian magnetosphere and were recognizable even in interplanetary space out to 2 AU from Jupiter [@gruen1993; @gruen1994a]. These dust grains originate from Io [@horanyi1993a; @horanyi1993b; @graps2000; @krueger2003a], their ultimate source probably being the most powerful of Io’s volcanic plumes. Bigger, micrometer-sized particles form a tenuous dust ring between the Galilean moons and further away from Jupiter. Many of these particles orbit Jupiter on prograde orbits whereas a population on retrograde orbits exists as well [@colwell1998a; @thiessenhusen2000; @krivov2002a; @krivov2002b]. An overview of the Galileo dust measurements at Jupiter including the dust instrument itself can be found in @krueger2003c.
A detailed analysis of the dust grains detected at Galileo’s four Ganymede flybys in 1996 and 1997 has been published earlier [@krueger1999d; @krueger2000a], showing that this Jovian moon is surrounded by a dust cloud formed by impact ejecta. Here, we analyze the dust impacts detected close to Io, Europa and Callisto, and compare our results with the measurements at Ganymede. Relevant physical properties of these moons are summarized in Table \[phys\_prop\]. In Section \[sec\_detection\] we give a brief overview of the most important aspects of the Galileo dust instrument, the Galileo spacecraft and the procedure to identify impacts of ejecta cloud particles in the Galileo dust data set. In Section \[sec\_data\] we analyze the dust detections at Io, Europa and Callisto. In Section \[sec\_comp\] we compare the properties of the dust clouds of all four Galilean moons and check them against modeling. Section \[sec\_conclusions\] lists our conclusions.
Dust impact detection {#sec_detection}
=====================
Galileo dust instrument {#sec_instrument}
-----------------------
The Galileo in-situ dust measurements at Ganymede provided the first in-situ detection of an impact-generated dust cloud in space [@krueger1999d]. Processing of these measurements has been described in detail by @krueger2000a. Here we apply the same analysis techniques to the dust measurements obtained in the close vicinities of Europa, Callisto and Io. Descriptions of the dust instrument, Galileo spacecraft, data transmission etc. have been published in previous papers. In what follows we recall only the most important aspects and give references to earlier publications where necessary.
Galileo is a dual-spinning spacecraft, with an antenna that points antiparallel to the positive spacecraft spin axis. During most of Galileo’s orbital mission about Jupiter, the antenna pointed towards Earth. The Galileo Dust Detector System (DDS), like its twin on-board Ulysses, is a multi-coincidence impact ionization detector [@gruen1992a] which measures submicrometer- and micrometer-sized dust impacts onto the detector target. The dust instrument is mounted on the spinning section of Galileo and its sensor axis is offset by an angle of $60^{\circ}$ from the positive spin axis (Fig. \[geometry\]). Thus, during one spin revolution of the spacecraft, the detector scans the entire anti-Earth hemisphere, whereas particles approaching from the Earth-ward direction remain undetectable.
For each dust grain hitting the detector target, three independent measurements of the impact-created plasma cloud are used to derive the impact speed $ v$ and the mass $ m$ of the particle. The charge $ Q_{\rm I}$ released upon impact onto the target is described by the relation [@goeller1989; @gruen1995a] $$Q_{\rm I} \propto m \cdot v\,^{3.5} .
\label{equ1}$$ The calibrated speed range of the dust instrument is 2 to $\rm 70\,km{\,\hbox{s}}^{-1}$. The coincidence times of the three charge signals together with the charges themselves are used to classify each impact into one of four categories. Class 3 impacts have three charge signals, two are required for class 2 and class 1 events, and only one for class 0 [@baguhl1993b; @gruen1995a; @krueger1999a]. Class 3 signals, our highest class, are real dust impacts and class 0 events are noise. Class 1 and class 2 events are true dust impacts in interplanetary space [@baguhl1993a; @krueger1999a]. However, in the Jovian system, within about 15$\rm R_J$ distance from Jupiter, energetic particles from the Jovian plasma environment cause an enhanced noise rate in class 2 and the lower quality classes. By analysing the properties of the Io stream particles and comparing them with the noise events, the noise could be eliminated from the class 2 data [@krueger1999c]. All class 0 and class 1 events detected in the Jovian environment are usually classified as noise.
Our noise identification scheme [@krueger1999c], however, was derived for the Jovian dust stream particles and, hence, its applicability to other populations of dust had to be verified. Since Europa orbits Jupiter within the region where the high noise rates occurred, a slightly modified scheme has been developed for the ejecta grains detected in the close vicinity of Europa [see @krueger2001a their Table 4]. It will be applied in this paper to remove noise events from the data sets obtained at the flybys at Europa and Io. For the Callisto data, no noise removal is necessary because Callisto orbits Jupiter outside the region where the high noise rates occurred. Noise removal was also not necessary in the earlier analysis of the Ganymede data [@krueger2000a]. It has to be noted that the noise removal technique uses statistical arguments and is applicable to large data sets only. Individual dust impacts may be erroneously classified as noise and vice versa.
Galileo has a very low data transmission capability because its high-gain antenna did not open completely. For the dust measurements this means that the full set of parameters measured during a dust particle impact (spacecraft rotation angle, impact charges, charge rise times, etc.) could only be transmitted to Earth for a limited number of impact events. During the close satellite flybys at the Galilean moons these limits were between one event per minute (record mode) and one event per 21 minutes (real time science mode) [@krueger2001a]. When event rates ([[*i.e.*]{}]{} dust impacts plus noise) exceeded these numbers, the full set of parameters was transmitted to Earth for only a fraction of all detected events. All events, however, were always counted with one of 24 accumulators [@gruen1995a]. This way, the data can be corrected for incomplete data transmission so that reliable impact rates can be determined for all satellite flybys [@krueger2000a].
Since its injection into an orbit about Jupiter, the Galileo spacecraft has been exposed to the harsh radiation environment of the Jovian magnetosphere with energetic particles of up to several MeV energies. The Galileo Jupiter mission was extended three times so that the spacecraft was exposed to a total radiation dosage five times higher than it was originally designed for. Especially high radiation dosages were acquired during orbit insertion in December 1995 and during several perijove passages after mid-1999 when Galileo’s perijove distance from Jupiter was about $\rm 6\, R_J$. It was anticipated that these high radiation levels would cause severe damages to the spacecraft electronics and the scientific instruments. Although degradation of the dust instrument was recognised in the dust data, no failure has occurred so far. The degradation effects include – amoung others – a drop of the channeltron amplification, shifts of the measured instrument current, charge rise times and amplitudes which reduced the sensitivity for dust impacts and noise events (Krüger et al., in prep.). The most important effect for our analysis here is shifts in the speed and mass calibration of the dust impacts.
Impact direction
----------------
The analysis of the dust measurements obtained at Ganymede showed that the impact direction of the particles could be used as one important parameter to identify ejecta particles belonging to a dust cloud surrounding this moon [@krueger2000a]: in particular, the impact direction of the grains could be used to separate cloud particles from Jovian dust stream particles [@gruen1998]. As rotation angle, $\Theta$, we define the viewing direction of the dust sensor at the time of particle impact. During one spin revolution of the spacecraft, the rotation angle scans through $360^{\circ}$. Rotation angles for the Galileo dust instrument, however, are reported opposite to that of the actual spacecraft rotation direction. This is done to easily compare Galileo results with the dust detector data taken on the Ulysses spacecraft, which, unlike Galileo, has the opposite spin direction. Zero degrees of rotation angle is taken when the dust sensor points close to the ecliptic north direction. At rotation angles of $90^{\circ}$ and $270^{\circ}$ the sensor axis lies nearly in the ecliptic plane (which is close to Jupiter’s equatorial plane).
The dust instrument itself has a $140^{\circ}$ wide field of view (FOV). Dust particles which arrive within $10^{\circ}$ of the positive spin axis can be sensed at all rotation angles, while those that arrive at angles from $10^{\circ}$ to $130^{\circ}$ from the positive spin axis can only be sensed over a limited range of rotation angles. A sketch of the detection geometry at close satellite flybys is shown in Fig. \[geometry\].
Impact velocity
---------------
Calibrated impact velocities are derived from the rise times of the impact charge signals by an empirically derived algorithm [@gruen1995a]. The analysis of the dust impacts detected close to Ganymede showed that their average impact velocity onto the detector target was close to the encounter velocity of Galileo with this moon ($\rm 8\,{\,\hbox{km}}{\,\hbox{s}}^{-1}$) [@krueger2000a]. It implied that the particles truly originated from Ganymede and that they belonged to a steady-state dust cloud surrounding this satellite. This good agreement of the measured mean impact velocity with the expected velocity also showed that the calibration of the dust instrument is reliable in this velocity range.
Two statistical subsets of particles could be separated: nearly all of the Ganymede particles had calibrated velocities below $\rm 10\,{\,\hbox{km}}{\,\hbox{s}}^{-1}$, whereas most of the stream particles had higher velocities. The calibrated impact velocity has been used as a parameter to separate both populations of dust at Galileo’s G8 Ganymede flyby when the Jovian stream particles approached the dust detector from the direction towards Ganymede and particles belonging to Ganymede’s steady-state dust cloud could not be identified by their impact direction alone. The true impact velocities of stream particles exceeded $\rm 200\,{\,\hbox{km}}{\,\hbox{s}}^{-1}$ [@zook1996] and were much faster than the velocity range of the dust instrument calibrated in the laboratory ($\rm 70\,{\,\hbox{km}}{\,\hbox{s}}^{-1}$). Thus, the velocities for the stream particles derived from the instrument calibration significantly underestimate the true particle velocities.
Data analysis {#sec_data}
=============
Europa
------
### Impact direction {#sec_rot_eu}
The antenna of Galileo usually pointed towards Earth for data transmission. This fixed the spin axis of the spacecraft so that the detector basically scanned the anti-Earth hemisphere. In addition, due to the orbital motion of Jupiter about the Sun, the geometry for dust detection with the dust instrument gradually changed with time, leading to the non-detectability of dust particles in the close vicinity of the Galilean moons after mid-1999. For this reason, ejecta particles were measurable during ten close Galileo flybys at Europa out of 11 flybys in total. No data could be collected during two of these encounters due to spacecraft anomalies (safings) so that data sets from eight Europa flybys are available (E4, E6, E11, E12, E14, E15, E17, E19; see also Tab \[tab\_sat\]). The labels of the encounters are: the first letter of the satellite encountered by Galileo plus the number of Galileo’s orbit about Jupiter.
In Fig. \[rot\_eu\] we show the impact direction (rotation angle) of the dust particles detected within about 2 h around closest approach to Europa whose complete set of measured impact parameters has been transmitted to Earth. During most flybys, particle impacts with $\rm 180^{\circ} < \Theta < 360^{\circ}$ were concentrated towards Europa. This is most obvious during encounters E4, E11, E12 and E19. Most of these impacts were detected at altitudes below $\rm 3\,R_E$ ($\rm \,R_E$ is the Europa radius, see Table \[phys\_prop\]).
To analyse the impact direction of the dust grains onto the detector we assumed that the speed of dust relative to Europa in the vicinity of the moon is low compared to Galileo’s flyby speed. Thus, the approach direction of the dust for an observer moving with the spacecraft is more or less parallel to the velocity vector of Europa relative to the spacecraft (the so-called ram direction). Since the orbital planes of Europa and Galileo about Jupiter coincide to within a few degrees, such particles approached the detector from a direction corresponding to $\rm \approx 270^{\circ}$ rotation angle during all eight encounters. Rotation angles of about $\rm 90^{\circ}$ are opposite to the direction towards Europa. With the sensor field of view of $\rm 140^{\circ}$, particles detected with rotation angles $\rm 180^{\circ} < \Theta < 360^{\circ}$ are compatible with an origin from Europa itself. In the following we will call them Europa particles. This detection geometry is very similar to the one at the majority of the Ganymede flybys [@krueger2000a].
The direction from which the Jovian dust stream particles were observed varied during Galileo’s path through the Jovian system: when Galileo approached the inner Jovian system, rotation angles around $\rm 270^{\circ}$ were observed. Between 1996 and early 1999 (the time span considered here) the rotation angle shifted to $\rm 90^{\circ}$ shortly before Galileo’s closest approach to Jupiter and the stream particles approached from this direction on the outbound portion of the spacecraft trajectory. Therefore, depending on when an individual satellite flyby occurred, stream particles approached the sensor from one or the other direction. In the cases of the Europa flybys considered here, the stream particles approached from rotation angles $\rm 0^{\circ} < \Theta < 180^{\circ}$ ([[*i.e.*]{}]{} opposite to that of the Europa particles) or stream particle impacts had already ceased because of the unfavourable detection geometry. It should be emphasized that during most of these eight Europa encounters more impacts were detected from the Europa direction than from the direction from which stream particles were to be expected. Only between zero and two impacts from the direction of the stream particles occurred during six encounters. Only encounters E6 and E17 showed the same number of impacts from $\rm 0^{\circ} < \Theta < 180^{\circ}$ (stream particles) as from the opposite direction (Europa particles). The statistics of particle detections is given in Table \[tab\_sat\].
A total number of 64 Europa particles have been identified below $\rm 8\,R_E$ altitude during these eight encounters purely by their impact direction (Table \[tab\_sat\]). For our further analysis we use a cut-off altitude of $\rm 8\,R_E$ because this is close to the extension of Europa’s Hill sphere (Table \[phys\_prop\]). We can minimize the potential contamination by particles belonging to other Jovian dust populations this way (see Sect. \[sec\_rate\_eu\] for a discussion of their impact rates). For instance, a cut-off altitude of $\rm 10\,R_E$ would increase the number of grains classified as Europa particles by only five.
For some flybys, the numbers of identified Europa particles are lower limits to the true numbers of detected grains because the complete set of parameters measured upon impact could be transmitted to Earth for only a fraction of all impacts (column 9 of Table \[tab\_sat\]). At E14, E15 and E19, however, the complete set of parameters was transmitted for all impacts within 2 h around closest approach.
In Table \[tab\_sat\] (columns 7 and 8) we compare the number of Europa particles with the number of all events (dust plus noise) detected by the instrument below $\rm 8\,R_E$ and from a direction $\rm 180^{\circ} < \Theta < 360^{\circ}$. This shows that the noise contribution to the total number of detected events in this altitude range is between 0 and 60% with an average of 23% (17 out of 81 events are classified as noise). Although Europa orbits Jupiter in the region where high noise rates occurred, the total number of noise events in the data set is relatively small. A plot similar to Fig. \[rot\_eu\] but with all detected events (column 8 of Table \[tab\_sat\]) also shows dust concentrations at the Europa closest approaches so that the derived densities would also peak towards the satellite (Sect. \[sec\_rate\_eu\]). Hence, our conclusion that the dust impacts are concentrated towards the surface of Europa does not depend upon the applied noise removal algorithm. We will come back to the noise problem in Sect. \[sec\_rate\_eu\] where we will determine the spatial distribution of dust surrounding Europa. For our analysis of the Europa dust cloud we will use class 3 and noise-removed class 2 data. It should be noted that there is no physical difference between dust impacts categorised into class 2 and class 3.
### Impact velocity {#sec_velocity_eu}
In Fig. \[velocities\_eu\] we show the velocity distribution of the Europa particles whose impact velocity has been determined with a velocity error factor $\rm VEF < 6$ [@gruen1995a]. 53 particles fulfill this criterion. During all eight encounters of Galileo with Europa, the flyby velocity was close to $\rm 6\,{\,\hbox{km}}{\,\hbox{s}}^{-1}$ (Table \[tab\_sat\]), well above the detection threshold of the dust instrument for micrometer-sized grains at $\rm 2\,km{\,\hbox{s}}^{-1}$. We can separate two subsets of particles from their velocity distribution, similar to the measurements at Ganymede: the Jovian stream particles with velocities typically above $\rm 10\,{\,\hbox{km}}{\,\hbox{s}}^{-1}$ and the slower Europa particles. The mean velocity of the 53 Europa particles in Fig. \[velocities\_eu\] is $\rm 5.5 \pm 3.5 \,{\,\hbox{km}}{\,\hbox{s}}^{-1}$ ($\rm 1\, \sigma$). Given a typical uncertainty for an individual velocity measurement of a factor of two, this value is in good agreement with the velocity of Galileo relative to Europa.
The velocity measurements – like the measurements at Ganymede – are in agreement with dust particles belonging to a dust cloud of Europa. They confirm that the empirical velocity calibration of the dust instrument can be applied to the relatively slow ejecta particles, although the calibration is wrong for the much smaller and faster Jovian dust stream particles. This result also confirms that the calibrated impact velocities can be used to identify particles belonging to a dust cloud when such grains cannot be identified by their impact direction. This will be applied to the data obtained in the vicinity of Callisto in Sect. \[sec\_ca\_velocity\]. It should be emphasized that the two velocity distributions in Fig. \[velocities\_eu\] overlap, leading to some ambiguity in the identification of individual grains. Hence, the velocity criterion can only be applied to a statistically large data set.
In Fig. \[rot\_eu\] we have marked particles according to their calibrated velocities: those with impact speeds below $\rm 10\,km{\,\hbox{s}}^{-1}$ are shown as circles, faster grains as crosses. During seven Europa encounters of Galileo, the majority of particles with $\rm 180^{\circ} \leq \Theta \leq 360^{\circ}$ had impact speeds below $\rm 10\,{\,\hbox{km}}{\,\hbox{s}}^{-1}$, consistent with a particle origin from Europa. Thus, for these seven flybys the identification of cloud particles with the velocity criterion is in agreement with the identificaton by the impact direction alone. Only the E19 encounter had 50% of particles (5 out of 10) with higher calibrated impact speeds so that the majority of grains would be classified as stream particles and be rejected if the impact direction were not applicable as the main identification parameter. Most of the particles detected at E19 with impact speeds above $\rm 10\,km{\,\hbox{s}}^{-1}$ (4 out of 6 particles with $\rm 0^{\circ} \leq \Theta \leq 360^{\circ}$), however, have a velocity error factor $\rm VEF > 6 $, which makes their speed calibration very uncertain, anyway. To summarize, the identification of Europa particles from their impact direction and impact speed is quite reliable for all eight Europa encounters.
### Impact rate and number density {#sec_rate_eu}
With 64 complete data sets of particles detected during eight Europa encounters we can calculate the dust impact rate in the close vicinity of this moon (Fig. \[rate\_eu\]). We have defined distance bins equally spaced on a logarithmic scale. Then we divided the number of particle impacts in each bin for which the complete set of measured impact parameters has been transmitted to Earth by the time Galileo has spent in that bin (dotted lines). Finally, to correct for incomplete data transmission (Sec. \[sec\_instrument\]), we have multiplied the impact rate bin by bin with the ratio between the number of counted particles and the number of particles for which the complete data set has been transmitted. These corrected impact rates are shown as solid lines.
Figure \[rate\_eu\] shows that for those Europa encounters where the number of detections is sufficiently large (at least ten particles; E4, E11, E12 and E19) the impact rate clearly increases towards Europa. This implies a concentration of dust particles at Europa. It was already obvious from Fig. \[rot\_eu\] and confirms the earlier results of a dust concentration at Ganymede [@krueger2000a]. On the other hand, at the remaining four encounters, the number of detections is so low that no statistically meaningful radial profile can be derived although the data are also compatible with particle concentrations at Europa. It should be noted that the correction for incomplete transmission is small in all bins and does not significantly affect the slopes of the power law fits. The slopes derived with correction for incomplete transmission are in the range $-1.4$ to $-2.7$ (Table \[tab\_sat\]). This is somewhat flatter than the most reliable slopes obtained for Ganymede [@krueger2000a].
An important question arises: is the slope of the dust distribution at Europa truly flatter than that at Ganymede or is it an artefact caused by the measurement process? Two effects may cause a flattening of the slope: 1) a background of dust particles in jovicentric space; and 2) incomplete removal of noise events by our noise removal technique. Both would lead to an artificial flattening of the derived dust impact rate profile.
To analyze the first hypothesis – a background dust population – we consider the complete data set of Galileo dust measurements in the Jovian system: the dust instrument has detected a number of micrometer-sized particles mostly in the region between the Galilean moons [@gruen1998]. At least two populations of grains can be distinguished: particles on bound prograde orbits about Jupiter and a population on retrograde orbits [@colwell1998a; @colwell1998b; @thiessenhusen2000; @krivov2002a]. Depending on the detection geometry of the dust instrument during a specific orbit of Galileo about Jupiter, impact rates of particles from both populations taken together were up to six per day in the region of Europa [@thiessenhusen2000 their Fig. 10]. Considering that one Galileo passage through Europa’s Hill sphere (Table \[phys\_prop\]) lasted about 75 min, only one dust impact from these populations has to be expected every third Europa flyby. We therefore conclude that a contamination by particles on jovicentric orbits is negligible for our analysis of the Europa dust cloud. A potential contamination caused by such grains would be even smaller at Ganymede and Callisto because their number density decreases further away from Jupiter.
In order to check the second possibility – incomplete removal of noise events from the data set – we use an alternative approach to calculate noise-free dust impact rates. The noise removal algorithm applied so far determines whether each individual class 2 event is most likely a true dust impact or a noise event. Instead, we calculate the average noise rate measured with the dust instrument and subtract it from the total counted rate to obtain the dust impact rate. We have first calculated the total event rate of dust impacts [*plus noise*]{} with the same technique as before, [[*i.e.*]{}]{} from the complete data sets, without applying our noise-removal scheme. This gives somewhat flatter power law slopes than those derived for the noise-removed data set (between $-1$ and $-2$). We have then calculated the fraction of noise events in the class 2 accumulator data with our noise identification scheme during a one-day interval centered around each Europa flyby and calculated the rate of noise events in the counter data by multiplying the total counted rate with the fraction of noise events. Typical noise rates are between 0.04 and 0.1 per minute. We have then subtracted this noise rate from the total event rate obtained from the entire data sets. The resulting radial density profiles have power law slopes between $-2$ and $-4$. The theoretically expected value is about $-2.5$, or more precisely it is steeper than $-2.5$ for $r\la 5\,R_{\rm sat}$ (bound grains dominate) and it becomes flatter, between $-2.5$ and $-2$ (escaping grains dominate), farther out [@krivov2003]. Thus, we conclude that the observed radial density profiles which are somewhat flatter than the expected values may be due to incomplete noise removal.
We do not investigate variations of the slopes between individual encounters because of the large statistical uncertainties and the potential unrecognised noise contamination of the data. Spatial variations with respect to the flyby position relative to the satellite will be addressed in a future investigation.
With the impact rate profiles derived for the individual Europa encounters we can now calculate the spatial density of dust in the environment of this satellite. We first divide bin by bin the impact rate by the spin-averaged detector area to obtain fluxes ($\rm m^{-2}\,s^{-1}$). Then we divide these fluxes by the mean impact velocity (spacecraft velocity relative to the moon) for a given flyby. This results in mean number densities ($\rm m^{-3}$) in the various distance bins. Note that the slope of the number density is the same as that of the impact rate, because both the spin-averaged detector area and the mean impact velocity are assumed to be constant (independent of distance) for any individual flyby. The result is shown in Fig. \[num\_dens\_eu\]. The number densities show a clear increase towards Europa and the average slope is $\rm -2.02\pm 0.63$. It is remarkable that the variation in the derived number densities from encounter to encounter is relatively small. Since the closest approaches of Galileo at Europa occurred at different longitudes and latitudes of Europa, it indicates that the dust distribution around this moon does not show strong spatial or temporal variations.
We have also checked alternatives for the origin of the dust impacts detected in the close vicinity of the Galilean satellites other than the impact-ejection mechanism [@krueger2000a]. Gravitational capture of the grains by the satellites can be dynamically ruled out. Electromagnetic interactions seem to be too weak, in particular at Europa and Callisto which do not have their own magnetic fields. Although the geysers on Io are the most likely source for the Jovian dust stream particles, no geyser activity has been observed on the other Galilean moons. The most plausible explanation for the origin of the dust grains was continuous bombardment of the satellites by interplanetary micrometeoroids. In this work we assume spherical symmetry of the clouds which is supported by our measurements: passages of Galileo at different latitudes and longitudes of the moons did not reveal strong asymmetries in the dust distribution, except, possibly, at Callisto (see Sect. \[sec\_rate\_ca\]). This implies a spherical structure of the dust distribution surrounding the satellites. Of course, it does not rule out the existence of asymmetries in the dust density which will be investigated in the future [@sremcevic2003]. We now look at the number density profile expected from theory. @krivov2003 developed a model of a spherically symmetric, stationary dust cloud around a satellite, maintained by impacts of interplanetary micrometeoroids. To the first approximation, the number density of dust grains ejected into ballistic orbits, which dominate the cloud at distances of several satellite radii, is $$n_{\mathrm{bound}}(x) \propto x^{-5/2},
\label{bound_cal}$$ where $x\equiv r/R_{\mathrm{sat}}$ is the distance measured in satellite radii. The contribution of escaping grains into the cloud is somewhat shallower: $$n_{\mathrm{unbound}}(x) \propto x^{-2}
\label{unbound_cal}$$ which slightly flattens the overall radial profile at larger distances from the moon, closer to its Hill’s sphere.
@krivov2003 have also constructed an algorithm to calculate the proportionality factors in Eqs. (\[bound\_cal\])–(\[unbound\_cal\]). The algorithm implies a chain of estimates: for the mass flux and typical speed of projectiles, for gravitational focussing of impactors by Jupiter, for the ejecta yield, ejecta mass and velocity distributions, etc. The values of the model parameters (both assumed and derived) for Europa are given in Table \[tab\_parms\]. Other parameters that have the same values for all Galilean satellites, are: slope of cumulative ejecta mass distribution $\alpha=0.83$, maximum mass of an ejected fragment $M_{\rm max}=10^{-5}{\,\hbox{g}}$, opening angle of cone into which particles are ejected $\psi_0 = 90^\circ$. For a detailed description of the parameters, the reader is referred to the original paper. We note that the model calculates number densities of particles with masses above the detection threshold of the Galileo dust detector. As the threshold is speed-dependent, the number densities are computed separately for each Galileo flyby, and the results for specific flybys generally differ even for the same moon and the same distance. A strong advantage of this approach is that it enables direct comparison of the number densities predicted by the model with those derived from the measurements. Using this model for Europa, we obtained theoretical curves superimposed on the data points in Fig. \[num\_dens\_eu\]. A comparison between the number densities derived from the Galileo measurements and those computed with the model will be given in Sect. \[sec\_comp\].
### Mass distribution
The charge released by an impact of a dust particle onto the detector target depends on the mass and the velocity of the grain (Equ. \[equ1\]). In particular, to calculate the particle mass one has to know its impact velocity. The calibration of velocity and mass from the measured charge rise times and charge amplitudes is usually performed based on laboratory measurements obtained at a dust accelerator.
In Fig. \[mass\_hist\_eu\] we show the mass distribution of the particles from all eight Europa flybys for which the velocity could be reliably determined ($\rm VEF < 6$; 53 particles in total). In the upper panel the complete instrument calibration has been used to obtain particle speed and mass. With this method the uncertainty of the impact velocity is typically a factor of 2 and that of the mass is a factor of 10.
The dust detector has a velocity-dependent detection threshold [@gruen1995a]. The threshold for particles approaching with $\rm 6 {\,\hbox{km}}{\,\hbox{s}}^{-1}$ is shown as a dashed line. The mass distribution is incomplete around this value.
The mass distribution is also affected by the low data transmission capability of Galileo and the data storage scheme in the instrument memory. As a result, nearly all data sets lost are in the lowest amplitude range AR1 which — for particle velocities of about $\rm 6\, km\,s^{-1}$ — corresponds to the mass range below $\rm \sim 3 \times 10^{-15}\,kg$. If we assume that the lost particles are equally distributed over the mass bins below this value, the maximum of the mass distribution is artificially too low by less than a factor of 1.2. Thus, incomplete data transmission does not significantly affect the mass distribution for Europa particles.
If the individual impact velocities of dust particles were known with a higher accuracy than the typical factor of 2 uncertainty from the instrument calibration, the uncertainty in the mass determination could be improved. The measured mean impact velocities of Europa particles are close to the velocity of Galileo relative to Europa during the individual encounters (Table \[tab\_sat\]). We therefore assume the latter ones as the particles’ impact velocities and show the recalculated particle mass in the lower panel of Fig. \[mass\_hist\_eu\]. The width of the mass distribution is significantly smaller than that derived from the calibrated impact velocities. This method has also been successfully applied to calculate the size distribution of Ganymede ejecta particles [@krueger2000a] and interstellar dust particles measured with Galileo and Ulysses [@landgraf2000a].
The mean mass of the Europa particles is $\rm 9.0 \times 10^{-15}\,kg$. Assuming spherical particles with a density of $\rm 1\,g\,cm^{-3}$ – the density of water ice – this corresponds to a particle radius of $ \rm \approx 1 \,\mu m$.
Degradation of the dust instrument caused by the high radiation dosages in the Jovian magnetosphere lead to shifts of the calibrated masses and impact speeds of the dust particles. All data collected after mid-1997 are affected by this shift, the later in the mission the data were collected, the stronger the shift. For our Europa measurements this means that masses are too low by a factor of about 1.5 beginning with the E11 encounter. We have corrected the calibrated masses for these data sets and constructed a corrected mass distribution (solid histogram in the bottom panel of Fig. \[mass\_hist\_eu\]). Since we have taken the speed of Galileo relative to Europa as the impact speed of the particles, we need to correct the masses only. The shift in the velocity calibration caused by the instrument degradation does not affect this mass distribution. The resulting mean mass of the Europa particles is $\rm 1.3 \times 10^{-14}\,kg$. It should be emphasized that the shift in the velocity calibration does not affect the identification of the Europa particles because the particles at the Europa encounters have been identified by their impact direction alone, without using the impact speed as an additional criterion.
Callisto
--------
### Impact direction {#impact-direction-1}
During Galileo’s prime mission about Jupiter in 1996 and 1997 the spacecraft had three close flybys at Callisto (C3, C9, C10). The spacecraft orientation during these encounters allowed the detection of ejecta cloud particles close to Callisto. During all Callisto flybys after mid-1999 the spacecraft orientation prevented the detection of ejecta cloud particles so that the measurements at Callisto are restricted to these three encounters. Unfortunately, all three of them occurred on the portion of the Galileo trajectory inbound to Jupiter where Jupiter stream particles and potential ejecta particles from Callisto approached the dust sensor from the same direction ($\rm 180^{\circ} \leq
\Theta\leq 360^{\circ}$; Sect. \[sec\_rot\_eu\]). This is shown in Fig. \[rot\_ca\]: almost the entire number of dust impacts measured close to Callisto occurred from this direction. Thus, Callisto particles could not be uniquely identified by their impact direction alone, and we had to use the impact velocity as an additional criterion to identify them.
The analysis of the ejecta particles detected at Europa (Sect. \[sec\_velocity\_eu\]) and Ganymede [@krueger2000a] showed that – on average – stream particles and ejecta cloud particles occupy different regimes in calibrated impact speed: cloud particles have typical speeds below $\rm 10\,km{\,\hbox{s}}^{-1}$ which are on average very close to the encounter velocity of Galileo with the satellite, whereas stream particles have significantly higher calibrated speeds. This has been successfully applied to identify Ganymede cloud particles from Galileo’s G8 encounter at Ganymede. We apply the same velocity criterion here to separate Callisto particles from the dust streams. The numbers of Callisto particles identified this way are listed in Table \[tab\_sat\] for each orbit. The total number of Callisto particles from all three encounters is 35.
Only particles detected at altitude below $\rm 6\,R_C$ are considered for further analysis in order to minimize the contamination by stream particles. The analysis of the Europa and Ganymede data showed that a few stream particles have calibrated velocities below $\rm 10\,{\,\hbox{km}}{\,\hbox{s}}^{-1}$ and would erroneously be classified as Callisto particles (Fig. \[rot\_ca\]). Most cloud particle impacts at Europa and Ganymede occurred below an altitude of about $\rm 6\,R_{\rm sat}$ (Fig. \[rot\_eu\] and [@krueger2000a]) so that the inclusion of particles detected further away would increase the probability that the particles are actually stream particles rather than dust cloud particles. An apparent concentration of stream particle impacts within $\rm 3\,R_C$ altitude at C3 and within $\rm 5\,R_C$ at C10 (Fig. \[rot\_ca\]) is due to a higher data transmission rate of Galileo in these periods [@krueger2001a].
Callisto orbits Jupiter outside the region within $\rm 15\,R_J$ where the high noise rates occurred. Thus, a potential noise contamination of the Callisto data is expected to be very low. For our analysis of the Callisto dust cloud we will use class 3 and class 2 data without noise removal.
### Impact velocity {#sec_ca_velocity}
A total number of 35 Callisto particles have been identified by their calibrated impact velocity below $\rm 10\,{\,\hbox{km}}{\,\hbox{s}}^{-1}$ and below an altitude of $\rm 6\,R_C$. Their mean impact velocity is $\rm 6.4 \pm 2.1\,{\,\hbox{km}}{\,\hbox{s}}^{-1}$ ($1\,\sigma$). This value is artificially too low because the velocity distribution is cut off at $\rm 10\,{\,\hbox{km}}{\,\hbox{s}}^{-1}$. The average flyby speed of Galileo at Callisto was $\rm 8.1\,{\,\hbox{km}}{\,\hbox{s}}^{-1}$ and both speeds agree within $1\,\sigma$. The velocity distribution of the Callisto particles is shown in Fig. \[velocities\_ca\].
Degradation of the dust instrument electronics does not significantly affect the impact velocities of the particles because most the Callisto encounters occurred relatively early during the Galileo mission. In particular, the identification of Callisto particles via their impact speed is not affected.
### Impact rate and number density {#sec_rate_ca}
With Callisto particles from all three Callisto encounters identified by their impact speed and impact direction, we can construct the radial profile of the dust impact rate in the same way as we did for the Europa flybys before. This is done in Fig. \[rate\_ca\]. For the C3 and C10 encounters the impact rate increased towards Callisto. At C9 the number of detected particles was only three. The derived radial profile — although being very uncertain — is compatible with an increase towards Ganymede. We conclude that the dust impact rates detected at Callisto are compatible with a dust concentration surrounding this moon.
In Fig. \[num\_dens\_ca\] we show the number densities for Callisto derived from the radial profiles of the impact rate. The data from the individual flybys show an increase towards the surface of the moon. The number densities derived from the three flybys show a large variation from flyby to flyby, much larger than those for Europa. In particular, the number densities derived from the C9 data are very low. Possible reasons for this variation will be discussed in Sect. \[sec\_comp\].
In the same figure, the superimposed curves show the number density profile calculated with the model [@krivov2003] and parameters listed in Table \[tab\_parms\]. A comparison between the data and the model will be given in Sect. \[sec\_comp\].
### Mass distribution
The mass distribution for the 35 Callisto particles with $\rm VEF < 6$ is shown in Fig. \[mass\_hist\_ca\]. As for Europa, we show the mass distribution with the calibrated impact velocities (upper panel) and that obtained by applying the velocity of Galileo relative to Callisto (bottom panel). Here the detection threshold is $\rm 8\,km{\,\hbox{s}}^{-1}$, and the mass distribution is incomplete around the threshold. Again, the mass distribution is incomplete in the bins below about $\rm 10^{-15}\,kg$ due to incomplete data transmission and the maximum of the mass distribution may be artificially too low by up to a factor of 1.3.
The mean calibrated mass of the Callisto particles is $\rm 3.7 \times 10^{-16}\,kg$, which is an order of magnitude lower than the value derived for the Europa cloud particles. Again, assuming spherical particles with a density of $\rm 1\,g\,cm^{-3}$, this corresponds to a particle radius of $ \rm 0.5 \,\mu m$. We have also corrected the masses of the particles for instrument degradation and constructed a corrected mass distribution (solid histogram in the bottom panel of Fig. \[mass\_hist\_ca\]). The corrected mean mass of the Callisto particles is $\rm 5.2 \times 10^{-16}\,kg$.
Io
--
### Impact direction {#impact-direction-2}
Galileo had a total of seven flybys at Io but only the initial one in December 1995 (I0; the orbit notation is I ”zero“) had a favourable detection geometry. During the other flybys at this satellite which occurred after mid-1999 the sensor orientation prevented the detection of ejecta dust particles. Figure \[rot\_io\] shows the sensor orientation at particle impact at the I0 encounter. The bottom panel shows class 3 and noise-removed class 2 data (only four impacts).
The noise identification criteria applied to the Europa data have been developed for the spatial region outside $\rm 10\,R_J$. The I0 data, however, have been collected closer to Jupiter at $\rm 6\,R_J$ where the noise characteristics may have been different [@krueger1999c]. We therefore show the full data set of classes 1 to 3 in the top panel of Fig. \[rot\_io\]. The noise fraction in class 2 derived with the algorithm for secondary ejecta grains is about 80%. The complete class 2 data set also shows a concentration of grains towards Io. Therefore, the noise rejection algorithm may be too restrictive, thus rejecting too many events. Class 1 events, which are usually classified as noise in the Jovian environment, show an interesting behaviour: the impacts cluster at rotation angles $\rm 0 \leq \Theta \leq 180^{\circ}$. This direction is compatible with the approach direction of plasma particles in the Io torus.
A few days before the flyby at Io, the channeltron high voltage was decreased and the charge detection thresholds were raised [@gruen1996c; @krueger1999a] to reduce the instrument sensitivity in the high radiation environment of the inner Jovian magnetosphere. This reduced the instrument sensitivity for class 3 impacts. Those impacts, however, that did not generate enough charge to become class 3 events should have shown up as class 2 impacts. Unfortunately, class 2 is contaminated with noise so that the identification of these particles is ambiguous. Fortunately, only a small fraction of the data sets of particles was lost due to incomplete data transmission (Table \[tab\_sat\]).
Figure \[rot\_io\] shows four particles within an altitude of about $\rm 6\,R_I$ (the Hill sphere of Io). Two of these particles approached the detector from a direction $\Theta \approx
270^{\circ}$. The other two particles were detected when the dust detector pointed $\approx 90^{\circ}$ away from this direction. A check of the approach direction of potential Io particles revealed that all four particles are compatible with an Io origin: the approach direction of Io particles as seen from Galileo was so close to the spin axis of the spacecraft (direction opposite to antenna axis) that they were detectable at almost all rotation angles [@krueger1999c]. We therefore consider all four particles as probable Io particles. Only one of these impacts occurred in ion amplitude range AR1, two were detected in AR2 and one in AR3. Io dust stream particles were detected in AR1 only [@gruen1998] so that the identified particles were unlikely stream particles. In addition, the impact rate of stream particles was reduced in the inner Jovian system because of reduced impact speeds [@graps2001a].
### Impact velocity {#impact-velocity-1}
The total number of identified potential ejecta particles from Io is only four. With this very low number of detections no reasonable velocity distribution can be constructed. We can, however, still check whether the average particle velocity is compatible with the hypothsis that the particles are basically at rest with respect to Io. The velocity of Galileo relative to Io at the I0 encounter was $\rm 15\,km{\,\hbox{s}}^{-1}$. The averaged particle speed derived for the four particles is $\rm 10.3 \pm 8.4\,km{\,\hbox{s}}^{-1}$. Although one has to keep in mind that the statistical uncertainty of this value is very large, it is compatible with particles being bound to Io.
### Impact rate and number density {#impact-rate-and-number-density}
The dust impact rate at Io derived from the four identified Io particles is shown in Fig. \[rate\_io\]. It shows a slight concentration towards Io. One has to keep in mind, however, that the radial profile of the impact rate is very uncertain because of the small number of dust detections in the close vicinity of this moon.
The number density of dust in the close vicinity of Io derived from the impact rate profiles is shown in Fig. \[num\_dens\_io\]. For comparison we show the data points for the noise-removed data set (solid lines) and for the complete class 2 and class 3 data set. The curves in the same figure depict the number density profile calculated with the model [@krivov2003] and parameters listed in Table \[tab\_parms\]. A comparison between the data and the model will be given in Sect. \[sec\_comp\].
### Mass distribution
No attempt has been made to construct a mass distribution because of the small number of detections at this moon. The mean mass of the particles taking the speed of Galileo relative to Io as the impact speed is $\rm 8.5 \times 10^{-16}\,kg$. Note that this value is not affected by the instrument degradation because this Io flyby occurred at the beginning of Galileo’s Jupiter mission.
Comparison of the dust clouds surrounding the Galilean moons {#sec_comp}
============================================================
In the previous section we have analyzed the dust impacts detected in the circumsatellite dust clouds individually for Io, Europa and Callisto. The dust cloud of Ganymede has been investigated in an earlier publication [@krueger2000a]. For each of the four moons we have identified impacts of probable ejecta cloud particles in the Galileo dust data set, determined their impact speeds and mass distributions and have derived impact rate and number density profiles. We now take the data sets for all four moons together to compare the properties of their surrounding dust clouds.
Mass distribution
-----------------
The mass distribution of the grains allows a simple check for the compatibility of the data with the hypothesis of the impact origin of the detected particles. We took the mass distributions for Europa, Callisto and Ganymede (Fig. \[mass\_hist\_eu\], Fig. \[mass\_hist\_ca\] and [@krueger2000a]) and show linear fits to the cumulative distributions in Fig. \[massdist\]. The slopes of the cumulative mass distributions assuming Galileo’s velocity relative to each moon as the particle impact speed are given in Table \[tab\_results\]. Only one data point is shown for Io because of the scarcity of the data. For the three other moons the slopes are in the range between 0.58 and 0.86, which is in good agreement with the typical slopes one expects for impact ejecta ($0.5 \la \alpha \la 1.0$; see, [[*e.g.*]{}]{}, @koschny2001b). It should be emphasized that even though the statistical uncertainties in the data sets are relatively large because of the small number of detections, the slopes derived for the three moons do not differ very much. The slopes derived for the mass distributions taking the calibrated impact speeds are in the range 0.5 – 0.6 and are thus even closer together, although they are flatter than those derived with the spacecraft speed relative to the moon.
It should be noted that the mass distributions of dust around Ganymede and Callisto agree very well (0.82 vs. 0.86) whereas the one for Europa is somewhat flatter (0.58). This might reflect differences in surface properties of the satellites: for instance, flatter mass distributions are typical of looser targets than of consolidated ones [see, [[*e.g.*]{}]{}, @koschny2001b and references therein].
Number densities {#subsec_numden}
----------------
The number densities derived for all four Galilean moons are shown in Fig. \[num\_dens\_all\]. Straight lines are least squares fits to the data for each moon. We list the slopes of these curves, which are averages of slopes for individual flybys at each moon, in Table \[tab\_results\] (col. 8). The average slopes for each moon are between $-1.6$ for Callisto and $-2.8$ for Ganymede, with Europa being in between ($-2.0$). The very uncertain slope for the Io data ($-2.0$) is close to the Europa value. The Ganymede data show the steepest slope but also the largest uncertainty. This is mainly caused by the incomplete data transmission which mostly affected the G2 and G8 measurements. The incompleteness affects the error via $\propto(N\pm\sqrt{N})\times{\rm correction}$, therefore giving larger limits compared to the $100\%$ transmission case.
Altogether, the slopes are roughly consistent with the one predicted in the framework of the spherically-symmetric cloud model [@krivov2003], between $-2.0$ and $-2.5$ \[see Eqs. (\[bound\_cal\])–(\[unbound\_cal\])\]. Only the slope derived for Callisto is somewhat flatter. A study of possible asymmetry effects in the clouds has shown that this “reference” slope may be substantially flatter or steeper, depending on the flyby geometry and the position of the satellite in its orbit at the time of flyby [@sremcevic2003]. This may account for a tangible scatter in the slopes that we derived from the data.
The measured absolute number densities of all four clouds at a given distance (measured in satellite radii) are similar — see Fig. \[num\_dens\_all\]. It is important to compare this result with the theory. The absolute number density of a dust cloud at a given distance from a satellite should depend in a non-trivial way on a number of factors: mass of the moon, its planetocentric distance (through a distance-dependent gravitational focussing of the impactor flux), as well as the satellite surface properties. All these dependences are taken into account in the model which was confined, however, to a solid ice-silicate surface [@krivov2003]. The number densities of the clouds around Europa, Ganymede and Callisto, computed with the model, turned out to be within a factor of several from each other (most notably, larger yields $Y$ for closer-in satellites are compensated by lower ejecta speeds $u_0$, because of an energy conservation requirement used in the model). The number densities of these three clouds are consistent with the densities derived from the data. Not so for the Io cloud: the same model [@krivov2003] predicts a much higher dust number density (mostly because of the larger flyby speed at Io, resulting in a possibility of detecting much smaller dust grains compared to the other Galileans). The number density in the Io cloud predicted by the model is at least an order of magnitude higher than observed. This could be due to the scarcity of the Io data (4 individual dust impacts only). Alternatively, this may be a real effect, caused by different surface properties of Io (volcanic deposits and condensed gases like $SO_2$ frost) compared to the other three Galilean moons (presumably “solid” ice with some contamination by non-volatile materials). Such a view seems to be indirectly supported by our preliminary analysis of the dust environment between the orbits of Io and Europa (work in progress): the density of the “Galilean ring” [@krivov2002a] does not seem to increase from the Europa orbit towards Jupiter, which might be compatible with Io being a weaker source of ejecta than it would be, if it were similar in surface properties to the other Galilean satellites.
A comparison of data from different flybys at the same moon shows that for Europa the derived dust densities do not vary significantly from flyby to flyby. This indicates little or no temporal variation and/or dust density variation between the leading and the trailing side of this moon with respect to the field of impactors. On the other hand, the three Callisto flybys showed a significant variation by more than an order of magnitude between the C9 and the C10 flybys. For Ganymede, the G2 and G8 flybys showed somewhat larger number densities than the G1 and G7 flybys. The former, however, have the largest uncertainties because of incomplete data transmission and particle identification via the impact speed criterion. At present, it is not clear whether the differences between the data from different flybys of Callisto and Ganymede can be attributed to asymmetries in the circumsatellite dust clouds modelled in @sremcevic2003. A comparison between the data and theory is hampered by poorly known directionality of impactors in the vicinity of Jupiter. This issue will be the subject of a future investigation.
Grain velocities in the clouds
------------------------------
The theoretical models of dust clouds predict that, on average, the constituent particles should have substantial velocities with respect to their parent satellites. Based on @krivov2003 formulas, about 40–50% of the grains at distances from 2 to 8 $R_{\mathrm{sat}}$ in the Ganymede and Callisto clouds should be faster that $2{\,\hbox{km}}{\,\hbox{s}}^{-1}$. For Europa the fraction is 30–40%.
The question that we address now is: can we find indications in the data that some grains have appreciable velocities relative to the respective moon? The most natural way would be to look at possible deviations of the actual impact speeds from the mean value, equal to the spacecraft speed with respect to the satellite. Unfortunately, this is not possible: as we have seen (Figs. \[velocities\_eu\] and \[velocities\_ca\]), the instrument calibration is by far not accurate enough to do that. Another possibility would be to look at the impact directions, [[*i.e.*]{}]{} at the rotation angles of impacts. For most of the flybys the cloud particles, if they were at rest with respect to the moon, could only be detected in the rotation angle range $\Theta = 270^{\circ} \pm x$, where the semi-width $x$ of the detectability range is a (known) function of the FOV opening angle and the angle between the Galileo ram direction and its spin axis, $\beta$. The semi-width $x$ is $180^\circ$ for $\beta \le 10^\circ$ and decreases to $67^\circ$ at $\beta = 90^\circ$. Should a particle, which we identify as a cloud particle, have had a $\Theta$ value somewhat outside the range $\Theta = 270^{\circ} \pm x$, this would be an indication that the particle had an appreciable velocity, so that the impact velocity deviated markedly from the anti-ram direction.
One should not expect the number of such events to be high: even very fast grains can only show up in this test if the direction of their velocity vector is appropriate, and only for some flyby geometries. We have checked all cloud particles and found four individual impacts of this kind: one in G7, one in E11 and two in E12.
Mass budget
-----------
As was the case for Ganymede earlier [@krueger2000a], we give some general estimates concerning the mass budget of the dust clouds of the other Galilean moons as derived from the model [@krivov2003]. The results, including new estimates for Ganymede, are collected in Table \[tab\_budget\]. Note that these are only crude estimates which are uncertain by at least one order of magnitude, perhaps even more. The expected steady-state masses of the clouds range from about 10 tons for Callisto to about 200 tons for Io; the Io cloud may, however, be lighter — see discussion in Sect. \[subsec\_numden\]. Interestingly, the mass injection rate of the material into circumjovian space is similar for all four moons and is, in turn, comparable to the mass flux of impactors onto respective satellites, $\sim
100{\,\hbox{g}}{\,\hbox{s}}^{-1}$. This means that each satellite “redirects” nearly as much dust into the circumplanetary space as it receives from the interplanetary one. Of course, the mass/size and velocity distributions of the “incoming” and “outgoing” matter are generally quite different.
Conclusions {#sec_conclusions}
===========
We have examined the dust impacts registered by the Galileo dust detector in the immediate vicinity of Io, Europa and Callisto during a total of 12 flybys at these Jovian moons. By analyzing impact directions and velocities and the mass distribution, as well as spatial locations of the dust impacts in comparison with model predictions [@krivov2003], we have shown that the particles originated from the moons. Our analysis technique was similar to an earlier investigation of dust data collected at Ganymede [@krueger2000a]. The dust impacts recorded at all four moons are compatible with impact debris produced by hypervelocity impacts onto the surfaces of these moons. For the icy moons Europa, Ganymede and Callisto, the mass distributions of the detected grains, as well as the spatial dust densities derived from the measurements are in fairly good agreement with the predictions from the model of hypervelocity impacts of interplanetary dust particles (IDPs), assuming contemporary models of IDP flux at a heliocentric distance of Jupiter and a low-temperature ice-silicate target. For Io, the number of dust detections is too small to derive a reliable mass distribution. The number density obtained for this moon is more than an order of magnitude lower than predicted by the model which assumes a solid ice surface. The lack of detections may be due to a softer, fluffier surface of Io compared to the three icy Galilean moons. Io’s surface is (at least partially) covered with volcanic deposits.
This work continues the analysis of the dust clouds surrounding the Galilean moons and confirms the previous scenario of ejecta dust clouds generated by hypervelocity impacts of micrometeoroids. Up to now, this had only been tested at Ganymede. Our theoretical description is based on the physical conditions in the Jovian system as well as available laboratory data of hypervelocity impacts. We have neglected any spherical asymmetries of the dust clouds surrounding the moons. To a first approximation, this is supported by the data, especially the measurements taken at Europa. On the other hand, the Callisto and possibly Ganymede measurements show a variation between different flybys which might be indicative of a leading-trailing cloud asymmetry caused by the motion of the moons through the field of impactors, as predicted by theory [@sremcevic2003]. This will be addressed in a future analysis.
Most of the dust ejected from the surface is launched into bound orbits and falls back to the moon. These short-lived, but continuously replenished grains form the ejecta dust clouds. A tiny fraction of impact debris is ejected at speeds sufficient to escape from the moon entirely. The ejected mass is comparable with the incoming flux of IDP impactors. The escaping grains go into orbit about Jupiter and most of them will eventually be swept up by one of the Galilean satellites. A tiny fraction of them forms a tenuous dust ring surrounding the planet [@krivov2002a]. This ring is by far too tenuous to be detected optically. By the impact ejecta mechanism, moons turn out to be efficient sources for dusty planetary rings. In particular, Jupiter’s gossamer ring and Saturn’s E ring are thought to be maintained by ejecta particles from smaller moons which orbit their parent planets within the rings. In November 2002 Galileo traversed the gossamer ring and had a close flyby of Amalthea, one of the small Jovian moons which orbits the planet within the ring region. The dust measurements collected during this passage may give new insights into the dynamics and feeding mechanism of this dusty ring and about the significance of small moons as sources of dust.
All celestial bodies without gaseous atmospheres (asteroids and planetary satellites of all sizes) should be surrounded by an ejecta dust cloud. The dust particles in the cloud are composed of surface material from the parent body and, hence, carry information about the surface from which they have been kicked up. Our analysis of the Galileo in-situ dust data has shown that spacecraft measurements near celestial bodies — which act as sources of dust — can be used as a new diagnostic tool to analyze the surface properties of these bodies. This is of particular interest for the Cassini mission which will investigate the Saturnian system beginning in 2004. The Cassini dust instrument will be able to measure the chemical composition of particles in the dust clouds surrounding the Saturnian moons. This way, the surface properties of the source moons can be investigated remotely. Interestingly, the in-situ dust measurements turn into a remote sensing technique where the dust instrument is used like a telescope for surface investigation.
The authors wish to thank Frank Spahn for many valuable discussions. This research has been supported by the German Bundesministerium für Bildung und Forschung through Deutsches Zentrum für Luft- und Raumfahrt e. V. (DLR, grants 50 QJ 9503 3 and 50 OH 0003) and by the Deutsche Forschungsgemeinschaft (DFG, grant No. Sp 384/12-3). We wish to thank the Galileo project at JPL for effective and successful mission operations.
paper.bbl
=0.7
=0.8
=0.8
=0.6
=0.6
=0.6
=0.6
=0.8
=0.7
=0.9
=0.8
=0.6
=0.6
=0.9
=0.9
=0.9
=0.7
=0.7
---------- ------------- --------------- ----------- ----------------- --------------------
Object Jovicentric Radius Symbol Hill radius Escape speed
distance $r_{\rm obj}$ $ r_{\rm Hill}$ $v_{\rm esc}$
($\rm R_J$) (km) ($\rm R_{obj}$) ($\rm km\,s^{-1}$)
Jupiter – 71,492 $\rm R_J$ – –
Io 5.9 1,818 $\rm R_I$ 5.8 2.56
Europa 9.4 1,560 $\rm R_E$ 8.7 2.03
Ganymede 15.0 2,634 $\rm R_G$ 12.0 2.74
Callisto 26.3 2,409 $\rm R_C$ 20.9 2.44
---------- ------------- --------------- ----------- ----------------- --------------------
: Physical properties of Jupiter and the Galilean satellites used in this paper. The Hill radius is defined as $r_{\rm Hill} = r (\frac{m}{3(M + m)})^{1/3}$ with $M, m$ being the masses of Jupiter and the moon, separated by distance $r$. []{data-label="phys_prop"}
-------------- ------------ ---------- ------------------------------------ ------------------------------------ -------------- ---------------- ---------------- ----------------- --------------------
Flyby Date Altitude Spacecraft Average Sensor Particles All events Complete- Slope of
velocity particle area with full with full ness of impact
velocity data sets data sets data set rate
(Year-Day) (km) (${\,\hbox{km}}{\,\hbox{s}}^{-1}$) (${\,\hbox{km}}{\,\hbox{s}}^{-1}$) ($\rm cm^2$) (%)
(1) (2) (3) (4) (5) (6) (7) (8) (9) (10)
\[-1.0ex\]
\[0.5ex\] E4 96-354.287 698 5.7 $6.8 \pm 4.0$ 233 18 23 86 $-2.21\pm 0.36$
E6 97-051.713 586 5.7 $5.9 \pm 4.0$ 224 3 7 88 ($-1.69\pm 0.58$)
E11 97-310.855 2,043 5.6 $5.5 \pm 2.2$ 235 10 11 90 $-1.43\pm 2.25$
E12 97-350.502 201 6.3 $4.5 \pm 1.7$ 142 12 13 61 $-2.72\pm 0.91$
E14 98-088.556 1,644 6.5 $3.4 \pm 1.3$ 61 3 4 100 ($-1.26\pm 0.44$)
E15 98-151.884 2,515 6.4 $5.5 \pm 1.5$ 155 4 6 100 ($-2.11\pm 0.59$)
E17 98-269.163 3,582 6.0 $2.8 \pm 0.8$ 180 4 4 54 ($+0.27\pm 1.56$)
E19 99-032.097 1,439 5.8 $5.7 \pm 4.7$ 152 $10^{\dagger}$ $13^{\dagger}$ $100^{\dagger}$ $-1.70\pm 0.62$
\[-0.5ex\]
\[0.5ex\] C3 96-309.566 1,118 8.0 $6.3 \pm 1.7$ 224 10 35 80 $-1.90 \pm 0.45$
C9 97-176.575 415 8.2 $7.6 \pm 0.2$ 197 3 5 100 ($-0.69 \pm 0.10$)
C10 97-260.013 538 8.2 $6.3 \pm 2.3$ 142 22 37 64 $-1.31 \pm 1.06$
\[-1.0ex\]
\[0.5ex\] I0 95-341.740 892 15.0 $10.3 \pm 8.4$ 133 4 13 87 ($-2.00 \pm 1.88$)
-------------- ------------ ---------- ------------------------------------ ------------------------------------ -------------- ---------------- ---------------- ----------------- --------------------
: Galileo flyby characteristics and parameters for the dust particles detected during Galileo’s flybys at Europa, Callisto and Io (for Europa, particles within $\rm 8\,R_E$ altitude have been included, and for Callisto and Io within $\rm 6\,R_C$ and $\rm 6\,R_I$, respectively): Flyby number (col. 1), time of flyby (col. 2), altitude at closest approach to satellite (col. 3), velocity of Galileo relative to satellite (col. 4), average measured particle velocity (velocity error factor $\rm VEF < 6$, @gruen1995a; col. 5), spin-averaged sensor area (maximum value $\rm 235\,cm^{2}$; col. 6), number of class 2 and class 3 satellite particles for which their complete data set has been transmitted to Earth ($\rm 180^{\circ} < \Theta < 360^{\circ}$; col. 7), number of all events (dust plus noise) detected with $\rm 180^{\circ} < \Theta < 360^{\circ}$ (both within $\rm 8\,R_E$, col. 8), completeness of data set due to incomplete data transmission of Galileo (col. 9) and slope of power law fit to the radial variation of the impact rate (Fig. \[rate\_eu\], \[rate\_ca\], \[rate\_io\]; col. 10). The slopes are weighted with the square root of the number of particles, and the uncertainty takes into account the error bar of each data point. Values given in parentheses have been derived from only four or fewer particles. []{data-label="tab_sat"}
\
$\dagger$: Data transmission ceased 1 Feb 99, 02:38:46h, [[*i.e.*]{}]{} 19 min after closest approach, at 6,767km altitude. All data sets of particles detected earlier during the E19 encounter were transmitted.
----------- ------------------------------------ ------ --------------- ----------------------- ------------------ ----------------------------------- ----------
Satellite $v_{\rm imp}$ $A$ $G_{\rm sil}$ $K_{\rm e}/K_{\rm i}$ $Y$ $u_0$ $\gamma$
$({\,\hbox{km}}{\,\hbox{s}}^{-1})$ (%) $({\,\hbox{m}}{\,\hbox{s}}^{-1})$
Io 26 0.61 0 30 $2.8\times 10^4$ 28 2.0
Europa 21 0.64 0 30 $1.6\times 10^4$ 30 2.0
Callisto 15 0.20 70 20 $7.1\times 10^2$ 51 1.4
----------- ------------------------------------ ------ --------------- ----------------------- ------------------ ----------------------------------- ----------
: Model parameters for different satellites: speed of impactors, geometric albedo, assumed silicate content, energy partitioning parameter, characteristic yield, parameters of the ejecta speed distribution (see Krivov et al. 2002b for description of the parameters).
\[tab\_parms\]
---------- ----------------------------- ------------ ----------------------------- -------------------------- --------------- ----------------- ------------------
Object Average Number Average Average Average Slope $\alpha$ Slope of
flyby of particle particle particle of mass radial number
speed detections speed mass radius distrib. density distrib.
($\rm km{\,\hbox{s}}^{-2}$) ($\rm km{\,\hbox{s}}^{-2}$) (kg) ($\rm \mu m$)
(2) (3) (4) (5) (6) (7) (8)
Io 15.0 4 $10.3 \pm 8.4$ $\rm 8.5\times 10^{-16}$ $\approx 0.6$ – ($ -2.00$)
Europa 6.0 64 $5.5\pm 3.5$ $\rm 1.3\times 10^{-14}$ $\approx 1.0$ $0.58 \pm 0.04$ $-2.02\pm 0.63$
Ganymede 8.2 38 $7.2\pm 4.9$ $\rm 9.5\times 10^{-16}$ $\approx 0.6$ $0.82 \pm 0.04$ $-2.82\pm 2.60$
Callisto 7.2 35 $6.4\pm 2.1$ $\rm 5.2\times 10^{-16}$ $\approx 0.5$ $0.86 \pm 0.14$ $-1.60\pm 0.39$
---------- ----------------------------- ------------ ----------------------------- -------------------------- --------------- ----------------- ------------------
: Derived properties of the impact-generated dust clouds (data for Ganymede taken from @krueger2000a): mean flyby speed of Galileo at the satellite (col. 2), number of detected cloud particles for which the complete data set has been transmitted (col. 3), mean measured impact speed of these particles ($\rm VEF < 6$; col. 4), mean particle mass (col. 5), corresponding particle radius assuming spherical particles with density $\rm 1\,g\,cm^{-3}$ (col. 6), slope of the power law fit to the cumulative mass distribution (col. 7) and average slope of radial number density distribution (col. 8; averages of slopes from individual flybys). The speed of Galileo relative to the satellite has been assumed for the values in col. 5 to 7. For the uncertainties in the number density slopes (col. 8) the uncertainties of the slopes from each individual flyby have been taken into account. []{data-label="tab_results"}
Satellite Io Europa Ganymede Callisto
----------------------------------------------------------------------------- --------------------- --------------------- --------------------- ---------------------
Mass flux of impactors \[${\,\hbox{g}}{\,\hbox{m}}^{-2}{\,\hbox{s}}^{-1}$\] $1 \times 10^{-11}$ $7 \times 10^{-12}$ $4 \times 10^{-12}$ $3 \times 10^{-12}$
Mass inflow of impactors \[${\,\hbox{g}}{\,\hbox{s}}^{-1}$\] 130 50 100 50
Yield $3 \times 10^4$ $2 \times 10^4$ $4 \times 10^3$ $7 \times 10^2$
Mass production rate of ejecta \[${\,\hbox{g}}{\,\hbox{s}}^{-1}$\] $4 \times 10^6$ $9 \times 10^5$ $4 \times 10^5$ $3 \times 10^4$
Mean lifetime of ejecta \[${\,\hbox{s}}$\]$^{a)}$ 50 70 120 250
Steady-state mass of the cloud \[tons\] 200 60 40 9
Fraction of escaping ejecta $1 \times 10^{-4}$ $2 \times 10^{-4}$ $8 \times 10^{-4}$ $4 \times 10^{-3}$
Ejection rate into circumjovian space \[${\,\hbox{g}}{\,\hbox{s}}^{-1}$\] 400 200 300 100
: Mass budgets of the dust clouds (model estimates, see Krivov et al. 2002b for description of the model).
$^{a)}$ Residence time within the Hill sphere before recollision or escape
\[tab\_budget\]
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
We report the first results of a light weakly interacting massive particles (WIMPs) search from the CDEX-10 experiment with a 10 kg germanium detector array immersed in liquid nitrogen at the China Jinping Underground Laboratory with a physics data size of 102.8 kg day. At an analysis threshold of 160 eVee, improved limits of 8 $\times 10^{-42}$ and 3 $\times 10^{-36}$ cm$^{2}$ at a 90% confidence level on spin-independent and spin-dependent WIMP-nucleon cross sections, respectively, at a WIMP mass ($m_{\chi}$) of 5 GeV/${c}^2$ are achieved. The lower reach of $m_{\chi}$ is extended to 2 GeV/${c}^2$.
PACS numbers
: [95.35.+d, 29.40.-n, 98.70.Vc]{}
author:
- 'H. Jiang'
- 'L. P. Jia'
- 'Q. Yue'
- 'K. J. Kang'
- 'J. P. Cheng'
- 'Y. J. Li'
- 'H. T. Wong'
- 'M. Agartioglu'
- 'H. P. An'
- 'J. P. Chang'
- 'J. H. Chen'
- 'Y. H. Chen'
- 'Z. Deng'
- 'Q. Du'
- 'H. Gong'
- 'L. He'
- 'J. W. Hu'
- 'Q. D. Hu'
- 'H. X. Huang'
- 'H. B. Li'
- 'H. Li'
- 'J. M. Li'
- 'J. Li'
- 'X. Li'
- 'X. Q. Li'
- 'Y. L. Li'
- 'B. Liao'
- 'F. K. Lin'
- 'S. T. Lin'
- 'S. K. Liu'
- 'Y. D. Liu'
- 'Y. Y. Liu'
- 'Z. Z. Liu'
- 'H. Ma'
- 'J. L. Ma'
- 'H. Pan'
- 'J. Ren'
- 'X. C. Ruan'
- 'B. Sevda'
- 'V. Sharma'
- 'M. B. Shen'
- 'L. Singh'
- 'M. K. Singh'
- 'T. X. Sun'
- 'C. J. Tang'
- 'W. Y. Tang'
- 'Y. Tian'
- 'G. F. Wang'
- 'J. M. Wang'
- 'L. Wang'
- 'Q. Wang'
- 'Y. Wang'
- 'S. Y. Wu'
- 'Y. C. Wu'
- 'H. Y. Xing'
- 'Y. Xu'
- 'T. Xue'
- 'L. T. Yang'
- 'S. W. Yang'
- 'N. Yi'
- 'C. X. Yu'
- 'H. J. Yu'
- 'J. F. Yue'
- 'X. H. Zeng'
- 'M. Zeng'
- 'Z. Zeng'
- 'F. S. Zhang'
- 'Y. H. Zhang'
- 'M. G. Zhao'
- 'J. F. Zhou'
- 'Z. Y. Zhou'
- 'J. J. Zhu'
- 'Z. H. Zhu'
bibliography:
- 'CDEX10.bib'
title: 'Limits on Light Weakly Interacting Massive Particles from the First 102.8 kg ${\times}$ day Data of the CDEX-10 Experiment'
---
Weakly interacting massive particles (WIMPs, denoted as $\chi$) have been extensively searched via elastic scattering with normal matter in underground direct detection experiments [@PDG2017; @tech] under ultralow background conditions. Liquid noble gas detectors are leading the sensitivities at WIMP mass ($m_\chi$) above 10 GeV/${c}^2$ [@lux; @pandax; @xenon], while solid state detectors are generally used for ${m}_\chi~<~10~$ GeV/$c^2$ [@cogent; @cdex0; @cdex12014; @cdex12016; @cdex12018; @supercdms; @cdmslite; @CRESST-II].
With excellent energy resolution and low energy threshold, *p*-type point contact germanium (*p*PCGe) detectors have been used and further developed for light WIMP searches by CDEX [@cdex0; @cdex12014; @cdex12016; @cdex12018]. Located in the China Jinping Underground Laboratory (CJPL) [@cjpl], the first generation experiments used 1-kg-scale single-element $p$PCGe cooled by a cold finger since 2010 [@cdex12014; @cdex12016; @cdex12018]. With an energy threshold of 160 eVee (“eVee" represents electron equivalent energy derived from a charge calibration) and an exposure of 737.1 kg day, CDEX-1B provided improved limits on WIMP-nucleon spin-independent (SI) and spin-dependent (SD) scattering down to $m_{\chi}$ of 2 GeV/$c^2$ [@cdex12018].
Toward a future ton-scale DM experiment, the second generation CDEX experiment with a total detector mass of about 10 kg, called CDEX-10, has used three triple-element $p$PCGe strings (C10A, B, C) directly immersed in liquid nitrogen (LN$_{2}$). Compared with cold finger cooling and high-$Z$ material shielding systems, low-$Z$ material shielding, such as with LN$_{2}$ or liquid argon, provides better control of radiation background. The concept of integrated shielding and cooling, first proposed in the GENIUS project [@GENIUS], is realized in the GERDA experiment with the lowest background among neutrinoless double beta decay (0$\nu\beta\beta$) experiments [@gerda] and will be expanded into the next generation LEGEND 0$\nu\beta\beta$ program [@LEGEND]. CDEX-10 focuses on the arraying technologies and background understanding of the prototype $p$PCGe detectors developed based on the CDEX-1 technique. The new CDEX-10 array detectors and dedicated data acquisition (DAQ) system started testing and data taking inside a LN$_{2}$ tank in 2016 at CJPL. C10A was returned to the CANBERRA factory in France for upgrades. Of the remaining six detectors, two had faulty cabling, and two others had a high level of noise. In this Letter, we report the results from a first physics data set of one of the two operational detectors C10B-Ge1, which had the lower threshold. The stainless steel LN$_{2}$ tank was located in the polyethylene room with 1 m thick walls at CJPL-I for cooling of the CDEX-10 detectors, which are surrounded by 20 cm thick high-purity oxygen-free copper immersed in LN$_{2}$ to shield the ambient radioactivities. The shielding configuration of CDEX-10 and the structure of a detector string are shown in Fig. \[fig::detector\_config\].
![ Configuration of CDEX-10 experimental setup (left) and C10B detector layout inside the string (right). C10B and C10C are running inside the LN$_{2}$ tank which has an outer diameter of 1.5 m and a height of 1.9 m. Each detector string consists of three PCGe detectors tagged as Ge1 to Ge3 from bottom to top. The size of each germanium crystal is approximately [$\Phi$]{}62 mm $\times$ H62 mm. []{data-label="fig::detector_config"}](Figure1.pdf){width="0.9\linewidth"}
The DAQ system received signals from the $p^{+}$ point contact electrode of C10B-Ge1 which were fed into a pulsed reset preamplifier. Five identical output signals of the preamplifier were further processed and digitized. Two of them were distributed into 6 $\mu$s ($S_{p6}$) and 12 $\mu$s ($S_{p12}$) shaping amplifiers for a 0-12 keVee energy range. These two channels were used for energy calibration and signal and noise discrimination. The third channel was loaded to a timing amplifier ($T_{p}$) to measure the rise time of signals within a 0-12 keVee energy range which can be used for bulk or surface events discrimination. The remaining two were loaded to a shaping amplifier and a timing amplifier with low gains aiming at a high energy range for background understanding. To estimate the dead time of the DAQ system and cut efficiencies uncorrelated with energies, random trigger (RT) events were recorded once every 20 seconds. The output signals of the above amplifiers were digitized by the 14-bit 100-MHz flash analog-to-digital converters. Data taking with C10B-Ge1 was performed from February 26, 2017 to November 7, 2017. The DAQ dead time fraction was measured by RT events to be 4.8%, giving a live time of 112.3 days.
The data analysis follows the procedures described in our earlier work [@cdex12014; @cdex12016; @cdex12018], starting from the parameters extracted from the digitized pulses. The optimal integrated area of the pulse from $S_{p12}$ is selected to define the energy for its excellent energy linearity at the low energy region. Energy calibration was done with the internal cosmogenic x-ray peaks: 10.37 keVee of $^{68}$Ge and 8.98 keVee of $^{65}$Zn, and the zero energy defined by the RT events. Analysis procedures follow those with similar detectors in CDEX-1B [@cdex12018]. Basic filtering algorithms are first applied to the $S_{p6,12}$ and $T_p$ pedestals to reject events with anomalous electronic noise profiles. These cuts are energy independent, and the efficiency is measured to be 97.4% by the survival of RT events, giving rise to a valid data sample of 109.4 days.
The second step is a physics-noise event (PN) cut to discriminate the signals from electronic noises near the energy threshold. The PN cut is based on the relationship between the energy and maximum amplitude of $S_{p12}$. The experimental data of a $^{137}$Cs source are used to derive the PN cut and the trigger efficiencies. The efficiency curves with 1$\sigma$ bands are shown in the inset of Fig. \[fig::residualspec\](a).
Events depositing energy in the *n$^{+}$* surface layer generate a slow rising pulse and an incomplete charge collection due to the weak electric field and severe recombination of electron-hole pairs in this region [@LiHB_2014a]. Since C10B-Ge1 and CDEX-1B detectors have the same crystal mass, crystal structure, and fabrication procedure, the same dead layer thickness of 0.88 $\pm$ 0.12 mm [@deadlayer] is taken for this analysis . This gives rise to a fiducial mass of 939 g and accordingly a physics data size of 102.8 kg day.
The bulk and surface events (BS) cut is carried out to select bulk events. WIMP candidate events in the bulk of the detector are then separated from the surface events via the rise-time differences of the $T_{p}$ signals. The rise-times ($\tau$) are measured by fitting the $T_{p}$ pulse to a hyperbolic tangent function [@LiHB_2014a; @cdex12014; @cdex12016; @cdex12018]. The log$_{10}(\tau$) distribution versus measured energy of [*in situ*]{} events is depicted in Fig. \[fig::bsplot\](a), showing a two-band structure of bulk and surface events well separated above 1.5 keVee. However, at lower energies the bulk and surface events infiltrate into each other, as a result of the electronic noise smearing effect. Multisite events are located off band and of negligible fraction at the keVee-range energy [@soma2016].
![ (a) Scatter plot of the rise time \[log$_{10}(\tau$)\] versus deposited energy of background events. \[$b_{0}$, $b_{1}$\] and \[$s_{0}$, $s_{1}$\] are the “pure" regions we chose to derive the count rates. Extremely-fast and extremely-slow events are with log$_{10}(\tau)<b_0$ and $>s_1$, respectively”. Comparison of the rise-time distribution of various sources and background at typical energies of 0.16-0.66 keVee (b),(c) and 1.66-2.16 keVee (d),(e) with the normalization related to the “pure" bulk and surface regions (yellow shadow), respectively. []{data-label="fig::bsplot"}](Figure2a.pdf "fig:"){width="0.95\linewidth"} ![ (a) Scatter plot of the rise time \[log$_{10}(\tau$)\] versus deposited energy of background events. \[$b_{0}$, $b_{1}$\] and \[$s_{0}$, $s_{1}$\] are the “pure" regions we chose to derive the count rates. Extremely-fast and extremely-slow events are with log$_{10}(\tau)<b_0$ and $>s_1$, respectively”. Comparison of the rise-time distribution of various sources and background at typical energies of 0.16-0.66 keVee (b),(c) and 1.66-2.16 keVee (d),(e) with the normalization related to the “pure" bulk and surface regions (yellow shadow), respectively. []{data-label="fig::bsplot"}](Figure2b-e.pdf "fig:"){width="0.92\linewidth"}
It has been shown that the background and calibration sources data share the common bulk or surface rise-time distribution probability density function (PDF) [@ratiomethod]. The ratio method has been developed accordingly to address the BS discrimination problem in $p$PCGe [@ratiomethod; @cdex12018]. In this analysis, the inputs of the ratio method include the background data and three calibration samples ($^{137}$Cs, $^{60}$Co, $^{109}$Cd), while $^{109}$Cd is a pure surface source. Considering that the low-energy gammas from the $^{109}$Cd source can hardly penetrate the *n*$^{+}$ surface layer, their rise-time distribution can describe the surface PDF. Four boundary parameters related to the approximately “pure" bulk and surface regions are depicted in Fig. \[fig::bsplot\](a). Two outside boundaries \[log$_{10}(\tau)$ = $b_0$ and log$_{10}(\tau)$ = $s_{1}$\] are derived by fitting the best normalization interval of each energy bin of 500 eVee from 160 eVee on, based on the selection principles of making the statistics as significant as possible while the rise-time distributions of those events remain as consistent as possible. As depicted in Figs. \[fig::bsplot\](b) and \[fig::bsplot\](c) and Figs. \[fig::bsplot\](d) and \[fig::bsplot\](e), the comparisons of the rise-time distributions of those samples at 0.16-0.66 keVee and 1.66-2.16 keVee demonstrate that they share common rise-time distribution PDFs when normalized to the “pure" bulk and surface regions.
[lcc]{} Energy bin & 0.16-0.26 keVee & 1.96-2.06 keVee\
I) Statistic errors & 1.14 & 0.50\
II) Systematic errors\
(i) Choice of \[$b_{0}$, $b_{1}$\] & &\
and \[$s_{0}$, $s_{1}$\]\
(ii) Choice of sources & 0.09 & 0.05\
(iii) $\tau$ rebin size & 0.63 & 0.06\
(iv) shift of $\tau$ & 0.06 & 0.01\
Combined & 1.37 & 0.13\
-----------------------------------
$B_{r}$ and Errors
(kg$^{-1}$keVee$^{-1}$day$^{-1}$)
-----------------------------------
: \[sys. err\] Main contribution to errors of the $B_{r}$ at the threshold bin and a typical high energy bin.
&
--------------------------
2.47 $\pm$ 1.14\[stat.\]
$\pm$ 1.37\[sys.\]
=2.47 $\pm$ 1.78
--------------------------
: \[sys. err\] Main contribution to errors of the $B_{r}$ at the threshold bin and a typical high energy bin.
&
--------------------------
2.15 $\pm$ 0.50\[stat.\]
$\pm$ 0.13\[sys.\]
=2.15 $\pm$ 0.52
--------------------------
: \[sys. err\] Main contribution to errors of the $B_{r}$ at the threshold bin and a typical high energy bin.
\
![ (a) Spectra after different event-selection cuts. The trigger efficiency and PN cut efficiency curves derived from $^{137}$Cs source events and fitted by an error function with a 1$\sigma$ band (yellow shadow) are shown in the inset. (b) $L$-$X$ and $M$-$X$ lines predicted by the $K$-$X$ intensities [@KLratio]. (c) Residual spectrum with the $L$-$X$ and $M$-$X$ contributions subtracted, together with the best-fit spectrum at $m_{\chi}$ = 5 GeV/${c}^2$ (red line), with an uncertainty band (yellow shadow) at the 90% confidence level. An excluded case at $m_{\chi}$ = 5 GeV/${c}^2$, $\sigma_{\chi N}^{SI} = 10^{-41} \rm{cm}^2$ is superimposed as a black dashed line for illustration. []{data-label="fig::residualspec"}](Figure3a.pdf "fig:"){width="0.95\linewidth"} ![ (a) Spectra after different event-selection cuts. The trigger efficiency and PN cut efficiency curves derived from $^{137}$Cs source events and fitted by an error function with a 1$\sigma$ band (yellow shadow) are shown in the inset. (b) $L$-$X$ and $M$-$X$ lines predicted by the $K$-$X$ intensities [@KLratio]. (c) Residual spectrum with the $L$-$X$ and $M$-$X$ contributions subtracted, together with the best-fit spectrum at $m_{\chi}$ = 5 GeV/${c}^2$ (red line), with an uncertainty band (yellow shadow) at the 90% confidence level. An excluded case at $m_{\chi}$ = 5 GeV/${c}^2$, $\sigma_{\chi N}^{SI} = 10^{-41} \rm{cm}^2$ is superimposed as a black dashed line for illustration. []{data-label="fig::residualspec"}](Figure3b.pdf "fig:"){width="0.45\linewidth"} ![ (a) Spectra after different event-selection cuts. The trigger efficiency and PN cut efficiency curves derived from $^{137}$Cs source events and fitted by an error function with a 1$\sigma$ band (yellow shadow) are shown in the inset. (b) $L$-$X$ and $M$-$X$ lines predicted by the $K$-$X$ intensities [@KLratio]. (c) Residual spectrum with the $L$-$X$ and $M$-$X$ contributions subtracted, together with the best-fit spectrum at $m_{\chi}$ = 5 GeV/${c}^2$ (red line), with an uncertainty band (yellow shadow) at the 90% confidence level. An excluded case at $m_{\chi}$ = 5 GeV/${c}^2$, $\sigma_{\chi N}^{SI} = 10^{-41} \rm{cm}^2$ is superimposed as a black dashed line for illustration. []{data-label="fig::residualspec"}](Figure3c.pdf "fig:"){width="0.45\linewidth"}
{width="0.5\linewidth"}{width="0.5\linewidth"}
There are extremely-fast events (EFEs) with a faster rise time in the bulk band due to better rise-time resolution in C10B-Ge1 than CDEX-1A and CDEX-1B [@cdex12016; @cdex12018]. It has been verified with simulations using realistic detector electric field that these EFEs mainly originate from the region in the vicinity of the $p^+$ point electrode. An additional convincing evidence is that x rays from Cu are observed only in the EFEs band at 8 keV energy; they can only enter the active area through the passivated surface layer near $p^+$ point. Unfortunately, EFEs can only be distinguished clearly from the bulk band above sub-keVee, while the differentiation is not possible at a low energy region due to the smearing from electronic noise. A cut \[log$_{10}$($\tau$) $<$ $b_0$\] was used to remove the EFEs, together with an extremely-slow events cut \[log$_{10}(\tau) > s_1$\] [@ratiomethod] to remove those events which are seriously attenuated by the $n^+$ surface layer. Both kinds of events are included to bulk and surface counts after the $B$ and $S$ correction procedures [@ratiomethod].
The corrected bulk or surface counts ($B_r$/$S_r$) can be derived by integrating the optimized PDFs which are derived by numerically minimizing the $\chi^2$ of Eq. (7) in Ref. [@ratiomethod]. The reconstructed $^{137}$Cs and $^{60}$Co spectra are consistent with GEANT4 [@Geant4] simulations. The $B_r$ of the background from C10B-Ge1 with the main contributions of errors at the first bin of 0.16-0.26 keVee and a typical high energy of 1.96-2.06 keVee are shown in Table 1. The systematic errors mainly come from the choices of $b_0$, $b_1$, $s_0$, and $s_1$, the errors of which are estimated by varying the more “pure" bulk and surface regions of Fig. \[fig::bsplot\](a). Further details of the $BS$ analysis and uncertainties derivations can be found in Ref. [@ratiomethod].
The spectra after different event-selection cuts are shown in Fig. \[fig::residualspec\](a). The physics analysis threshold is 160 eVee. From the spectra in Fig. \[fig::residualspec\](a), characteristic $K$-shell x ray peaks from internal cosmogenic radionuclides like $^{68,71}$Ge, $^{68}$Ga, $^{65}$Zn, $^{57}$Co, $^{55}$Fe, $^{54}$Mn and $^{49}$V can be identified. In addition, x-ray peaks from Cu and Zn isotopes excited by high energy $\gamma$ rays are observed in the extremely-fast events region of the background spectrum. Their intensities are derived by best fit from the spectrum [@cdex12014; @cdex12016; @cdex12018]. The contributions of $L$- or $M$-shell x-ray peaks are derived from corresponding $K$-shell peaks and subtracted from the $B_{r}$ spectrum, shown in Fig. \[fig::residualspec\](b) [@KLratio]. A minimum-$\chi^2$ analysis [@cdex12014] is applied to the residual spectrum, using two free parameters characterizing the flat background and the possible $\chi$-$N$ SI cross-section ($\csnospin$). The best-fit spectrum at $m_{\chi}$ = 5 GeV/${c}^2$ where $\csnospin = (-0.61 \pm 4.3)\times10^{-42}$ cm$^2$ at $\chi^2/\rm{DOF} = 11.2/22$ ($p$ value = 0.97), is also depicted in Fig. \[fig::residualspec\](c). A standard WIMP galactic halo assumption and conventional astrophysical models [@astropara] are used to describe WIMP-induced interactions, with the local WIMP density of 0.3 GeV/cm$^{3}$, the Maxwellian velocity distribution of $v_0 = 220$ km/s, and the escape velocity of $v_{\rm esc}$ = 544 km/s. The quenching factor in Ge is calculated by the TRIM software package [@trim1; @trim2; @soma2016; @trim3] with a 10% systematic error adopted for the analysis [@cdex12016].
Upper limits are derived following standard procedures [@stat_method; @cdex12016]. The exclusion plots of SI and SD at a 90% confidence level (C.L.) are depicted in Figs. \[fig::exclusionplot\](a) and \[fig::exclusionplot\](b), respectively, with several selected benchmark direct search experiments superimposed [@cdex12016; @cdex12018; @supercdms; @cdmslite; @CRESST-II; @DAMA; @cdmssi; @lux; @lux2016; @pandax; @cdmslite2018]. The most stringent accelerator bounds on SI from the LHC experiments [@cms2015; @cms3; @atlas1; @atlas2; @atlas3] are more constraining in SI $-$ with $\csnospin < 10^{-48} ~ {\rm cm^2}$ for $m_{\chi} \sim 5 ~ {\rm GeV}$ $-$ than the scale displayed in Fig. \[fig::exclusionplot\](a). They are, however, extremely sensitive to particle physics models and the choice of parameters. The LHC results are derived with $\chi$-proton cross sections and hence unrelated to the SD constraints on $\chi$-neutron cross sections. This study achieves the lowest threshold and background among the various CDEX data set to date, which brings forth almost an order of magnitude improvement over our previous bounds [@cdex12016; @cdex12018]. New regions on SI for $m_{\chi}$ at 4-5 GeV/${c}^{2}$ are probed and excluded. The CDEX-10 detector array will be installed in a new large LN$_2$ cryotank with a volume of about 1700 m$^3$ at Hall-C of CJPL-II [@cjpl] by the end of 2018, where shielding from ambient radioactivity is provided by the 6 m-thick LN$_2$. The projected parameter space to be probed with a reduced background comparable to the best achieved in germanium detectors [@MJD] is also shown in Fig. \[fig::exclusionplot\].
This work was supported by the National Key Research and Development Program of China (Grant No. 2017YFA0402201) and the National Natural Science Foundation of China (Grants No.11475092, No. 11475099, No. 11675088, No. 11725522).
H. J. and L.P. J. contributed equally to this work.
*Note added.* $-$ We are aware of stronger light WIMPs constraints on $\csnospin$ reported in a preprint by the DarkSide-50 experiment [@darkside].
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In this article left invariant measures and functionals on locally compact nonassociative fan loops are investigated. For this purpose necessary properties of topological fan loops, estimates and approximations of functions on them are studied. An existence of nontrivial left invariant measures on locally compact fan loops is proved. Abundant families of fan loops are provided with the help of different types of their products. [^1]'
author:
- 'S.V. Ludkowski'
date: 10 December 2018
title: 'Left invariant measures on locally compact fan loops.'
---
Address: Dep. Appl. Mathematics, Moscow State Techn. Univ. MIREA,
av. Vernadsky 78, Moscow 11944, Russia; e-mail: [email protected]
Introduction.
=============
Left invariant measures or Haar measures on locally compact groups play very important role in measure theory, harmonic analysis, representation theory, geometry, mathematical physics, etc. (see, for example, [@boloktodb; @fell; @hew] and references therein). On the other hand, in nonassociative algebra, in noncommutative geometry, field theory, topological algebra there frequently appear binary systems which are nonassociative generalizations of groups and related with loops, quasi-groups, Moufang loops, IP-loops, etc. (see [@bruckb; @kakkar; @kiechlb; @pickert; @razm; @smithb; @vojtech] and references therein). An arbitrary IP-loop $Y$ is a loop with a restriction: for each $x\in Y$ there exist elements $x_1$ and $x_2$ in $Y$ such that for each $y$ in $Y$ the identities are satisfied $x_1(xy)=y$ and $(yx)x_2=y$, where $x_1$ and $x_2$ are also denoted by $\mbox{}^{-1}x$ and $x^{-1}$ and called left and right inverses of $x$ respectively. It was investigated and proved in the 20-th century that a nontrivial geometry exists if and only if there exists a corresponding loop.
Very important role in mathematics and quantum field theory play octonions and generalized Cayley-Dickson algebras [@albmajja99; @allcja98; @baez; @dickson]. A multiplicative law of their canonical bases is nonassociative and leads to a more general notion of a metagroup instead of a group [@ludlmla18]. They are used not only in algebra and geometry, but also in noncommutative analysis and PDEs, particle physics, mathematical physics (see [@baez; @castdoyfior; @dickson; @frenludkfejms18]-[@guetze; @kansol]-[@ludkcvee13] and references therein). The preposition “meta” is used to emphasize that such an algebraic object has properties milder than a group. By their axiomatic metagroups are loops with weak relations. They were used in [@ludlmla18] for investigations of automorphisms and derivations of nonassociative algebras.
In this article more general binary systems such as fan loops are studied (see Definition 2.1). They also are more general than IP-loops, because in fan loops $G$ left and right inverses $\mbox{}^{-1}x$ and $x^{-1}$ of nonunit elements $x$ in $G$ may not exist.
This article is devoted to left invariant measures (see Definition 3.18) on locally compact fan loops. Necessary preliminary results about fan loops are given in Section 2. Specific algebraic and topological features of fan loops are studied in Lemmas 2.2-2.6, 2.12 and Propositions 2.9, 2.11. A quotient of a fan loop by its fan is investigated in Theorem 2.8. A uniform continuity of maps on topological fan loops is studied in Theorem 2.14 and Corollary 2.15.
Left invariant functionals and measures are investigated in Section 3. These properties are more complicated than for groups and IP-loops, because of the nonassociativity of fan loops and absence of left and right inverses in general. Main results are theorems 3.15, 3.16, 3.19, 3.20. For their proofs estimates of nonnegative functions with compact supports in fan loops are investigated in Lemmas 3.2, 3.4, 3.6. Functionals on a space of nonnegative functions with compact supports in a fan loop are studied in Lemmas 3.7, 3.8, 3.10, 3.13 and Theorem 3.9. In Theorem 3.11 approximations of nonnegative functions with compact supports in the fan loop are described.
In an appendix abundant families of fan loops are provided with the help of a direct product and smashing products (see Remark 4.3 and Definition 4.5). For this purpose Theorems 4.1 and 4.4 are proved.
All main results of this paper are obtained for the first time. They can be used in harmonic analysis on nonassociative algebras and metagroups and loops, representation theory, geometry, mathematical physics, quantum field theory, particle physics, PDEs, etc.
Fan loops.
==========
To avoid misunderstandings we give necessary definitions. For short it will be written fan loop instead of nonassociative fan loop.
[**2.1. Definition.**]{} Let $G$ be a set with a multiplication (that is a single-valued binary operation) $G^2\ni (a,b)\mapsto ab
\in G$ defined on $G$ satisfying the conditions:
$(2.1.1)$ for each $a$ and $b$ in $G$ there is a unique $x\in
G$ with $ax=b$ and
$(2.1.2)$ a unique $y\in G$ exists satisfying $ya=b$, which are denoted by $x=a\setminus b=Div_l(a,b)$ and $y=b/a=Div_r(a,b)$ correspondingly,
$(2.1.3)$ there exists a neutral (i.e. unit) element $e_G=e\in G$: $~eg=ge=g$ for each $g\in G$.
We consider subsets in $G$:
$(2.1.4)$ $Com (G) := \{ a\in G: ~ \forall b\in G, ~ ab=ba \} $;
$(2.1.5)$ $N_l(G) := \{a\in G: ~ \forall b\in G, ~ \forall c\in G, ~ (ab)c=a(bc) \}
$;
$(2.1.6)$ $N_m(G) := \{a\in G: ~ \forall b\in G, ~ \forall c\in G, ~ (ba)c=b(ac)
\} $;
$(2.1.7)$ $N_r(G) := \{a\in G: ~ \forall b\in G, ~ \forall c\in G, ~ (bc)a=b(ca)
\} $;
$(2.1.8)$ $N(G) := N_l(G)\cap N_m(G)\cap N_r(G)$;
$Z(G) := Com (G)\cap N(G)$.
Then $N(G)$ is called a nucleus of $G$ and $Z(G)$ is called the center of $G$.
We call $G$ a fan loop if a set $G$ possesses a multiplication and satisfies conditions $(2.1.1)$-$(2.1.3)$ and
$(2.1.9)$ $(ab)c=t(a,b,c)a(bc)$ and $(ab)c=a(bc)p(a,b,c)$\
for each $a$, $b$ and $c$ in $G$, where
$t(a,b,c)=t_G(a,b,c)\in N(G)$ and $p(a,b,c)=p_G(a,b,c)\in N(G)$.
Then $G$ will be called a central fan loop if in addition to $(2.1.9)$ it satisfies the condition:
$(2.1.10)$ $ab={\sf t}_2(a,b)ba$\
for each $a$ and $b$ in $G$, where ${\sf t}_2(a,b)\in Z(G)$.
Let $\tau $ be a topology on $G$ such that the multiplication $G\times G\ni (a,b)\mapsto ab\in G$ and the mappings $Div_l(a,b)$ and $Div_r(a,b)$ are jointly continuous relative to $\tau $, then $(G, \tau )$ will be called a topological fan loop. Henceforth it will be assumed that $\tau $ is the $T_1\cap
T_{3.5}$ topology, if something other will not be specified.
A minimal closed subgroup $N_0(G)$ in the topological fan loop $G$ containing $t(a,b,c)$ and $p(a,b,c)$ for each $a$, $b$ and $c$ in $G$ will be called a fan of $G$.
Elements of the fan loop $G$ will be denoted by small letters, subsets of $G$ will be denoted by capital letters. If $A$ and $B$ are subsets in $G$, then $A-B$ means the difference of them $A-B=\{
a\in A: ~a \notin B \} $. Henceforward, maps and functions on fan loops are supposed to be single-valued if something other will not be specified.
[**2.2. Lemma.**]{}
*If $G$ is a fan loop, then for each $a$, $b$ and $c$ in $G$ the following identities are fulfilled:*
$(2.2.1)$ $ ~ b\setminus e=t(e/b,b,b\setminus e)(e/b)$;
$(2.2.1')$ $b\setminus e=(e/b)p(e/b,b,b\setminus e)$;
$(2.2.2)$ $(a\setminus e)b=t(e/a,a,a\setminus e)[t(e/a,a,a\setminus b)]^{-1}(a\setminus b)$;
$(a\setminus b)= (a\setminus e)bp(a,a\setminus e,b)$;
$(2.2.2')$ $(bc)\setminus a=(c\setminus (b\setminus
a))[p(b,c,(bc)\setminus a)]^{-1}$;
$(2.2.2'')$ $(a\setminus b)c=(a\setminus (bc))[p(a,a\setminus b,c)]^{-1}$;
$(2.2.2''')$ $(ab)\setminus e = (b\setminus e)(a\setminus
e)[t(a,b,b\setminus e)]^{-1}t(ab,b\setminus e,a\setminus e)$;
$(2.2.3)$ $b(e/a)=(b/a)p(b/a,a,a\setminus e)[p(e/a,a,a\setminus e)]^{-1} $;
$(b/a)= [t(b,e/a,a)]^{-1}b(e/a)$;
$(2.2.3')$ $a/(bc)=t(a/(bc),b,c)((a/c)/b)$;
$(2.2.3'')$ $c(b/a)=t(c,b/a,a)(cb)/a$;
$(2.2.3''')$ $e/(ab)=[p(e/b,e/a,ab)]^{-1}p(e/a,a,b)(e/b)(e/a)$.
[**Proof.**]{} Note that $N(G)$ is a subgroup in $G$ due to Conditions $(2.1.5)$-$(2.1.8)$ (see also [@bruckb]). Then Conditions $(2.1.1)$-$(2.1.3)$ imply that
$(2.2.4)$ $b(b\setminus a)=a$, $~b\setminus (ba)=a$;
$(2.2.5)$ $(a/b)b=a$, $~(ab)/b=a$\
for each $a$ and $b$ in any loop $G$ (see also [@bruckb; @smithb]). Using Condition $(2.1.9)$ and Identities $(2.2.4)$ and $(2.2.5)$ we deduce that
$e/b=(e/b)(b(b\setminus e)) = [t(e/b,b,b\setminus e)]^{-1}(b\setminus e)
$\
which leads to $(2.2.1)$.
Let $c=a\setminus b$, then from Identities $(2.2.1)$ and $(2.2.4)$ it follows that
$(a\setminus e)b=t(e/a,a,a\setminus
e)(e/a)(ac)$
$=t(e/a,a,a\setminus e)[t(e/a,a,a\setminus b)]^{-1}
((e/a)a)(a\setminus b)$\
which taking into account $(2.2.5)$ provides $(2.2.2)$.
On the other hand, $b\setminus e=((e/b)b)(b\setminus e) =
(e/b)(b(b\setminus e))p(e/b,b,b\setminus e)$ that gives $(2.2.1')$.
Let now $d=b/a$, then Identities $(2.2.1')$ and $(2.2.5)$ imply that
$b(e/a)=(da)(a\setminus e)[p(e/a,a,a\setminus
e)]^{-1}$
$=(b/a)p(b/a,a,a\setminus e)[p(e/a,a,a\setminus e)]^{-1} $\
which demonstrates $(2.2.3)$.
Next we infer from $(2.1.9)$ and $(2.2.4)$ that
$b(c((bc)\setminus a))=(bc)((bc)\setminus a)[p(b,c,(bc)\setminus
a)]^{-1}= a[p(b,c,(bc)\setminus a)]^{-1}$, hence $c((bc)\setminus
a)=(b\setminus a)[p(b,c,(bc)\setminus a)]^{-1}$ that implies $(2.2.2')$.
Symmetrically it is deduced that $(a/(bc))b)c=t(a/(bc),b,c)a$, consequently, $(a/(bc))b=t(a/(bc),b,c)(a/c)$. From the latter identity it follows $(2.2.3')$.
Evidently, formulas
$a((a\setminus
b)c)=(a(a\setminus b))c[p(a,a\setminus b,c)]^{-1}=bc[p(a,a\setminus
b,c)]^{-1}$ and
$(c(b/a))a=t(c,b/a,a)cb$\
imply $(2.2.2'')$ and $(2.2.3'')$ correspondingly.
From $(2.1.9)$ we infer that
$(ab)((b\setminus
e)(a\setminus e))=[t(ab,b\setminus e,a\setminus
e)]^{-1}t(a,b,b\setminus e)$, since by $(2.2.4)$
$(a(b(b\setminus e)))(a\setminus e)=e$.\
This together with $(2.1.1)$ and $(2.1.2)$ implies $(2.2.2''')$.
Analogously form $(2.1.9)$ we deduce that
$((e/b)(e/a))(ab)=[p(e/a,a,b)]^{-1}p(e/b,e/a,ab)$, since by $(2.2.5)$
$(e/b)(((e/a)a)b)=e$.\
Finally applying $(2.1.1)$ and $(2.1.2)$ we get Identity $(2.2.3''')$.
[**2.3. Lemma.**]{}
*Assume that $G$ is a fan loop. Then for every $a$, $a_1$, $a_2$, $a_3$ in $G$ and $z_1$, $z_2$, $z_3$ in $Z(G)$, $b\in N(G)$:*
$(2.3.1)$ $t(z_1a_1,z_2a_2,z_3a_3)=t(a_1,a_2,a_3)$;
$(2.3.1')$ $p(z_1a_1,z_2a_2,z_3a_3)=p(a_1,a_2,a_3)$;
$(2.3.2)$ $t(a,a\setminus e,a)a=ap(a,a\setminus e,a)$;
$(2.3.2')$ $t(a,e/a,a)a=ap(a,e/a,a)$;
$(2.3.2'')$ $p(a,a\setminus e,a)t(e/a,a,a\setminus e)=e$;
$(2.3.3)$ $t(a_1,a_2,a_3b)=t(a_1,a_2,a_3)$;
$(2.3.3')$ $p(ba_1,a_2,a_3)=p(a_1,a_2,a_3)$;
$(2.3.4)$ $t(ba_1,a_2,a_3)=bt(a_1,a_2,a_3)b^{-1}$;
$(2.3.4')$ $p(a_1,a_2,a_3b)=b^{-1}p(a_1,a_2,a_3)b$.
[**Proof.**]{} Since $(a_1a_2)a_3=t(a_1,a_2,a_3)a_1(a_2a_3)$ and $t(a_1,a_2,a_3)\in N(G)$ for every $a_1$, $a_2$, $a_3$ in $G$, then
$(2.3.5)$ $t(a_1,a_2,a_3)=((a_1a_2)a_3)/(a_1(a_2a_3))$.
Therefore, for every $a_1$, $a_2$, $a_3$ in $G$ and $z_1$, $z_2$, $z_3$ in $Z(G)$ we infer that
$t(z_1a_1,z_2a_2,z_3a_3)=
(((z_1a_1)(z_2a_2))(z_3a_3))/((z_1a_1)((z_2a_2)(z_3a_3)))$
$=
((z_1z_2z_3)((a_1a_2)a_3))/((z_1z_2z_3)(a_1(a_2a_3)))=
((a_1a_2)a_3)/(a_1(a_2a_3))$, since
$(2.3.6)$ $b/(qa)=q^{-1}b/a$ and $b/q=q\setminus b=bq^{-1}$\
for each $q\in
Z(G)$, $ ~ a$ and $b$ in $G$, because $Z(G)$ is the commutative group satisfying Conditions $(2.1.4)$ and $(2.1.8)$. Thus $t(z_1a_1,z_2a_2,z_3a_3)=t(a_1,a_2,a_3)$.
Symmetrically we get
$(2.3.7)$ $p(a_1,a_2,a_3)=(a_1(a_2a_3))\setminus ((a_1a_2)a_3)$ and
$p(z_1a_1,z_2a_2,z_3a_3)=
((z_1a_1)((z_2a_2)(z_3a_3)))\setminus
(((z_1a_1)(z_2a_2))(z_3a_3))$
$=((z_1z_2z_3)(a_1(a_2a_3)))\setminus ((z_1z_2z_3)((a_1a_2)a_3))=
(a_1(a_2a_3))\setminus ((a_1a_2)a_3)$\
that provides $(2.3.1')$.
From Formulas $(2.3.5)$ and $(2.2.1)$ it follows that
$t(a,a\setminus
e,a)=((a(a\setminus e))a)/(a((a\setminus
e)a))=a/[at(e/a,a,a\setminus e)]$, consequently,
$(2.3.8)$ $t(a,a\setminus e,a)at(e/a,a,a\setminus e)=a$.\
Then from Formulas $(2.3.7)$, $(2.2.4)$ and Condition $(2.1.9)$ we deduce that
$p(a,a\setminus e,a)=(a((a\setminus e)a))\setminus
((a(a\setminus e))a)= \{ [t(a,a\setminus e,a)]^{-1}a \} \setminus
a$,\
which implies $(2.3.2)$. Identities $(2.3.2)$ and $(2.3.8)$ lead to $(2.3.2'')$. Next using $(2.3.7)$ and $(2.1.9)$ we deduce that
$p(a,e/a,a)=[a((e/a)a)]\setminus [(a(e/a))a]=a\setminus [t(a,e/a,a)a
]$\
that implies $(2.3.2')$. From $(2.1.9)$ we get that
$((a_1a_2)a_3)b=(a_1a_2)(a_3b)=(t(a_1,a_2,a_3b)a_1(a_2a_3))b$,\
from which and $(2.2.5)$ and $(2.3.5)$ Identity $(2.3.3)$ follows, because $b\in N(G)$. Then
$b((a_1a_2)a_3)=((ba_1)a_2)a_3=b(a_1(a_2a_3)p(ba_1,a_2,a_3))$\
and $(2.2.4)$ and $(2.3.7)$ imply Identity $(2.3.3')$. Symmetrically we deduce
$b((a_1a_2)a_3)=t(ba_1,a_2,a_3))b(a_1(a_2a_3))$ and
$((a_1a_2)a_3)b=(a_1(a_2a_3))bp(a_1,a_2,a_3b)$\
that together with $(2.3.5)$ and $(2.3.7)$ imply Identities $(2.3.4)$ and $(2.3.4')$.
[**2.4. Lemma.**]{} [*If $(G, \tau )$ is a topological loop, then the functions $t(a_1,a_2,a_3)$ and $p(a_1,a_2,a_3)$ are jointly continuous in $a_1$, $a_2$, $a_3$ in $G$.*]{}
[**Proof.**]{} This follows immediately from Formulas $(2.3.5)$, $(2.3.7)$ and Definition 2.1.
[**2.5. Lemma.**]{} [*Assume that $(G,\tau )$ is a topological loop and $U$ is an open subsets in $G$, then for each $b\in G$ sets $Ub$ and $bU$ are open in $G$.*]{}
[**Proof.**]{} Take any $c\in Ub$ and consider the equation
$(2.5.1)$ $xb=c$.
Then from $(2.1.2)$ it follows that
$(2.5.2)$ $x=c/b$.
Thus $x=\psi _b(c)$, where $\psi _b(c)=c/b$ is a continuous bijective function in the variable $c$ due to Identity $(2.2.3)$ and Lemma 2.4. On the other hand, the right shift mapping
$(2.5.3)$ $R_bu:=ub$\
from $G$ into $G$ is continuous and bijective in $u$ (see Definition 1). Moreover, $\psi _b(R_bu)=u$ and $R_b(\psi _b(c))=c$ for each fixed $b\in G$ and all $u\in G$ and $c\in G$ by Identities $(2.2.5)$. Thus $R_b$ and $\psi _b$ are open mappings, consequently, $Ub$ is open in $G$.
Similarly for the equation
$(2.5.4)$ $by=c$ the unique solution is
$(2.5.5)$ $y=b\setminus c$ by Condition $(2.1.1)$.
Therefore, $y=\theta _b(c)$, where $\theta _b(c)=b\setminus c$ is a continuous bijective function in $c$ according to Lemma 2.4 and Formula $(2.2.2)$. Next we consider the left shift mapping
$(2.5.6)$ $L_bu=bu$\
for each fixed $b\in G$ and any $u\in G$. This mapping $L_b$ is continuous, since the multiplication on $G$ is continuous. Then $L_b(\theta _b(c))=c$ and $\theta _b(L_bu)=u$ for every fixed $b\in G$ and all $u\in G$ and $c\in G$ by Identities $(2.2.4)$. Therefore $\theta _b$ and $L_b$ are open mappings. Thus the subset $bU$ is open in $G$.
[**2.6. Lemma.**]{}
*Let $(G,\tau )$ be a topological loop.*
$(i)$. Let also $U$ and $V$ be subsets in $G$ such that either $U$ or $V$ is open, then $UV$ is open in $G$.
$(ii)$. If $A$ and $B$ are compact subsets in $G$, then $AB$ is compact.
$(iii)$. For each open neighborhood $U$ of $e$ in $G$ there exists an open neighborhood $V$ of $e$ such that
$(2.6.1)$ $\check{V}\subseteq U$, where
$(2.6.2)$ $\check{V} =V\cup Inv_l(V)\cup Inv_r(V)$,\
where $Inv_l(a)=Div_l(a,e)$, $~Inv_r(a)=Div_r(a,e)$ for each $a\in G$,
$(2.6.3)$ $DQ=\{ x=ab: ~ a\in D, ~ b \in Q \} $,
$(2.6.4)$ $Inv_l(D) = \{ x=a\setminus e: ~ a \in D \} $,
$(2.6.5)$ $Inv_r(D) = \{ x=e/a: ~ a \in D \} $ for any subsets $D$ and $Q$ in $G$.
[**Proof.**]{} $(i)$. In view of Lemma 2.5 $Ub$ and $aV$ are open in $G$ for each $a\in U$ and $b\in V$, consequently, $UV=\{ x: ~ x=uv, ~ u\in U, ~ v\in V
\}=\bigcup_{b\in V}Ub= \bigcup_{a\in U}aV$ is open in $G$.
$(ii)$. A subset $AB = \{ c: ~ c=ab, ~ a\in A, ~ b\in B \} $ is a continuous image of a compact subset $A\times B$ in $G\times G$, where $G\times G$ is supplied with the product (i.e. Tychonoff) topology, consequently, $AB$ is a compact subset in $G$ (see Theorem 3.1.10 and the Tychonoff Theorem 3.2.4 in [@eng]).
$(iii)$. The mappings $Inv_l$ and $Inv_r$ are homeomorphisms of $G$ onto itself as the topological space, since they are bijective, continuous and
$(2.6.6)$ $Inv_l(Inv_r(b))=b$ and $Inv_r(Inv_l(b))=b$\
for each $b$ in $G$ by $(2.2.4)$, $(2.2.5)$. Therefore for each open neighborhood $U$ of $e$ there exists an open neighborhood of $e$ of the form
$(2.6.7)$ $V:=\hat{U}$, where $\hat{U}:=U\cap
Inv_l(U)\cap Inv_r(U)$.
From $(2.6.6)$ we infer that $Inv_r(Inv_l(U))=U$ and $Inv_l(Inv_r(U))=U$, hence $Inv_l(V)\subseteq U\cap Inv_l(U)\cap
Inv_l(Inv_l(U))\subseteq U\cap Inv_l(U)$ and $Inv_r(V)\subset U\cap
Inv_r(U)$, consequently, $V\cup Inv_l(V)\cup Inv_r(V)\subseteq U$.
[**2.7. Definition.**]{} A subloop $H$ of a loop $G$ is called normal if it satisfies
$(2.7.1)$ $xH=Hx$ and
$(2.7.2)$ $(xy)H=x(yH)$ and $(xH)y=x(Hy)$ and $H(xy)=(Hx)y$\
for each $x$ and $y$ in $G$.
A family of cosets $\{ bH: ~ b\in G \} $ will be denoted by $G/\cdot
/N_0$.
[**2.8. Theorem.**]{} [*If $G$ is a $T_1$ topological fan loop, then its fan $N_0$ is a normal subgroup and its quotient $G/\cdot
/N_0$ is a $T_1\cap T_{3.5}$ topological group.*]{}
[**Proof.**]{} Let $\tau $ be a $T_1$ topology on $G$ relative to which $G$ is a topological loop. Then each point $x$ in $G$ is closed, since $G$ is the $T_1$ topological space (see Section 1.5 in [@eng]). From the joint continuity of the multiplication and the mappings $Div_l$ and $Div_r$ it follows that the nucleus $N=N(G)$ is closed in $G$. Therefore the subgroup $N_0$ is the closure of a subgroup $N_{0,0}(G)$ in $N$ generated by elements $t(a,b,c)$ and $p(a,b,c)$ for all $a$, $b$ and $c$ in $G$ (see Definition 2.1). According to $(2.1.5)$-$(2.1.8)$ one gets that $N$ and hence $N_0$ are subgroups in $G$ satisfying Conditions $(2.7.2)$, because $N_0\subseteq N$ (see also [@bruckb; @smithb]).
Let $a$ and $b$ belong to $N$ and $x\in G$. Then $x(x\setminus (ab))=ab$ and\
$x((x\setminus
a)b)=(x(x\setminus a))b=ab$, consequently,
$(2.8.1)$ $x\setminus (ab)=(x\setminus a)b$ for each $a$ and $b$ in $N(G)$, $x\in G$.\
Similarly it is deduced
$(2.8.2)$ $(ab)/x=a(b/x)$ for each $a$ and $b$ in $N(G)$, $x\in G$.
Therefore from $(2.1.9)$ and $(2.2.4)$ and $(2.8.1)$ it follows that
$((x\setminus a)x)((x\setminus b)x)=(x\setminus
a)(x((x\setminus b)x))p(x\setminus a,x,(x\setminus b)x)$
$=(x\setminus (ab))x[p(x,x\setminus b,x)]^{-1}p(x\setminus
a,x,(x\setminus b)x)$,\
since $(x\setminus a)(bx)=((x\setminus a)b)x=(x\setminus (ab))x$. Thus
$(2.8.3)$ $(x\setminus (ab))x= ((x\setminus
a)x)((x\setminus b)x)[p(x\setminus a,x,(x\setminus b)x)]^{-1}
p(x,x\setminus b,x)$ for each $a$ and $b$ in $N(G)$, $x\in G$.
From Identities $(2.2.2')$ and $(2.2.2'')$ it follows that
$(2.8.4)$ $x\setminus ((u\setminus v)y)=((ux)\setminus (vy))p(u,x,(ux)\setminus (vy)) [p(u,u\setminus v,x)]^{-1}$\
for each $u$, $v$, $x$ and $y$ in $G$, since
$x\setminus ((u\setminus v)y)=x\setminus (u\setminus (vy))[p(u,u\setminus
v,y)]^{-1}$.
In particular for $u=a(bc)$ and $v=(ab)c$ with any $a$, $b$ and $c$ in $G$ we infer using $(2.1.9)$ that $ux=(a(b(cx)))p(b,c,x)p(a,bc,x)$ and $vx=(ab)(cx)p(ab,c,x)$, hence from $(2.8.4)$ and $(2.3.7)$ it follows that
$(2.8.5)$ $x\setminus
(p(a,b,c)x)=[p(b,c,x)p(a,bc,x)]^{-1}p(a,b,cx)p(u,x,(ux)\setminus
(vx))$, since
$x\setminus
(p(a,b,c)x)=[(a(b(cx)))p(b,c,x)p(a,bc,x)]\setminus
[(ab)(cx)p(ab,c,x)]$
$p(u,x,(ux)\setminus (vx))[p(u,u\setminus
v,x)]^{-1}$,\
because $u\setminus v=p(a,b,c)\in N(G)$ and $p(u,u\setminus v,x)=e$.
Notice that $(2.1.1)$, $(2.1.2)$ and $(2.1.9)$ imply $u\setminus (tu)=p$, where $t=t(a,b,c)$, $p=p(a,b,c)$, $u=a(bc)$ for any $a$, $b$ and $c$ in $G$. Let $z\in G$, then there exists $x\in
G$ such that $z=ux$, that is $x=u\setminus z$. Therefore we deduce that
$(2.8.6)$ $z\setminus (tz)=[x\setminus (px)]p(u,u\setminus
(tu),x)[p(u,x,(ux)\setminus (tux))]^{-1}$,\
since $t\in N(G)$, $p\in N(G)$, $(u\setminus (tu))x=(u\setminus (tux))[p(u,u\setminus
(tu),x)]^{-1}$ by $(2.2.2'')$; $~x\setminus (u\setminus
(tux))=[(ux)\setminus (tux))]p(u,x,(ux)\setminus (tux))$ by $(2.2.2')$. Thus from Identities $(2.8.3)$, $(2.8.5)$ and $(2.8.6)$ it follows that a group $N_{0,0}=N_{0,0}(G)$ generated by $ \{
p(a,b,c), ~ t(a,b,c): a\in G, ~ b\in G, ~ c\in G \} $ satisfies Condition $(2.7.1)$. From the joint continuity of the multiplication and the mappings $Div_l$ and $Div_r$ it follows that the closure $N_0$ of $N_{0,0}$ also satisfies $(2.7.1)$. Thus $N_0$ is a closed normal subgroup in $G$. In view of Theorem 1.1 in Ch. IV, Section 1 in [@bruckb] a quotient loop $G/\cdot / N_0$ exists consisting of all cosets $aN_0$, where $a\in G$.
Then from Conditions $(2.1.9)$, $(2.7.1)$ and $(2.7.2)$ it follows that for each $a$, $b$, $c$ in $G$ the identities take place
$(aN_0)(bN_0)=(ab)N_0$ and
$((aN_0)(bN_0))(cN_0)=(aN_0)((bN_0)(cN_0))$ and $eN_0=N_0$,\
since $p(a,b,c)\in N_0$ and $t(a,b,c)\in N_0$ for all $a$, $b$ and $c$ in $G$.
In view of Lemmas 2.2 and 2.3 $(aN_0)\setminus e= e/(aN_0)$, consequently, for each $aN_0\in G/\cdot / N_0$ a unique inverse $(aN_0)^{-1}$ exists. Thus the quotient $G/\cdot / N_0$ of $G$ by $N_0$ is a group. Since the topology $\tau $ on $G$ is $T_1$ and $N_0$ is closed in $G$, then the quotient topology $\tau _q$ on $G/\cdot / N_0$ is also $T_1$. By virtue of Theorem 8.4 in [@hew] this implies that $\tau _q$ is a $T_1\cap T_{3.5}$ topology on $G/\cdot / N_0$.
[**2.9. Proposition.**]{}
*Assume that $G$ is a $T_1$ topological fan loop and functions $t$ and $p$ on $G$ are defined by Formulas $(2.1.9)$. Then for each compact subset $S$ in $G$ and each open neighborhood $V$ of $e$ there exists an open neighborhood $U$ of $e$ in $G$ such that*
$(2.9.1)$ $t((u_1a)v_1,(u_2b)v_2,(u_3c)v_3)\in (Vt(a,b,c))\cap (t(a,b,c)V)$ and
$(2.9.2)$ $p((u_1a)v_1,(u_2b)v_2,(u_3c)v_3)\in (Vp(a,b,c))\cap (p(a,b,c)V)$\
for every $a$, $b$, $c$ in $S$ and $u_j$, $v_j$ in $\check{U}$ for each $j\in \{ 1, 2, 3 \} $.
[**Proof.**]{} Take arbitrary fixed elements $f$, $g$ and $h$ in $S$. From the joint continuity of the maps $t(a,b,c)$ and $p(a,b,c)$ in the variables $a$, $b$ and $c$ in $G$ it follows that there exists an open neighborhood $U_{f,g,h}$ of $e$ in $G$ and an open neighborhood $W_{f,g,h}$ of $(f,g,h)\in S\times S\times S$ in $G\times G\times G$ such that $(2.9.1)$ and $(2.9.2)$ are valid for each $u_j$, $v_j$ in $\check{U}_{f,g,h}$, $j\in \{ 1, 2, 3 \} $, and $(a,b,c)\in W_{f,g,h}$ (see Lemmas 2.4 and 2.6). Notice that $S\times S\times S$ is compact in the Tychonoff product $G\times
G\times G$ of $G$ as the topological space (see Section 2.3 and Theorem 3.2.4 in [@eng]). Hence an open covering $ \{ W_{f,g,h}:
~ f\in S, ~ g\in S, ~ h\in S \} $ of $S\times S\times S$ has a finite subcovering $ \{ W_{f_i,g_i,h_i}: ~ i=1,...,n \} $, where $n$ is a natural number, $n\ge 1$. That is $S\times S\times S\subseteq
\bigcup_{i=1}^n W_{f_i,g_i,h_i}$. Then $\bigcap_{i=1}^n
U_{f_i,g_i,h_i}=:U$ is an open neighborhood of $e$ in $G$. Therefore, Properties $(2.9.1)$ and $(2.9.2)$ are satisfied for every $a$, $b$, $c$ in $S$ and $u_j$, $v_j$ in $\check{U}$ for each $j\in \{ 1, 2, 3 \} $
We remind the following.
[**2.10. Definition.**]{} Let $G$ be a topological loop. For a subset $U$ in $G$ it is put:
$(2.10.1)$ ${\cal L}_{U,G} := \{ (x,y)\in G\times G: ~ x\setminus y\in U \}
$ and
$(2.10.2)$ ${\cal R}_{U,G} := \{ (x,y)\in G\times G: ~ y/x \in U \}
$.
A family of all subsets ${\cal L}_{U,G}$ (or ${\cal
R}_{U,G}$) with $U$ being an open neighborhood of $e$ will be denoted by ${\cal L}_G$ (or ${\cal R}_G$ correspondingly).
[**2.11. Proposition.**]{} [*Let $G$ be a $T_1$ topological locally compact fan loop. Then a family ${\cal L}_G$ (or ${\cal R}_G$) induces a uniform structure on $G$. A topology $\tau _1$ on $G$ provided by ${\cal
L}_G$ (or ${\cal R}_G$ respectively) is $T_1\cap T_{3.5}$ and equivalent to the initial topology $\tau $ on $G$.*]{}
[**Proof.**]{} Let $(G, \tau )$ be a topological loop and let ${\cal
B}_e$ denote a base of its open neighborhoods at $e$. In view of Lemma 2.5 ${\cal C}_l(U):= \{ xU: ~ x\in G \} $ is an open covering of $G$ for each $U\in {\cal B}_e$. We put ${\cal C}_l^0 = \{ {\cal
C}_l(U): ~ U\in {\cal B}_e \} $ and ${\cal C}_l$ to be a family of all coverings for each of which there exists a refinement of the type ${\cal C}_l^0$.
Below it is verified, that the family ${\cal C}_l$ satisfies Conditions $(UC1)$-$(UC4)$ of Section 8.1 in [@eng]. If ${\cal A}\in {\cal C}_l$, ${\cal E}$ is a covering of $G$ and ${\cal A}$ refines ${\cal E}$, then there exists $U\in {\cal
B}_e$ such that ${\cal C}_l(U)$ refines ${\cal A}$ and hence ${\cal
C}_l(U)$ refines ${\cal E}$. Thus $(UC1)$ is satisfied.
Let ${\cal A}_1$ and ${\cal A}_2$ belong to ${\cal C}_l$. There are $U_1$ and $U_2$ in ${\cal B}_e$ such that ${\cal
C}_l(U_j)$ refines ${\cal A}_j$ for each $j\in \{ 1, 2 \} $. We put $U=U_1\cap U_2$, consequently, $U\in {\cal B}_e$ and hence ${\cal
C}_l(U)$ refines both ${\cal C}_l(U_1)$ and ${\cal C}_l(U_2)$. Therefore ${\cal C}_l(U)$ refines ${\cal A}_1$ and ${\cal A}_2$. Thus $(UC2)$ also is satisfied.
Condition $(UC3)$ means that for each ${\cal A}\in {\cal C}_l$ there exists ${\cal E}\in {\cal C}_l$ such that ${\cal E}$ is a star refinement of ${\cal A}$. In order to prove it, it evidently is sufficient to prove that for each $U\in {\cal B}_e$ there exists $U_1\in {\cal B}_e$ such that
$(2.11.1)$ $St(xU_1,{\cal
C}_l(U_1))\subset xU$ for each $x\in G$,\
where $St (M,{\cal A})$ denotes a star of a set $M$ with respect to ${\cal A}$ (see Section 5.1 in [@eng]).
Note that a map $f(x_1,x_2,x_3)=(x_1/x_2)x_3$ is the composition of jointly continuous maps $G\times G\ni (x_1,x_2)\mapsto x_1/x_2\in G$ and $G\times G\ni (y,x_3)\mapsto yx_3\in G$, hence it is jointly continuous from $G\times G\times G$ into $G$ and $f(e,e,e)=e$, because $G$ is the topological loop (see Definition 2.1). The loop $G$ is locally compact. Notice that for each open neighborhood $Q_1$ of $e$ in $G$ there exists an open neighborhood $Q_2$ of $e$ such that its closure $cl_G(Q_2)$ is compact and $cl_G(Q_2)\subset Q_1$ by the corresponding Theorem 3.3.2 in [@eng] for topological spaces. Hence for each open neighborhood $W$ of $e$ in $G$ there exists an open neighborhood $U_0$ of $e$ in $G$ with the compact closure $cl_G\check{U}_0$ such that $cl_G\check{U}_0$ is contained in $W$ (see Lemma 2.6).
Therefore for each $U\in {\cal B}_e$ there exists $V_1\in {\cal B}_e$ such that $f(V_1,V_1,V_1)\subset U$ and $cl_G(V_1)$ is compact. If for an arbitrary fixed element $x\in
G$ and some $x_1\in G$ the intersection $xV_1\cap x_1V_1\ne
\emptyset $ is non void, then there are $h_0$ and $h_1$ in $V_1$ such that $x_1=(xh_0)/h_1$. On the other hand, $x_1h\in x_1V_1$ for each $h\in V_1$ and for each $y\in x_1V_1$ there exists $h\in V_1$ with $y=x_1h$, consequently, $x_1h=((xh_0)/h_1)h\in
((xV_1)/V_1)V_1$.
Using Identities $(2.2.3)$ and $(2.1.9)$ we get that
$(2.11.2)$ $x_1h=(x(h_0(e/h_1))p(x,h_0,e/h_1)$
$p(e/h_1,h_1,h_1\setminus
e)[p((xh_0)/h_1,h_1,h_1\setminus e)]^{-1})h$.
We choose open neighborhoods $V$ and $W$ of $e$ in $G$ such that $\check{V}^2\subset W$ and $\check{W}^2\subset V_1$ by Lemma 2.6. In view of the inclusion $(2.9.2)$ of Proposition 2.9 and Formula $(2.11.2)$ there exists $U_1\in {\cal B}_e$ such that $\check{U}_1\subset V$ and
$(2.11.3)$ $p((u_1a)v_1,(u_2b)v_2,(u_3c)v_3)\in (Vp(a,b,c))\cap (p(a,b,c)V)$\
for every $a$, $b$, $c$ in $cl_G(V_1)$ and $u_j$, $v_j$ in $\check{U}_1$ for each $j\in \{ 1, 2, 3 \} $. This implies $(2.11.1)$ and hence $(UC3)$, since $p(a,b,c)=e$ if either $a=e$ or $b=e$ or $c=e$.
It remains to prove that ${\cal C}_l$ also has the property $(UC4)$. That is for each $x\ne y$ in $G$ there exists ${\cal A}\in
{\cal C}_l$ such that $\{ x, y \} \cap V\ne \{ x , y \} $ for each $V\in {\cal A}$. It is sufficient to find an open neighborhood $U$ of $e$ in $G$ such that $x/U\cap y/U=\emptyset $, because this implies $x_0U\cap \{ x, y \} \ne \{ x, y \} $ for each $x_0\in G$. The loop $G$ is $T_1$. By virtue of Lemmas 2.5 and 2.6 and the joint continuity of the multiplication and $Div_r$ in $G$ there is $U_1\in
{\cal B}_e$ such that $y\notin (xU_1)/U_1$, that is $xU_1\cap
yU_1=\emptyset $ by $(2.2.5)$. In view of Proposition 2.9 there exists $U\in {\cal B}_e$ such that $(e/U)p(e/U,U,U\setminus
e)[p(a/U,U,U\setminus e)]^{-1}\subset U_1$ for each $a\in \{ x , y
\} $, since the two-point set $ \{ x, y \} $ is compact in $G$, for each $W\in {\cal B}_e$ there exists $W_1\in {\cal B}_e$ such that $e/W_1\subset W$. From $(2.2.3)$ it follows that $x/U\cap
y/U=\emptyset $. Therefore $\{ x, y \} \cap V\ne \{ x , y \} $ for each $V\in {\cal C}_l(U)$.
By virtue of Theorem 8.1.1 in [@eng] the uniformity ${\cal C}_l$ induces a $T_1$ topology $\tau _1$ on $G$. Note that the family ${\cal C}_l$ consists of open coverings of $G$ and that for each $x\in G$ and each open neighborhood $V$ of $x$ in the initial topology $\tau $ there exists $U\in {\cal B}_e$ such that $xU\subset V$. Therefore from the latter inclusion and $(2.11.1)$ it follows that the topology $\tau _1$ induced by ${\cal C}_l$ coincides with the initial topology $\tau $ on $G$. In view of Corollary 8.1.13 in [@eng] $(G, \tau )$ is a Tychonoff space, that is $(G, \tau )$ is a completely regular space, $T_1\cap T_{3.5}$. Finally note that ${\cal C}_l^0={\cal L}_G$. Symmetrically the case ${\cal
C}_r^0={\cal R}_G$ is proved.
[**2.12. Lemma.**]{}
*Suppose that $(G, \tau )$ is a $T_1$ topological loop, $S$ is a compact subset in $G$, $q$ is a fixed element in $G$, $V$ is an open neighborhood of the unit element $e$. Then there are elements $b_1,...,b_m$ in $G$ and an open neighborhood $U$ of $e$ such that*
$(2.12.1)$ $\check{U}\subset
V$ and
$(2.12.2)$ $ \{ b_1\setminus (qU),...,b_m\setminus (qU) \} $ is an open covering of $S$.
[**Proof.**]{} The multiplication is continuous on $G$, hence the left shift mapping $L_b(x)=bx$ is continuous on $G$ in the variable $x$. On the other hand, the mapping $Inv_l$ is continuous on $G$.
In view of $(2.1.1)$, $(2.1.2)$, Lemmas 2.5 and 2.6 and the compactness of $S$ for each open neighborhood $U$ of $e$ in $G$ with $\check{U}\subset V$ there are $b_1,...,b_m$ in $G$ such that $\{
b_1\setminus (qU),...,b_m\setminus (qU) \} $ is an open covering of $S$.
[**2.13. Corollary.**]{}
*Let $G$ be a $T_1$ topological loop. Then for each open neighborhood $W$ of $e$ in $G$ there exists an open neighborhood $U$ of $e$ such that $\check{U}\subset W$ and*
$(2.13.1)$ $(\forall x$ $\forall
y$ $((x\in G) \& (y\in G)\& (x\setminus y \in U))) \Rightarrow (y\in
xW)$ and
$(2.13.2)$ $(\forall x$ $\forall y$ $((x\in G) \&
(y\in G)\& (y/x \in U))) \Rightarrow (y\in Wx)$.
[**Proof.**]{} This follows from Lemmas 2.6 and 2.12, $(2.1.1)$ and $(2.1.2)$.
[**2.14. Theorem.**]{} [*Let $G$ and $H$ be $T_1$ topological fan loops (see Definition 2.1) and let $f: G\to H$ be a continuous map so that for each open neighborhood $V$ of a unit element $e_H$ in $H$ a compact subset $K_V$ in $G$ exists such that $f(G- K_V)\subset V$. Then $f$ is uniformly $({\cal L}_G, {\cal
L}_H)$ continuous and uniformly $({\cal R}_G, {\cal R}_H)$ continuous (see also Definition 2.10).*]{}
[**Proof.**]{} Since the multiplication in $H$ is continuous, then for each open neighborhood $Y$ of $e_H$ there exists an open neighborhood $X$ of $e_H$ such that $X^2\subset Y$. In view of Lemma 2.6 there exists an open neighborhood $V_1$ of $e_H$ in $H$ such that $\check{V}_1^2\subset V$, where $A^2=AA$ for a subset $A$ in $H$. By the conditions of this theorem a compact subset $K_{V_1}$ in $G$ exists such that $f(G- K_{V_1})\subset V_1$.
For a subset $A$ of the loop $G$ let
$(2.14.1)$ $P(A)=(P_0(A)\cup \{ e \} )(P_0(A)\cup \{ e \} )$,
where $P_0(A)=A\cup Inv_l(A)\cup Inv_r(A)$,\
hence $A\subset P_0(A)$ and $P_0(A)\cup \{ e \} \subset P(A)$. Then $S_1=P(K_{V_1})$ is a compact subset in $G$, since the mappings $Inv_l$ and $Inv_r$ are continuous on $G$ and the multiplication is jointly continuous on $G\times G$ (see Theorems 3.1.10, 8.3.13-8.3.15 in [@eng]), hence $R_1=P(f(S_1))$ is compact in $H$.
By virtue of Proposition 2.9 there exists an open neighborhood ${V'}_2$ of $e_H$ in $H$ such that
$(2.14.2)$ $[t_H((V_2a)V_2,(V_2b)V_2,(V_2c)V_2)V_2]\cup
[V_2t_H((V_2a)V_2,(V_2b)V_2,(V_2c)V_2)]$
$\subset
(V_3t_H(a,b,c))\cap (t_H(a,b,c)V_3)$ and
$[p_H((V_2a)V_2,(V_2b)V_2,(V_2c)V_2)V_2]\cup
[V_2p_H((V_2a)V_2,(V_2b)V_2,(V_2c)V_2)]$
$\subset
(V_3p_H(a,b,c))\cap (p_H(a,b,c)V_3)$\
for every $a$, $b$, $c$ in $R_1$, where $\check{V}_3^2\subset V_1$, $ ~ V_2=\check{V'}_2$, $V_3$ is an open neighborhood of $e$ in $H$. For $V_2$ there exists a compact subset $K_{V_2}$ in $G$ such that $f(G- K_{V_2})\subset
V_2$ by the conditions of this theorem. If $A$ and $B$ are compact subsets in $G$, then their union $A\cup B$ is also compact. Therefore it is possible to choose $K_{V_2}$ such that $K_{V_1}\subset K_{V_2}$, since $V_2\subset V_1$ and $(G- A)- B=G-
(A\cup B)\subset G- A$. We take $S_2=P(K_{V_2})$ by Formula $(2.14.1)$, consequently, $S_1\subset S_2$, since $K_{V_1}\subset
K_{V_2}$.
From the continuity of the map $f$ and Lemmas 2.5, 2.6 it follows that for each $x\in G$ open neighborhoods $W_{x,l}$ and $W_{x,r}$ of $e$ in $G$ exist such that
$f(x\check{W}_{x,l}^2)\subset (f(x)V_2)$ and $f(\check{W}_{x,r}^2x)\subset
(V_2f(x))$, consequently,
$(2.14.3)$ $f(x\check{W}_x^2)\subset (f(x)V_2)$ and $f(\check{W}_x^2x)\subset (V_2f(x))$\
for an open neighborhood $W_x=W_{x,l}\cap W_{x,r}$ of $e$ in $G$. The compactness of $S_2$ imply that coverings $ \{ xW_x: ~ x\in S_2 \} $ and $ \{ W_yy: ~
y\in S_2 \} $ of $S_2$ have finite subcoverings $ \{ x_jW_{x_j}: ~
x_j\in S_2, ~ j=1,...,n \} $ and $ \{ W_{y_i}y_i: ~ y_i\in S_2, ~
i=1,...,m \} $. Hence
$(2.14.4)$ $W=\bigcap_{j=1}^nW_{x_j}\cap
\bigcap_{i=1}^m W_{y_i}$\
is an open neighborhood of $e$ in $G$. Therefore according to Proposition 2.9 there exists an open neighborhood $U'$ of the unit element $e$ in $G$ such that
$(2.14.5)$ $[t_G((Ua)U,(Ub)U,(Uc)U)U]\cup [Ut_G((Ua)U,(Ub)U,(Uc)U)]
$
$\subset [W_3t_G(a,b,c)]\cap [t_G(a,b,c)W_3]$ and
$[p_G((Ua)U,(Ub)U,(Uc)U)U]\cup [Up_G((Ua)U,(Ub)U,(Uc)U)]
$
$\subset [W_3p_G(a,b,c)]\cap [p_G(a,b,c)W_3]$\
for every $a$, $b$, $c$ in $S_2$, where $U=\check{U'}$, $~ W_0$ and $W_3$ are open neighborhoods of $e$ in $G$ such that $\check{W}_3^2\subset
W_0$ and $\check{W}_0^2\subset W$.
Let now $x$ and $y$ in $G$ be such that $x\setminus y\in U$. Then Formula $(2.2.4)$ imply that
$(2.14.6)$ $y\in xU$.
There are several options. Consider at first the case $x\in K_{V_2}$. From Formulas $(2.14.4)$-$(2.14.6)$ and Corollary $(2.13)$ it follows that there exists $j\in \{ 1,...,n \}$ such that $x\in x_jW_{x_j}$ and $y\in x_jW_{x_j}^2$. Therefore, Formulas $(2.14.2)$ and $(2.14.3)$ imply that $f(x)\setminus f(y)\in
V$.
From $x\setminus y\in U$ and Identities $(2.2.4)$ it follows that $y=xu$ for a unique $u\in U$. Hence
$(2.14.7)$ $x=[t(y,e/u,u)]^{-1} y(e/u)$\
according to Identities $(2.2.3)$, $(2.2.5)$.
If $y\in K_{V_2}$, then similarly from Formulas $(2.14.4)$- $(2.14.7)$ and Corollary $(2.13)$ it follows that there exists $k\in
\{ 1,...,n \} $ such that $y\in x_kW_{x_k}$ and $x \in
x_kW_{x_k}^2$, since $t(a,b,e)=t(a,e,b)=t(e,a,b)=e$ for each $a$ and $b$ in $G$. Therefore, $f(x)\setminus f(y)\in V$ by Formulas $(2.14.2)$ and $(2.14.3)$, since $S_2=P(K_{V_2})$ (see Formula $(2.14.1)$).
It remains the case $x\in G- K_{V_2}$ and $y\in
G- K_{V_2}$. Therefore $f(x)\in V_2$ and $f(y)\in V_2$. According to the choice of $R_1$ we have $e_H\in R_1$. From Condition $(2.14.2)$, Identity $(2.2.4)$ and the inclusion $\check{V}_1^2\subset V$, it follows that $f(x)\setminus f(y)\in V$. Taking into account the inclusion $K_{V_1}\subset K_{V_2}$ we get that $f$ is uniformly $({\cal L}_G, {\cal L}_H)$ continuous.
The uniform $({\cal R}_G, {\cal R}_H)$ continuity is proved analogously using the finite subcovering $ \{ W_{y_i}y_i: ~ y_i\in
S_2, ~ i=1,...,m \} $ and Corollary 2.13.
[**2.15. Corollary.**]{} [*Let $G$ be a $T_1$ topological locally compact fan loop and let $f\in C_0(G)$ and let $H=({\bf C}, +)$ be the complex field $\bf C$ considered as an additive group. Then $f$ is uniformly $({\cal L}_G, {\cal L}_H)$ continuous and uniformly $({\cal R}_G,
{\cal R}_H)$ continuous.*]{}
Left invariant measures.
========================
[**3.1. Notation.**]{} For a completely regular topological space $X$ by $C_b(X)$ is denoted the Banach space of all continuous bounded functions $f$ from $X$ into the complex field $\bf C$ supplied with the norm
$(3.1.1)$ $\| f \|_X = \sup_{x\in X} |f(x)|<\infty $.
We put
$(3.1.2)$ $C_0(X) := \{ f \in C_b(X): ~ \forall ~ \epsilon
>0, ~ \exists ~ S\subset X, ~ S \mbox{ is compact, }$
$ ~ \forall ~ x\in
X- S, ~ |f(x)|<\epsilon \} $ and
$(3.1.3)$ $C_{0,0}(X) := \{ f \in C_b(X):
~ \exists S\subset X, ~ S \mbox{ is compact, }$
$ ~ \forall x\in
X- S, ~ f(x)=0 \} $ and
$(3.1.4)$ $C_{0,0}^+(X) = \{ f\in C_{0,0}(X): ~ \forall x\in X, ~ f(x)\ge 0 \}
$.
Let $G$ be a loop. For a function $f: G\to \bf C$ and an element $b\in G$ let $L_bf(x)=\mbox{}_bf(x)=f(bx)$ and $R_bf(x)=f_b(x)=f(xb)$ for each $x\in G$. Consider a support $S_f := cl_G \{ x\in G: ~ f(x)\ne
0 \} $ of $f\in C_b(G)$, where $cl_G(A)$ denotes the closure of a subset $A$ in $G$.
[**3.2. Lemma.**]{}
*Let $(G, \tau )$ be a $T_1$ topological locally compact fan loop, let also $f$ and $\phi
$ belong to $C_{0,0}^+(G)$ and $\phi $ be not identically zero (see Notation 3.1). Then there exist a natural number $m>0$, elements $b_1,...,b_m$ in $G$ and positive constants $c_1,...,c_m$ such that*
$(3.2.1)$ $\forall x\in G$, $ ~f(x)\le \sum_{j=1}^m c_j L_{b_j}\phi
(x).$
[**Proof.**]{} Since $f\in C_{0,0}^+(G)$, then the support $S_f$ is compact. The function $\phi $ is not null, hence there exists $q\in G$ such that $\phi (q)>0$. From Lemma 2.5 and from the continuity of the function $\phi $ it follows that there exists an open neighborhood $qV$ of $q$ such that $\phi (x)> \phi (q)/2$ for reach $x\in qV$, where $V$ is an open neighborhood of the unit element $e$. By virtue of Lemma 2.12 there exists an open neighborhood $U$ of $e$ and elements $b_1,...,b_m$ in $G$ such that $\check{U}\subset V$ and for each $x\in S_f$ there exists $j\in \{ 1,...,m \} $ such that $x\in
b_j\setminus (qU)$.
Therefore, $$f(x)\le \| f \| _G (2/\phi (q))\sum_{j=1}^m \phi
(b_jx)$$ for each $x\in G$ according to $(2.2.4)$, so it is sufficient to take $c_j\ge \| f \| _G (2/\phi (q))$ for each $j=1,...,m$.
[**3.3. Corollary.**]{} [*Let the conditions of Lemma 3.2 be satisfied and let $$(3.3.1) \quad (f:\phi ) := \inf \{ \sum_{j=1}^m c_j: ~ \exists
~ \{ b_1,...,b_m \} \subset G, ~ \exists ~ \{ c_1,...,c_m \} \subset
(0, \infty ),$$ $$~ \forall ~ x\in G, ~ f(x)\le \sum_{j=1}^m c_j
L_{b_j}\phi (x) \} .$$ Then $(f: \phi )\le 2m\| f \| _G /\phi (q)$ in the notation of Lemma 3.2.*]{}
[**3.4. Lemma.**]{}
*Assume that the conditions of Lemma 3.2 are fulfilled, then for each $b\in G$*
$(3.4.1)$ $(\mbox{}_b f: \phi
)=(f: \phi ^b)$;
$(3.4.2)$ $(f: \mbox{}_b\phi
)=(f^b: \phi )$,\
where $f^b(x)=f(b\setminus
x)$ for each $x\in G$; particularly,
$(3.4.1')$ $(\mbox{}_{\gamma } f: \phi )=(f: \phi )$ and
$(3.4.2')$ $(f: \mbox{}_{\gamma } \phi )=(f: \phi )$ for each $\gamma \in
N(G)$;
$(3.4.3)$ $(\alpha f: \phi )=\alpha (f: \phi )$ for each $\alpha \ge 0$;
$(3.4.4)$ $((f_1+f_2): \phi )\le (f_1: \phi ) +(f_2: \phi )$ for every $f_1$ and $f_2$ in $C_{0,0}^+(G)$.
$(3.4.5)$ If $f(x)\le f_1(x)$ for each $x\in G$, then $(f: \phi
)\le (f_1: \phi )$.
[**Proof.**]{} Let $c_1,...,c_m$ in $(0, \infty )$ and $b_1,...,b_ m$ in $G$ be such that $$(3.4.6)\quad \mbox{}_bf(x)\le
\sum_{j=1}^mc_jL_{b_j}\phi (x)$$ for each $x\in G$. From Formulas $(2.2.4)$ and $(3.4.6)$ by changing of a variable $y=bx$ it follows that $$(3.4.7)\quad f(y)\le
\sum_{j=1}^mc_jL_{b_j}\phi (b\setminus y)$$ for each $y\in G$. From $(3.4.7)$ it follows $(3.4.1)$. Similarly from the inequality $$(3.4.8)\quad f(x)\le
\sum_{j=1}^mc_jL_{b_j}(L_b\phi (x))$$ for each $x\in G$ we infer that $$(3.4.9)\quad f(b\setminus y)\le \sum_{j=1}^mc_jL_{b_j}\phi (y)$$ for each $y\in
G$. Thus $(3.4.9)$ implies Equality $(3.4.2)$.
In particular, if $\gamma \in N(G)$, then $b_j(\gamma \setminus y)=(b_j\gamma ^{-1})y$ and $b_j(\gamma
y)=(b_j\gamma )y$ for each $y$ and $b_j$ in $G$ by Condition $(2.1.8)$ and Formulas $(2.2.2)$ and $(2.3.1)$. Hence $(3.4.7)$ transforms into to $$f(y)\le \sum_{j=1}^mc_jL_{b_j\gamma ^{-1}}\phi (y)$$ and $(3.4.8)$ into $$f(x)\le \sum_{j=1}^mc_jL_{b_j\gamma }\phi (x)$$ with $\gamma \in N(G)$ instead of $b$. This implies Equalities $(3.4.1')$, $(3.4.2')$.
Properties $(3.4.3)$ and $(3.4.4)$ evidently follow from Formula $(3.3.1)$.
For proving Property $(3.4.5)$ note that if $f(x)\le f_1(x)$ for each $x\in
G$, then from $f_1(x)\le \sum_{j=1}^mc_jL_{b_j}\phi (x)$ for each $x\in G$ it follows that $f(x)\le \sum_{j=1}^mc_jL_{b_j}\phi (x)$ for each $x\in G$, consequently, $(f: \phi )\le (f_1: \phi )$.
[**3.5. Notation.**]{} Let $\phi $, $f_0$ and $f$ belong to $C_{0,0}^+(G)$ and $\phi $ and $f_0$ be not null, where $G$ is a $T_1$ topological locally compact fan loop. We consider a functional $$(3.5.1)\quad J_{\phi ,f_0}(f):= \frac{(f: \phi )}{(f_0: \phi )}.$$
Assume that
$(3.5.2)$ there exists a compact subgroup $N_0=N_0(G)$ in $N(G)$ such that
$t(a,b,c)\in N_0$ and $p(a,b,c)\in N_0$ for every $a$, $b$ and $c$ in $G$.
Then we denote by $\Upsilon (G,N_0)$ a family of all non null functions $h$ in $C_{0,0}^+(G)$ such that
$(3.5.3)$ $h(\gamma a)=h(a)$ for each $a\in G$ and $\gamma \in N_0$.
Evidently Condition $(3.5.3)$ for $h\in C_{0,0}^+(G)$ is equivalent to
$(3.5.4)$ $h(a\gamma )=h(a)$ for each $a\in G$ and $\gamma \in N_0$,\
since $aN_0=N_0a$ for each $a\in G$ according to Theorem 2.8.
[**3.6. Lemma.**]{}
*Let $G$ be a $T_1$ topological locally compact fan loop satisfying Condition $(3.5.2)$, $f$ and $\phi $ be in $C^+_{0,0}(G)$ and $\omega \in \Upsilon (G,N_0)$ (see Condition $(3.5.3)$), $\phi $ be non null. Then*
$(3.6.1)$ $(f: \phi ) \le (f: \omega )
(\omega : \phi )$.
[**Proof.**]{} If $b$ is a fixed element in $G$ and there are elements $b_1,...,b_m$ in $G$ and positive constants $c_1,...,c_m$ such that $$(3.6.2)\quad \mbox{}_b\omega(x)\le \sum_{j=1}^mc_j\phi (b_jx)$$ for each $x\in G$, then $$(3.6.3)\quad \mbox{}_b\omega(x)\le
\sum_{j=1}^mc_j\phi (b_jx\gamma )$$ for each $x\in G$ and $\gamma
\in N_0$, since $N_0\subset N(G)$ and $\mbox{}_b\omega(x\gamma
)=\mbox{}_b\omega(x)$ for each $x\in G$ and $\gamma \in N_0$ by $(3.5.4)$ equivalent to $(3.5.3)$.
By the conditions of this lemma $N_0$ is a compact group. Therefore there exists a Haar measure $\lambda $ on the Borel $\sigma $-algebra ${\cal B}(N_0)$ of $N_0$ and with values in the unit segment $[0,1]$ such that $\lambda (N_0)=1$, $\lambda
(sA)=\lambda (A)$ and $\lambda (As)=\lambda (A)$ for each $s\in N_0$ and $A\in {\cal B}(N_0)$ (see Theorems 15.5, 15.9 and 15.13 and Subsection 15.8 in [@hew]). In view of this, Conditions $(3.1.3)$ and $(3.1.4)$ and Corollary 2.15 a function $$(3.6.4)\quad \phi ^{[\lambda ]}(x):=\int_{N_0} \phi (\gamma x)\lambda (d\gamma )$$ on $G$ is nonzero and belongs to $C_{0,0}^+(G)$, since $N_0S_{\phi
}$ is a compact subset in $G$ by Lemma 2.6, where $S_{\phi }$ is a compact support of $\phi $. From Formula $(3.6.4)$ it follows that
$(3.6.5)$ $\phi ^{[\lambda ]}(\beta
x)=\phi ^{[\lambda ]}(x)$ for each $\beta \in N_0$ and $x\in G$,\
since the measure $\lambda $ is left and right invariant $\lambda
(\beta A)=\lambda (A)= \lambda (A\beta )$ for each $\beta \in N_0$ and each Borel subset $A$ in $N_0$. Hence $\phi ^{[\lambda ]}\in
\Upsilon (G,N_0)$, since $S_{\phi }N_0$ is compact, and since Conditions $(3.5.3)$ and $(3.5.4)$ are equivalent, where $S_{\phi }$ is the support of $\phi $ (see Subsection 3.2). From $(3.6.4)$, $(3.6.5)$, $(3.5.3)$, $(3.5.4)$ and the Fubini theorem it follows that $$(3.6.4')\quad \phi ^{[\lambda ]}(x)=\int_{N_0} \phi (x\beta )\lambda (d\beta )\mbox{,
since}$$ $$\phi ^{[\lambda ]}(x)=\int_{N_0}(\int_{N_0} \phi (\gamma x\beta )\lambda (d\gamma
))\lambda (d\beta )$$ $$=\int_{N_0}(\int_{N_0} \phi (\gamma x\beta )\lambda (d\beta ))\lambda (d\gamma
)=\int_{N_0} \phi (x\beta )\lambda (d\beta )),$$ because $\int_{N_0}
\phi (x\gamma \beta )\lambda (d\beta ) =\int_{N_0} \phi (x\beta
)\lambda (d\beta )$ for each $\gamma \in N_0(G)$.
Integrating both sides of Inequality $(3.6.3)$ and utilizing Formulas $(3.6.4)$, $(3.6.4')$ we infer that $$(3.6.6)\quad \mbox{}_b\omega(x)\le
\sum_{j=1}^mc_j\phi ^{[\lambda ]}(b_jx)$$ for each $x\in G$. On the other hand, $$\int_{N_0} (\sum_{j=1}^mc_j\mbox{ }_{b_j}\phi )(x\gamma )\lambda
(d\gamma )=(\sum_{j=1}^mc_j\mbox{ }_{b_j}\phi )^{[\lambda ]
}(x)=\sum_{j=1}^mc_j\mbox{ }_{b_j}\phi ^{[\lambda ]}(x),$$ hence for each $x\in G$ there exists $\gamma \in N_0$ such that $$(\sum_{j=1}^mc_j\mbox{ }_{b_j}\phi )(x\gamma )\ge
\sum_{j=1}^mc_j\mbox{ }_{b_j}\phi ^{[\lambda ]}(x).$$ Thus vice versa from $\omega \in \Upsilon (G,N_0)$ and $(3.6.6)$ it follows $(3.6.3)$ and hence $(3.6.2)$, consequently,
$(3.6.7)$ $(\mbox{}_b\omega: \phi ^{[\lambda ]})= (\mbox{}_b\omega: \phi )$.
Let $a_1,...,a_n$ in $G$ and positive constants $q_1,...,q_n$ be such that $$(3.6.8)\quad \mbox{}_b\omega(x)\le
\sum_{j=1}^nq_j\phi ^{[\lambda ]}(a_jx)$$ for each $x\in G$ (see Lemma 3.2). From Formulas $(2.2.2)$, $(3.6.5)$, $(3.6.8)$ and Conditions $(3.5.2)$, $(3.5.3)$, $(3.5.4)$ we deduce that $$(3.6.9)\quad \omega(y)\le
\sum_{j=1}^nq_j\phi ^{[\lambda ]}((a_j(b\setminus e))y
[p(a_j,b\setminus e,y)]^{-1} p(b,b\setminus e, y) )$$ $$=\sum_{j=1}^nq_j\phi ^{[\lambda ]} (d_jy)$$ for each $y\in G$, where $d_j=a_j(b\setminus e)$ for each $j$. Therefore
$(\mbox{}_b\omega: \phi ^{[\lambda ]})\le (\omega: \phi ^{[\lambda ]})$ for each $b\in G$. Notice that
$(3.6.10)$ $L_cL_{c\setminus
e}\omega(x)=\omega(x)$ for each $c$ and $x$ in $G$ by Lemmas 2.2, 2.3 and Condition $(3.5.3)$. Therefore we analogously get
$(\omega: \phi ^{[\lambda ]}) \le (\mbox{}_c\omega : \phi ^{[\lambda
]})$ for each $c\in G$. Thus
$(3.6.11)$ $(\mbox{}_b\omega: \phi ^{[\lambda ]})= (\omega: \phi
^{[\lambda ]})$ for each $b\in G$.
From $(3.6.7)$ and $(3.6.11)$ it follows that
$(3.6.12)$ $(\mbox{}_b\omega: \phi )= (\omega: \phi )$ for each $b\in
G$.
If $c_1,...,c_n$, $h_1,...,h_k$ in $(0,\infty )$ and $a_1,..,a_k$, $g_1,...,g_n$ in $G$ are such that $$(3.6.13)\quad f(x)\le \sum_{j=1}^kh_jL_{a_j}\omega (x) \mbox{ and}$$ $$(3.6.14)\quad \omega (x)\le \sum_{i=1}^n c_i L_{g_i}\phi (x)$$ for each $x\in G$ (see Lemma 3.2). Then from $(3.5.3)$, $(3.6.7)$, $(3.6.12)$-$(3.6.14)$ and Lemma 2.2 we infer that $$(3.6.15)\quad f(x)\le \sum_{j=1}^k h_j \sum_{i=1}^n c_i
L_{g_i}L_{a_j}\phi (x)= \sum_{j=1}^k h_j \sum_{i=1}^n c_i \phi
((g_ia_j)x).$$ Apparently $(3.6.15)$ implies $(3.6.1)$.
[**3.7. Lemma.**]{}
*Let $G$ be a $T_1$ topological locally compact fan loop, $\phi $, $f_0$ be nonzero functions belonging to $C_{0,0}^+(G)$. Then for each functions $f$, $f_1$ in $C_{0,0}^+(G)$ and $\alpha \ge 0$*
$(3.7.1)$ $J_{\phi ,f_0}(\alpha f)=\alpha J_{\phi ,f_0}(f)$ and
$(3.7.2)$ $J_{\phi ,f_0}(f+f_1)\le J_{\phi ,f_0}(f) + J_{\phi
,f_0}(f_1)$ and
$(3.7.3)$ if $f(x)\le f_1(x)$ for each $x\in G$, then $J_{\phi ,f_0}(f)\le J_{\phi ,f_0}(f_1)$.
Moreover, if $G$ satisfies Condition $(3.5.2)$ and $f_0\in \Upsilon (G,N_0)$ (see Condition $(3.5.3)$), then
$(3.7.4)\quad J_{\phi ,f_0}(f)
\le (f: f_0).$
[**Proof.**]{} Properties $(3.7.1)$ and $(3.7.2)$ follow immediately from $(3.4.3)$ and $(3.4.4)$. Property $(3.7.3)$ follows from Property $(3.4.5)$.
Applying Inequality $(3.6.1)$ and Formula $(3.5.1)$ we infer Inequality $(3.7.4)$, since $J_{\phi ,
f_0}(f_0)=1$.
[**3.8. Lemma.**]{} [*Assume that $G$ is a $T_1$ topological locally compact fan loop, functions $\phi $, $f_0$ and $f$ belong to $C_{0,0}^+(G)$ and $\phi $ and $f_0$ are not null. Then mappings $J_{\phi ,f_0}(\mbox{}_bf)$ and $J_{\phi ,f_0}(f_b)$ are continuous in the variable $b$ in $G$.*]{}
[**Proof.**]{} For each $x$, $b_1$ and $b_2$ in $G$ we have $\mbox{}_{b_1}f(x)-\mbox{}_{b_2}f(x)=f(b_1x)-f(b_2x)$. In view of Corollary 2.15 for each $\epsilon >0$ there exists an open of the form $(2.6.1)$ neighborhood $U$ of $e$ in $G$ with a compact closure $cl_G(U)$ for which
$(3.8.1)$ $|f(b_1x)-f(b_2x)|<\epsilon $ for each $x$, $b_1$ and $b_2$ in $G$ such that $(b_2x)\setminus
(b_1x)\in U$.
On the other hand, a support $S_f$ of $f$ is compact, consequently, $bS_f=L_bS_f$ is compact for each $b\in G$. Let $b_1$ be fixed. For each $x\in G$ with $b_1x\in S_f$ there exists an open of the form $(2.6.1)$ neighborhood $W_x$ of $e$ in $G$ such that $(b_2x)\setminus (b_1x)\in U$ for each $b_2x\in (b_1W_x)x\cap
b_1(xW_x)$ according to Lemmas 2.2, 2.4, 2.5, Proposition 2.9 and Formula $(2.14.3)$. For an open covering $\{ (b_1W_x)x\cap
b_1(xW_x): ~ b_1x\in S_f, ~ x\in G \} $ of $S_f$ there exists a finite subcovering $\{ (b_1W_{x_j})x_j\cap b_1(x_jW_{x_j}): ~
b_1x_j\in S_f, ~ x_j\in G, ~ j=1,...,m \} $ (see also Lemma 2.5), since the subset $S_f$ is compact.
We take $W_0=U\cap \bigcap_{j=1}^m W_{x_j}$ and choose an open of the form $(2.6.1)$ neighborhood $W$ of $e$ in $G$ with a compact closure $cl_G(W)$ contained in $W_0$ (see Theorem 3.3.2 in [@eng] and Formula $(2.14.3)$), because $G$ is locally compact.
In view of Proposition 2.9 and Lemma 2.6 there exists an open neighborhood $V'$ of $e$ in $G$ with $V=\check{V'}$ and a compact closure $cl_G(V)$ such that
$(3.8.2)$ $[t((Va)V,(Vb)V,(Vc)V)V]\cup [Vt((Va)V,(Vb)V,(Vc)V)]
$
$\subset [t(a,b,c)W_1]\cap [W_1t(a,b,c)]$ and
$[p((Va)V,(Vb)V,(Vc)V)V]\cup [Vp((Va)V,(Vb)V,(Vc)V)]
$
$\subset [p(a,b,c)W_1]\cap [W_1p(a,b,c)]$\
for each $a$, $b$ and $c$ in $S$, where $\check{W}_1^2\subset W$, $
~ W_1$ is an open neighborhood of $e$ in $G$,
$S=P(S_1)$, $S_1=S_2 \cup cl_G(U)$,
$S_2=\{ y\in G: ~
y=(b_1u)x, ~ u\in cl_G(U), b_1x\in S_f \} $\
(see Formula $(2.14.1)$), since $S$ is compact, $t(a,b,c)=e$ and $p(a,b,c)=e$ if $e\in \{ a, b, c \} $. For $b_1x\notin S_f$ and $b_2x\notin S_f$ certainly $f(b_1x)-f(b_2x)=0$. So remain two cases either $b_1x\in
S_f$ or $b_2x\in S_f$ which are similar to each other up to a notation. From Formulas $(2.2.5)$ it follows that $b_2x\in (b_1V)x$ is equivalent to $b_2\in b_1V$. Hence Lemma 2.2 and Inclusion $(3.8.2)$ provide that $(b_2x)\setminus (b_1x)\in U$ for each $b_2\in b_1V$ and $b_1x\in S_f$.
Let $w\in C_{0,0}^+(G)$ be a function such that $w(y)=1$ for each $y\in (cl_G(U)S_f)cl_G(U)$. Then we deduce that $|f(b_1x)-f(b_2x)|<\epsilon w(x)$ for each $x$, $b_1$ and $b_2$ in $G$ such that $b_2\in b_1V$ and with $b_1x\in
S_f$.
Therefore for each $\epsilon
>0$ there exists and open neighborhood $V$ of $e$ in $G$ such that $|(\mbox{}_{b_1}f: \phi )- (\mbox{}_{b_2}f: \phi
)|<\epsilon (w:\phi )$ for each $b_2\in b_1V$,\
consequently,
$(3.8.3)$ $|J_{\phi ,f_0}(\mbox{}_{b_1}f)- J_{\phi
,f_0}(\mbox{}_{b_2}f)|<\epsilon J_{\phi ,f_0}(w)$\
according to Formula $(3.5.1)$, since $(f_0: \phi )>0$. Thus the mapping $J_{\phi
,f_0}(\mbox{}_{b}f)$ is continuous in the parameter $b\in G$, since $0<J_{\phi ,f_0}(w)<\infty $ (see Lemmas 3.2, 3.7 and Corollary 3.3).
The case $J_{\phi ,f_0}(f_b)$ is proved symmetrically.
[**3.9. Theorem.**]{}
*Assume that $G$ is a $T_1$ topological locally compact fan loop satisfying Condition $(3.5.2)$, $\phi $, $f$ and $f_1$ are nonzero functions belonging to $C^+_{0,0}(G)$ and $f_0\in \Upsilon (G,N_0)$ (see $(3.5.3)$), then*
$(3.9.1)\quad (f_0:f ) ^{-1}\le J_{\phi ,f_0}(f) \le (f:f_0)$ and
$(3.9.2)\quad (f_1: f_0)^{-1}(f_0: f ) ^{-1}\le J_{\phi ,f_1}(f)
\le (f: f_0) (f_0: f_1)$.
[**Proof.**]{} The right inequality in $(3.9.1)$ follows from the inequality $(3.7.4)$.
Formulas $(3.6.7)$ and $(3.6.12)$ imply that
$(3.9.3)$ $(\mbox{}_bf_0: f)=
(f_0: f^{[\lambda ]})$ and $(\mbox{}_bf^{[\lambda ]}: \phi )=
(f^{[\lambda ]}: \phi ^{[\lambda ]})$ for each $b\in G$.
Let $c_1,...,c_k$, $h_1,...,h_n$ in $(0,\infty )$ and $a_1,..,a_k$, $g_1,...,g_n$ in $G$ be such that $$(3.9.4)\quad f_0(x)\le \sum_{j=1}^kc_jf^{[\lambda ]}(a_jx) \mbox{ and}$$ $$(3.9.5)\quad f^{[\lambda ]}(x)\le \sum_{i=1}^n h_i \phi ^{[\lambda ]}(g_ix)$$ for each $x\in G$ (see Lemma 3.2). Then from Identity $(2.1.9)$, Inequalities $(3.9.4)$, $(3.9.5)$ and Conditions $(3.5.3)$, $(3.5.4)$ we deduce that $$(3.9.6)\quad f_0(x)\le \sum_{j=1}^k c_j \sum_{i=1}^n h_i
\phi ^{[\lambda ]}((g_ia_j)x[p(g_i,a_j,x)]^{-1})=\sum_{j=1}^k c_j
\sum_{i=1}^n h_i \phi ^{[\lambda ]}((g_ia_j)x).$$ Suppose that there are $y_1,...,y_k\in G$ and $q_1,...,q_k\in ( 0, \infty )$ such that $$(3.9.7)\quad f(x)\le \sum_{i=1}^k q_i \phi (y_ix)$$ for each $x\in G$. Taking the integral $\int_{N_0}f(x\gamma )\lambda (d\gamma
)$ and similarly for the right side (see Formulas $(3.6.4)$ and $(3.6.4')$), we get from Inequality $(3.9.7)$ that $$f^{[\lambda ]}(x)\le \sum_{i=1}^k q_i
\phi ^{[\lambda ]}(y_ix)$$ for each $x\in G$ (see Lemma 3.2). Hence
$(3.9.8)$ $(f^{[\lambda ]}: \phi ^{[\lambda ]})\le (f: \phi )$.
Utilizing Formulas $(3.6.1)$, $(3.9.3)$ and $(3.9.8)$ we infer that
$(3.9.9)$ $(f_0: \phi )\le (f_0: f)(f^{[\lambda ]}: \phi ^{[\lambda ]})
\le (f_0: f)(f: \phi )$ for each $f_0\in \Upsilon (G,N_0)$ and nonzero functions $f$ and $\phi $ in $C_{0,0}^+(G)$.
Using $(3.5.1)$ and $(3.9.9)$ we infer that $$(f_0: f)J_{\phi ,f_0}(f)=\frac{(f_0: f)(f:\phi )}{(f_0:\phi )
}\ge \frac{(f_0: \phi )}{(f_0:\phi ) }=1,$$ consequently, $J_{\phi
,f_0}(f)\ge (f_0: f)^{-1}$. Thus the left inequality in $(3.9.1)$ also is proved.
From Inequalities $(3.9.1)$ for $J_{\phi ,f_0}(f)$ and $J_{\phi ,f_0}(f_1)$ and Formula $(3.5.1)$ it follows $(3.9.2)$.
[**3.10. Lemma.**]{} [*Let $G$ be a $T_1$ topological locally compact fan loop satisfying Condition $(3.5.2)$, let $f_0\in \Upsilon (G,N_0)$ (see Condition $(3.5.3)$) and let $f_1$,...,$f_m$ be nonzero functions belonging to $C_{0,0}^+(G)$, let also $0<\delta <\infty $, $0<\delta
_1<\infty $. Then there exists an open neighborhood $V$ of $e$ in $G$ such that for each nonzero function $\phi $ in $C_{0,0}^+(G)$ with a support $S_{\phi }$ contained in $V$ and $0\le q_j\le \delta
_1$ for each $j=1,...,m$ the following inequality is satisfied: $$(3.10.1)\quad \sum_{j=1}^mq_jJ_{\phi ,f_0}(f_j)\le J_{\phi
,f_0}(\sum_{j=1}^mq_jf_j)+\delta .$$*]{}
[**Proof.**]{} The loop $G$ is locally compact. Let $S_{f_0,...,f_m}=\bigcup_{j=0}^m S_{f_j}$ be a common compact support of these functions, where $S_{f_j}$ denotes a closed support of $f_j$ (see also Subsection 3.1). We choose any function $g_1$ in $C_{0,0}^+(G)$ such that $g_1: G\to [0,1]$ and $g_1(S_{f_0,...,f_m}cl_G(Y_1))= \{ 1 \} $, where ${Y'}_1$ is an open neighborhood of $e$ in $G$ with $Y_1=\check{Y'}_1$ and a compact closure $cl_G(Y_1)$ (see Lemma 2.6). Consider arbitrary fixed positive numbers $0<\delta <\infty $, $0<\delta _1<\infty $ and $0<\epsilon <M$ such that
$\epsilon \delta _1 \sum_{j=1}^m(f_j:
~ f_0 ) +\epsilon (1+\epsilon )(g_1: ~ f_0)\le \delta $, where $M=\delta _1 m \max _{j=1,...m} \| f _j \|_G$. By virtue of Corollary 2.15 the functions $f_0,...,f_m$ are uniformly $({\cal
L}_G, {\cal L}_H)$ continuous, where $H=({\bf C}, +)$. Therefore there exists an open neighborhood $W'$ of $e$ with $W=\check{W'}$ and a compact closure $cl_G(W)$ in $G$ and $W\subset Y_1$, since $G$ is locally compact, such that
$(3.10.2)$ $|f_j(s)-f_j(x)|<\epsilon ^3[4Mm\delta _1]^{-1}$\
for each $s\setminus x\in W$. Next we take a function $g\in C^+_{0,0}(G)$ such that $g: G\to
[0,1]$ and $g(S_{f_0,...,f_m}cl_G(W))= \{ 1 \} $ and $g(x)\le
g_1(x)$ for each $x\in G$, because $W\subset Y_1$. Hence $(g:
f_0)\le (g_1: f_0)$ by Inequality $(3.4.5)$.
Let $S=P((S_{f_0,...,f_m}\cup
S_g)cl_G(W))$ (see Formula $(2.14.1)$). Since $cl_G(V)$, $S_{f_0,...,f_m}$ and $S_g$ are compact, then $S$ is a compact subset in $G$. For each open neighborhood $Y$ of $e$ in $G$ there exists an open neighborhood $X$ of $e$ in $G$ such that $X^2\subset
Y$, since the multiplication in $G$ is continuous. In view of Proposition 2.9 and Corollary 2.15 there exist open neighborhoods ${U'}_k$ of $e$ in $G$ such that $U_k=\check{U'}_k$ and
$(3.10.3)$ $[t((U_ka)U_k,(U_kb)U_k,(U_kc)U_k)U_k]
\cup [U_kt((U_ka)U_k,(U_kb)U_k,(U_kc)U_k)]$
$\subset [t(a,b,c)
W_{k-1}]\cap [ W_{k-1}t(a,b,c)]$
$[p((U_ka)U_k,(U_kb)U_k,(U_kc)U_k)U_k]
\cup [U_kp((U_ka)U_k,(U_kb)U_k,(U_kc)U_k)]$
$\subset [p(a,b,c)
W_{k-1}]\cap [ W_{k-1}p(a,b,c)]$\
for every $a$, $b$, $c$ in $S$ and $k\in \{ 1, 2 \} $ with $U_0=W$ and an open of the form $(2.6.1)$ neighborhood $W_{k-1}$ of $e$ in $G$ such that $\check{W}_{k-1}^2\subset U_{k-1}$ and
$(3.10.4)$ $|g(s)-g(x)|<\epsilon ^2[4M]^{-1}$\
for each $s$ and $x$ in $G$ such that $s\setminus x\in U_1$, where $t=t_G$.
Take any $0\le q_j\le \delta _1$ for each $j=1,...,m$ and put
$(3.10.5)$ $\Psi = \epsilon g + \sum_{j=1}^m q_jf_j$ and
$(3.10.6)$ $h_j(x)=q_jf_j(x)[\Psi (x)]^{-1}$ for each $x\in S_{f_1,...,f_m}$ and
$h_j(x)=0$ for each $x\in G-
S_{f_1,...,f_m}$,\
where $S_{f_1,...,f_m}=\bigcup_{j=1}^m
S_{f_j}$. Therefore the function $\Psi $ belongs to $C^+_{0,0}(G)$ and $\sum_{j=1}^m h_j(x)\le 1$ for each $x\in G$.
From Inequalities $(3.10.2)$ and $(3.10.4)$ it follows that
$(3.10.7)$ $|\Psi (s)-\Psi (x)| \le \epsilon ^3[2M]^{-1}$\
for each $s$ and $x$ in $G$ such that $s\setminus x\in U_1$. Moreover, $\| \Psi \|_G \le M+\epsilon <2M$.
Let $s$ and $x$ belong to $S_{f_1,...,f_m}cl_G(W)$ and $s\setminus x\in U_1$. The latter inclusion is equivalent to $x\in sU_1$ and also to $s\in x/U_1$. Then from $(3.10.2)$ and $(3.10.7)$ we deduce that
$(3.10.8)$ $|h_j(s)-h_j(x)|\le \epsilon /m$.
Next we consider the following case: $s\setminus x\in U_1$ and $x\notin
S_{f_1,...,f_m}cl_G(W)$. Suppose that $s\in S_{f_1,...,f_m}$, then Condition $(3.10.3)$, Lemmas 2.2, 2.3 imply that $x\in
S_{f_1,...,f_m}cl_G(W)$ contradicting an assumption $x\notin
S_{f_1,...,f_m}cl_G(W)$. Hence $s\notin S_{f_1,...,f_m}$ and consequently, $h_j(s)=0$ and $h_j(x)=0$. Thus Inequality $(3.10.8)$ takes place in this case as well.
In the case $s\setminus x\in
U_1$ and $s\notin S_{f_1,...,f_m}cl_G(W)$ Condition $(3.10.3)$, Lemmas 2.2, 2.3 imply that $x\notin S_{f_1,...,f_m}$. Therefore the inequality $(3.10.8)$ is fulfilled in this case also. Thus the estimate $(3.10.8)$ is satisfied for each $s$ and $x$ in $G$ such that $s\setminus x\in U_1$.
Next we choose any fixed function $\phi \in C^+_{0,0}(G)$ such that $\phi $ is not identically zero on $G$ and $\phi (y)=0$ for each $y\in G- {U'}_2$. By virtue of Lemma 3.2 there are $m\in \bf
N$, $c_j>0$ and $b_j\in G$ for each $j\in \{ 1,...,m \} $ such that $$(3.10.9) \quad \Psi (x)\le \sum_{j=1}^mc_j\phi (b_jx)$$ for each $x\in G$ and $$(3.10.10) \quad -\epsilon +\sum_{j=1}^mc_j\le (\Psi : ~ \phi )\le
\sum_{j=1}^mc_j.$$ Then Formulas $(3.10.3)$, $(3.10.8)$, $(3.10.9)$ and Lemma 2.2 imply that for each $x\in G$ $$\Psi (x)h_l(x)\le \sum_{j=1}^m c_j \phi (b_jx)[h_l(b_j\setminus e
) +\epsilon /m ]$$ for each $l$. Hence for each $x\in G$ we get $$q_lf_l(x)=\Psi (x)h_l(x)\le \sum_{j=1}^m c_j
[h_l(b_j\setminus e ) +\epsilon /m ]\phi (b_jx)$$ and consequently, $(q_lf_l: ~ \phi )\le \sum_{j=1}^m c_j [h_l(b_j\setminus e )
+\epsilon /m ]$. From $\sum_{l=1}^m h_l\le 1$ we deduce that $\sum_{l=1}^m(q_lf_l: ~ \phi )\le (1+\epsilon ) \sum_{j=1}^m c_j$. Together with Inequalities $(3.10.10)$ this leads to the following estimate: $$\sum_{j=1}^m (q_if_j: ~ \phi )\le (1+\epsilon ) (\Psi : ~ \phi ).$$ Dividing both sides of it on $(f_0: ~ \phi )$ we get the inequality $$(3.10.11)\quad \sum_{j=1}^mq_jJ_{\phi ,f_0}(f_j)\le (1+\epsilon )
J_{\phi ,f_0}(\Psi ).$$ Then from $(3.7.1)$, $(3.7.2)$, $(3.10.5)$ and $(3.10.11)$ we infer that $$(3.10.12)\quad
\sum_{j=1}^mq_jJ_{\phi ,f_0}(f_j)\le J_{\phi
,f_0}(\sum_{j=1}^mq_jf_j)+\epsilon \sum_{j=1}^mq_jJ_{\phi ,f_0}(f_j)
+\epsilon (1+\epsilon )J_{\phi ,f_0}(g).$$ Therefore from Inequalities $(3.9.1)$, $(3.10.12)$, $(3.4.5)$ and for $\epsilon $ as above it follows that $$\sum_{j=1}^mq_jJ_{\phi
,f_0}(f_j)\le J_{\phi ,f_0}(\sum_{j=1}^mq_jf_j)+\epsilon \delta _1
\sum_{j=1}^m(f_j: ~ f_0 )$$ $$+\epsilon (1+\epsilon )(g: ~ f_0) \le
J_{\phi ,f_0}(\sum_{j=1}^mq_jf_j)+ \delta .$$ This implies the estimate $(3.10.1)$ with $V={U'}_2$.
[**3.11. Theorem.**]{}
*Let $G$ be a $T_1$ topological locally compact fan loop, let $0<\epsilon $ and $f$ in $C_{0,0}^+(G)$ be a nonzero function, $S_f=cl_G \{ x\in G: ~ f(x)\ne 0 \} $. Let also $V'$ be an open neighborhood of $e$ in $G$ such that $V=\check{V'}$ and*
$(3.11.1)$ $|f(x)-f(y)|<\epsilon $ for each $x$ and $y$ in $G$ with $x\setminus y\in V$. Let $g\in C_{0,0}^+(G)$ be a nonzero function such that $g(x)=0$ for each $x\in G- V'$. Then for each $\delta >
\epsilon $ and each open neighborhood ${W'}_e$ of $e$ in $G$ with $W_e=\check{W'}_e$ and a compact closure $cl_G(W_e)$ contained in $V$ there is an open neighborhood $U'$ of $e$ in $G$ such that $U=\check{U'}$ and for each nonzero function $\phi $ in $C_{0,0}^+(G)$ with a support $S_{\phi }$ contained in $U'$ there are positive constants $c_1,...,c_n$ and elements $b_1,...,b_n$ in $S_fcl_G(W_e)$ such that for each $x\in G$ and $\gamma \in N(G)$: $$(3.11.2)\quad |f(\gamma x) - \sum_{j=1}^n \frac{c_j}{J^v_{\phi
,f_0}(g(v\setminus x)) } g(b_j\setminus \gamma x)|\le \delta ,$$ where an expression $J^v_{\phi ,f_0}(g(v\setminus x))$ means that a functional $J_{\phi ,f_0}$ is taken in the $v$ variable.
[**Proof.**]{} The continuous functions $f$ and $g$ are with compact supports, hence they are uniformly $({\cal
L}_G, {\cal L}_H)$ continuous and uniformly $({\cal R}_G, {\cal
R}_H)$ continuous on $G$ by Corollary 2.15, where $H=({\bf C}, +)$. For each $y\in G$ the right translation operator $R_y$ is the homeomorphism of $G$ as the topological space onto itself (see also Subsection 2.5). Therefore a function $\nu (y) := (f(x): ~
g(x\setminus y))$ is continuous on the loop $G$ and consequently, uniformly continuous on the compact subset $S_f$, hence $\sup_{y\in
S_f} \nu (y)< \infty $, where $(f(x): ~ g(x\setminus y))=(f: z)$ is calculated in the $x$ variable with $z(x)=g(x\setminus y)$ for a fixed parameter $y$. We take any fixed $\delta $ such that $\epsilon
<\delta <\infty $. Evidently there exists $0<\eta $ such that
$(3.11.3)$ $\eta \sup_{y\in S_f} \nu (y)<\delta - \epsilon $.
Therefore take any fixed open neighborhood ${W'}_e$ of $e$ in $G$ such that $W_e=\check{W'}_e$ and $cl_G(W_e)$ is compact and $cl_G(W_e)\subset V$ (see Lemma 2.6). By virtue of Corollary 2.15 the functions $g$ and $h$ are uniformly $({\cal L}_G, {\cal L}_H)$ continuous and uniformly $({\cal R}_G,
{\cal R}_H)$ continuous. Hence there exists an open neighborhood ${W'}_1$ of $e$ in $G$ such that $W_1=\check{W'}_1$ and $cl_G(W_1)$ is compact and $cl_G(W_1)\subset {W'}_e$ and for each $x$ and $y$ in $G$ with $x\setminus y\in W_1$:
$(3.11.4)$ $|g(x)-g(y)|<\eta $.
Therefore, a subset $S_f cl_G(W_1)$ is compact in $G$ (see Theorems 3.1.10, 8.3.13-8.3.15 in [@eng], Lemma 2.6). Then we take compact subsets $S_1=S_f cl_G(W_1)$ and $S=P(S_f cl_G(W_1))$ in $G$ (see Formula $(2.14.1)$). In view of Lemma 2.6 they contain open subsets $S_fW_1$ and $P(S_fW_1)$ respectively, since $W_1$ is open in $G$. Mention that the topological spaces $S_1$ and $S$ are normal, since they are compact and $T_1\cap T_{3.5}$ (see Theorem 3.1.9 in [@eng]). Using Proposition 2.9 we take an open neighborhood ${W'}_2$ of $e$ in $G$ with $W_2=\check{W'}_2$ such that\
$(3.11.5)$ $[t((W_2a)W_2,(W_2b)W_2,(W_2c)W_2)W_2]\cup
[W_2t((W_2a)W_2,(W_2b)W_2,(W_2c)W_2)]$
$ \subset
[t(a,b,c)W_3]\cap [W_3t(a,b,c)]$ and
$[p((W_2a)W_2,(W_2b)W_2,(W_2c)W_2)W_2]\cup
[W_2p((W_2a)W_2,(W_2b)W_2,(W_2c)W_2)] $
$\subset
[p(a,b,c)W_3]\cap
[W_3p(a,b,c)]$\
for every $a$, $b$, $c$ in $S$, where $W_3$ is an open neighborhood of $e$ in $G$ such that $\check{W}_3^2\subset W_1$.
In view of the Dieudonné theorem 3.1 in [@hew] there exists a partition of unity on $S_1$. Together with Theorem 3.3.2 in [@eng] and Lemma 2.5 this implies that there are functions $q_1,...,q_n$ in $C_{0,0}^+(G)$ and elements $w_1,...,w_n$ in $S_1$ such that $S_1\subset \bigcup_{j=1}^n w_jW_2$ and
$(3.11.6)$ $\sum_{j=1}^nq_j(x)=1$ for each $x\in S_1$ and
$(3.11.7)$ $q_j(y)=0$ for each $y\in G- (w_jW_2)$.
The conditions of this theorem imply that for each $x$ and $y$ in $G$ with $y\setminus x\in V$ the following inequalities are satisfied:
$(3.11.8)$ $[f(x)-\epsilon ]g(y\setminus x)\le
f(y)g(y\setminus x)\le [f(x) + \epsilon ]
g(y\setminus x)$,\
since for $y\setminus x\in V$ Inequality $(3.11.1)$ is fulfilled; for $u=y\setminus x\notin V$ the function $g$ is nil, $g(u)=0$.
Certainly $y\in w_jW_2$ if and only if there exists $b\in W_2$ such that $y=w_jb$. Then $(y\setminus x)\setminus (w_j\setminus
x)\in W_1$ if and only if there exists $c\in W_1$ such that $w_j\setminus x=((w_jb)\setminus x)c$. For $w_j\setminus x=v\in V$ this gives $c=((w_jb)\setminus (w_jv))\setminus v$. In view of $(2.2.2')$, $(2.2.4)$, $(2.1.8)$ and $(2.1.9)$
$((w_jb)\setminus (w_jv))\setminus v=p(w_j,b,(w_jb)\setminus
(w_jv))((b\setminus v)\setminus v)$.
Therefore, from Conditions $(3.11.5)$-$(3.11.7)$ it follows that for each $x$ and $y$ in $G$ and $j=1,...,n$:
$(3.11.9)$ $q_j(y)f(y)[g(y\setminus x)-\eta ] \le q_j(y)
f(y)g(w_j\setminus x) $
$\le q_j(y)f(y)[g(y\setminus x)+\eta ]$.\
Summing by $j$ in $(3.11.9)$, using $(3.11.8)$ we infer that for each $x$ and $y$ in $G$:
$(3.11.10)$ $[f(x)-\epsilon
]g(y\setminus x)- \eta f(y)$
$\le \sum_{j=1}^n
q_j(y)f(y)g(w_j\setminus x)\le [f(x)+\epsilon ]g(y\setminus x)+ \eta
f(y)$.
Next we take any $\phi $ and $f_0$ in $C_{0,0}^+(G)$ such that $\phi $ and $f_0$ are not identically zero. From Inequalities $(3.11.10)$ after dividing on $J^y_{\phi ,f_0}(g(y\setminus x))$ and Lemma 3.7 it follows that for each $x$ in $G$: $$(3.11.11)\quad [f(x)-\epsilon ] -\eta \frac{ J_{\phi
,f_0}(f)}{J^y_{\phi ,f_0}(g(y\setminus x))}\le J^y_{\phi
,f_0}(\frac{\sum_{j=1}^ng(w_j\setminus x)q_j(y)f(y)}{J^v_{\phi
,f_0}(g(v\setminus x)) })$$ $$\le [f(x)+\epsilon ] +\eta \frac{
J_{\phi ,f_0}(f)}{J^y_{\phi ,f_0}(g(y\setminus x))},$$ where $J^y_{\phi ,f_0}(g(y\setminus u))=J_{\phi ,f_0}(z)$ means that the functional $J^y_{\phi ,f_0}$ is taken in the $y$ variable in $G$, where $z(y)=g(y\setminus u)$ for each $y\in G$ and a fixed parameter $u$ in $G$.
Notice that the function $g(y\setminus x)$ is jointly continuous in $(x,y)\in G\times G$. On the other hand, in view of Lemmas 2.2, 2.4, 2.6
$\{ u=y\setminus x: ~ x\in S_f, ~ u\in S_g
\} $ is a compact subset in $G$,\
since $Inv_l(S_f)$, $S_g$, $S_fS_g$ and $t(S_f,Inv_l(S_f),S_fS_g)$ are compact subsets in $G$. By virtue of Lemma 3.8 a mapping $\psi (x):=J^y_{\phi
,f_0}(g(y\setminus x))$ is continuous in the variable $x\in S_f$, $
~ \psi : S_f \to (0,\infty )$. Hence
$(3.11.12)$ $0<K_0=\inf_{x\in S_f}\psi (x) \le \sup_{x\in S_f}\psi (x)=K_1<\infty
$.
Apparently in Formula $(3.11.3)$ the parameter $\eta >0$ can be taken sufficiently small, because Inequalities $(3.11.3)$ and $(3.11.12)$ are independent. Then from $(3.11.11)$ and $(3.11.12)$ we deduce that for each $\beta
>\epsilon $ there exist $q_j$ and $w_j$ (see above) such that
$\eta J_{\phi ,f_0}(f) <(\beta - \epsilon )\min (1, K_0)$, consequently, $$(3.11.13)\quad f(x)-\beta \le J^y_{\phi
,f_0}(\frac{\sum_{j=1}^ng(w_j\setminus x)q_j(y)f(y)}{J^v_{\phi
,f_0}(g(v\setminus x)) })\le f(x)+\beta$$ for each $x\in G$.
In view of Lemmas 3.7 and 3.10 for each $\delta >\delta _1>\beta >\epsilon $ there exists an open of the form $(2.6.1)$ neighborhood $U$ of $e$ in $G$ such that $U\subset W_2$ and $$(3.11.14)\quad | J^y_{\phi
,f_0}(\frac{\sum_{j=1}^ng(w_j\setminus x)q_j(y)f(y)}{J^v_{\phi
,f_0}(g(v\setminus x)) }) -\sum_{j=1}^n \frac{J_{\phi
,f_0}(q_jf)}{J^v_{\phi ,f_0}(g(v\setminus x)) }g(w_j\setminus
x)|<\delta _1 - \beta$$ for each $x\in S_f$. We put $c_j =J_{\phi ,f_0}(q_jf)$ and $b_j=w_j$ for each $j=1,...,n$. Thus the estimates $(3.11.13)$ and $(3.11.14)$ and Formula $(3.4.1')$ imply the assertion of this theorem.
[**3.12. Definition.**]{} Let $W$ be an open neighborhood of $e$ in a locally compact loop $G$ and a nonzero function $\phi
_{W}\in C^+_{0,0}(G)$ be such that $\phi _{W}(x)=0$ for each $x\in
G- W$. A family $ \{ \phi _W \} $ of these functions will be directed by:
$(3.12.1)$ $\phi _{W_1}\preceq \phi _{W_2}$ if and only if $W_2\subseteq W_1$ and $\phi _{W_2}(x)=0$ implies $\phi
_{W_1}(x)=0$. If $\phi _{W_1}\preceq \phi _{W_2}$ and $\phi _{W_1}$ and $\phi _{W_2}$ are different functions, then it will be written $\phi _{W_1}\prec \phi _{W_2}$.
[**3.13. Lemma.**]{}
*Let $G$ be a $T_1$ topological locally compact fan loop satisfying Condition $(3.5.2)$ and let a family of nonzero functions $ \{ \phi _U \} $ in $C^+_{0,0}(G)$ be directed by Condition $(3.12.1)$. Let also $f_0\in \Upsilon (G,N_0)$ (see $(3.5.3)$) and $f \in C^+_{0,0}(G)$. Then the limit exists:*
$(3.13.1)$ $\lim_{\{ \phi _U \} } J_{\phi _U, f_0}(f)=:J_{f_0}(f)$.
[**Proof.**]{} It is sufficient to prove that a net $\{ J_{\phi _U, f_0}(f): \phi _U \} $ is fundamental (i.e. Cauchy) in $\bf R$, where a net $ \{ \phi _U \} $ is directed by Condition $(3.12.1)$. We take any fixed open neighborhood ${U'}_0$ of $e$ in $G$ with $U_0=\check{U'}_0$ and a compact closure $cl_G(U_0)$. Let $A=S_{f+f_0}cl_G(U_0)$, where $S_{f+f_0}=cl_G \{ x\in G: ~
f(x)+f_0(x)\ne 0 \} $. Therefore, a subset $S=P(A)$ is compact (see Formula $(2.14.1)$ and Lemma 2.6), since $S_{f+f_0}$ is compact.
We choose any function $z\in C_{0,0}^+(G)$ such that $z|_A=1$. Let $0<\epsilon <1$ and $\xi _1 = \epsilon (16[1+(z:f_0)]
[1+(f:f_0)])^{-1}$. From Corollary $(2.15)$ it follows that there exists an open neighborhood $W'$ of $e$ in $G$ with $W=\check{W'}$ such that
$(3.13.2)$ $|f(x)-f(y)|<\xi _1 /2$ and
$(3.13.3)$ $|f_0(x)-f_0(y)|<\xi _1 /2$\
for each $x$ and $y$ in $G$ with $x\setminus y\in W$.
In view of Proposition 2.9 there exists an open neighborhood ${U'}_2$ of $e$ in $G$ with $U_2=\check{U'}_2$ such that
$(3.13.4)$ $[t((U_2a)U_2,(U_2b)U_2,(U_2c)U_2)U_2]\cup
[U_2t((U_2a)U_2,(U_2b)U_2,(U_2c)U_2)]$
$ \subset
[t(a,b,c)B_1]\cap [B_1t(a,b,c)]$ and
$[p((U_2a)U_2,(U_2b)U_2,(U_2c)U_2)U_2]\cup
[U_2p((U_2a)U_2,(U_2b)U_2,(U_2c)U_2)]$
$ \subset
[p(a,b,c)B_1]\cap [B_1p(a,b,c)]$\
for every $a$, $b$, $c$ in $S$, where $B_1$ is an open neighborhood of $e$ in $G$ such that $\check{B}_1^2\subset U_1$, $U_1={U'}_0\cap
W'$ (see Lemma 2.6). Next we take a nonzero function $g\in
C_{0,0}^+(G)$ such that $g(x)=0$ for each $x\in G- {U'}_2$.
By virtue of Theorem 3.11 for any fixed $0<\delta <\xi _1 $ and each open neighborhood ${W'}_e$ of $e$ in $G$ with $W_e=\check{W'}_e$ and a compact closure $cl_G(W_e)$ contained in ${U'}_2$ there is an open neighborhood ${U'}_{3,f}$ of $e$ in $G$ with $U_{3,f}=\check{U'}_{3,f}$ such that for each nonzero function $\phi $ in $C_{0,0}^+(G)$ with a support $S_{\phi }$ contained in ${U'}_{3,f}$ there are positive constants $c_1,...,c_n$ and elements $b_1,...,b_n$ in $S_fcl_G(W_e)$ such that for each $x\in G$ and $\gamma \in N(G)$: $$(3.13.5)\quad |f(\gamma x) - \sum_{j=1}^n \frac{c_j}{J^v_{\phi
,f_0}(g(v\setminus x)) } g(b_j\setminus \gamma x)|\le \delta .$$ Taking $U_{3,f}\subset {U'}_2$ we get $f(x)=0$ and $g(b_j\setminus
x)=0$ for each $x\in G- A$ according to the choice of $b_j$ in the proof of Theorem 3.11, consequently, $$(3.13.6)\quad |f(\gamma x) - \sum_{j=1}^n \frac{c_j}{J^v_{\phi
,f_0}(g(v\setminus x)) } g(b_j\setminus \gamma x)|\le \delta
z(\gamma x)$$ for each $x\in G$ and $\gamma \in N(G)$. From the latter estimate and Lemma 3.7 we infer that $$(3.13.7)\quad |J_{\phi ,f_0}(f) - K_{\phi ,f_0}(f;g)|\le
\delta J_{\phi ,f_0}(z)\le \delta (z: f_0), \mbox{ where}$$ $$K_{\phi ,f_0}(f;g)= J^x_{\phi ,f_0} (\sum_{j=1}^n
\frac{c_j}{J^v_{\phi ,f_0}(g(v\setminus x)) } g(b_j\setminus x)).$$ From Estimate $(3.13.7)$ and the right Inequality $(3.9.1)$ it follows that $$(3.13.7')\quad \sup_{\{ \phi _U \} }K_{\phi _U,f_0}(f;g)\le (1+\delta ) (f:f_0)+\delta
(z:f_0)<\infty .$$ Applying the proof above to $f_0$ instead of $f$ we get and open neighborhood ${U'}_{3,f_0}$ of $e$ with $U_{3,f_0}=\check{U'}_{3,f_0}$ and $U_{3,f_0}\subset {U'}_2$ such that for each nonzero function $\phi $ in $C_{0,0}^+(G)$ with a support $S_{\phi }$ contained in ${U'}_{3,f_0}$ there are positive constants $d_1,...,d_m$ and elements $v_1,...,v_m$ in $S_{f_0}cl_G(W_e)$ such that $$(3.13.8)\quad |f_0(\gamma x) - \sum_{j=1}^m \frac{d_j}{J^v_{\phi
,f_0}(g(v\setminus x)) } g(v_j\setminus \gamma x)|\le \delta
z(\gamma x)$$ for each $x\in G$ and $\gamma \in N(G)$, consequently, $$(3.13.9)\quad |1- K_{\phi ,f_0}(f_0;g)|\le
\delta (z: f_0), \mbox{ where}$$ $$K_{\phi ,f_0}(f_0;g)= J^x_{\phi ,f_0} (\sum_{j=1}^m
\frac{d_j}{J^v_{\phi ,f_0}(g(v\setminus x)) } g(v_j\setminus x)),$$ since $J_{\phi ,f_0}(f_0)=1$. Moreover, $$(3.13.10)\quad
\sup_{\{ \phi _U \} }K_{\phi _U,f_0}(f_0;g)\le (1+\delta )+ \delta
(z:f_0)<\infty .$$ Then ${U'}_3={U'}_{3,f}\cap {U'}_{3,f_0}$ is an open neighborhood of $e$ in $G$. From $(3.13.7)$, $(3.13.9)$ and $(3.13.10)$ we deduce that $$(3.13.11)\quad |J_{\phi ,f_0}(f) -
\frac{K_{\phi ,f_0}(f;g)}{K_{\phi ,f_0}(f_0;g)}| \le \delta
_2+[1+\delta +\delta _2]\delta _2(1-\delta _2)^{-1},$$ where $\delta
_2=\delta (z: f_0)<\xi _1 (z: f_0)<1/16$. In view of Lemmas 3.7 and 3.10, Formulas $(3.13.5)$ and $(3.13.6)$ there exists an open neighborhood ${U'}_4$ of $e$ with $U_4=\check{U'}_4$ and $U_4$ contained in ${U'}_3$ such that for each nonzero $\phi $ in $C_{0,0}^+(G)$ with $S_{\phi }\subset {U'}_4$ there are inequalities: $$(3.13.12)\quad |K_{\phi ,f_0}(f;g)- \sum_{j=1}^nc_jJ^x_{\phi ,f_0}
(\frac{g(b_j\setminus x)}{J^v_{\phi ,f_0}(g(v\setminus \gamma x))})
| \le \delta \mbox{ and}$$ $$(3.13.13)\quad |K_{\phi ,f_0}(f_0;g)-
\sum_{j=1}^md_jJ^x_{\phi ,f_0}(\frac{g(v_j\setminus x)}{J^v_{\phi
,f_0}(g(v\setminus \gamma x))}) | \le \delta$$ for each $\gamma \in
N(G)$. On the other hand, Formulas $(3.5.1)$, $(3.6.4)$, $(3.6.4')$, $(3.6.12)$ and $(3.4.3)$ imply that $$(3.13.14)\quad J^x_{\phi ,f_0}(\frac{g(b_j\setminus x)}{J^v_{\phi ,f_0}(g(v\setminus \gamma x))})=
\int_{N_0}(\frac{g(b_j\setminus x)}{(g(v\setminus \gamma x):\phi
(v))}: \phi (x))\lambda (d\gamma )$$ $$=
\frac{(g(b_j\setminus x): \phi (x))}{(g^{[\lambda ]}(v\setminus e): \phi
(v))}.$$ Then from Proposition $(2.9.1)$ and Formulas $(3.4.2)$, $(3.4.2')$ it follows that for each $b\in G$ and each $0<\delta
_3\le \delta $ there exists an open neighborhood ${U'}_{5,b}$ of $e$ in $G$ with $U_{5,b}=\check{U'}_{5,b}$ such that for each nonzero $\phi _U\in C_{0,0}^+(G)$ with $S_{\phi _U}\subset U\subset
{U'}_{5,b}$ $$(3.13.15)\quad |\frac{(g(b\setminus x):\phi
_U(x))}{(g^{[\lambda ]}(v\setminus e):\phi _U(v))}-\frac{(g(x): \phi
_U(x))}{(g^{[\lambda ]}(v\setminus e): \phi _U(v))}|<\delta _3,$$ since $S_{\phi _U}\subset U$ and $t(a,b,e)=t(a,e,b)=t(e,a,b)=e$ and $p(a,b,e)=p(a,e,b)=p(e,a,b)=e$ for each $a$ and $b$ in $G$. Therefore we take ${U'}_5=\bigcap_{j=1}^n {U'}_{5,b_j}\cap
\bigcap_{k=1}^m{U'}_{5,v_k}\cap {U'}_4$ and $\phi =\phi _Y$ with $Y={U'}_5$. We put $c=\sum_{j=1}^nc_j$ and $d=\sum_{k=1}^md_k$. From $(3.13.11)-(3.13.15)$ and $(3.9.1)$ it follows that $$\frac{c}{d}< K_1, \mbox{ where }K_1=3[1+(f:f_0)] (1+\delta
)(1-\delta )^{-1}<4[1+(f:f_0)].$$ Then we deduce from Formulas $(3.13.11)$-$(3.13.15)$ for each $\phi _U$ with an open neighborhood $U$ of $e$ in $G$ such that $U\subset {U'}_5$: $$|J_{\phi _U,f_0}(f) - \frac{c}{d}|< \delta (1-\delta
)^{-1}4[1+(f:f_0)]+\delta _2+[1+\delta +\delta _2]\delta _2
(1-\delta _2)^{-1},$$ consequently,
$(3.13.16)$ $|J_{\phi
_{V_1},f_0}(f) - J_{\phi _{V_2},f_0}(f) |<8 \delta (1-\delta
)^{-1}[1+(f:f_0)]$
$+2\delta _2+2[1+\delta +\delta _2]\delta _2(1-\delta _2)^{-1}<\epsilon $\
for each open neighborhoods $V_1$ and $V_2$ of $e$ in $G$ such that $V_1 \subset {U'}_5$ and $V_2\subset {U'}_5$. Thus the net $ \{
J_{\phi _U,f_0}(f) : \phi _U \} $ is fundamental, where the net $ \{
\phi _U \} $ is directed by Condition $(3.12.1)$.
[**3.14. Remark.**]{} Suppose that $G$ is a $T_1$ topological locally compact fan loop and Condition $(3.5.2)$ is fulfilled and $f_0\in \Upsilon (G,N_0)$ (see $(3.5.3)$), functions $f$ and $g$ belong to $C^+_{0,0}(G)$ and $g$ is nonzero. Then in view of Lemma 3.13 a functional exists
$(3.14.1) \quad
J_{g}(f)=J_{f_0}(f)/J_{f_0}(g)$.
As a consequence of Lemma 3.13 and Formulas $(3.5.1)$ and $(3.14.1)$ we get that
$(3.14.2)$ the functional $J_{g}(f)$ is independent of $f_0$.
Then Formula $(3.9.2)$ and Lemma 3.13 imply that
$(3.14.3) \quad (g: f_0)^{-1}(f_0: f )
^{-1}\le J_{g}(f) \le (f: f_0) (f_0: g)$\
for each $f_0 \in
\Upsilon (G,N_0)$ and a nonzero function $f\in C^+_{0,0}(G)$.
[**3.15. Theorem.**]{}
*Let $G$ be a $T_1$ topological locally compact fan loop fulfilling Condition $(3.5.2)$ and a functional $J=J_g$ be defined by Formula $(3.14.1)$. Then $J$ possesses the following properties:*
$(3.15.1)$ $J(f)\ge 0$ for each $f\in C^+_{0,0}(G)$; and if a function $f$ is nonzero, then $J(f)>0$;
$(3.15.2)$ $J(\alpha _1f_1+...+\alpha _nf_n)=
\alpha _1J(f_1)+...+\alpha _nJ(f_n)$ for each $f_1,...,f_n$ in $C^+_{0,0}(G)$ and $\alpha _1\ge
0$,...,$\alpha _n\ge 0$;
$(3.15.3)$ $J(\mbox{}_bf)=J(f)$ for each $b\in G$ and $f\in C^+_{0,0}(G)$.
[**Proof.**]{} Property $(3.15.1)$ follows from Formula $(3.14.3)$. On the other hand, Lemmas 3.7, 3.10, 3.13 imply Equality $(3.15.2)$.
Then Formulas $(3.6.4)$, $(3.6.4')$, $(3.6.12)$, $(3.14.1)$ and Lemma 3.13 imply
$(3.15.4)$ $J(\mbox{}_bf^{[\lambda ]})=J(f^{[\lambda ]})$ for each $b\in G$ and $f$ in $C^+_{0,0}(G)$.
As a topological space $G$ is locally compact. According to the measure theory on locally compact spaces (see Chapter 3, Section 11 in [@hew]) a functional $J$ on $C^+_{0,0}(G)$ satisfying Conditions $(3.15.1)$ and $(3.15.2)$ induces
$(3.15.5)$ a regular $\sigma $-additive measure $\mu $ on a Borel $\sigma $-algebra ${\cal B}(G)$ of $G$ such that $\mu (U) =
\sup \{ \mu (X): ~ X \mbox{ is compact}, ~ X\subset U \} $\
for each open subset $U$ in $G$ and
$(3.15.6)$ $\mu (A) = \inf \{ \mu (V): ~ V \mbox{ is open},
~ A\subset V\subset G \} $\
for each $A\in {\cal B}(G)$ and
$(3.15.7)$ $J(f)=\int_G f(x) \mu (dx)$ for each $f\in
C^+_{0,0}(G)$ and
$(3.15.8)$ the functional $J$ has an extension $\bar{\bar{J}}$ such that $\bar{\bar{J}}(f)=\int_Gf(x)\mu
(dx)$ for each nonnegative $\mu $-measurable function $f$ on $G$, where
$\bar{\bar{J}}(f)=\inf \{ \bar{J}(h): ~ h\ge f, ~ h \mbox{
is lower semicontinuous } \} ,$
$\bar{J}(h)= \sup \{ J(p): ~
p\in C^+_{0,0}(G), ~ p\le h \} $\
(see Theorems 11.22, 11.23, 11.36 and Corollary 11.37 in [@hew]).
On the other hand, for each $\gamma \in N(G)$ Formulas $(3.4.1')$ and $(3.4.2')$ give
$(3.15.9)$ $(\mbox{}_{\gamma }f: \phi _U)= (f: \mbox{}_{\gamma }\phi _U)=(f:
\phi _U)$.
From Lemma 3.13, Formulas $(3.14.1)$ and $(3.15.9)$ we deduce that
$(3.15.10)$ $J(\mbox{}_{\gamma }f)=J(f)$ for each $\gamma \in N_0(G)$.
By virtue of the Fubini theorem 13.8 in [@hew], $(3.5.3)$, $(3.5.4)$, Formulas $(3.15.4)$, $(3.15.7)$ and $(3.15.10)$ above we infer that $$J(\mbox{}_bf)=\int_{N_0} J(\mbox{}_{b\gamma }f) \lambda (d\gamma
)= \int_G \int_{N_0} \mbox{ }_bf(\gamma x)\lambda (d\gamma )\mu
(dx)$$ $= J(\mbox{}_bf^{[\lambda ]})=J(f^{[\lambda
]})=\int_{N_0}J(\mbox{}_{\gamma }f)\lambda (d\gamma )=J(f)$,\
since $\lambda (N_0)=1$ and $N_0\subset N(G)$. Thus the last assertion of this theorem also is proved.
[**3.16. Theorem.**]{}
*If $G$ is a $T_1$ topological locally compact fan loop fulfilling Condition $(3.5.2)$, then there exists*
$(3.16.1)$ a regular $\sigma $-additive measure $\mu $ on a Borel $\sigma $-algebra ${\cal B}(G)$ of $G$, $\mu : {\cal B}(G)\to
[0, \infty ]$ such that
$(3.16.2)$ $\mu (U)>0$ for each open subset in $G$;
$(3.16.3)$ $\mu (A)<\infty $ for each compact subset $A$ in $G$;
$(3.16.4)$ $\mu (bB)=\mu (B)$ for each $B\in {\cal B}(G)$ and $b\in G$.
Such $\mu $ can be chosen corresponding to a functional $J$ satisfying Conditions $(3.15.1)$-$(3.15.3)$.
[**Proof.**]{} This is an immediate consequence of $(3.15.1)$-$(3.15.3)$, $(3.15.5)$-$(3.15.8)$. In particular $\mu
(A)=\bar{\bar{J}}(\chi _A)$ for the characteristic function $\chi
_A$ of a Borel subset $A$ in $G$, where $\chi _A(x)=1$ for each $x\in A$, $\chi _A(y)=0$ for each $y\in G-A$.
[**3.17. Remark.**]{} Each function $f$ in $C_{0,0}(G)$ can be represented as $f=f^+ - f^-$, where $f^+(x)=\max (0, f(x))$, $f^+$ and $f^-$ belong to $C^+_{0,0}(G)$. Therefore, a functional $J$ satisfying Conditions $(15.1)$ and $(15.2)$ can be extended to a linear functional on $C_{0,0}(G)$ such that $J(f)=J(f^+)-J(f^-)$. Hence Property $(3.15.3)$ extends onto $C_{0,0}(G)$.
[**3.18. Definition.**]{} A linear functional $J$ on $C_{0,0}(G)$ satisfying Property $(3.15.3)$ is called left invariant.
A measure $\mu $ on the Borel $\sigma $-algebra ${\cal B}(G)$ of a topological fan loop $G$ such that $\mu $ satisfies Condition $(3.16.4)$ is called left invariant.
[**3.19. Theorem.**]{} [*Let $G$ be a $T_1$ topological locally compact fan loop fulfilling Condition $(3.5.2)$ and let $\mu $ be a measure possessing Properties $(3.16.1)$-$(3.16.4)$. Then $\mu (G)<\infty $ if and only if $G$ is compact.*]{}
[**Proof.**]{} If $G$ is compact, then by $(3.16.3)$ $\mu (G)<\infty $.
Vice versa suppose that $\mu (G)<\infty $ and consider the variant that $G$ is not compact and take an open neighborhood $U'$ of $e$ in $G$ with $U=\check{U}'$ such that $U=N_0U$ and its closure $cl_G(U)$ is compact, hence $0<\mu
(U)<\infty $ (see also Condition $(3.5.2)$). By virtue of Theorem 2.8 there exists an open neighborhood $V'$ of $e$ in $G$ with $V=\check{V}'$ such that $V=N_0V$ and $[cl_G(V)]^2\subset U'$. In view of Lemma 2.5 a subset $xU$ is open in $G$ for each $x\in G$.
At first we take some fixed $x_1\in G$. Then we construct a sequence $\{ x_j: ~ j \in {\bf N} \} $ by induction. Let $x_1,...,x_n$ be constructed such that if $n\ge 2$, then $x_jV\cap
x_kV=\emptyset $ for each $1\le j<k\le n$. There exists
$y\in
G-\bigcup_{j=1}^nU_j$, where $U_j:=x_jUp(x_jU,V,V)p(V,V,V)[p(x_jU,V,V)]^{-1}$,\
since $G$ is not compact and $U_j$ is open by Lemma 2.6 and $cl_G(U_j)$ is compact by Theorem 3.1.10 in [@eng] and Lemmas 2.4, 2.6. Put $x_{n+1}=y$ with this $y$.
Suppose that there is $z\in x_jV\cap
x_{n+1}V$ for some $1\le j\le n$. Therefore there would be $v$ and $u$ in $V$ for which $z=x_jv=x_{n+1}u$, consequently, $(x_jv)/u=(x_{n+1}u)/u=x_{n+1}$ by Condition $(2.1.2)$ and Formula $(2.2.5)$. Therefore by Formulas $(2.2.3)$, $(2.3.3')$ and Condition $(2.1.9)$
$x_{n+1}=x_j(v(e/u))p(x_j,v,e/u)p(e/u,u,u\setminus e)
[p(x_j(v(e/u)),u,u\setminus e)]^{-1}$\
contradicting the choice of $x_{n+1}$, since $[cl_G(V)]^2\subset U'$. Thus $x_jV\cap
x_kV=\emptyset $ for each $1\le j<k\le n+1$. This would mean by $(3.16.4)$ that $\mu (G)\ge \sum_{j=1}^n \mu (x_jV)=n\mu (V)$ for each $n$, contradicting $0<\mu (G)<\infty $.
[**3.20. Theorem.**]{}
*Assume that $G$ is a $T_1$ topological locally compact fan loop satisfying Condition $(3.5.2)$ and functionals $J$ and $H$ on $C^+_{0,0}(G)$ satisfy Conditions $(3.15.1)$-$(3.15.3)$.*
Then a positive constant $\kappa $ exists such that
$(3.20.1)$ $H(f)=\kappa J(f)$ for each $f\in C^+_{0,0}(G)$.
[**Proof.**]{} By virtue of Theorem 3.16 there exist two measures $\mu _1$ and $\mu _2$ corresponding to $J$ and $H$. We consider a subalgebra ${\cal C}(G):=\theta ^{-1}({\cal B}(G/\cdot /N_0))$ in ${\cal B}(G)$, where $\theta : G\to G/\cdot /N_0$ is the quotient homomorphism, ${\cal B}(G)$ denotes the Borel $\sigma $-algebra on $G$. Put $\nu _j(A)= \mu _j(\theta ^{-1}(A))$ for each $j$ and $A\in
{\cal B}(G/\cdot /N_0)$.
From Theorems 2.8 and 3.16 it follows that the measure $\nu _j$ on the group $G/\cdot /N_0$ is such that $\nu _j(V)>0$ for each open subset $V$ in $G/\cdot /N_0$, $\nu
_j(A)<\infty $ for each compact subset $A$ in $G/\cdot /N_0$, $\nu
_j(cB)=\nu _j(B)$ for each $c\in G/\cdot /N_0$ and $B\in {\cal
B}(G/\cdot /N_0)$, $j \in \{ 1, 2 \} $. By virtue of Theorem 15.6 in [@hew] there are positive constants $p_j$ such that $\nu
_j=p_j\eta $, where $\eta $ is a left invariant Haar measure on $G/\cdot /N_0$. Thus $J(f^{[\lambda ]})=p_1H(f^{[\lambda ]})/p_2$ for each $f\in C^+_{0,0}(G)$.
We consider $\eta _1(b,f)=J(\mbox{}_bf)/J(f^{[\lambda ]})$ and $\eta _2(b,f)=H(\mbox{}_bf)/H(f^{[\lambda ]})$ for each $b\in G$ and a nonzero function $f$ in $C^+_{0,0}(G)$. According to Property $(3.15.3)$ we get the identities $\eta _j(b,f)=\eta _j(e,f^{[\lambda
]})=1$ for each $j\in \{ 1, 2 \} $. This implies that for each nonzero function $f\in C^+_{0,0}(G)$ and $b\in G$:
$(3.20.2)$ $J(\mbox{}_bf)/H(\mbox{}_bf)=p_1/p_2$.
The measures $\mu _1$ and $\mu _2$ possess Properties $(3.16.1)$-$(3.16.4)$. In view of the Lebesgue-Radon-Nikodym theorem (see Theorem (12.17) in [@hew] or see [@bogachmtb]) there exists a $\mu _1$ measurable nonnegative function $h(x)$ such that $\int_Gg(x)\mu _2(dx)=\int_Gg(x)h(x)\mu _1(dx)$ for each $g\in
C^+_{0,0}(G)$. Therefore from Formulas $(3.15.8)$ and $(3.20.2)$ it follows that $h(x)$ is a positive constant. Thus $(3.20.1)$ is proved.
Appendix. Products of fan loops.
================================
The main subject of this paper are measures on fan loops. Nevertheless, in this section it is shortly demonstrated that there are abundant families of fan loops besides those which appear in areas described in the introduction.
[**4.1. Theorem.**]{}
*Let $(G_j,\tau _j)$ be a family of topological $T_1$ fan loops (see Definition 2.1), where $j\in J$, $J$ is a set. Then their direct product $G=\prod_{j\in J}G_j$ relative to the Tychonoff product topology $\tau $ is a topological $T_1$ fan loop and*
$(4.1.1)$ $Z(G)=\prod_{j\in J}Z(G_j)$ and $N(G)=\prod_{j\in J}N(G_j)$.
[**Proof.**]{} The direct product of topological loops is a topological loop (see [@bruckb; @eng; @kakkar]). Thus conditions $(2.1.1)$-$(2.1.3)$ are satisfied.
Each element $a\in G$ is written as $a= \{ a_j: ~ \forall j\in J, ~ a_j\in G_j
\}$. From $(2.1.4)$-$(2.1.7)$ we infer that
$(4.1.2)$ $Com (G)
:= \{ a\in G: \forall b\in G, ~ ab=ba \} =$
$ \{ a\in G: ~ a= \{
a_j: \forall j\in J, a_j\in G_j \}; \forall b\in G, ~ b= \{ b_j:
\forall j\in J, b_j\in G_j \} ; \forall j\in J, ~ a_jb_j=b_ja_j \}
=\prod_{j\in J} Com (G_j)$,
$(4.1.3)$ $N_l(G) := \{a\in G: ~ \forall b\in G, ~ \forall c\in G, ~ (ab)c=a(bc)
\} = \{a\in G: ~ a= \{ a_j: \forall j\in J, a_j\in G_j \}; ~ \forall
b\in G, ~ b= \{ b_j: \forall j\in J, b_j\in G_j \}; ~ \forall c\in
G, ~ c= \{ c_j: \forall j\in J, c_j\in G_j \}; ~ \forall j\in J,~
(a_jb_j)c_j=a_j(b_jc_j) \}= \prod_{j\in J} N_l(G_j)$\
and similarly
$(4.1.4)$ $N_m(G)=\prod_{j\in J} N_m(G_j)$ and
$(4.1.5)$ $N_r(G)=\prod_{j\in J} N_r(G_j)$.\
Therefore $(4.1.3)$-$(4.1.5)$ and $(2.1.8)$ imply that
$(4.1.6)$ $N(G)=\prod_{j\in J}N(G_j)$. Thus
$(4.1.7)$ $Z(G) := Com (G)\cap N(G)=\prod_{j\in J}Z(G_j)$.
Let $a$, $b$ and $c$ be in $G$, then
$(ab)c=\{ (a_jb_j)c_j: ~ \forall j \in J, ~ a_j\in G_j, b_j\in
G_j, c_j\in G_j \} $
$= \{ t_{G_j}(a_j,b_j,c_j) a_j(b_jc_j): ~
\forall j \in J, ~ a_j\in G_j, b_j\in G_j, c_j\in G_j \}$
$ =
t_G(a,b,c) a(bc)$
and analogously $(ab)c = a(bc)p_G(a,b,c) $, where
$(4.1.8)$ $t_G(a,b,c) = \{ t_{G_j}(a_j,b_j,c_j): ~ \forall j \in J, ~
a_j\in G_j, b_j\in G_j, c_j\in G_j \} $ and
$(4.1.9)$ $p_G(a,b,c) = \{ p_{G_j}(a_j,b_j,c_j): ~ \forall j \in J, ~
a_j\in G_j, b_j\in G_j, c_j\in G_j \} $.
Therefore, Formulas $(4.1.7)$-$(4.1.9)$ imply that Conditions $(2.1.9)$ also are satisfied. Thus $G$ is a topological fan loop. By virtue of Theorem 2.3.11 in [@eng] a product of $T_1$ spaces is a $T_1$ space, hence $G$ is the $T_1$ topological fan loop.
[**4.2. Corollary.**]{}
*$(1)$. Let conditions of Theorem 4.1 be satisfied and for each $j\in J$ a fan loop $G_j$ satisfies Condition $(3.5.2)$. Then the product fan loop $G$ satisfies Condition $(3.5.2)$.*
$(2)$. Moreover, if $G_j$ is compact for all $j\in
J_0$ and locally compact for each $j\in J\setminus J_0$, where $J_0\subset J$ and $J\setminus J_0$ is a finite set, then $G$ is locally compact.
[**Proof.**]{} Using Formulas $(4.1.8)$ and $(4.1.9)$ it is sufficient to take $N_0(G)=\prod_{j\in J} N_0(G_j)$, since the direct product of compact groups $N_0(G_j)$ is a compact group $N_0(G)$ (see the Tychonoff theorem 3.2.4 in [@eng] or [@hew]). The last assertion $(2)$ follows from the known fact that $G$ as a topological space is locally compact under the imposed above conditions (see Theorem 3.3.13 in [@eng]).
[**4.3. Remark.**]{}
$(4.3.1)$. Let $A$ and $B$ be two fan loops and let $N$ be a group such that $N_0(A)\hookrightarrow N$, $N_0(B)\hookrightarrow N$, $N\hookrightarrow N(A)$ and $N\hookrightarrow N(B)$ and let $N$ be normal in $A$ and in $B$ (see also Sections 2.1, 2.7 and 3.5).
Using direct products it is always possible to extend either $A$ or $B$ to get such a case. In particular, either $A$ or $B$ may be a group. On $A\times B$ an equivalence relation $\Xi $ is considered such that
$(4.3.2)$ $(v\gamma ,b)\Xi (v,\gamma b)$\
for every $v$ in $A$, $b$ in $B$ and $\gamma $ in $N$.
$(4.3.3)$. Let $\phi : A\to {\cal A}(B)$ be a single-valued mapping, where ${\cal A}(B)$ denotes a family of all bijective surjective single-valued mappings of $B$ onto $B$ subjected to the conditions given below. If $a\in A$ and $b\in B$, then it will be written shortly $b^a$ instead of $\phi (a)b$, where $\phi (a) : B\to B$. Let also
$\eta _{\phi }: A\times A\times
B\to N$, $\kappa _{\phi } : A\times B\times B\to N$
and $\xi
_{\phi } : ((A\times B)/\Xi ) \times ((A\times B)/\Xi )\to N$\
be single-valued mappings written shortly as $\eta $, $\kappa $ and $\xi $ correspondingly such that
$(4.3.4)$ $(b^u)^v=b^{vu}\eta (v,u,b)$, $~ {\gamma }^u=\gamma $, $~b^{\gamma }=b$;
$(4.3.5)$ $\eta (v,u,(\gamma _1b)\gamma _2)=\eta (v,u,b)$;
if $\gamma \in \{ v, u, b \} $ then $\eta (v,u,b)=e$;
$(4.3.6)$ $(cb)^u=c^ub^u\kappa (u,c,b)$;
$(4.3.7)$ $\kappa (u,(\gamma _1c)\gamma _2,(\gamma _3b)\gamma _4)=\kappa
(u,c,b)$ and
if $\gamma \in \{ u, c, b)$ then $\kappa (u,c
,b)=e$;
$(4.3.8)$ $\xi (((\gamma u)\gamma _1,(\gamma _2c)\gamma _3),((\gamma _4v)\gamma _5,
(\gamma _6b)\gamma _7))= \xi ((u,c),(v,b))$ and
$\xi ((e,e), (v,b))=e$ and $\xi ((u,c),(e,e))=e$\
for every $u$ and $v$ in $A$, $b$, $c$ in $B$, $\gamma $, $\gamma _1$,...,$\gamma _7$ in $N$, where $e$ denotes the neutral element in $N$ and in $A$ and $B$.
We put
$(4.3.9)$ $(a_1,b_1)(a_2,b_2)=(a_1a_2,b_1b_2^{a_1}\xi
((a_1,b_1),(a_2,b_2)))$\
for each $a_1$, $a_2$ in $A$, $b_1$ and $b_2$ in $B$.
The Cartesian product $A\times B$ supplied with such a binary operation $(4.3.9)$ will be denoted by $A\bigotimes
^{\phi , \eta , \kappa , \xi }B$.
[**4.4. Theorem.**]{} [*Let the conditions of Remark 4.3 be fulfilled. Then the Cartesian product $A\times B$ supplied with a binary operation $(4.3.9)$ is a fan loop.*]{}
[**Proof.**]{} From the conditions of Remark 4.3 it follows that the binary operation $(4.3.9)$ is single-valued. The group $N$ is normal in the loops $A$ and $B$ by Conditions $(4.3.1)$. Hence for each $a\in A$ and $\beta \in N$ there exists $(a\beta )/a \in N$ and $a\setminus (\beta a)\in N$, since $aN=Na$ for each $a\in A$. Similarly it is for $B$. Thus there are single-valued mappings
$r_{A,a}(\beta )=(a\beta )/a$, $ ~ \check{r}_{A,a}(\beta
)=a\setminus (\beta a)$,
$r_{B,b}(\beta )=(b\beta )/b$, $ ~ \check{r}_{B,b}(\beta
)=b\setminus (\beta b)$,
$r_{A,a}: N\to N$, $ ~
\check{r}_{A,a}: N\to N$, $~ r_{B,b}: N\to N$, $ ~ \check{r}_{B,b}:
N\to N$\
for each $a\in A$ and $b\in B$. Evidently
$r_{A,a}(
\check{r}_{A,a}(\beta ))=\beta $ and $\check{r}_{A,a}(r_{A,a}(\beta
))=\beta $\
for each $a\in A$ and $\beta \in N$, and similarly for $B$.
Let $I_1=((a_1,b_1)(a_2,b_2))(a_3,b_3)$ and $I_2=
(a_1,b_1)((a_2,b_2)(a_3,b_3))$, where $a_1$, $a_2$, $a_3$ belong to $A$, $b_1$, $b_2$, $b_3$ belong to $B$. Then we infer that
$I_1= ((a_1a_2)a_3,(b_1b_2^{a_1})\xi ((a_1,b_1),(a_2,b_2))
b_3^{a_1a_2}\xi ((a_1a_2,b_1b_2^{a_1}),(a_3,b_3)))$ and
$I_2= (a_1(a_2a_3), b_1(b_2^{a_1}b_3^{a_1a_2})\beta )$ with
$\beta =\eta
(a_1,a_2,b_3)\kappa (a_1,b_2,b_3^{a_2})[\xi
((a_2,b_2),(a_3,b_3))]^{a_1} \xi ((a_1,b_1),
(a_2a_3,b_2b_3^{a_2}))$. Hence
$I_1=(a,b\alpha )$ and $I_2=(a,b\beta )$, where $a=a_1(a_2a_3)$ and $b=b_1(b_2^{a_1}b_3^{a_1a_2})$,
$\alpha =\check{r}_{B,b}(
p_A(a_1,a_2,a_3))$
$p_B(b_1,b_2^{a_1},b_3^{a_1a_2}) ~
\check{r}_{B,b_3^{a_1a_2}}(\xi ((a_1,b_1),(a_2,b_2))) ~ \xi
((a_1a_2,b_1b_2^{a_1}),(a_3,b_3)))$.
Therefore
$(4.4.1)$ $I_1=I_2p$ with $p=p_{A\bigotimes ^{\phi , \eta , \kappa , \xi }B}((a_1,b_1),(a_2,b_2),(a_3,b_3))$ and
$I_1=tI_2$ with $t=t_{A\bigotimes ^{\phi , \eta , \kappa ,
\xi }B}((a_1,b_1),(a_2,b_2),(a_3,b_3))$;
$(4.4.2)$ $p=\beta ^{-1}\alpha $ and $t=r_{A,a}(r_{B,b}(p))$.\
Apparently $t_{A\bigotimes ^{\phi , \eta , \kappa , \xi }B}
((a_1,b_1),(a_2,b_2),(a_3,b_3))\in N$ and\
$p_{A\bigotimes ^{\phi
, \eta , \kappa , \xi }B} ((a_1,b_1),(a_2,b_2),(a_3,b_3))\in N$ for each $a_j\in A$, $b_j\in B$, $j\in \{ 1, 2, 3 \} $, since $\alpha $ and $\beta $ belong to the group $N$.
If $\gamma \in N$ and either $(\gamma ,e)$ or $(e,\gamma )$ belongs to $\{ (a_1,b_1),(a_2,b_2),(a_3,b_3) \} $, then from the conditions of Section 4.3 and Formulas $(4.4.1)$ and $(4.4.2)$ it follows that
$p_{A\bigotimes ^{\phi , \eta , \kappa , \xi }B}((a_1,b_1),(a_2,b_2),(a_3,b_3))=e$ and
$t_{A\bigotimes ^{\phi , \eta , \kappa , \xi
}B}((a_1,b_1),(a_2,b_2),(a_3,b_3))=e$,\
consequently, $(N,e)\cup
(e,N)\subset N(A\bigotimes ^{\phi , \eta , \kappa , \xi }B)$.
Apparently $(2.1.3)$ follows from $(4.3.8)$ and $(4.3.9)$.
Next we consider the following equation
$(4.4.3)$ $(a_1,b_1)(a,b)=(e,e)$, where $a\in A$, $b\in B$.
From $(2.1.2)$ for fan loops $A$ and $B$, $(4.3.8)$ and $(4.3.9)$ we deduce that
$(4.4.4)$ $a_1=e/a$,\
consequently, $b_1b^{(e/a)}\xi ((e/a,b_1),(a,b))=e$ and hence
$(4.4.5)$ $b_1=e/[b^{(e/a)}\xi
((e/a,b^{(e/a)}),(a,b))]$.\
Thus $a_1\in A$ and $b_1\in B$ given by $(4.4.4)$ and $(4.4.5)$ provide a unique solution of $(4.4.3)$.
Similarly from the following equation
$(4.4.6)$ $(a,b)(a_2,b_2)=(e,e)$, where $a\in A$, $b\in B$ we infer that
$(4.4.7)$ $a_2=a\setminus e$,\
consequently, $bb_2^a\xi ((a,b),(a\setminus e,b_2))=e$ and hence
$b_2^a=
b\setminus [\xi ((a,b),(a\setminus e,b_2))]^{-1}$\
by Conditions $(2.1.1)$, $(2.1.2)$ and $(4.3.3)$ for fan loops $A$ and $B$. On the other hand, $(b_2^a)^{e/a}=b_2\eta (e/a,a,b_2)$, consequently, by Lemmas 2.2, 2.3 and the conditions of Section 4.3
$(4.4.8)$ $b_2= (b\setminus [\xi ((a,b),(a\setminus
e,(b\setminus e)^{e/a}))]^{-1})^{e/a})/\eta (e/a,a,(b\setminus
e)^{e/a}) $.\
Thus Formulas $(4.4.7)$ and $(4.4.8)$ provide a unique solution of $(4.4.6)$.
Next we put $(a_1,b_1)=(e,e)/(a,b)$ and $(a_2,b_2)=(a,b)\setminus (e,e)$ and
$(4.4.9)$ $(a,b)\setminus (c,d)=((a,b)\setminus (e,e))(c,d)
p((a,b),(a,b)\setminus (e,e), (c,d))$;
$(4.4.10)$ $(c,d)/(a,b)=[t((c,d),(e,e)/(a,b),(a,b))]^{-1}(c,d)((e,e)/(a,b))$
and $e_G=(e,e)$, where $G=A\bigotimes ^{\phi , \eta , \kappa , \xi }B$. Therefore Properties $(2.1.1)$-$(2.1.3)$ and $(2.1.9)$ are fulfilled for $A\bigotimes ^{\phi , \eta , \kappa , \xi }B$.
[**4.5. Definition.**]{} The fan loop $A\bigotimes ^{\phi , \eta , \kappa , \xi }B$ provided by Theorem 4.4 we call a smashed product of fan loops $A$ and $B$ with smashing factors $\phi $, $\eta $, $\kappa $ and $\xi $.
[**4.6. Corollary.**]{} [*Suppose that the conditions of Remark 4.3 are fulfilled and $A$ and $B$ are topological $T_1$ fan loops and smashing factors $\phi $, $\eta $, $\kappa $, $\xi $ are jointly continuous by their variables. Suppose also that $A\bigotimes ^{\phi
, \eta , \kappa , \xi }B$ is supplied with a topology induced from the Tychonoff product topology on $A\times B$. Then $A\bigotimes
^{\phi , \eta , \kappa , \xi }B$ is a topological $T_1$ fan loop.*]{}
[**4.7. Corollary.**]{} [*If the conditions of Corollary 4.6 are satisfied and loops $A$ and $B$ are locally compact, then $A\bigotimes ^{\phi , \eta , \kappa , \xi }B$ is locally compact. Moreover, if $A$ and $B$ satisfy Condition $(3.5.2)$ and ranges of $\eta $, $\kappa $, $\xi $ are contained in $N_0(A)N_0(B)$, then $A\bigotimes ^{\phi , \eta , \kappa , \xi }B$ satisfies Condition $(3.5.2)$.*]{}
[**Proof.**]{} Corollaries 4.6 and 4.7 follow immediately from Theorems 2.3.11, 3.2.4, 3.3.13 in [@eng] and Theorem 4.4.
[**4.8. Remark.**]{} From Theorems 4.1, 4.4 and Corollaries 4.2, 4.6, 4.7 it follows that taking nontrivial $\phi $, $\eta $, $\kappa $ and $\xi $ and starting even from groups with nontrivial $N(G_j)$ or $N(A)$ and $G_j/\cdot/N(G_j)$ or $A/\cdot /N(A)$ it is possible to construct new fan loops with nontrivial $N_0(G)$ and ranges $t_G(G,G,G)$ and $p_G(G,G,G)$ of $t_G$ and $p_G$ may be infinite and nondiscrete. With suitable smashing factors $\phi $, $\eta $, $\kappa $ and $\xi $ and with nontrivial fan loops or groups $A$ and $B$ it is easy to get examples of fan loops in which $e/a\ne a\setminus e$ for an infinite family of elements $a$ in $A\bigotimes ^{\phi , \eta , \kappa , \xi }B$.
[**4.9. Conclusion.**]{} The results of this article can be used for further studies of measures on homogeneous spaces and noncommutative manifolds related with loops. Besides applications of left invariant measures on loops outlined in the introduction it is interesting to mention possible applications in mathematical coding theory and its technical applications [@blautrctb; @petbagsychrtj; @srwseabm14], because frequently codes are based on topological-algebraic binary systems and measures. Another very important applications are in representation theory of loops and harmonic analysis on loops, mathematical physics, quantum field theory, quantum gravity, gauge theory, etc.
[199]{}
H. Albuquerque, S. Majid. “Quasialgebra structure of the octonions”. J. of Algebra [**220: 1**]{} (1999), 188-224.
D. Allcock. “Reflection groups and octave hyperbolic plane”. J. of Algebra [**213: 2**]{} (1998), 467-498.
J.C. Baez. “The octonions”. Bull. Amer. Mathem. Soc. [**39: 2**]{} (2002), 145-205.
R.E. Blahut. “Algebraic codes for data transmission” (Cambridge: Cambridge Univ. Press, 2003).
V.I. Bogachev. “Measure theory”. V. 1, 2 (Berlin: Springer-Verlag, 2007).
N.N. Bogolubov, A.A. Logunov, A.I. Oksak, I.T. Todorov. “General principles of quantum field theory” (Moscow: Nauka, 1987).
N. Bourbaki. “Algebra” (Berlin: Springer, 1989).
R.H. Bruck. “A survey of binary systems” (Berlin: Springer-Verlag, 1971).
O.A. Castro-Alvaredo, B. Doyon, D. Fioravanti. “Conical twist fields and null polygonal Wilson loops”. Nuclear Physics [**B931**]{} (2018), 146-178.
L.E. Dickson. “The collected mathematical papers”. Volumes 1-5 (New York: Chelsea Publishing Co., 1975).
R. Engelking. “General topology”. 2-nd ed., Sigma Series in Pure Mathematics, V. 6 (Berlin: Heldermann Verlag, 1989).
J.M.G. Fell, R.S. Doran. “Representations of $*$-algebras, locally compact groups, and Banach $*$-algebraic bundles”. V. 1 and V. 2 (Boston: Acad. Press, 1988).
E. Frenod, S. V. Ludkowski. “Integral operator approach over octonions to solution of nonlinear PDE”. Far East J. of Mathem. Sci. (FJMS). [**103: 5**]{} (2018), 831-876; DOI: 10.17654/MS103050831.
J.E. Gilbert, M.A.M. Murray. “Clifford algebras and Dirac operators in harmonic analysis”. Cambr. studies in advanced Mathem. [**26**]{} (Cambridge: Cambr. Univ. Press, 1991).
P.R. Girard. “Quaternions, Clifford algebras and relativistic Physics” (Basel: Birkhäuser, 2007).
K. Gürlebeck, W. Sprössig. “Quaternionic and Clifford calculus for physicists and engineers” (Chichester: John Wiley and Sons, 1997).
F. Gürsey, C.-H. Tze. “On the role of division, Jordan and related algebras in particle physics” (Singapore: World Scientific Publ. Co., 1996).
E. Hewitt, K.A. Ross. “Abstract harmonic analysis” (Berlin: Springer-Verlag, 1979).
V. Kakkar. “Boolean loops with compact left inner mapping groups are profinite.” Topology and Its Appl. [**244**]{} (2018), 51-54.
I.L. Kantor, A.S. Solodovnikov. “Hypercomplex numbers” (Berlin: Springer-Verlag, 1989).
H. Kiechle. “Theory of K-loops”, (Berlin: Springer-Verlag, 2002).
S.V. Ludkowski. “Decompositions of PDE over Cayley-Dickson algebras”. Rendic. dell’Istit. di Matem. dell’Univer. di Trieste. Nuova Ser. [**46**]{} (2014), 1-23.
S.V. Ludkowski. “Integration of vector Sobolev type PDE over octonions”. Complex Variab. and Elliptic Equat. [**61: 7**]{} (2016), 1014-1035.
S.V. Ludkowski. “Manifolds over Cayley-Dickson algebras and their immersions”. Rendic. dell’Istit. di Matem. dell’Univer. di Trieste. Nuova Ser. [**45**]{} (2013); 11-22.
S.V. Ludkovsky. “Normal families of functions and groups of pseudoconformal diffeomorphisms of quaternion and octonion variables”. J. Mathem. Sci., N.Y. (Springer) [**150: 4**]{} (2008), 2224-2287.
S.V. Ludkovsky. “Functions of several Cayley-Dickson variables and manifolds over them”, J. Mathem. Sci.; N.Y. (Springer) [**141: 3**]{} (2007), 1299-1330.
S.V. Ludkovsky, W. Sprössig. “Ordered representations of normal and super-differential operators in quaternion and octonion Hilbert spaces”. Adv. Appl. Clifford Alg. [**20: 2**]{} (2010), 321-342.
S.V. Ludkovsky, W. Sprössig. “Spectral theory of super-differential operators of quaternion and octonion variables”. Adv. Appl. Clifford Alg. [**21: 1**]{} (2011), 165-191.
S.V. Ludkovsky. “Integration of vector hydrodynamical partial differential equations over octonions”. Complex Variab. and Elliptic Equat. [**58: 5**]{} (2013); 579-609.
S.V. Ludkowski. “Automorphisms and derivations of nonassociative $C^*$ algebras”. Linear and Multilinear Algebra, (2018), 1-8, DOI: 10.1080/03081087.2018.1460794.
A.B. Petrov, S.V. Bagrov, A.I. Sycheva. “Approaches to sustainable operation of complex information systems for government and corporate purpose”. Russian Technological J. [**2: 4**]{} (2015), 175-183.
G. Pickert. “Projektive Ebenen” (Berlin: Springer-Verlag, 1955).
Yu.P. Razmyslov. “Identities of algebras and their representations”. Series “Modern Algebra”. [**14**]{} (Moscow: Nauka, 1989).
J.D.H. Smith. “An introduction to quasigroups and their representations” (Boca Raton: Chapman and Hall/CRC, Taylor and Francis Group, 2007).
K.P. Shum, X. Ren, Y. Wang. “Semigroups on semilattice and the constructions of generalized cryptogroups”. Southeast Asian Bull. of Mathem. [**38**]{} (2014), 719-730.
P. Vojt$\check{e}$choivsk$\acute{y}$. “Bol loops and Bruch loops of order $pq$ up to isotopism”. Finite Fields and Their Appl. [**52**]{} (2018), 1-9.
[^1]: key words and phrases: measure; left invariant; loop; locally compact\
Mathematics Subject Classification 2010: 43A05; 28C10; 20N05; 22A30
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We generalize the Beckner’s type Poincaré inequality [@Beckner] to a large class of probability measures on an abstract Wiener space of the form $\mu\star\nu$, where $\mu$ is the reference Gaussian measure and $\nu$ is a probability measure satisfying a certain integrability condition. As the Beckner inequality interpolates between the Poincaré and logarithmic Sobolev inequalities, we utilize a family of products for functions which interpolates between the usual point-wise multiplication and the Wick product. Our approach is based on the positivity of a quadratic form involving Wick powers and integration with respect to those convolution measures. Our dimension-independent results are compared with some very recent findings in the literature. In addition, we prove that in the finite dimensional case the class of densities of convolutions measures satisfies a point-wise covariance inequality.'
---
\
**Keywords:** Beckner’s type Poincaré inequality, Ornstein-Uhlenbeck semigroup, Wick product, convolution measures.
**Mathematics Subject Classification (2000):** 60H07, 60H30.
Introduction
============
In 1989 Beckner [@Beckner] proved the following inequality: $$\begin{aligned}
\label{beckner}
\int_{W}|f(w)|^2d\mu(w)-\int_{W}\big|e^{-\tau N}f(w)\big|^2d\mu(w)\leq (2-p)\int_{W}|Df(w)|^2d\mu(w)\end{aligned}$$ where $1\leq p\leq 2$, $e^{-\tau}=\sqrt{p-1}$, $\mu$ is a standard Gaussian probability measure on the (possibly infinite dimensional) space $W$, $Df$ denotes a suitable gradient of $f$ and $N$ stands for the *number* or *Ornstein-Uhlenbeck* operator. Observe that when $p=1$ or equivalently $\tau=+\infty$ then (\[beckner\]) coincides with the classic Poincaré inequality ([@Chernoff], [@Nash]): $$\begin{aligned}
\int_{W}|f(w)|^2d\mu(w)-\Big(\int_{W}f(w)d\mu(w)\Big)^2\leq \int_{W}|Df(w)|^2d\mu(w).\end{aligned}$$ Moreover, utilizing the Nelson’s hyper-contractive estimate [@Nelson]: $$\begin{aligned}
\int_{W}\big|e^{-\tau N}f(w)\big|^2d\mu(w)\leq\Big(\int_{W}|f(w)|^pd\mu(w)\Big)^{\frac{2}{p}}\end{aligned}$$ one can rewrite (\[beckner\]) as $$\begin{aligned}
\label{beckner 2}
\int_{W}|f(w)|^2d\mu(w)-\Big(\int_{W}|f(w)|^pd\mu(w)\Big)^{\frac{2}{p}}\leq (2-p)\int_{W}|Df(w)|^2d\mu(w).\end{aligned}$$ Dividing both sides of (\[beckner 2\]) by $2-p$ and letting $p\to 2^-$ one obtains the logarithmic Sobolev inequality ([@Gross]): $$\begin{aligned}
&&\int_{W}|f(w)|^2\ln\big(|f(w)|^2\big)d\mu(w)-\int_{W}|f(w)|^2d\mu(w)\cdot\ln\Big(\int_{W}|f(w)|^2d\mu(w)\Big)\nonumber\\
&&\leq 2\int_{W}|Df(w)|^2d\mu(w).\end{aligned}$$ Inequalities (\[beckner\])-(\[beckner 2\]), viewed as an interpolation between the Poincaré and logarithmic Sobolev inequalites, have attracted the attention of several authors; generalizations to log-concave measures, search for best constants and applications to partial differential equations have been the main topics of investigation. We refer the reader to the papers [@ABD], [@AMTU], [@LO], [@Wang] and the references quoted there.\
In the paper [@DLS; @2] the authors introduced and studied a family of products for functions defined on Gaussian spaces: $$\begin{aligned}
\label{alpha intro}
f\circ_{\alpha} g:=e^{\tau N}\Big(e^{-\tau N}f\cdot e^{-\tau N}g\Big)\end{aligned}$$ where $\sqrt{\alpha}:=e^{-\tau}$ and $f,g$ belong to some suitable function space (see Section 2.2 below for precise conditions). This family of products interpolates between the usual point-wise multiplication, when $\alpha=1$, and the Wick product, when $\alpha =0$ (this is obtained in [@DLS; @2] through a limit argument). The crucial role of this family of products is in connection with the theory of stochastic integration and stochastic differential equations; in fact, one can prove that the following Wong-Zakai-type theorem holds:\
If for $k\geq1$, $\{W^k_t\}_{0\leq t\leq T}$ is a smooth approximation of the white noise $W_t:=\frac{dB_t}{dt}$ ($B_t$ being a one dimensional Brownian motion) then the solution of $$\begin{aligned}
\frac{dX_t^k}{dt}=b(X_t^k)+X_t^k\circ_{\alpha} W_t^k,\quad
X_0^k=x\end{aligned}$$ converges in $\mathcal{L}^p(W,\mu)$, as $k$ goes to infinity, to the solution of $$\begin{aligned}
dX_t=b(X_t)dt+X_td^{\alpha}B_t,\quad X_0=x\end{aligned}$$ where $$\begin{aligned}
\int_0^TX_td^{\alpha}B_t:=\lim_{n\to+\infty}\sum_{k=1}^nX_{(1-\frac{\alpha}{2})t_{k-1}+\frac{\alpha}{2} t_{k}}\cdot(B_{t_{k}}-B_{t_{k-1}}).\end{aligned}$$ Observe that when $\alpha=0$ or $1$ we obtain the Itô and Stratonovich integrals, respectively.\
The aim of the present paper is to show that the Beckner’s inequality (\[beckner\]) can be generalized in a natural way to convolution measures on abstract Wiener spaces. This generalization passes through the use of the products $\circ_{\alpha}$ defined in (\[alpha intro\]) and contains as a particular case the Poincaré-type inequality obtained in [@L]. More precisely, we will prove the following inequality: $$\begin{aligned}
\label{main result intro}
\int_{W}|f(w)|^2d\rho(w)-\int_{W} (f\circ_{\alpha}f)(w)d\rho(w)\leq (1-\alpha)\int_{W}\Vert Df(w)\Vert_H^2d\rho(w)\end{aligned}$$ where $(W,H,\mu)$ is an abstract Wiener space, $\rho=\mu\star\nu$ and $\nu$ is a probability measure on $(W,\mathcal{B}(W))$. To see how (\[main result intro\]) reduces to (\[beckner\]) when $\rho=\mu$ observe that by definition $\alpha=p-1$ and that we can write $$\begin{aligned}
\int_{W} (f\circ_{\alpha}f)(w)d\mu(w)&=&\int_{W}e^{\tau N}\big(e^{-\tau N}f\cdot e^{-\tau N}f\big) (w)d\mu(w)\\
&=&\int_{W}e^{-\tau N}f(w)\cdot e^{-\tau N}f (w)d\mu(w)\\
&=&\int_{W}\big|e^{-\tau N}f(w)\big|^2 d\mu(w).\end{aligned}$$ Moreover, since $\circ_{\alpha}$ approaches the Wick product $\diamond$ as $\alpha\to 0^+$, inequality (\[main result intro\]) becomes in that limit $$\begin{aligned}
\int_{W}|f(w)|^2d\rho(w)-\int_{W} (f\diamond f)(w)d\rho(w)\leq\int_{W}\Vert Df(w)\Vert_H^2d\rho(w).\end{aligned}$$ This last inequality, obtained in [@L], is weaker than the classic Poincaré inequality for the measure $\rho$ since in general we have $$\begin{aligned}
\Big(\int_{W}f(w)d\rho(w)\Big)^2\leq\int_{W} (f\diamond f)(w)d\rho(w).\end{aligned}$$ Our approach is based on a novel idea whose crucial ingredient is the positive definiteness of a certain quadratic form involving Wick powers and integration with respect to convolution measures (see Proposition \[convolution measures\] below). We mention that in the very recent papers [@Z], [@WW] and [@CZ] Poincaré, weak Poincaré and logarithmic Sobolev inequalities for convolution measures on finite dimensional Euclidean spaces have been investigated: here the reference Gaussian (or log-concave) measure is convolved with compactly supported measures. We work in a dimension free framework and once we specify our assumptions for the finite dimensional case (see Corollary \[corollary\] below) we get an exponential integrability condition on the measure $\nu$ (see (\[integrability condition\]) below), which is clearly satisfied for compactly supported measures. However, as we mentioned above, inequality (\[main result intro\]) is weaker, at least for $\alpha=0$, than the Poincaré inequality (and hence the logarithmic Sobolev inequality) studied in the above mentioned papers.\
We first prove inequality (\[main result intro\]) for $f$ being a linear combination of stochastic exponentials and then, under an additional condition on the integrating measure $\rho$, we extend the validity of the result by density to suitable Sobolev spaces, which clearly contain the class of smooth cylindrical functions (that usually represent the class for testing functional inequalities on infinite dimensional domains).\
The paper is organized as follows: Section 2 collects definitions, notations and the necessary background material while in Section 3, after some preliminary results, we state and prove the main theorem of the paper followed by some refinements for the finite dimensional case and a point-wise covariance inequality (see (\[a\]) below) satisfied by the densities of convolution measures with respect to the reference Gaussian measure.
Framework
=========
The aim of this section is to collect the necessary background material and fix the notation. For the sake of clarity the topics will not be treated in their greatest generality. For more details the interested reader is referred to the books of Bogachev [@Bogachev], Janson [@J], Nualart [@Nualart] and to the paper by Potthoff and Timpel [@PT] (the latter reference is suggested, among other things, for the theory of the spaces $\mathcal{G}_{\lambda}$ and the notion of Wick product).
The spaces $\mathbb{D}^{k,p}$ and $\mathcal{G}_{\lambda}$
---------------------------------------------------------
Let $(H,W,\mu)$ be an *abstract Wiener space*, that means $(H,\langle\cdot,\cdot\rangle_H)$ is a separable Hilbert space which is continuously and densely embedded in the Banach space $(W,\Vert\cdot\Vert_W)$ and $\mu$ is a Gaussian probability measure on the Borel sets $\mathcal{B}(W)$ of $W$ such that $$\begin{aligned}
\label{Gaussian characteristic}
\int_{W}e^{i\langle w,w^*\rangle}d\mu(w)=e^{-\frac{1}{2}\Vert
w^*\Vert_H^2},\quad\mbox{ for all }w^*\in W^*.\end{aligned}$$ Here $W^*\subset H$ denotes the dual space of $W$ (which in turn is dense in $H$) and $\langle\cdot,\cdot\rangle$ stands for the dual pairing between $W$ and $W^*$. We will refer to $H$ as the *Cameron-Martin* space of $W$. Set for $p\geq 1$, $$\begin{aligned}
\mathcal{L}^p(W,\mu):=\Big\{f:W\to\mathbb{R}\mbox{ such that }\Vert
f\Vert_p:=\Big(\int_W|f(w)|^pd\mu(w)\Big)^{\frac{1}{p}}<+\infty\Big\}.\end{aligned}$$ It follows from (\[Gaussian characteristic\]) that the map $$\begin{aligned}
W^*&\to&\mathcal{L}^2(W,\mu)\\
w^*&\mapsto&\langle w,w^*\rangle\end{aligned}$$ is an isometry; we can therefore define for $\mu$-almost all $w\in
W$ the quantity $\langle w,h\rangle$ for $h\in H$ as an element of $\mathcal{L}^2(W,\mu)$.\
We now introduce the gradient operator and a class of functions of Sobolev type. On the set $$\begin{aligned}
\mathcal{S}:=\{f(w)=\varphi(\langle w,h_1\rangle,...,\langle w,h_n\rangle)\mbox{ where
}n\in\mathbb{N}, h_1,...,h_n\in H\mbox{ and }\varphi\in
C_0^{\infty}(\mathbb{R}^n)\}\end{aligned}$$ define $$\begin{aligned}
D(\varphi(\langle w,h_1\rangle,...,\langle w,h_n\rangle)):=\sum_{j=1}^n\frac{\partial\varphi}{\partial
x_j}(\langle w,h_1\rangle,...,\langle w,h_n\rangle)h_j.\end{aligned}$$ The operator $D$ maps $\mathcal{S}$ into $\mathcal{L}^p(W,\mu;H)$; moreover by means of the integration by parts formula $$\begin{aligned}
\int_W\langle Df(w),h\rangle_Hd\mu(w)=\int_Wf(w)\cdot\langle w,h\rangle d\mu(w),\quad f\in\mathcal{S}, h\in H\end{aligned}$$ one can prove that $D$ is closable in $\mathcal{L}^p(W,\mu)$; we therefore define the space $\mathbb{D}^{1,p}$ to be closure of $\mathcal{S}$ under the norm $$\begin{aligned}
\Vert f\Vert_{1,p}:=\Big(\int_W|f(w)|^pd\mu(w)+\int_W\Vert
Df(w)\Vert_H^pd\mu(w)\Big)^{\frac{1}{p}}.\end{aligned}$$ In a similar way, iterating the definition of $D$ and introducing for any $k\in\mathbb{N}$ the norms $$\begin{aligned}
\Vert f\Vert_{k,p}:=\Big(\int_W|f(w)|^pd\mu(w)+\sum_{j=1}^k\int_W\Vert
D^jf(w)\Vert_{H^{\otimes j}}^pd\mu(w)\Big)^{\frac{1}{p}}.\end{aligned}$$ one constructs the spaces $\mathbb{D}^{k,p}$.\
In order to prove our main results we need to introduce an additional class of functions. To this aim recall that by the Wiener-Itô chaos decomposition theorem any element $f$ in $\mathcal{L}^2(W,\mu)$ has an infinite orthogonal expansion $$\begin{aligned}
f=\sum_{n\geq 0} \delta^n(f_n),\end{aligned}$$ where $f_n\in H^{\hat{\otimes}n}$, the space of symmetric elements of $H^{\otimes n}$, and $\delta^n(f_n)$ stands for the multiple Wiener-Itô integral of $f_n$. We remark that $\delta^1(f_1)$ coincides with the element $\langle w,f_1\rangle$ mentioned above. Moreover, one has $$\begin{aligned}
\Vert f\Vert_2^2=\sum_{n\geq 0}n!\Vert f_n\Vert^2_{H^{\otimes n}}.\end{aligned}$$ It is useful to observe that if $f$ happens to be in $\mathbb{D}^{1,2}$ then $$\begin{aligned}
\int_W\Vert Df(w)\Vert_H^2d\mu(w)=\sum_{n\geq 1}n n!\Vert f_n\Vert^2_{H^{\otimes
n}}.\end{aligned}$$ For any $\lambda\geq 0$ define the operator $\Gamma(\lambda)$ acting on $\mathcal{L}^2(W,\mu)$ as $$\begin{aligned}
\Gamma(\lambda)\Big(\sum_{n\geq 0}\delta^n(f_n)\Big):=\sum_{n\geq 0}
\lambda^n\delta^n(f_n).\end{aligned}$$ Observe that with $\lambda=e^{-\tau}$, $\tau\geq 0$ then the operator $\Gamma(\lambda)$ coincides with the Ornstein-Uhlenbeck semigroup $$\begin{aligned}
(P_{\tau}f)(w):=\int_Wf\big(e^{-\tau}w+\sqrt{1-e^{-2\tau}}\tilde{w}\big)d\mu(\tilde{w}),\quad
w\in W, \tau\geq 0\end{aligned}$$ which is a bounded operator. Otherwise, $\Gamma(\lambda)$ is an unbounded operator with domain in $\mathcal{L}^2(W,\mu)$ given by $$\begin{aligned}
\mathcal{G}_{\lambda}:=\Big\{f=\sum_{n\geq 0}
\delta^n(f_n)\in\mathcal{L}^2(W,\mu)\mbox{ such that }\Vert
f\Vert_{\mathcal{G}_{\lambda}}^2:=\sum_{n\geq 0}n!\lambda^{2n}\Vert
f_n\Vert^2_{H^{\otimes n}}<+\infty\Big\}.\end{aligned}$$ The family $\{\mathcal{G}_{\lambda}\}_{\lambda\geq 1}$ is a collection of Hilbert spaces with the property that $$\begin{aligned}
\mathcal{G}_{\lambda_2}\subset\mathcal{G}_{\lambda_1}\subset\mathcal{L}^2(W,\mu)\end{aligned}$$ for $1<\lambda_1<\lambda_2$. Define $\mathcal{G}:=\bigcap_{\lambda\geq 1}\mathcal{G}_{\lambda}$ endowed with the projective limit topology; the space $\mathcal{G}$ turns out to be a reflexive Fréchet space. Its dual $\mathcal{G}^*$ is a space of generalized functions that can be represented as $\mathcal{G}^*=\bigcup_{\lambda> 0}\mathcal{G}_{\lambda}$. We remark that for $f\in\mathcal{L}^2(W,\mu)$ and $g\in\mathcal{G}$ one has $$\begin{aligned}
\langle\langle f,g\rangle\rangle=\int_Wf(w)g(w)d\mu(w)\end{aligned}$$ where $\langle\langle\cdot,\cdot\rangle\rangle$ stands for the dual pairing between $\mathcal{G}^*$ and $\mathcal{G}$.\
One of the most representative elements of $\mathcal{G}$ is the so called *stochastic exponential* $$\begin{aligned}
w\in W\mapsto\mathcal{E}(h)(w):=\exp\Big\{\langle w,h\rangle-\frac{\Vert
h\Vert_H^2}{2}\Big\},\quad h\in H.\end{aligned}$$ We recall that stochastic exponentials correspond among other things to Radon-Nikodym derivatives, with respect to the underlying Gaussian measure $\mu$, of probability measures on $(W,\mathcal{B}(W))$ obtained through shifted copies of $\mu$ along Cameron-Martin directions. Its membership to $\mathcal{G}$ can be easily verified since the Wiener-Itô chaos decomposition of $\mathcal{E}(h)$ is obtained with $f_n=\frac{h^{\otimes n}}{n!}$. Moreover the linear span of the stochastic exponentials, that we denote with $\mathcal{E}$, is dense in $\mathcal{L}^p(W,\mu)$, $\mathbb{D}^{k,p}$, for any $p\geq 1$ and $k\in\mathbb{N}$, and $\mathcal{G}$.
The Wick and $\alpha$-products
------------------------------
For $h,k\in H$ define $$\begin{aligned}
\mathcal{E}(h)\diamond\mathcal{E}(k):=\mathcal{E}(h+k).\end{aligned}$$ This is called the *Wick product* of $\mathcal{E}(h)$ and $\mathcal{E}(k)$. Extend this operation by linearity to $\mathcal{E}$ to get a commutative, associative and distributive (with respect to the sum) multiplication. The Wick product is easily seen to be an unbounded bilinear operator on the $\mathcal{L}^p(W,\mu)$ spaces; for instance, the Wick product $f\diamond g$ of the two square integrable elements $f$ and $g$ lives in the distributional space $\mathcal{G}^*$.\
Now, let $f,g\in\mathcal{L}^p(W,\mu)$ for some $p> 1$. For $\alpha\in ]0,1]$ define $$\begin{aligned}
\label{alpha}
(f\circ_{\alpha}g)(w):=\Gamma(1/\sqrt{\alpha})(\Gamma(\sqrt{\alpha})f\cdot\Gamma(\sqrt{\alpha})g)(w),\quad w\in W.\end{aligned}$$ This is called the *$\alpha$-product* of $f$ and $g$; it was introduced for the first time in [@DLS; @2] in connection with stochastic integrals and stochastic differential equations. This family of products provides an interpolation between the usual point-wise multiplication (obtained trivially with $\alpha$=1) and the Wick product (obtained in the limit as $\alpha\to 0^+$). A simple calculation shows that $$\begin{aligned}
\label{exponential alpha}
\mathcal{E}(h_1)\circ_{\alpha}\mathcal{E}(h_2)=\mathcal{E}(h_1+h_2)e^{\alpha\langle h_1,h_2\rangle_H}.\end{aligned}$$ The reader is referred to Theorem \[aurel\] below for a sharp Hölder inequality for the family of products $\circ_{\alpha}$.
Main results
============
In the sequel we will call *convolution measure* on $(W,\mathcal{B}(W))$ any probability measure of the form $\mu\star\nu$ where $\mu$ is the reference Gaussian measure on $(W,\mathcal{B}(W))$, $\nu$ is a probability measure on $(W,\mathcal{B}(W))$ and $$\begin{aligned}
(\mu\star\nu)(A):=\int_W\mu(A-w)d\nu(w),\quad A\in\mathcal{B}(W).\end{aligned}$$ We begin this section with a simple but crucial result: the description of the interplay between the Wick product and convolution measures.
\[convolution measures\] Let $\nu$ be a probability measure on $(W,\mathcal{B}(W))$ and define $\rho:=\mu\ast\nu$. Then for every $z_1,...,z_n\in\mathbb{C}$ and $h_1,...,h_n\in H$ one has $$\begin{aligned}
\int_W\Big(\sum_{j=1}^nz_j\exp\Big\{i\langle w,h_j\rangle+\frac{\Vert h_j\Vert_H^2}{2}\Big\}\Big)\diamond\overline{\Big(\sum_{j=1}^nz_j\exp\Big\{i\langle w,h_j\rangle+\frac{\Vert h_j\Vert_H^2}{2}\Big\}\Big)}d\rho(w)\geq 0\end{aligned}$$ where $i$ is the imaginary unit and $\overline{u}$ stands for the complex conjugate of $u$.
We simply need to utilize the definition of Wick product and the Fourier transform characterization (\[Gaussian characteristic\]) of the underlying Gaussian measure $\mu$: $$\begin{aligned}
&&\int_W\Big(\sum_{j=1}^nz_j\exp\Big\{i\langle w,h_j\rangle+\frac{\Vert h_j\Vert_H^2}{2}\Big\}\Big)\diamond\overline{\Big(\sum_{j=1}^nz_j\exp\Big\{i\langle w,h_j\rangle+\frac{\Vert h_j\Vert_H^2}{2}\Big\}\Big)}d\rho(w)\\
&=&\int_W\sum_{j,k=1}^nz_j\bar{z}_k\exp\Big\{i\langle w,h_j\rangle+\frac{\Vert h_j\Vert_H^2}{2}\Big\}\diamond\exp\Big\{-i\langle w,h_k\rangle+\frac{\Vert h_k\Vert_H^2}{2}\Big\}d\rho(w)\\
&=&\int_W\sum_{j,k=1}^nz_j\bar{z}_k\exp\Big\{i\langle w,h_j-h_k\rangle+\frac{\Vert h_j-h_k\Vert_H^2}{2}\Big\}d\rho(w)\\
&=&\sum_{j,k=1}^nz_j\bar{z}_k\exp\Big\{\frac{\Vert h_j-h_k\Vert_H^2}{2}\Big\}\int_W\exp\Big\{i\langle w,h_j-h_k\rangle\Big\}d\rho(w)\\
&=&\sum_{j,k=1}^nz_j\bar{z}_k\exp\Big\{\frac{\Vert h_j-h_k\Vert_H^2}{2}\Big\}\int_W\exp\Big\{i\langle w,h_j-h_k\rangle\Big\}d\mu(w)\int_W\exp\Big\{i\langle w,h_j-h_k\rangle\Big\}d\nu(w)\\
&=&\sum_{j,k=1}^nz_j\bar{z}_k\int_W\exp\Big\{i\langle w,h_j-h_k\rangle\Big\}d\nu(w)\\
&=&\int_W\Big(\sum_{j=1}^nz_j\exp\Big\{i\langle w,h_j\rangle\Big\}\Big)\cdot\overline{\Big(\sum_{j=1}^nz_j\exp\Big\{i\langle w,h_j\rangle\Big\}\Big)}d\nu(w)\\
&=&\int_W\Big|\sum_{j=1}^nz_j\exp\Big\{i\langle w,h_j\rangle\Big\}\Big|^2d\nu(w)\\
&\geq&0.\end{aligned}$$
\[strong positivity\] Assume the measure $\rho$ from the previous proposition to be absolutely continuous with respect to $\mu$ with a density $\xi$ belonging to $\mathcal{L}^p(W,\mu)$ for some $p>1$. In this case we can write $$\begin{aligned}
\label{positive definite}
&&\int_W\Big(\sum_{j=1}^nz_j\exp\Big\{i\langle w,h_j\rangle+\frac{\Vert h_j\Vert_H^2}{2}\Big\}\Big)\diamond\overline{\Big(\sum_{j=1}^nz_j\exp\Big\{i\langle w,h_j\rangle+\frac{\Vert h_j\Vert_H^2}{2}\Big\}\Big)}d\rho(w)\nonumber\\
&=&\int_W\Big(\sum_{j=1}^nz_j\exp\Big\{i\langle w,h_j\rangle+\frac{\Vert h_j\Vert_H^2}{2}\Big\}\Big)\diamond\overline{\Big(\sum_{j=1}^nz_j\exp\Big\{i\langle w,h_j\rangle+\frac{\Vert h_j\Vert_H^2}{2}\Big\}\Big)}\cdot\xi(w)d\mu(w)\nonumber\\
&=&\sum_{j,k=1}^nz_j\bar{z}_k\int_W\exp\Big\{i\langle w,h_j-h_k\rangle+\frac{\Vert h_j-h_k\Vert_H^2}{2}\Big\}\cdot\xi(w)d\mu(w)\nonumber\\
&=&\sum_{j,k=1}^nz_j\bar{z}_k\tau_{\xi}(h_j-h_k)\end{aligned}$$ where $$\begin{aligned}
h\in H\mapsto \tau_{\xi}(h):=\int_W\exp\Big\{i\langle w,h\rangle+\frac{\Vert h\Vert_H^2}{2}\Big\}\cdot\xi(w)d\mu(w).\end{aligned}$$ With this notation the statement of Proposition \[convolution measures\] reads $$\begin{aligned}
\sum_{j,k=1}^nz_j\bar{z}_k\tau_{\xi}(h_j-h_k)\geq 0\end{aligned}$$ which means that the function $\tau_{\xi}$ is positive definite; the latter is in turn equivalent, according to Proposition 5.1 in [@NZ], to the property $$\begin{aligned}
\label{def. strong pos.}
\langle\langle \Gamma(1/\sqrt{\alpha})\xi,\varphi\rangle\rangle\geq 0\mbox{ for each non negative }\varphi\in\mathcal{G}\mbox{ and }\alpha>0.\end{aligned}$$ Here $\langle\langle\cdot,\cdot\rangle\rangle$ denotes the dual pairing between the distributional space $\mathcal{G}^*$ and the test function space $\mathcal{G}$. We mention that elements satisfying condition (\[def. strong pos.\]) are referred in [@NZ] as *strongly positive*.
Another connection between convolution measures and Wick product is the following.
Let $\rho_1:=\mu\star\nu_1$ and $\rho_2:=\mu\star\nu_2$ be convolution measures on $(W,\mathcal{B}(W))$ and assume the existence of $\xi_1,\xi_2\in\mathcal{L}^1(W,\mu)$ such that $d\rho_1=\xi_1d\mu$ and $d\rho_2=\xi_2d\mu$. Then for $\rho_3:=\mu\star\nu_1\star\nu_2$ one has $d\rho_3=\xi_1\diamond\xi_2 d\mu$.
Let $h\in H$; then $$\begin{aligned}
\int_W\exp\{i\langle w,h\rangle\}d\rho_3(w)&=&\int_W\exp\{i\langle w,h\rangle\}d(\mu\star\nu_1\star\nu_2)(w)\\
&=&\int_W\exp\{i\langle w,h\rangle\}d\mu(w)\cdot\int_W\exp\{i\langle w,h\rangle\}d\nu_1(w)\\
&&\times\int_W\exp\{i\langle w,h\rangle\}d\nu_2(w)\\
&=&\exp\Big\{-\frac{\Vert h\Vert_H^2}{2}\Big\}\cdot\int_W\exp\{i\langle w,h\rangle\}d\nu_1(w)\\
&&\times\int_W\exp\{i\langle w,h\rangle\}d\nu_2(w)\\
&=&\exp\Big\{\frac{\Vert h\Vert_H^2}{2}\Big\}\cdot\int_W\exp\{i\langle w,h\rangle\}d\rho_1(w)\\
&&\times\int_W\exp\{i\langle w,h\rangle\}d\rho_2(w)\\
&=&\exp\Big\{\frac{\Vert h\Vert_H^2}{2}\Big\}\cdot\int_W\exp\{i\langle w,h\rangle\}\xi_1(w)d\mu(w)\\
&&\times\int_W\exp\{i\langle w,h\rangle\}\xi_2(w)d\mu(w)\\
&=&\exp\Big\{-\frac{\Vert h\Vert_H^2}{2}\Big\}\cdot\int_W\exp\Big\{i\langle w,h\rangle+\frac{\Vert h\Vert_H^2}{2}\Big\}\xi_1(w)d\mu(w)\\
&&\times\int_W\exp\Big\{i\langle w,h\rangle+\frac{\Vert h\Vert_H^2}{2}\Big\}\xi_2(w)d\mu(w)\\
&=&\exp\Big\{-\frac{\Vert h\Vert_H^2}{2}\Big\}\cdot\int_W\exp\Big\{i\langle w,h\rangle+\frac{\Vert h\Vert_H^2}{2}\Big\}(\xi_1\diamond\xi_2)(w)d\mu(w)\\
&=&\int_W\exp\{i\langle w,h\rangle\}(\xi_1\diamond\xi_2)(w)d\mu(w)\end{aligned}$$ where we utilized the characterizing property of the Wick product $$\begin{aligned}
\int_W(f\diamond g)(w)\mathcal{E}(h)(w)d\mu(w)=\int_Wf(w)\mathcal{E}(h)(w)d\mu(w)\cdot\int_Wg(w)\mathcal{E}(h)(w)d\mu(w)\end{aligned}$$ which holds for any $h\in H$.
The next theorem is a particular case of a more general result proved in [@Stan] where the reader is referred for the proof (the link between the theorem presented below and the results in the reference mentioned before is: $\Gamma(\lambda)(f\circ_{\alpha}g)=\Gamma(\lambda)f\circ_{\frac{\alpha}{\lambda^2}}\Gamma(\lambda)g$). It provides a Hölder inequality for the family of $\alpha$-products $\circ_{\alpha}$ which we will utilize to find the right function spaces for our extension of the Beckner’s type Poincaré inequality.
\[aurel\] Let $p,q,r>1$ and $\alpha\in [0,1]$ be such that $$\begin{aligned}
\label{condition on parameters}
\frac{1}{r-\frac{1-\alpha}{1+\alpha}}=\frac{1+\alpha}{2(p-1)+2\alpha}+\frac{1+\alpha}{2(q-1)+2\alpha}.\end{aligned}$$ Then for any $f\in\mathcal{L}^p(W,\mu)$ and $g\in\mathcal{L}^q(W,\mu)$ one has $\Gamma\big(\sqrt{(1+\alpha)/2}\big)(f\circ_{\alpha}g)\in\mathcal{L}^r(W,\mu)$. More precisely, $$\begin{aligned}
\label{holder}
\Big\Vert\Gamma\big(\sqrt{(1+\alpha)/2}\big)(f\circ_{\alpha}g)\Big\Vert_r\leq\Vert f\Vert_p\cdot\Vert g\Vert_q.\end{aligned}$$
Observe that when $\alpha=1$ then $\circ_{\alpha}$ coincides with the usual point-wise product and (\[condition on parameters\])-(\[holder\]) become the classic Hölder inequality. On the other hand, when $\alpha=0$ then $\circ_{\alpha}$ coincides with the Wick product and (\[condition on parameters\])-(\[holder\]) reduce to the Hölder-Young-Lieb inequality proved in [@DLS].
We now make the first step towards the main result of the present paper. We are going to show that the left hand side of our main inequality (see (\[main inequality\]) below) is non negative.
\[left inequality theorem\] Let $\nu$ be a probability measure on $(W,\mathcal{B}(W))$ and choose $\alpha\in ]0,1]$. Assume that $\rho:=\mu\ast\nu$ is absolutely continuous with respect to $\mu$ with density $\xi$ belonging to $\mathcal{G}_{\sqrt{2/(1+\alpha)}}$. Then for any $f\in\mathcal{L}^{3+\alpha}(W,\mu)$ one gets $$\begin{aligned}
\label{left inequality}
\int_W|f(w)|^2d\rho(w)-\int_W (f\circ_{\alpha}f)(w)d\rho(w)\geq 0.\end{aligned}$$
First of all observe that the integrals appearing in the left hand side of (\[left inequality\]) are finite. In fact, by the Nelson hyper-contractive inequality we deduce that $$\begin{aligned}
\Vert\xi\Vert_{\frac{3+\alpha}{1+\alpha}}&=&\Big\Vert\Gamma\big(\sqrt{(1+\alpha)/2}\big)\Gamma\big(\sqrt{2/(1+\alpha)}\big)\xi\Big\Vert_{\frac{3+\alpha}{1+\alpha}}\\
&\leq&\Big\Vert\Gamma\big(\sqrt{2/(1+\alpha)}\big)\xi\Big\Vert_2\\
&<&+\infty\end{aligned}$$ which implies that $\xi\in\mathcal{L}^{\frac{3+\alpha}{1+\alpha}}(W,\mu)$. Therefore, using Hölder inequality we get $$\begin{aligned}
\int_W|f(w)|^2d\rho(w)&=&\int_W|f(w)|^2\cdot\xi(w) d\mu(w)\\
&\leq&\Big(\int_W|f(w)|^{3+\alpha}d\rho(w)\Big)^{\frac{2}{3+\alpha}}\cdot\Vert\xi\Vert_{\frac{3+\alpha}{1+\alpha}}\\
&=&\Vert f\Vert^2_{3+\alpha}\cdot\Vert\xi\Vert_{\frac{3+\alpha}{1+\alpha}}\end{aligned}$$ where $\frac{3+\alpha}{2}$ is the conjugate exponent of $\frac{3+\alpha}{1+\alpha}$. This shows the finiteness of the first integral in (\[left inequality\]). Concerning the second integral, note that for $\alpha\leq1$ one has $2(1+\alpha)\leq 3+\alpha$ which implies $\mathcal{L}^{3+\alpha}(W,\mu)\subset\mathcal{L}^{2(1+\alpha)}(W,\mu)$. Now choosing $p=q=2(1+\alpha)$ and $r=2$ in (\[condition on parameters\]) we get from (\[holder\]) that $$\begin{aligned}
\Big\Vert\Gamma\big(\sqrt{(1+\alpha)/2}\big)(f\circ_{\alpha}f)\Big\Vert_2\leq\Vert f\Vert^2_{2(1+\alpha)}\end{aligned}$$ which implies that $f\circ_{\alpha}f\in\mathcal{G}_{\sqrt{(1+\alpha)/2}}$ (under our assumption on $f$). Therefore, the integral $$\begin{aligned}
\int_W (f\circ_{\alpha}f)(w)d\rho(w)=\int_W (f\circ_{\alpha}f)(w)\cdot\xi(w)d\mu(w) \end{aligned}$$ is finite if $\xi\in\mathcal{G}_{\sqrt{2/(1+\alpha)}}$.\
To prove inequality (\[left inequality\]) we recall (see Remark \[strong positivity\] above) that the function $\xi$, being the density of a convolution measure, is strongly positive, i.e. $\Gamma(1/\sqrt{\alpha})\xi\geq 0$ (in distributional sense) for any $\alpha>0$. Hence using the definition of $f\circ_{\alpha}f$ we can write $$\begin{aligned}
\int_W (f\circ_{\alpha}f)(w)d\rho(w)&=&\int_W (f\circ_{\alpha}f)(w)\xi(w)d\mu(w)\\
&=&\int_W \Gamma(1/\sqrt{\alpha})(\Gamma(\sqrt{\alpha})f)^2(w)\cdot\xi(w)d\mu(w)\\
&=&\int_W(\Gamma(\sqrt{\alpha})f)^2(w)\cdot(\Gamma(1/\sqrt{\alpha})\xi)(w)d\mu(w)\\
&\leq&\int_W\Gamma(\sqrt{\alpha})f^2(w)\cdot(\Gamma(1/\sqrt{\alpha})\xi)(w)d\mu(w)\\
&=&\int_Wf^2(w)\cdot\xi(w)d\mu(w)\\
&=&\int_W|f(w)|^2d\rho(w)\end{aligned}$$ where in the inequality we utilized the Jensen inequality for the bounded operator $\Gamma(\sqrt{\alpha})$ and the convex function $x\mapsto x^2$.
We are now ready to prove the main theorem of the present paper.
\[main theorem\] Let $\nu$ be a probability measure on $(W,\mathcal{B}(W))$ and choose $\alpha\in ]0,1]$. Assume that $\rho:=\mu\ast\nu$ is absolutely continuous with respect to $\mu$ with density $\xi$ belonging to $\mathcal{G}_{\sqrt{2/(1+\alpha)}}$. Then for every $f\in\mathbb{D}^{1,3+\alpha}$ one has $$\begin{aligned}
\label{main inequality}
\int_W|f(w)|^2d\rho(w)-\int_W(f\circ_{\alpha}f)(w)d\rho(w)\leq (1-\alpha)\int_W\Vert Df(w)\Vert_H^2d\rho(w)\end{aligned}$$ or equivalently, $$\begin{aligned}
\label{main inequality 2}
\int_W|f(w)|^2d\rho(w)-\int_W|(\Gamma(\sqrt{\alpha})f)(w)|^2\cdot(\Gamma(1/\sqrt{\alpha})\xi)(w)d\mu(w)\nonumber\\
\leq (1-\alpha)\int_W\Vert Df(w)\Vert_H^2d\rho(w).\end{aligned}$$
Observe that for $\nu=\delta_0$, the Dirac measure concentrated at $0\in W$, the measure $\rho$ coincides with $\mu$ implying that $\xi\equiv 1$ and in particular $\Gamma(1/\sqrt{\alpha})\xi\equiv 1$. Inserting these quantities in (\[main inequality 2\]) we recover the Beckner’s type Poincaré inequality (\[beckner\]).
For any $\alpha\in ]0,1]$ define the map $$\begin{aligned}
\label{T}
T_{\alpha}:\mathcal{E}&\to&\mathcal{E}\nonumber\\
f&\mapsto& T_{\alpha}(f):=f\circ_{\alpha} f-|f|^2+(1-\alpha)\Vert Df\Vert_H^2.\end{aligned}$$ Since $f\in\mathcal{E}$ we can write $f=\sum_{j=1}^n\lambda_j\mathcal{E}(h_j)$ for some $\lambda_1,...,\lambda_n\in\mathbb{R}$ and $h_1,...,h_n\in H$. Now substitute this expression into (\[T\]) to obtain (recall identity (\[exponential alpha\])), $$\begin{aligned}
T_{\alpha}(f)&=&\sum_{j,k=1}^n\lambda_j\lambda_k\mathcal{E}(h_j)\circ_{\alpha}\mathcal{E}(h_k)-
\sum_{j,k=1}^n\lambda_j\lambda_k\mathcal{E}(h_j)\cdot\mathcal{E}(h_k)\\
&&+(1-\alpha)\sum_{j,k=1}^n\lambda_j\lambda_k\mathcal{E}(h_j)\cdot\mathcal{E}(h_k)\langle
h_j,h_k\rangle_H\\
&=&\sum_{j,k=1}^n\lambda_j\lambda_k\mathcal{E}(h_j)\diamond\mathcal{E}(h_k)e^{\alpha\langle h_j,h_k\rangle_H}-
\sum_{j,k=1}^n\lambda_j\lambda_k\mathcal{E}(h_j)\diamond\mathcal{E}(h_k)e^{\langle h_j,h_k\rangle_H}\\
&&+(1-\alpha)\sum_{j,k=1}^n\lambda_j\lambda_k\mathcal{E}(h_j)\diamond\mathcal{E}(h_k)e^{\langle
h_j,h_k\rangle_H}\langle
h_j,h_k\rangle_H\\
&=&\sum_{j,k=1}^n\lambda_j\lambda_k\mathcal{E}(h_j)\diamond\mathcal{E}(h_k)\Big(e^{\alpha\langle h_j,h_k\rangle_H}-e^{\langle
h_j,h_k\rangle_H}+(1-\alpha)e^{\langle h_j,h_k\rangle_H}\langle
h_j,h_k\rangle_H\Big).\end{aligned}$$ We now integrate with respect to the measure $\rho$ the first and last terms of the previous chain of equalities to obtain $$\begin{aligned}
\label{equality}
&&\int_WT_{\alpha}(f)(w)d\rho(w)\nonumber\\
&=&\sum_{j,k=1}^n\lambda_j\lambda_k\Big(e^{\alpha\langle h_j,h_k\rangle_H}-e^{\langle
h_j,h_k\rangle_H}+(1-\alpha)e^{\langle h_j,h_k\rangle_H}\langle
h_j,h_k\rangle_H\Big)\int_W(\mathcal{E}(h_j)\diamond\mathcal{E}(h_k))(w)d\rho(w)\nonumber\\
&=&\sum_{j,k=1}^n\lambda_j\lambda_ka_{jk}b_{jk},\end{aligned}$$ where for $j,k\in\{1,...,n\}$ we set $$\begin{aligned}
a_{jk}:=e^{\alpha\langle h_j,h_k\rangle_H}-e^{\langle h_j,h_k\rangle_H}+(1-\alpha)e^{\langle
h_j,h_k\rangle_H}\langle h_j,h_k\rangle_H\end{aligned}$$ and $$\begin{aligned}
b_{jk}:=\int_W(\mathcal{E}(h_j)\diamond\mathcal{E}(h_k))(w)d\rho(w).\end{aligned}$$ Observe that the matrix $A=\{a_{jk}\}_{1\leq j,k\leq n}$ is positive semi-definite; in fact, if in the Beckner’s type Poincaré inequality $$\begin{aligned}
\int_W|f(w)|^2d\mu(w)-\int_W|(\Gamma(\sqrt{\alpha})f)(w)|^2d\mu(w)\leq (1-\alpha)\int_W\Vert Df(w)\Vert_H^2d\mu(w)\end{aligned}$$ we take $f$ to be $\sum_{j=1}^n\lambda_j\mathcal{E}(h_j)$ one gets $$\begin{aligned}
\sum_{j,k=1}^n\lambda_j\lambda_k(e^{\langle h_j,h_k\rangle_H}-e^{\alpha\langle h_j,h_k\rangle_H})\leq
(1-\alpha)\sum_{j,k=1}^n\lambda_j\lambda_ke^{\langle h_j,h_k\rangle_H}\langle
h_j,h_k\rangle_H,\end{aligned}$$ which corresponds exactly to what we are claiming. On the other hand, from Proposition \[convolution measures\] the matrix $B=\{b_{jk}\}_{1\leq j,k\leq n}$ is positive semi-definite . Therefore the matrix $A\Box B:=\{a_{jk}\cdot b_{jk}\}_{1\leq
j,k\leq n}$ (which corresponds to the Hadamard product of the matrix $A$ with the matrix $B$) is also positive semi-definite (see for instance Styan [@S]), that means $$\begin{aligned}
\sum_{j,k=1}^n\lambda_j\lambda_ka_{jk}b_{jk}\geq 0,\end{aligned}$$ for any $\lambda_1,...,\lambda_n\in\mathbb{R}$. From (\[equality\]) this corresponds to $$\begin{aligned}
\int_WT_{\alpha}(f)(w)d\rho(w)\geq 0\quad\mbox{ for all }f\in\mathcal{E}.\end{aligned}$$ Recalling the definition of $T_{\alpha}$ this is equivalent to $$\begin{aligned}
\int_W(f\circ_{\alpha}f)(w)d\rho(w)-\int_W|f(w)|^2d\rho(w)+(1-\alpha)\int_W\Vert Df(w)\Vert_H^2d\rho(w)\geq 0.\end{aligned}$$ which proves inequality (\[main inequality\]) for $f\in\mathcal{E}$.\
The next step is to extend the validity of the last inequality to the whole $\mathbb{D}^{1,3+\alpha}$.\
Since the measure $\rho$ is assumed to be absolutely continuous with respect to $\mu$ with density $\xi$ belonging to $\mathcal{G}_{\sqrt{2/(1+\alpha)}}\subset\mathcal{L}^{\frac{3+\alpha}{1+\alpha}}(W,\mu)$ we can control, via the Hölder inequality, the quantities $$\begin{aligned}
\int_W|f(w)|^2d\rho(w)\quad\mbox{ and }\quad\int_W\Vert Df(w)\Vert_H^2d\rho(w)\end{aligned}$$ with $$\begin{aligned}
\Vert f\Vert_{3+\alpha}\quad\mbox{ and }\quad\Vert \Vert Df\Vert_H\Vert_{3+\alpha}\end{aligned}$$ respectively, and exploit the density of the set $\mathcal{E}$ in $\mathbb{D}^{1,3+\alpha}$. Moreover, Theorem \[aurel\] guarantees that for any $\alpha\in [0,1]$ the bilinear map $$\begin{aligned}
(f,g)\mapsto f\circ_{\alpha}g\end{aligned}$$ is continuous from $\mathcal{L}^{3+\alpha}(W,\mu)\times\mathcal{L}^{3+\alpha}(W,\mu)$ into $\mathcal{G}_{\sqrt{(1+\alpha)/2}}$. This fact, together with the density of $\mathcal{E}$ in $\mathcal{L}^{3+\alpha}(W,\mu)$, completes the proof of (\[main inequality\]).\
Inequality (\[main inequality 2\]) follows in the same manner through the self-adjointness of the operator $\Gamma(1/\sqrt{\alpha})$.
The finite dimensional case and a point-wise covariance inequality
------------------------------------------------------------------
In the previous section we proved Theorem \[main theorem\] under the assumptions that $\rho$ is a probability measure of convolution type, i.e. of the form $\rho=\mu\star\nu$, on a general abstract Wiener space with reference Gaussian measure $\mu$ and that $\rho$ is absolutely continuous with respect to $\mu$ with a density $\xi$ belonging to $\mathcal{G}_{\sqrt{2/(1+\alpha)}}$.\
We now want to focus on finite dimensional abstract Wiener spaces and give easy-to-check sufficient conditions on $\nu$ which guarantee the existence of the above mentioned smooth density.\
To this aim, consider the abstract Wiener space $W=H=\mathbb{R}^n$ with $$\begin{aligned}
\label{finite dimensional gaussian}
\mu(A)=\int_A(2\pi)^{-\frac{n}{2}}\exp\Big\{-\frac{|w|^2}{2}\Big\}dw,\quad A\in\mathcal{B}(\mathbb{R}^n)\end{aligned}$$ where $|\cdot|$ denotes the Euclidean norm on $\mathbb{R}^n$. Let $\nu$ be a probability measure on $(\mathbb{R}^n,\mathcal{B}(\mathbb{R}^n))$ and define $\rho:=\mu\star\nu$. It is easy to see that the assumption of absolute continuity of $\rho$ with respect to $\mu$ is automatically verified in this finite dimensional framework and that $$\begin{aligned}
\xi(w)&:=&\frac{d\rho}{d\mu}(w)\\
&=&\int_{\mathbb{R}^n}\exp\Big\{\langle w,y\rangle-\frac{|y|^2}{2}\Big\}d\nu(y),\quad w\in\mathbb{R}^n.\end{aligned}$$ Observe in addition that for each $y\in\mathbb{R}^n$ the function $$\begin{aligned}
w\mapsto \exp\Big\{\langle w,y\rangle-\frac{|y|^2}{2}\Big\}\end{aligned}$$ plays the role of stochastic exponential in the abstract Wiener space under consideration. We have the following.
Let $\nu$ be a probability measure on $(\mathbb{R}^n,\mathcal{B}(\mathbb{R}^n))$ and assume that $$\begin{aligned}
\int_{\mathbb{R}^n}\exp\Big\{\frac{\lambda^2|y|^2}{2}\Big\}d\nu(y)<+\infty,\quad\mbox{ for some }\lambda>1.\end{aligned}$$ Then the probability measure $\rho:=\mu\star\nu$ is absolutely continuous with respect to $\mu$ with density belonging to $\mathcal{G}_{\lambda}$.
We have only to check the membership of $$\begin{aligned}
\xi(w)=\int_{\mathbb{R}^n}\exp\Big\{\langle w,y\rangle-\frac{|y|^2}{2}\Big\}d\nu(y),\quad w\in\mathbb{R}^n\end{aligned}$$ to the space $\mathcal{G}_{\lambda}$. Using the Minkowski inequality we get that $$\begin{aligned}
\Vert\xi\Vert_{\mathcal{G}_{\lambda}}&=&\Vert\Gamma(\lambda)\xi\Vert_2\\
&=&\Big\Vert\Gamma(\lambda)\int_{\mathbb{R}^n}\exp\Big\{\langle \cdot,y\rangle-\frac{|y|^2}{2}\Big\}d\nu(y)\Big\Vert_2\\
&=&\Big\Vert\int_{\mathbb{R}^n}\exp\Big\{\langle \cdot,\lambda y\rangle-\frac{\lambda^2|y|^2}{2}\Big\}d\nu(y)\Big\Vert_2\\
&\leq&\int_{\mathbb{R}^n}\Big\Vert\exp\Big\{\langle \cdot,\lambda y\rangle-\frac{\lambda^2|y|^2}{2}\Big\}\Big\Vert_2 d\nu(y)\\
&=&\int_{\mathbb{R}^n}\exp\Big\{\frac{\lambda^2|y|^2}{2}\Big\}d\nu(y)\\
&<&+\infty.\end{aligned}$$
We can therefore rephrase our main theorem with more transparent conditions.
\[corollary\] Let $\alpha\in ]0,1]$ and consider a probability measure $\nu$ on $(\mathbb{R}^n,\mathcal{B}(\mathbb{R}^n))$ such that $$\begin{aligned}
\label{integrability condition}
\int_{\mathbb{R}^n}\exp\Big\{\frac{|y|^2}{1+\alpha}\Big\}d\nu(y)<+\infty.\end{aligned}$$ Define in addition $\rho:=\mu\star\nu$. Then for every $f\in\mathbb{D}^{1,3+\alpha}$ one has $$\begin{aligned}
\int_{\mathbb{R}^n}|f(w)|^2d\rho(w)-\int_{\mathbb{R}^n}(f\circ_{\alpha}f)(w)d\rho(w)\leq (1-\alpha)\int_{\mathbb{R}^n}\Vert Df(w)\Vert_H^2d\rho(w)\end{aligned}$$ or equivalently, $$\begin{aligned}
\int_{\mathbb{R}^n}|f(w)|^2d\rho(w)-\int_{\mathbb{R}^n}|(\Gamma(\sqrt{\alpha})f)(w)|^2\cdot(\Gamma(1/\sqrt{\alpha})\xi)(w)d\mu(w)\nonumber\\
\leq (1-\alpha)\int_{\mathbb{R}^n}\Vert Df(w)\Vert_H^2d\rho(w)\end{aligned}$$ where $$\begin{aligned}
(\Gamma(1/\sqrt{\alpha})\xi)(w):=\int_{\mathbb{R}^n}\exp\Big\{\langle w,\frac{y}{\sqrt{\alpha}}\rangle-\frac{|y|^2}{2\alpha}\Big\}d\nu(y),\quad w\in\mathbb{R}^n.\end{aligned}$$
We conclude the paper with an additional result on convolution measures on $\mathbb{R}^n$. We know from before that, if $\mu$ is the measure defined in (\[finite dimensional gaussian\]) and $\nu$ is a probability measure on $(\mathbb{R}^n, \mathcal{B}(\mathbb{R}^n))$, then $\rho:=\mu\star\nu$ is absolutely continuous with respect to $\mu$ with density $$\begin{aligned}
\xi(w):=\int_{\mathbb{R}^n}\exp\Big\{\langle w,y\rangle-\frac{|y|^2}{2}\Big\}d\nu(y),\quad w\in\mathbb{R}^n.\end{aligned}$$ We are going to show that functions of this type satisfy a point-wise covariance inequality, that means a point-wise inequality for functions which becomes after integration a covariance inequality in Gaussian spaces.
Let $\nu_1$ and $\nu_2$ be two probability measures on $(\mathbb{R}^n, \mathcal{B}(\mathbb{R}^n))$ and define for $w\in\mathbb{R}^n$, $$\begin{aligned}
\xi_1(w):=\int_{\mathbb{R}^n}\exp\Big\{\langle w,y\rangle-\frac{|y|^2}{2}\Big\}d\nu_1(y)\quad\mbox{ and }\quad\xi_2(w):=\int_{\mathbb{R}^n}\exp\Big\{\langle w,y\rangle-\frac{|y|^2}{2}\Big\}d\nu_2(y).\end{aligned}$$ Assume that for some $p>2$ the integrals $$\begin{aligned}
\label{1}
\int_{\mathbb{R}^n}\exp\Big\{(p-1)\frac{|y|^2}{2}\Big\}d\nu_i(y),\quad i=1,2\end{aligned}$$ are finite. Then $$\begin{aligned}
\label{a}
\xi_1\cdot\xi_2\geq \xi_1\diamond\xi_2+\sum_{k=1}^n\partial_{x_k}\xi_1\diamond\partial_{x_k}\xi_2\quad\mbox{ in}\quad\mathcal{G}^*\end{aligned}$$ i.e., for any non negative $\varphi\in\mathcal{G}$ one has $$\begin{aligned}
\label{b}
\langle\langle\xi_1\cdot\xi_2,\varphi\rangle\rangle-\langle\langle\xi_1\diamond\xi_2,\varphi\rangle\rangle-\sum_{k=1}^n\langle\langle\partial_{x_k}\xi_1\diamond\partial_{x_k}\xi_2,\varphi\rangle\rangle\geq 0.\end{aligned}$$
First of all observe that condition (\[1\]), due to Minkoski integral inequality, guarantees that $\xi_1$ and $\xi_2$ belong to $\mathcal{L}^p(\mathbb{R}^n,\mu)$ with $p>2$ and hence that all the terms appearing in (\[a\]) live in the distributional space $\mathcal{G}^*$. Now, for a non negative $\varphi\in\mathcal{G}$ we can write $$\begin{aligned}
\langle\langle\xi_1\cdot\xi_2,\varphi\rangle\rangle&=&\langle\langle \int_{\mathbb{R}^n}\exp\Big\{\langle \cdot,y\rangle-\frac{|y|^2}{2}\Big\}d\nu_1(y)\cdot\int_{\mathbb{R}^n}\exp\Big\{\langle \cdot,y\rangle-\frac{|y|^2}{2}\Big\}d\nu_2(y),\varphi\rangle\rangle\\
&=&\langle\langle \int_{\mathbb{R}^n}\exp\Big\{\langle \cdot,y+z\rangle-\frac{|y+z|^2}{2}\Big\}\exp{\langle y,z\rangle}d\nu_1(y)d\nu_2(z),\varphi\rangle\rangle\\
&=&\int_{\mathbb{R}^n}\langle\langle\exp\Big\{\langle \cdot,y+z\rangle-\frac{|y+z|^2}{2}\Big\},\varphi\rangle\rangle\exp{\langle y,z\rangle}d\nu_1(y)d\nu_2(z)\\
&\geq&\int_{\mathbb{R}^n}\langle\langle\exp\Big\{\langle \cdot,y+z\rangle-\frac{|y+z|^2}{2}\Big\},\varphi\rangle\rangle(1+\langle y,z\rangle)d\nu_1(y)d\nu_2(z)\\
&=&\int_{\mathbb{R}^n}\langle\langle\exp\Big\{\langle \cdot,y+z\rangle-\frac{|y+z|^2}{2}\Big\},\varphi\rangle\rangle d\nu_1(y)d\nu_2(z)\\
&&+\int_{\mathbb{R}^n}\langle\langle\exp\Big\{\langle \cdot,y+z\rangle-\frac{|y+z|^2}{2}\Big\},\varphi\rangle\rangle\langle y,z\rangle d\nu_1(y)d\nu_2(z)\\
&=&\langle\langle \int_{\mathbb{R}^n}\exp\Big\{\langle \cdot,y\rangle-\frac{|y|^2}{2}\Big\}d\nu_1(y)\diamond\int_{\mathbb{R}^n}\exp\Big\{\langle \cdot,y\rangle-\frac{|y|^2}{2}\Big\}d\nu_2(y),\varphi\rangle\rangle\\
&&\sum_{k=1}^n\langle\langle\partial_{x_k}\int_{\mathbb{R}^n}\exp\Big\{\langle \cdot,y\rangle-\frac{|y|^2}{2}\Big\}d\nu_1(y)\diamond\partial_{x_k}\int_{\mathbb{R}^n}\exp\Big\{\langle \cdot,y\rangle-\frac{|y|^2}{2}\Big\}d\nu_2(y),\varphi\rangle\rangle\\
&=&\langle\langle\xi_1\diamond\xi_2,\varphi\rangle\rangle-\sum_{k=1}^n\langle\langle\partial_{x_k}\xi_1\diamond\partial_{x_k}\xi_2,\varphi\rangle\rangle.\end{aligned}$$ The proof is complete.
If in (\[b\]) we take $\varphi\equiv 1$ and assume $\xi_1$ and $\xi_2$ belonging to $\mathbb{D}^{1,p}$ then we obtain $$\begin{aligned}
\int_{\mathbb{R}^n}\xi_1(w)\cdot\xi_2(w)d\mu(w)&\geq&\int_{\mathbb{R}^n}\xi_1(w)d\mu(w)\cdot\int_{\mathbb{R}^n}\xi_2(w)d\mu(w)\\
&&+\sum_{k=1}^n\int_{\mathbb{R}^n}\partial_{x_k}\xi_1(w)d\mu(w)\cdot\int_{\mathbb{R}^n}\partial_{x_k}\xi_2(w)d\mu(w)\end{aligned}$$ which is the finite dimensional version of the covariance inequality obtained in [@Hu] for convex functions. Here we utilized the properties $$\begin{aligned}
\langle\langle f,1\rangle\rangle=\int_{\mathbb{R}^n}f(w)d\mu(w)\quad\mbox{ when }\quad f\in\mathcal{L}^{2}(\mathbb{R}^n,\mu)\end{aligned}$$ and $$\begin{aligned}
\int_{\mathbb{R}^n}(f\diamond g)(w)d\mu(w)=\int_{\mathbb{R}^n}f(w)d\mu(w)\cdot\int_{\mathbb{R}^n}g(w)d\mu(w).\end{aligned}$$
[99]{}
A. Arnold, J.-P. Bartier and J. Dolbeault, Interpolation between logarithmic Sobolev and Poincaré inequalities, *Commun. Math. Sci.* [**5**]{} (2007) 971-979.
A. Arnold, P. Markowich, G. Toscani and A. Unterreiter, On logarithmic Sobolev inequalities and the rate of convergence to equilibrium for Fokker-Planck type equations, *Comm. Partial Differential Equations* [**26**]{} (2001) 43-100.
W. Beckner, A generalized Poincaré inequality for Gaussian measures, *Proc. Amer. Math. Soc.* [**105**]{} (1989) 397-400.
V.I. Bogachev, *Gaussian Measures*, American Mathematical Society, Providence, 1998.
L.-J. Cheng and S.-Q. Zhang, Weak Poincaré Inequality for Convolution Probability Measures, *arXiv:1407.4910v1*.
H. Chernoff, A note on an inequality involving the normal distribution, *Ann. Probab* **9** (1981) 533-535.
P. Da Pelo, A. Lanconelli and A. I. Stan, A Hölder-Young-Lieb inequality for norms of Gaussian Wick products, *Inf. Dim. Anal. Quantum Prob. Related Topics* **14** (2011) 375-407.
P. Da Pelo, A. Lanconelli and A. I. Stan, An Itô formula for a family of stochastic integrals and related Wong-Zakai theorems, *Stochastic Processes and their Applications* **123** (2013) 3183-3200.
P. Da Pelo, A. Lanconelli and A.I. Stan, A Hölder-Young inequality for norms of generalized Gaussian Wick products, *Preprint* (2014).
L. Gross, Logarithmic Sobolev inequality, *Amer. J. Math.* **97** (1975) 1061-1083.
Y. Hu, Itô-Wiener chaos expansion with exact residual and correlation, variance inequalities, *J. Theor. Probab.* [**10**]{} (1997) 835-848.
S. Janson, *Gaussian Hilbert spaces*, Cambridge Tracts in Mathematics, 129. Cambridge University Press, Cambridge, 1997.
A. Lanconelli, A new approach to Poincaré-type inequalities on the Wiener space, *Preprint* (2014).
R. Latala and K. Oleszkiewicz, Between Sobolev and Poincaré, *Lecture Notes in Math.* [**1745**]{} (2000) 147–168.
J. Nash, Continuity of solutions of partial and elliptic equations, *Amer. J. Math.* **80** 931-954.
E. Nelson, The free Markoff field, *J . Funct. Anal.* [**12**]{} (1973) 211-227.
D. Nualart, *Malliavin calculus and Related Topics, II edition*, Springer, New York 2006.
D. Nualart and M. Zakai, Positive and strongly positive Wiener functionals, [*Barcelona Seminar on Stochastic Analysis Progr. Probab.*]{} [**32**]{} (1991) 132–146.
J. Potthoff and M. Timpel, On a dual pair of spaces of smooth and generalized random variables, [*Potential Analysis*]{} [**4**]{} (1995) 637–654.
G. Styan, Hadamard product and multivariate statistical analysis, *Linear algebra and its applications* **6** (1973) 217-240.
F. Y. Wang, A generalization of Poincaré and log-Sobolev inequalities, *Potential Analysis* [**22**]{} (2005) 1-15.
F. Y. Wang and J. Wang, Functional inequalities for convolution probability measures, *arXiv:1308.1713*.
D. Zimmermann, Logarithmic Sobolev inequalities for mollified compactly supported measures, *J. Funct. Anal.* [**265**]{} (2013) 1064-1083.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Nous montrons comment associer [à]{} une gerbe d[é]{}finie sur un corps de nombres une obstruction de Brauer-Manin mesurant, comme dans le cas des vari[é]{}t[é]{}s, le d[é]{}faut d’existence d’une section globale. Ceci nous conduit [à]{} une g[é]{}n[é]{}ralisation de la dualit[é]{} de Tate-Poitou au cas non-ab[é]{}lien.'
author:
- 'Jean-Claude Douai - Michel Emsalem - St[é]{}phane Zahnd $^{\left(\ast\right)}$'
date: 18 mars 2003
title: 'Vari[é]{}t[é]{}s de descente, gerbes et obstruction de Brauer-Manin'
---
[^1]
Introduction {#introduction .unnumbered}
============
Soient $k$ un corps de nombres, $\bar{f}$ un $\bar{k}$-rev[ê]{}tement de corps des modules $k$, et $\mathcal{G}\left(\bar{f}\right)$ la $k$-gerbe associ[é]{}e [à]{} $\bar{f}$ (*i.e* la gerbe des mod[è]{}les de $\bar{f}$, cf [@DD1]). Dans [@DDM], nous avons introduit la notion de vari[é]{}t[é]{} de descente associ[é]{}e [à]{} $\bar{f}$. Si $V$ est une telle vari[é]{}t[é]{}, la gerbe $\mathcal{G}\left(\bar{f}\right)$ est alors isomorphe au champ quotient $\left[V/GL_{n}\right]$, pour un $n$ idoine. En fait, il existe une infinit[é]{} possible de telles vari[é]{}t[é]{}s de descente $V$ correspondant [à]{} une infinit[é]{} de choix possibles pour l’entier $n$. Soit maintenant $K$ une extension de $k$: tout $K$-point de $V$ d[é]{}finit un $K$-point[^2] de $\mathcal{G}$, et r[é]{}ciproquement tout $K$-point de $\mathcal{G}$ se rel[è]{}ve en un $K$-point de $V$. Il s’ensuit que si $V$ et $V'$ sont deux vari[é]{}t[é]{}s de descente correspondant [à]{} la m[ê]{}me $k$-gerbe $\mathcal{G}$, alors: $$V\left(K\right)\neq\emptyset\Leftrightarrow V'\left(K\right)\neq\emptyset$$ Forts de ces observations, on veut comparer les invariants de $V$ et $V'$; ils ne d[é]{}pendent que de $\mathcal{G}$. En particulier: $$Br_{a}V\approx Br_{a}V'\approx Br_{a}\mathcal{G},$$ et $$Pic\:V\approx Pic\:V'\approx Pic\:\mathcal{G}.$$ Ceci nous am[è]{}ne [à]{} calculer l’invariant de Brauer-Manin[^3] $m_{\EuScript{H}}\left(V\right)$ de $V$, [à]{} introduire l’invariant de Brauer-Manin $m_{\EuScript{H}}\left(\mathcal{G}\right)$ de la gerbe $\mathcal{G}$, puis [à]{} prouver que $m_{\EuScript{H}}\left(V\right)=m_{\EuScript{H}}\left(\mathcal{G}\right)$, l’int[é]{}r[ê]{}t de cette [é]{}galit[é]{} [é]{}tant sa validit[é]{} pour toute vari[é]{}t[é]{} de descente $V$ correspondant [à]{} $\mathcal{G}$ (plus loin, nous dirons que $V$ est une *pr[é]{}sentation* de $\mathcal{G}$).
Tout ce qui pr[é]{}c[è]{}de s’[é]{}tend aux $k$-gerbes quelconques localement li[é]{}es par un groupe fini (pour des raisons [é]{}videntes, de telles gerbes seront appel[é]{}es *gerbes de Deligne-Mumford*). L’application $m_{\EuScript{H}}$ qui [à]{} une classe de $k$-gerbes $\left[\mathcal{G}\right]$ associe l’invariant de Brauer-Manin d’un de ses repr[é]{}sentants peut alors [ê]{}tre vue comme une g[é]{}n[é]{}ralisation de la dualit[é]{} de Tate-Poitou dans le cas ab[é]{}lien (nous renvoyons au th[é]{}or[è]{}me 4.1 pour un [é]{}nonc[é]{} pr[é]{}cis); cet invariant vit dans le groupe de Tate-Shafarevich $$\mathcyr{SH}^{1}\left(k,\widehat{\bar{H}}\right)^{D}$$ o[ù]{} $\bar{H}$ est le groupe d’automorphismes d’un objet de $\mathcal{G}\left(Spec\:\bar{k}\right)$.
Notations {#notations .unnumbered}
---------
$k$ d[é]{}signe dans ces lignes un corps de caract[é]{}ristique nulle (souvent un corps de nombres), dont on fixe une cl[ô]{}ture alg[é]{}brique $\bar{k}$. Pour toute $k$-vari[é]{}t[é]{} alg[é]{}brique $V$, nous appelons groupe de Brauer cohomologique (ou simplement groupe de Brauer, lorsqu’aucune confusion n’est possible) de $V$ et nous noterons $Br\:V$ le groupe $H^{2}_{\acute{e}t}\left(V,\mathbb{G}_{m}\right)$. Nous noterons $\bar{V}$ la $\bar{k}$-vari[é]{}t[é]{} obtenue [à]{} partir de $V$ par extension des scalaires, le carr[é]{} suivant [é]{}tant alors commutatif: $$\xymatrix{\bar{V}=V\otimes_{k}\bar{k} \ar[d] \ar@//[r]^{\epsilon} & V \ar[d]^{p}\\ Spec\:\bar{k} \ar[r]& Spec\:{k}}$$ De ce diagramme, on d[é]{}duit deux morphismes: $$Br\:k\stackrel{\widetilde{p}}{\longrightarrow}Br\:V\stackrel{\widetilde{\epsilon}}{\longrightarrow}Br\:\bar{V}$$ On note $Br_{0}V$ (resp. $Br_{1}V$) l’image de $\widetilde{p}$ (resp. le noyau de $\widetilde{\epsilon}$), et $Br_{a}V$ le quotient $Br_{1}V/Br_{0}V$. Pour tout groupe alg[é]{}brique $G$, $\widehat{G}$ d[é]{}signe le groupe des caract[è]{}res de $G$; lorsque $A$ est un groupe ab[é]{}lien, nous noterons $A^{D}$ le dual (de Pontrjagin) de $A$: $Hom\left(A,\mathbb{Q}/\mathbb{Z}\right)$. Si $\bar{H}$ est un $Gal\left(\bar{k}/k\right)$-module, on note: $$H^{i}\left(k,\bar{H}\right)=H^{i}\left(Gal\left(\bar{k}/k\right),\bar{H}\right)\ \ \ \left(i=1,2\right)$$ Si de plus $k$ est un corps de nombres, on d[é]{}finit: $$\mathcyr{SH}^{i}\left(k,\bar{H}\right)=\ker\left\{H^{i}\left(k,\bar{H}\right)\longrightarrow\prod_{all\:v}{H^{i}\left(k_{v},\bar{H}\right)}\right\}\ \ \ \left(i=1,2\right)$$
[Rappels]{}
[Calcul de $Br_{a}V$ dans le cas o[ù]{} $V$ est un espace homog[è]{}ne de $SL_{n}$ avec isotropie $H$]{} $\ $
Soient $k$ un corps de caract[é]{}ristique nulle, et $V$ une $k$-vari[é]{}t[é]{} alg[é]{}brique lisse, g[é]{}om[é]{}triquement irr[é]{}ductible. De la suite spectrale $$H^{p}\left(k,H^{q}_{\acute{e}t}\left(\bar{V},\mathbb{G}_{m}\right)\right)\Longrightarrow H^{p+q}_{\acute{e}t}\left(V,\mathbb{G}_{m}\right)$$ on d[é]{}duit la suite exacte longue $$\xymatrix{0 \ar[r] & H^{1}\left(k,\bar{k}\left[V\right]^{\ast}\right) \ar[r] & Pic\:V \ar[r] & Pic\:\bar{V}^{Gal\left(\bar{k}/k\right)} \ar`dr_l[ll]`^dr[ll] [dl] & \\ &&Br_{1}V \ar[r]&H^{1}\left(k,Pic\:\bar{V}\right) \ar[r] & H^{3}\left(k,\bar{k}\left[V\right]^{\ast}\right)}$$ Posons: $$U\left(\bar{V}\right)=\frac{\bar{k}\left[V\right]}{\bar{k}^{\ast}}$$ La suite exacte $\left(1\right)$ fournit une nouvelle suite exacte: $$Pic\:\bar{V}^{Gal\left(\bar{k}/k\right)}\rightarrow H^{2}\left(k,U\left(\bar{V}\right)\right) \rightarrow Br_{a}V \rightarrow H^{1}\left(k,Pic\:\bar{V}\right) \rightarrow H^{3}\left(k,U\left(\bar{V}\right)\right)$$ Supposons que $V$ est un $k$-espace homog[è]{}ne d’un $k$-groupe alg[é]{}brique semi-simple simplement connexe $\widetilde{G}$ (*e.g* $SL_{n}$) avec isotropie un groupe fini, c’est-[à]{}-dire: il existe un $\bar{k}$-groupe fini $\bar{H}$ tel que: $$\bar{V}=\widetilde{G}\left(\bar{k}\right)/\bar{H}$$ Nous avons alors la suite exacte $$0\longrightarrow U\left(\bar{V}\right) \longrightarrow U\left(\widetilde{G}\left(\bar{k}\right)\right)$$ provenant de la fibration $\widetilde{G}\left(\bar{k}\right)\rightarrow \bar{V}$. Or on sait (cf le lemme 6.5 (iii) de [@Sa]) que:
$$U\left(\widetilde{G}\left(\bar{k}\right)\right)=\widehat{\widetilde{G}\left(\bar{k}\right)}=0$$ La suite exacte $\left(2\right)$ se r[é]{}duit alors [à]{} l’isomorphisme: $$Br_{a}V\stackrel{\sim}{\longrightarrow}H^{1}\left(k,Pic\:\bar{V}\right)$$
[Exemples]{}
(i) Si $H=0$, alors $\bar{V}\approx\widetilde{G}\left(\bar{k}\right)$, et $Pic\:\bar{V}=0$ (car $Pic\:\widetilde{G}\left(\bar{k}\right)=0$ par le corollaire 4.5 de [@FI]).
(ii) Si $H=\mu$ est un $k$-sous-groupe central de $\widetilde{G}$, alors $V=G=\widetilde{G}/\mu$ est semi-simple, et $Pic\:\widetilde{G}\left(\bar{k}\right)=\widehat{\mu\left(\bar{k}\right)}$ (par le corollaire 4.6 de [@FI]), d’o[ù]{}:
$$Br_{a}V=Br_{a}G=H^{1}\left(\bar{k}/k,\widehat{\mu\left(\bar{k}\right)}\right)=H^{1}\left(k,\widehat{\mu}\right)$$ $Pic\:G$ et $Br_{a}G$ sont justiciables de la philosophie de Kottwitz: ce sont des invariants des groupes semi-simples qui sont nuls lorsque $G=\widetilde{G}$ est simplement connexe. Ils peuvent donc s’exprimer en fonction du centre $Z\left(^{L}G\right)$ du dual de Langlands de $G$ [@K]; lorsque $G$ est semi-simple, ce dernier coïncide avec le dual du noyau du rev[ê]{}tement universel de $G$.
Cette remarque vaut encore pour $$\mathcyr{B}\left(G\right)=\ker\left\{Br_{a}G\rightarrow{\prod_{all\ v}Br_{a}G_{v}}\right\}$$ dans le cas o[ù]{} $k$ est un corps de nombres (le produit [é]{}tant pris sur toutes les places $v$ de $k$).
(iii) Soient $k$ un corps de nombres et prenons pour $V$ un $k$-tore $T$. Alors (cf le lemme 6.9 de [@Sa]): $$Pic\:\bar{T}=H^{1}\left(k,\widehat{T}\right)\ \ et\ \ Br_{a}T=H^{2}\left(k,\widehat{T}\right)$$ En outre: $$\mathcyr{B}\left(T\right)\approx\mathcyr{SH}^{2}\left(k,\widehat{T}\right)\approx\mathcyr{SH}^{1}\left(k,\widehat{T}\right)^{D}$$ le deuxi[è]{}me isomorphisme [é]{}tant fourni par la dualit[é]{} de Kottwitz [@K] qui [é]{}tend aux tores celle de Tate-Poitou.
(iv) Consid[é]{}rons maintenant un $k$-espace homog[è]{}ne de $SL_{n}$ avec isotropie un groupe fini; on suppose donc qu’il existe un groupe fini $\bar{H}$ tel que: $$\bar{V}\approx SL_{n,\bar{k}}/\bar{H}$$ Alors $Pic\:\bar{V}\approx\widehat{\bar{H}}$ (cf [@BK2]). On dispose en effet de la $\bar{k}$-fibration: $$SL_{n,\bar{k}}\longrightarrow SL_{n,\bar{k}}/\bar{H}$$ [à]{} laquelle est attach[é]{}e la suite spectrale $$E^{p,q}_{2}=H^{p}\left(\bar{H},H^{q}\left(SL_{n},\mathbb{G}_{m}\right)\right)\Longrightarrow H^{p+q}_{\acute{e}t}\left(\bar{V},\mathbb{G}_{m}\right)$$ Dans cette derni[è]{}re, le terme $E_{2}^{0,1}$ est nul[^4], donc: $$H^{1}_{\acute{e}t}\left(\bar{V},\mathbb{G}_{m}\right)=H^{1}\left(\bar{H},\mathbb{G}_{m}\right)=Hom\left(\bar{H},\mathbb{G}_{m}\right)=\widehat{\bar{H}}$$ D’o[ù]{} la:
Soit $V$ un $k$-espace homog[è]{}ne d’un groupe semi-simple simplement connexe $\widetilde{G}$ avec isotropie un groupe fini $\bar{H}$. Alors [@BK2]: $$Br_{a}V\approx H^{1}\left(k,\widehat{\bar{H}}\right)$$ En outre, si $k$ est un corps de nombres, et si on suppose que $V$ a des points localement partout (i.e si $V_{v}=V\otimes_{k}k_{v}$ a un $k_{v}$-point, pour toute place $v$ de $k$), alors: $$\mathcyr{B}\left(V\right)=\ker\left\{Br_{a}V\rightarrow{\prod_{all\ v}Br_{a}V_{v}}\right\}\approx \mathcyr{SH}^{1}\left(k,\widehat{\bar{H}}\right)$$
Sous les hypoth[è]{}ses et notations de la proposition pr[é]{}c[é]{}dente, si $\bar{H}$ est sans caract[è]{}re, alors $Br_{a}V=\mathcyr{B}\left(V\right)=0$.
[Interpr[é]{}tation comme champ quotient des $k$-gerbes localement li[é]{}es par un groupe alg[é]{}brique fini]{} On s’int[é]{}resse donc dans cette section aux $k$-gerbes qui sont des champs de Deligne-Mumford [@LMB]. Rappelons d’abord la proposition 5.1 de [@DDM]:
Soient $k$ un corps, et $\mathcal{G}$ une $k$-gerbe (pour la topologie [é]{}tale) qui est un champ de Deligne-Mumford. Alors:
1. Il existe une $k$-alg[è]{}bre $L$ avec action [à]{} gauche d’un groupe fini $\Gamma$ admettant $k$ comme anneau des invariants telle que $\mathcal{G}$ soit isomorphe au champ quotient $\left[Spec\:L/\Gamma\right]$;
2. Il existe un $k$-sch[é]{}ma affine $V$, un entier $n\geq0$, une action [à]{} droite de $GL_{n,k}$ sur $V$ et un $1$-morphisme $\pi:V\rightarrow\mathcal{G}$ avec les propri[é]{}t[é]{}s suivantes:
(i) $\pi$ induit un isomorphisme du champ quotient $\left[V\right/GL_{n,k}]$ vers $\mathcal{G}$;
(ii) $V$ est lisse et g[é]{}om[é]{}triquement irr[é]{}ductible;
(iii) l’action de $GL_{n,k}$ sur $V$ est transitive et [à]{} stabilisateurs finis;
(iv) pour chaque extension $K$ de $k$, chaque objet de $\mathcal{G}\left(K\right)$ se rel[è]{}ve en un point de $V\left(K\right)$ via $\pi$.
En particulier, [à]{} cause de (iii) et (iv), si $K$ est une extension de $k$ telle que $\mathcal{G}\left(K\right)\neq\emptyset$,[^5] la $K$-vari[é]{}t[é]{} $V\otimes_{k}K$ est isomorphe au quotient de $GL_{n,K}$ par un groupe fini.
$\ $
(a) (b)
$\ $
$\ $ Partons de l’extension $\left(\EuScript{E}\right)$ de la remarque pr[é]{}c[é]{}dente: $$\left(\EuScript{E}\right):1\rightarrow{H}\rightarrow\Gamma\rightarrow{Gal\left(L/k\right)}\rightarrow{1}$$ $\Gamma$ est un groupe fini; on peut donc le plonger dans $SL_{n}$ pour un certain $n$, ce qui conduit au diagramme suivant: $$\left(D\right):\xymatrix{1\ar[r] & H \ar[r] \ar@{=}[d] & \Gamma \ar[r] \ar[d] & Gal\left(L/k\right) \ar[r] \ar@{-->}[d] & 1\\ 1 \ar[r] & H \ar[r] & SL_{n,\bar{k}} \ar@{-->}[r] & SL_{n,\bar{k}}/H \ar[r] & 1}$$ $SL_{n,\bar{k}}/H$ n’est pas un groupe, puisque $H$ n’est pas n[é]{}cessairement normal dans $SL_{n,\bar{k}}$. C’est seulement un $k$-espace homog[è]{}ne (toujours au sens de Springer [@Sp]), d’o[ù]{} la pr[é]{}sence des pointill[é]{}s dans le diagrammme pr[é]{}c[é]{}dent. La fl[è]{}che verticale $$\xymatrix{Gal\left(L/k\right) \ar@{-->}[d]\\ SL_{n,\bar{k}}/H }$$ donne lieu [à]{} un $1$-cocycle dans $Z^{1}\left(L/k;SL_{n},H\right)$, qui repr[é]{}sente pr[é]{}cis[é]{}ment la classe du $k$-espace homog[è]{}ne $V$ du (2) de la proposition 2.1. La $k$-gerbe $\mathcal{G}\approx\left[V/SL_{n}\right]$ (associ[é]{}e [à]{} $\left(\EuScript{E}\right)$) s’interpr[è]{}te alors comme la gerbe des rel[è]{}vements du $k$-espace homog[è]{}ne $V$ [à]{} $SL_{n}$. En d’autres termes, $\left[\mathcal{G}\right]$ est l’image de $V$ par le cobord (cf [@Sp], [@D1]) $$\delta^{1}:Z^{1}\left(L/k;SL_{n},H\right)\longrightarrow H^{2}\left(k,\mathcal{L}_{H}\right)$$ Dans la suite, nous appellerons **pr[é]{}sentation de $\mathcal{G}=\left[V/SL_{n}\right]$** un couple $\left(V,\pi\right)$ comme dans la proposition 2.1.
[Invariant de Brauer-Manin d’une $k$-gerbe localement li[é]{}e par un groupe fini]{} Les $k$-champs alg[é]{}briques (en particulier les $k$-gerbes qui sont des champs de Deligne-Mumford) sont des g[é]{}n[é]{}ralisations de la notion de sch[é]{}ma[^6]. Par suite, il est tout-[à]{}-fait naturel de d[é]{}finir l’obstruction de Brauer-Manin d’une $k$-gerbe de mani[è]{}re analogue [à]{} celle d’un $k$-sch[é]{}ma.
Soit donc $\mathcal{G}$ une $k$-gerbe, qui est un champ de Deligne-Mumford; on suppose son $k$-lien localement repr[é]{}sentable par un groupe fini $H$. Le site [é]{}tale de $\mathcal{G}$ est d[é]{}fini au chapitre 12 de [@LMB]. Le groupe de Brauer cohomologique $Br\:\mathcal{G}=H^{2}_{\acute{e}t}\left(\mathcal{G},\mathbb{G}_{m}\right)$ est d[é]{}fini dans [@Ve] (o[ù]{}, d’une mani[è]{}re plus g[é]{}n[é]{}rale, la cohomologie d’un topos localement annel[é]{} est d[é]{}finie). Il existe une suite spectrale (cf [@Ve], prop. 5.3): $$H^{p}\left(k,H^{q}_{\acute{e}t}\left(\bar{\mathcal{G}},\mathbb{G}_{m}\right)\right)\Longrightarrow H^{p+q}_{\acute{e}t}\left(\mathcal{G},\mathbb{G}_{m}\right)$$ qui permet de d[é]{}finir $Br_{0}\mathcal{G}$, $Br_{1}\mathcal{G}$ et $Br_{a}\mathcal{G}$ de mani[è]{}re analogue [à]{} celle de la section 1.
Soit $\left(V,\pi\right)$ une pr[é]{}sentation de $\mathcal{G}$ (d’apr[è]{}s nos conventions, on a donc $\mathcal{G}\approx\left[V/SL_{n}\right]$). Posons: $$U\left(\bar{\mathcal{G}}\right)=\frac{\bar{k}\left[\mathcal{G}\right]^{\ast}}{\bar{k}^{\ast}}$$ On a: $$U\left(\bar{\mathcal{G}}\right)\subset U\left(\bar{V}\right) \subset U\left(SL_{n,\bar{k}}\right)=\widehat{SL_{n,\bar{k}}}=0$$ L’analogue de la suite exacte $\left(2\right)$ associ[é]{}e [à]{} la suite spectrale pr[é]{}c[é]{}dente implique alors: $$\begin{array}{rl}Br_{a}\mathcal{G}&\approx H^{1}\left(k,Pic\:\bar{\mathcal{G}}\right)\\ &\approx H^{1}\left(k,Hom\left(\Pi_{1}\left(\bar{\mathcal{G}}\right),\mathbb{G}_{m}\right)\right)\\&\approx H^{1}\left(k,Hom\left(\bar{H},\mathbb{G}_{m}\right)\right)\\\end{array}$$ car il est bien connu que $\Pi_{1}\left(\bar{\mathcal{G}}\right)=\bar{H}$ (cf [@No]), et finalement: $$Br_{a}\mathcal{G}\approx H^{1}\left(k,\widehat{\bar{H}}\right)$$
Si $K$ est un corps qui est une extension quelconque de $k$, nous pouvons d[é]{}finir un accouplement: $$\begin{array}{ccl}\mathcal{G}\left(K\right)\times Br\:\mathcal{G}& \rightarrow & Br\:K\\ \left(x,b\right) & \longmapsto & b\left(x\right)\end{array}$$ o[ù]{} l’image $b\left(x\right)$ peut [ê]{}tre interpr[é]{}t[é]{}e de diff[é]{}rentes façons ($\mathcal{B}$ d[é]{}signe ci-dessous un repr[é]{}sentant de $b$):
(i) $b\left(x\right)$ est la gerbe r[é]{}siduelle de $\mathcal{B}$ au point $x$ du champ alg[é]{}brique $\mathcal{G}=\left[V/SL_{n}\right]$;
(ii) $x$ peut [ê]{}tre vu comme une section au dessus de $Spec\:K$ du (1-)morphisme structural $\mathcal{G}\rightarrow{Spec\:k}$; autrement dit, c’est un (1-)morphisme rendant commutatif le diagramme (de morphismes de champs) suivant: $$\xymatrix{&\mathcal{G} \ar[d]\\Spec\:K \ar[r] \ar[ur]^{x}&Spec\:k}$$ $\mathcal{B}$ [é]{}tant un champ (puisque c’est une gerbe) sur $\mathcal{G}$, on peut consid[é]{}rer le champ image inverse $x^{\ast}\mathcal{B}$ de $\mathcal{B}$ par le morphisme $x$; la gerbe $x^{\ast}\mathcal{B}$ ainsi obtenue[^7] correspond exactement [à]{} $b\left(x\right)$; elle est donc obtenue par pull-back [à]{} partir de $x$ et de $\mathcal{B}$: $$\xymatrix{&\mathcal{B} \ar@{-}[d]\\x^{\ast}\mathcal{B} \ar@{-}[d] \ar@{--}[ur]^{x^{\ast}}&\mathcal{G} \ar[d]\\Spec\:K \ar[r] \ar[ur]^{x}&Spec\:k}$$
(iii) on peut interpr[é]{}ter $\mathcal{B}$ comme une alg[è]{}bre d’Azumaya sur $\mathcal{G}$ (*i.e* un torseur sous $PGL_{n}\left(\mathcal{O}_{\mathcal{G}}\right)$ pour la topologie [é]{}tale sur $\mathcal{G}$); ainsi, la fibre de $\mathcal{B}$ en $x$ est simplement une alg[è]{}bre simple centrale sur $K$, et c’est pr[é]{}cis[é]{}ment l’[é]{}l[é]{}ment $b\left(x\right)$ de $Br\:K$ recherch[é]{}. Notons que dans le cas g[é]{}n[é]{}ral, on ne peut utiliser cette interpr[é]{}tation puisque les deux groupes de Brauer (cohomologique et ) ne coïncident pas n[é]{}cessairement; on sait cependant qu’ils sont [é]{}gaux dans de nombreuses situations (cf [@G2], [@Ga], [@Sc]…) et que cette [é]{}galit[é]{} tient en particulier dans le cas des sch[é]{}mas affines [@Ga]. Or le groupe de Brauer de $\mathcal{G}$ est reli[é]{} au groupe de Brauer d’une quelconque de ses pr[é]{}sentations $V$; comme un tel $V$ est un espace homog[è]{}ne de $SL_{n}$, $V$ est en particulier affine, ce qui rend l[é]{}gitime notre interpr[é]{}tation.
De plus, toute $K$-section de la gerbe $\mathcal{G}$ (*i.e* tout objet de $\mathcal{G}\left(K\right)$) est un $H$-torseur sur $Spec\:K$ ($\mathcal{G}$ est par d[é]{}finition localement [é]{}quivalente [à]{} la gerbe $Tors\:H$; l’existence d’une $K$-section implique que $\mathcal{G}_{\left|K\right.}$ est [é]{}quivalente [à]{} la gerbe des $H$-torseurs sur $K$). Si on suppose que $K$ est un corps local, on obtient alors l’[é]{}nonc[é]{} suivant:
Soient $K$ un corps local, $\mathcal{G}$ une $K$-gerbe li[é]{}e par un groupe ab[é]{}lien fini $H$, et $\left(V,\pi\right)$ une pr[é]{}sentation de $\mathcal{G}$. Le diagramme suivant est commutatif: $$\xymatrix@C=2pt@R=25pt{\left(Acc.1\right)&V\left(K\right)\ar@{->>}[d]&\times&Br_{a}V\ar[dr] \\\left(Acc.2\right)&\mathcal{G}\left(K\right)\ar@{->>}[d]&\times&Br_{a}\mathcal{G}\ar @<1ex> [u]^{\approx}\ar[r]&\mathbb{Q}/\mathbb{Z} \\\left(Acc.3\right)& H^{1}\left(K,\bar{H}\right)&\times&H^{1}\left(K,\widehat{\bar{H}}\right) \ar @<1ex> [u]^{\approx} \ar[ur]}$$ o[ù]{}:
- l’accouplement $\left(Acc.1\right):V\left(K\right)\times Br_{a}V\longrightarrow\mathbb{Q}/\mathbb{Z}$ est d[é]{}fini comme suit: [à]{} un point $x$ de $V\left(K\right)$, et [à]{} une classe $b$ dans $Br_{a}V$, on associe: $$\left\langle x,b\right\rangle=\left[s_{x}\left(b\right)\right]_{x}$$ $s_{x}$ d[é]{}signant la section[^8] induite par $x$ de la projection canonique $p$ (cf [@BK1]): $$\xymatrix{Br_{1}V \ar[r]_{p}&Br_{a}V \ar@/_12pt/[l]_{s_{x}}}$$
- l’accouplement $\left(Acc.2\right):\mathcal{G}\left(K\right)\times Br_{a}\mathcal{G}\longrightarrow\mathbb{Q}/\mathbb{Z}$ est d[é]{}fini de la m[ê]{}me mani[è]{}re que $\left(Acc.1\right)$;
- l’accouplement $\left(Acc.3\right):H^{1}\left(K,\bar{H}\right)\times H^{1}\left(K,\widehat{\bar{H}}\right)\longrightarrow\mathbb{Q}/\mathbb{Z}$ est l’accouplement de Tate pour les corps locaux.
$\ $
Lorsque $k$ est un corps de nombres, ce que nous supposons [à]{} partir de maintenant, nous pouvons d[é]{}finir pour toute place $v$ de $k$ l’accouplement: $$\begin{array}{ccl} \mathcal{G}\left(k_{v}\right)\times Br_{1}\:\mathcal{G}& \rightarrow & \mathbb{Q}/\mathbb{Z}\\ \left(x,b\right) & \longmapsto & inv_{v}\left(b\left(x\right)\right)\end{array}$$ o[ù]{} comme d’habitude $inv_{v}$ est l’invariant donn[é]{} par la th[é]{}orie du corps de classes, et $b\left(x\right)$ est la classe dans $Br\:k_{v}$ de $\mathcal{B}_{x}$, o[ù]{} $\mathcal{B}$ est un repr[é]{}sentant de $b$. Supposons maintenant que $\mathcal{G}\left(k_{v}\right)$ soit non-vide pour toute place $v$ de $k$, et restreignons nous au sous-groupe $\mathcyr{B}\left(\mathcal{G}\right)$ de $Br_{a}\mathcal{G}$ d[é]{}fini par: $$\mathcyr{B}\left(\mathcal{G}\right)=\ker\left\{Br_{a}\mathcal{G}\longrightarrow \prod_{all\:v}{Br_{a}\:\mathcal{G}_{\left|k_{v}\right.}}\right\}$$ Nous d[é]{}finissons ainsi un accouplement: $$\begin{array}{rccl}\left\langle.\:,\:.\right\rangle: &
$\(\displaystyle\prod_{all\:v}{\mathcal{G}\left(k_{v}\right)}\)$ \times
\mathcyr{B}\left(\mathcal{G}\right)& \longrightarrow & \mathbb{Q}/\mathbb{Z}\\ &
\left(\left(x_{v}\right)_{v},b\right) & \longmapsto &
$\(\displaystyle\sum_{all\:v}{inv_{v}\left(\widetilde{b}\left(x_{v}\right)\right)}\)$\end{array}$$ o[ù]{} $\widetilde{b}$ est un relev[é]{} de $b$ dans $Br_{1}\mathcal{G}$. Par analogie avec la d[é]{}finition usuelle de cet accouplement (cf [@Bo1]), $\left\langle
\left(x_{v}\right)_{v},b\right\rangle$ ne d[é]{}pend pas de $\left(x_{v}\right)_{v}$, et $\left\langle
\left(x_{v}\right)_{v},b\right\rangle\neq0$ est une obstruction [à]{} l’existence d’une section $k$-rationnelle de $Spec\:k$ (*i.e* d’un objet de la cat[é]{}gorie fibre $\mathcal{G}\left(k\right)$). Nous obtenons de cette façon un [é]{}l[é]{}ment bien d[é]{}fini: $$m_{\mathcal{H}}\left(\mathcal{G}\right)\in \mathcyr{B}\left(\mathcal{G}\right)^{D}=Hom\left(\mathcyr{B}\left(\mathcal{G}\right),\mathbb{Q}/\mathbb{Z}\right)=Hom\left(\mathcyr{SH}^{1}\left(k,\widehat{\bar{H}}\right),\mathbb{Q}/\mathbb{Z}\right)$$ puisque: $\mathcyr{B}\left(\mathcal{G}\right)\approx\mathcyr{SH}^{1}\left(k,\widehat{\bar{H}}\right)$ (c’est une cons[é]{}quence imm[é]{}diate de l’isomorphisme $\left(3\right)$).
Soient $k$ un corps de nombres et $\mathcal{G}$ une $k$-gerbe. Pour toute pr[é]{}sentation $\left(V,\pi\right)$ de $\mathcal{G}$: $$\mathcyr{B}\left(V\right)\stackrel{\sim}{\longleftarrow}\mathcyr{B}\left(\mathcal{G}\right)$$ et $m_{\mathcal{H}}\left(\mathcal{G}\right)$ est [é]{}gale [à]{} l’image de $m_{\mathcal{H}}\left(V\right)$ par l’isomorphisme: $$\mathcyr{B}\left(V\right)^{D}\stackrel{\sim}{\longrightarrow}\mathcyr{B}\left(\mathcal{G}\right)^{D}$$
$\ $
L’isomorphisme $\mathcyr{B}\left(V\right)\stackrel{\sim}{\longleftarrow}\mathcyr{B}\left(\mathcal{G}\right)$ r[é]{}sulte de l’isomorphisme $$\mathcyr{SH}^{1}\left(k,Pic\:\bar{V}\right)\stackrel{\sim}{\longrightarrow}\mathcyr{SH}^{1}\left(k,Pic\:\bar{\mathcal{G}}\right)$$ ce dernier [é]{}tant induit par les isomorphismes compos[é]{}s $$H^{1}\left(k,Pic\:\bar{V}\right)\approx H^{1}\left(k,\widehat{\bar{H}}\right)\approx H^{1}\left(k,Pic\:\bar{\mathcal{G}}\right)$$ Nous avons le diagramme commutatif: $$\xymatrix@C=2pt@R=6pt{\prod{V\left(k_{v}\right)}\ar @<-1ex> @{->>}[dd]\ \ \times & \mathcyr{B}\left(V\right) \ar[dr] \\&&& \mathbb{Q}/\mathbb{Z}\\\prod{\mathcal{G}\left(k_{v}\right)}\ \ \times & \mathcyr{B}\left(\mathcal{G}\right) \ar@<1ex>[uu]^{\approx} \ar[ur]}$$ dans lequel la surjectivit[é]{} de la fl[è]{}che de gauche provient de la proposition 2.1(iv). On a vu que le calcul de $m_{\mathcal{H}}\left(\mathcal{G}\right)$ ne d[é]{}pendait pas de la famille $\left(x_{v}\right)_{v}$ choisie dans $\displaystyle\prod_{all\:v}{\mathcal{G}\left(k_{v}\right)}$. De la m[ê]{}me mani[è]{}re, on sait que le calcul de $m_{\mathcal{H}}\left(V\right)$ ne d[é]{}pend pas non plus de la famille $\left(y_{v}\right)_{v}$ choisie dans $\displaystyle\prod_{all\:v}{V\left(k_{v}\right)}$. Pour calculer $m_{\mathcal{H}}\left(V\right)$, on peut donc prendre pour $\left(y_{v}\right)_{v}$ n’importe quel rel[è]{}vement de la famille $\left(x_{v}\right)_{v}$. On en d[é]{}duit que $m_{\mathcal{H}}\left(\mathcal{G}\right)$ n’est autre que l’application compos[é]{}e: $$\mathcyr{B}\left(\mathcal{G}\right)\stackrel{\sim}{\longrightarrow}\mathcyr{B}\left(V\right)\xrightarrow{m_{\mathcal{H}}\left(V\right)}\mathbb{Q}/\mathbb{Z}$$ o[ù]{} $m_{\mathcal{H}}\left(V\right)\in \mathcyr{B}\left(V\right)^{D}$ est donn[é]{}e par: $b\longmapsto \left\langle
\left(y_{v}\right)_{v},b\right\rangle$.
Dans la suite, nous verrons l’[é]{}l[é]{}ment $m_{\mathcal{H}}\left(V\right)$ comme un [é]{}l[é]{}ment de $\mathcyr{SH}^{1}\left(k,\widehat{\bar{H}}\right)^{D}$.
[$1/2$-th[é]{}or[è]{}me de Tate-Poitou pour les groupes non-ab[é]{}liens]{} $\ $
Soient $k$ un corps de nombres, $H$ un $k$-groupe fini, $\mathcal{L}_{H}$ un $k$-lien localement repr[é]{}sentable par $H$. L’application $$\begin{array}{rccc}m_{\mathcal{H}}: & \mathcyr{SH}^{2}\left(k,\mathcal{L}_{H}\right) & \longrightarrow & \mathcyr{SH}^{1}\left(k,\widehat{\bar{H}}\right)^{D}\\ & \left[\mathcal{G}\right] & \longmapsto & m_{\mathcal{H}}\left(\mathcal{G}\right)\end{array}$$ o[ù]{} $\mathcyr{SH}^{2}\left(k,\mathcal{L}_{H}\right)$ d[é]{}signe l’ensemble des classes d’[é]{}quivalence de gerbes localement li[é]{}es par $H$ admettant partout localement une section (i.e qui sont partout localement neutres), se factorise par $$\xymatrix{\mathcyr{SH}^{2}\left(k,\mathcal{L}_{H}\right) \ar[r]^{ab} \ar[dr]_{m_{\mathcal{H}}} & \mathcyr{SH}^{2}\left(k,\frac{\bar{H}}{\left[\bar{H},\bar{H}\right]}\right) \ar[d]_{\approx}\\ & \mathcyr{SH}^{1}\left(k,\widehat{\frac{\bar{H}}{\left[\bar{H},\bar{H}\right]}}\right)^{D}}$$ o[ù]{} $ab$ est l’application d’ab[é]{}lianisation naturelle, et l’isomorphisme vertical est fourni par la dualit[é]{} de Tate-Poitou.
$\ $
1. 2.
[breitestes Label]{} M. Borovoi, *The Brauer-Manin obstructions for homogeneous spaces with connected or abelian stabilizer*, J. reine angew. Math. **473**, 181-194, 1996. M. Borovoi, *Abelianization of the second non-abelian Galois cohomology*, Duke Mathematical J. **72**, 217-239 , 1993. M. Borovoi, B. Kunyavsky, *Brauer equivalence in a homogeneous space with connected stabilizer*, Michigan Math. J. 49, 197-205, 2000. M. Borovoi, B. Kunyavsky, *On the Hasse Principle for homogeneous spaces with finite stabilizersx*, Ann. Fac. Sci. Toulouse 6 (3), 481-497, 1997. P. D[è]{}bes, J-C. Douai, *Gerbes and covers*, Communications in Algebra, 27(2), 577-594, 1999. P. D[è]{}bes, J-C. Douai, *Algebraic covers: field of moduli versus field of definition*, Ann. Scient. ENS, t.30, 303-338, 1997. P. D[è]{}bes, J-C. Douai, L. Moret-Bailly, *Descent varieties for algebraic covers*. J-C. Douai, *2-cohomologie galoisienne des groupes semi-simples*, Th[è]{}se de Doctorat, Universit[é]{} des Sciences et Techniques de Lille, 1976. D. Edidin, B. Hassett, A. Kresch, A. Vistoli,*Brauer groups and quotient stacks*, arXiv:mathAG/9905049 v3, fev 2001. R. Fossum, B. Iversen, *On Picard Groups of algebraic fibre spaces*, Journal of Pure and Applied Algebra 3, 269-280, 1973. O. Gabber, *Some theorems on Azumaya algebras*, LNM **844**, 129-209, Springer, Berlin, 1970. J. Giraud, *Cohomologie non-ab[é]{}lienne*, Springer-Verlag, 1971. A. Grothendieck, *Sites et topos [é]{}tales d’un sch[é]{}ma*, expos[é]{} VII, SGA 4, LNM **270**, Springer-Verlag, 1972. A. Grothendieck, *Le groupe de Brauer I, II et III*, dans *10 Expos[é]{}s sur la cohomologie des sch[é]{}mas*, Advanced Studies in Pure Mathematics, Masson et Cie, Paris, 1968. R.E. Kottwitz, *Stable trace formula: cuspidal tempered terms*, Duke Math. J. 51, 3, 611-650, 1984. G. Laumon, L. Moret-Bailly, *Champs alg[é]{}briques*, Springer-Verlag, 2000. A.S. Merkurjev, *Kaplansky conjecture in the theory of quadratic forms*, Journal of Soviet Math., 57, 3489-3497, 1991. B. Noohi, *Fundamental groups of algebraic stacks*, arXiv:mathAG/0201021 v1, jan 2002. J-J. Sansuc, *Groupe de Brauer et arithm[é]{}tique des groupes alg[é]{}briques lin[é]{}aires sur un corps de nombres*, J. reine angew. Math. **327**, 12-80, 1981. S. Schroer, *There are enough Azumaya algebras on surfaces*, arXiv:mathAG/0003229 v2, avr 2001. T.A. Springer, *Nonabelian $H^{2}$ in Galois cohomology*, Proc. Symp. Pure Mathematics, AMS, IX, 1966, pp 164-182. J-L. Verdier, *Cohomologie dans les topos*, expos[é]{} V, SGA 4, LNM **270**, Springer-Verlag, 1972.
$$\begin{array}{lll}\textup{\footnotesize{[email protected]}}\\\textup{\footnotesize{[email protected]}}\\\textup{\footnotesize{[email protected]}}\end{array}$$
[^1]: $^{\left(\ast\right)}$ Universit[é]{} des Sciences et Techniques de Lille I, Laboratoire AGAT, B[â]{}timent M2, 59655 Villeneuve d’Ascq, France.
[^2]: Par $K$-point de $\mathcal{G}$, on entend section de la gerbe au-dessus de $Spec\:K$, ou encore objet de la cat[é]{}gorie fibre de $\mathcal{G}$ au-dessus de $Spec\:K$ (que nous noterons $\mathcal{G}\left(Spec\:k\right)$).
[^3]: Le de est mis pour faire r[é]{}f[é]{}rence au Principe de Hasse.
[^4]: En effet, $Pic\:SL_{n}=H^{1}\left(SL_{n,\bar{k}},\mathbb{G}_{m}\right)=Hom\left(\Pi_{1}\left(SL_{n}\right),\mathbb{G}_{m}\right)=0$ puisque $SL_{n}$ est simplement connexe.
[^5]: Ce qui signifie que l’ensemble d’objets $Ob\left(\mathcal{G}\left(K\right)\right)$ de la cat[é]{}gorie fibre de $\mathcal{G}$ au-dessus de $Spec\:K$ est non-vide; par la suite, nous fairons syst[é]{}matiquement cet abus de langage.
[^6]: Plus rigoureusement, il existe un foncteur pleinement fid[è]{}le de la cat[é]{}gorie des $k$-sch[é]{}mas dans celle des $k$-champs alg[é]{}briques, d[é]{}fini en associant [à]{} un $k$-sch[é]{}ma $S$ le champ discret $S^{ch}$ qu’il repr[é]{}sente. Les objets de $S^{ch}$ n’ont pas d’automorphisme non-trivial, et r[é]{}ciproquement tout $k$-champ alg[é]{}brique dont les objets n’ont pas d’automorphisme non-trivial provient d’un $k$-espace alg[é]{}brique via ce foncteur. Nous renvoyons [à]{} [@LMB] pour plus de d[é]{}tails.
[^7]: L’image inverse d’une gerbe par un morphisme de champs est toujours une gerbe [@Gi].
[^8]: En effet, l’existence d’un point $K$-rationnel entra[î]{}ne que la suite:$$0\longrightarrow Br\:K\longrightarrow Br_{1}V\longrightarrow Br_{a}V\longrightarrow 0$$ est scind[é]{}e.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'In this paper we show that for a Koszul Calabi-Yau algebra, there is a shifted bi-symplectic structure in the sense of Crawley-Boevey-Etingof-Ginzburg [@CBEG], on the cobar construction of its co-unitalized Koszul dual coalgebra, and hence its DG representation schemes, in the sense of Berest-Khachatryan-Ramadoss [@BKR], have a shifted symplectic structure in the sense of Pantev-Toën-Vaquié-Vezzosi [@PTVV].'
address: 'Department of Mathematics, Sichuan University, Chengdu, Sichuan Province, 610064 P. R. China'
author:
- Xiaojun Chen
- Farkhod Eshmatov
date: 'February 8, 2018'
title: 'Calabi-Yau algebras and the shifted noncommutative symplectic structure'
---
Introduction {#sect:intro}
============
The notion of Calabi-Yau algebras was introduced by Ginzburg in 2007. It is a noncommutative generalization of affine Calabi-Yau varieties, and due to Van den Bergh, admits the so-called “noncommutative Poincaré duality". In a joint paper with Berest and Ramadoss [@BCER] we showed that for a Koszul Calabi-Yau algebra, say $A$, there is a version of noncommutative Poisson structure, called the [*shifted double Poisson structure*]{}, on its cofibrant resolution, and hence induces a shifted Poisson structure on the [*derived representation schemes*]{} of $A$, a notion introduced by Berest, Khachatryan and Ramadoss in [@BKR]. Here “shifted" means there is a degree shifting, depending on the dimension of the Calabi-Yau algebra, on the Poisson bracket. Such shifted Poisson structure was later further studied in [@CEEY] in much detail.
A natural question that arises now is, given a Koszul Calabi-Yau algebra, whether or not the noncommutative Poisson structure associated to it is noncommutative symplectic. To justify this question, let us remind a version of noncommutative symplectic structure introduced by Crawley-Boevey, Etingof and Ginzburg in [@CBEG], which they called the [*bi-symplectic structure*]{} in the paper. For an associative algebra, a bi-symplectic structure on it is a closed 2-form in its Karoubi-de Rham complex that induces an isomorphism between the space of noncommutative vector fields (more precisely, the space of [*double derivations*]{}) and the space of noncommutative 1-forms. In [@VdB Appendix], Van den Bergh showed that any bi-symplectic structure naturally gives rise to a double Poisson structure, which is completely analogous to the classical case. However, the reverse is in general not true. Nevertheless, it is still interesting to ask if this is true in the special case of Calabi-Yau algebras.
The second motivation of the paper comes from the 2012 paper [@PTVV] of Pantev, Toën, Vaquié and Vezzosi, where they introduced the notion of [*shifted symplectic structure*]{} for derived stacks. This not only generalizes the classical symplectic geometry to a much broader context, but also reveals many new features on a lot of geometric spaces, especially on the various moduli spaces that mathematicians are now studying. As remarked by the authors, the shifted symplectic structure, if it exists, always comes from the Poincaré duality of the corresponding source spaces; they also outlined how to generalize the shifted symplectic structure to noncommutative spaces, such as Calabi-Yau categories.
Note that Calabi-Yau algebras are highly related to Calabi-Yau categories. For example, a theorem of Keller (see [@Keller1 Lemma 4.1]) says that the bounded derived category of a Calabi-Yau algebra is a Calabi-Yau category. Applying the idea of [@PTVV], we would expect that the noncommutative Poincaré duality of a Calabi-Yau algebra shall also play a role in the corresponding shifted noncommutative symplectic structure if it exists.
The main results of the current paper may be summarized as follows. Let $A$ be a Calabi-Yau algebra of dimension $n$. Assume $A$ is also Koszul, and denote its Koszul dual coalegbra by $A^{{\textup{!`}}}$. Let $\tilde R=\Omega(\tilde A^{{\textup{!`}}})$ be the cobar construction of the co-unitalization $\tilde A^{{\textup{!`}}}$ of $A^{{\textup{!`}}}$. In this paper, we show that $\tilde R$, rather than $\Omega(A^{{\textup{!`}}})$ as studied in [@BCER; @CEEY], has a $(2-n)$-shifted bi-symplectic structure. Such bi-symplectic structure comes from the volume form of the noncommutative Poincaré duality of $A$, and naturally induces a $(2-n)$-shifted symplectic structure on the DG representation schemes $\mathrm{Rep}_V(\tilde R)$, for all vector spaces $V$ (see Theorem \[thm:mainthm1\]). By taking the corresponding trace maps we obtain a commutative diagram (see for more details) $$\label{digram:PDandshiftedsymplectic}
\xymatrixcolsep{4pc}
\xymatrix{
\mathrm{HH}^\bullet(A)\ar[r]^-{\mathrm{Tr}}\ar[d]^{\cong}
&\mathrm H^\bullet(\mathrm{Der}(\mathrm{Rep}_V(\tilde R))^{\mathrm{GL}})\ar[d]^{\cong}\\
\mathrm{HH}_{n-\bullet}(A)\ar[r]^-{\mathrm{Tr}}
&\mathrm H_{n-\bullet}(\Omega^1_{\mathrm{com}}(\mathrm{Rep}_V(\tilde R))^{\mathrm{GL}}),
}$$ where the left hand side are the Hochschild cohomology and homology of $A$, with the isomorphism being Van den Bergh’s noncommutative Poincaré duality, and the right hand side are the cohomology and homology of the $\mathrm{GL}(V)$-invariant complexes of vector fields and 1-forms on $\mathrm{Rep}_V(\tilde R)$ respectively.
By Van den Bergh’s result mentioned above, the shifted bi-symplectic structure induces a shifted double Poisson structure on $\tilde R$; therefore there is a shifted Poisson structure on $\mathrm{Rep}_V(\tilde R)$. In this paper, we will study the deformation quantization of such shifted Poisson structure, and show that it comes from the quantization of the “functions" $\tilde R_\natural:=\tilde R/[\tilde R,\tilde R]$ of $\tilde R$ (see Theorem \[thm:existenceofquantization\]). By Koszul duality, the homology of $\tilde R_{\natural}$ minus the unit is isomorphic to the cyclic homology of $A$, and thus we also obtain a quantization of the latter. This construction is inspired by quantization of quiver representations.
This paper is organized as follows. In §\[sect:basicsofNCgeometry\] we collect some basics of noncommutative geometry, such as the noncommutative 1-forms and vector fields, and some operations, such as the contraction and Lie derivative between them, then we recall the definition of bi-symplectic structure introduced by Crawley-Boevey et. al. in [@CBEG].
In §\[sect:Koszulduality\] we recall the notion of Koszul algebras and some of their basic properties. Let $A$ be a Koszul algebra over a field $k$. We give explicit formulas for the double derivations and 1-forms of $\tilde R$. The commutator quotient spaces of them are identified with the Hochschild cohomology and homology of $A$ respectively.
In §\[sect:KoszulCY\] we first recall the definition of Calabi-Yau algebras, and then show that if the Calabi-Yau algebra, say $A$, is Koszul, then its noncommutative volume form gives a shifted bi-symplectic structure on $\tilde R$, where $\tilde R$ is given as above.
In §\[sect:defofsss\] we first recall the DG representation schemes of a DG algebra, which was introduced by Berest et. al. in [@BKR]. Following the works [@BKR; @CBEG], we see that if a DG algebra admits a shifted bi-symplectic structure, then its DG representation schemes have a shifted symplectic structure. We then apply it to the Koszul Calabi-Yau algebra case.
In §\[sect:shiftedPoisson\] and §\[sect:quantization\] we study the shifted Poisson structure on the representation schemes of $\tilde R$ and its quantization. We show that such quantization is induced by the quantization of $\tilde R_{\natural}$ as a Lie bialgebra. This is completely analogous to the papers of Schedler [@Schedler] and Ginzburg-Schedler [@GS], where the quantization of the representation spaces of doubled quivers is constructed, which is compatible with the quantization of the necklace Lie bialgebra of the quivers via the canonical trace map.
In §\[sect:DRep\], we briefly discuss some relationships of the current paper with the series of papers by Berest and his collaborators [@BCER; @BFR; @BKR; @BR], where the derived representation schemes of associative algebras were introduced and studied.
In the last section, §\[sect:example\], we give the two examples of Calabi-Yau alegebras, namely, the 3- and 4-dimensional Sklyanin algebras, and study the corresponding shifted bi-symplectic structure in some detail.
This paper is a sequel to [@BCER; @CEEY], where the shifted noncommutative Poisson structure associated to Calabi-Yau algebras was studied; however, in the current paper we try to be as self-contained as possible. When we were in the final stage of the paper, we learned that Y. T. Lam has done some similar work in his Ph.D. Thesis [@Lam]. In particular, Theorem \[conj:Lam\] is a verification of his [@Lam Conjecture 7.4.1].
Throughout the paper, $k$ is a field of characteristic zero, though in a lot of cases it is not necessarily to be so. All morphisms and tensors are graded over $k$ unless otherwise specified. DG algebras (respectively DG coalgebras) are unital and augmented (respectively co-unital and co-augmented), with the degree of the differential being $-1$. For a chain complex, its homology is denoted by $\mathrm H_\bullet(-)$, and its cohomology is given by $\mathrm H^\bullet(-):=\mathrm H_{-\bullet}(-)$.
Some basics of noncommutative geometry {#sect:basicsofNCgeometry}
======================================
In this section, we collect some basic notions in noncommutative geometry. They are mostly taken from [@CBEG; @VdB; @VdB2]; here we work in the differential graded (DG for short) setting.
Noncommutative differential forms
---------------------------------
Suppose $(R, \partial)$ is a DG associative algebra over $k$, where $\partial$ is the differential. The [*noncommutative 1-forms*]{} of $R$ is a DG $R$-bimodule generated by symbols $d x$ of degree $|x|$, linear in $x$ for all $x\in R$ and subject to the relations $$\begin{aligned}
d (xy) &=& (d x)y+x (d y),\\
\partial(dx)&=&d(\partial x),\end{aligned}$$ for all $x, y\in R$. In this paper, we also assume $d 1 =0$.
Alternatively, $\Omega^1_{\mathrm{nc}}R$ is the kernel of the multiplication $R\otimes R\to R$, which is a subcomplex of $R$-bimodules in $R\otimes R$, generated by $1\otimes x-x\otimes 1$ for all $x\in R$, with differential induced from $\partial$. The identification of the above two constructed $R$-bimodules is given by $dx\mapsto 1\otimes x-x\otimes 1$. Let $\Omega^1_{\mathrm{nc}}R[-1]$ be the suspension of $\Omega^1_{\mathrm{nc}}R$, i.e. the degrees of $\Omega^1_{\mathrm{nc}}R$ are shifted up by one. Sometimes we also write it in the form $\Sigma\Omega^1_{\mathrm{nc}}R$, which means the degree-shifting operator applies from the left. (In the following, for a graded vector space $V_\bullet$, $V[n]$ is a graded vector space with $(V[n])_i=V_{i+n}$.)
Let $\Omega^\bullet_{\mathrm{nc}}R:=T_R(\Omega^1_{\mathrm{nc}} R[-1])$ be the free tensor algebra generated by $\Omega^1_{\mathrm{nc}}R[-1]$ over $R$. Let $$d:R\to \Omega^1_{\mathrm{nc}}R[-1],\;
x\mapsto d(x)= \Sigma dx,$$ then by our sign convention, it is direct to check[^1] $$\begin{aligned}
d(xy)&=&d(x) \cdot y+(-1)^{|x|}x\cdot d(y),\\
d\circ \partial(x)&=&-\partial \circ d(x),\end{aligned}$$ for all $x, y\in R$.
Now let $d(d(x))=0$ for all $x\in R$. Extend $d$ to be a map $d:\Omega^\bullet_{\mathrm{nc}}R\to \Omega^{\bullet+1}_{\mathrm{nc}}R$ by derivation. Also, extend $\partial$ to $\Omega^\bullet_{\mathrm{nc}}R$ by derivation. Then we have $$d^2=0,\quad \partial^2=0,\quad\mbox{and}\quad d\circ \partial+\partial \circ d=0.$$ In general, $d$ is called the [*de Rham differential*]{} of $\Omega^\bullet_{\mathrm{nc}}R$, and for convenience, $\partial$ is called the [*internal differential*]{} of $\Omega^\bullet_{\mathrm{nc}}R$.
Note that $\partial$ and $d$ have degrees $-1$ and $1$ respectively, which make $\Omega^\bullet_{\mathrm{nc}}R$ into a mixed DG algebra and is called the set of [*noncommutative differential forms*]{} of $R$. Let $$\mathrm{DR}^\bullet_{\mathrm{nc}}R:=\big(\Omega^\bullet_{\mathrm{nc}}R\big)_{\natural}
=\Omega^\bullet_{\mathrm{nc}}R/[\Omega^\bullet_{\mathrm{nc}}R,\Omega^\bullet_{\mathrm{nc}}R]$$ be the graded commutator quotient space. Since $\partial$ and $d$ are both derivations with respect to the product on $\Omega^\bullet_{\mathrm{nc}}R$, $\mathrm{DR}^\bullet_{\mathrm{nc}}R$ with the induced differentials, still denoted by $(\partial,d)$, is a mixed complex, and is called the [*Karoubi-de Rham complex*]{} of $R$.
Noncommutative polyvector fields
--------------------------------
Following [@CBEG; @VdB], the noncommutative vector fields on an associative algebra are given by the [*double derivations*]{}. By definition, the space of double derivations $\mathbb D\mathrm{er}\; R$ of $R$ is the set of derivations $\mathrm{Der}(R, R\otimes R)$. Since the map $R\to \Omega^1_{\mathrm{nc}}R, x\mapsto dx=1\otimes x-x\otimes 1$ is a universal derivation, meaning that every derivation of $R$ factors through $\Omega^1_{\mathrm{nc}}R$ ([*c.f.*]{} [@Quillen Proposition 3.3]), we have that $$\mathbb D\mathrm{er}\; R\cong\mathrm{Hom}_{R^e}(\Omega^1_{\mathrm{nc}} R, R\otimes R),$$ where $R^e$ is the enveloping algebra $R\otimes R^{\mathrm{op}}$. In the above notation, we have used the [*outer $R$-bimodule*]{} structure on $R\otimes R$; namely, for any $x, y, u, v\in R$, $$u\cdot (x\otimes y)\cdot v:=ux\otimes yv.$$ $R\otimes R$ has also an [*inner $R$-bimoule*]{} structure, which is given by $$u * (x\otimes y)* v:=(-1)^{|u||x|+|v||y|+|u||v|}xv\otimes uy.$$ With the inner $R$-bimodule structure on $R\otimes R$, $\mathbb D\mathrm{er}\; R$ is a DG $R$-bimodule. Now let $
T_R(\mathbb D\mathrm{er}\; R[1])
$ be the free DG algebra generated by $\mathbb D\mathrm{er}\; R[1]$ over $R$, which is called the space of [*(noncommutative) polyvector fields*]{} of $R$.
Actions of noncommutative vectors on noncommutative forms
---------------------------------------------------------
Analogously to the classical case, the noncommutative vectors act on noncommutative forms by contraction and by Lie derivative, which together form the noncommutative version of the Cartan identity (see Lemma \[Cartanformula\]).
### Contraction and Lie derivative
For any $\Theta\in\mathbb D\mathrm{er}\; R[1]$, the [*contraction*]{} operator $$i_{\Theta}:\Omega^\bullet_{\mathrm{nc}}R\longrightarrow
\Omega^{\bullet-1}_{\mathrm{nc}}R.$$ is given as follows: first, let $$i_{\Theta}(a)=0,\quad\mbox{for all}\; a\in R$$ and $$i_{\Theta}: \Omega^1_{\mathrm{nc}} R[-1]\to R\otimes R,\;
d(a)\mapsto \Theta(a),$$ where we have used the fact that $\mathbb D\mathrm{er}\; R[1]=\mathrm{Hom}_{R^e}(\Omega^{1}_{\mathrm{nc}} R[-1], R\otimes R).$ Second, extend $i_{\Theta}$ to be an $R$-linear operator on $\Omega^\bullet_{\mathrm{nc}}R$ in the natural way, that is, if we write $$i_{\Theta} a=\sum i_{\Theta}'a\otimes i_{\Theta}'' a, \quad\mbox{for}\; a\in\Omega^1_{\mathrm{nc}} R,$$ then $$i_{\Theta}(a_1\cdots a_n)=\sum_{1\le k\le n}(-1)^{\sigma_k}
a_1\cdots a_{k-1}(i_{\Theta}'a_k)\otimes (i_{\Theta}'' a_k) a_{k+1}\cdots a_n,$$ for $a_1,\cdots, a_n\in \Omega^1_{\mathrm{nc}} R$, where $(-1)^{\sigma_k}$ is the Koszul sign. Remind that in general, the Koszul sign comes from switching the positions of graded elements: for two graded vector spaces $V, W$, the isomorphism $V\otimes W\to W\otimes V$ is given by $v\otimes w\mapsto (-1)^{|v||w|}w\otimes v$; for example, in the above, $(-1)^{\sigma_k}=(-1)^{|\Theta|(|a_1|+\cdots+|a_{k-1}|)}$. In the following, we will sometimes omit its precise value if it is clear from the context.
Next, given $\Theta\in\mathbb D\mathrm{er}\; R[1]$ the [*Lie derivative*]{} of $\Theta$ on the noncommutative differential forms is given by $$\begin{aligned}
L_{\Theta}(x_0dx_1\cdots dx_n)
&:=&
\Theta'(x_0)\otimes\Theta''(x_0)dx_1\cdots dx_n\\
&&+\sum_{1\le k\le n}(-1)^{\sigma_k}\big(x_0dx_1\cdots dx_{k-1}d\Theta'(x_k)\otimes\Theta''(x_k)dx_{k+1}\cdots dx_n\\
&&\quad\quad+(-1)^{|\Theta'(x_k)|}x_0dx_1\cdots dx_{k-1} \Theta'(x_k)\otimes d\Theta''(x_k) dx_{k+1}\cdots dx_n\big).\end{aligned}$$
### Reduced contraction and Lie derivative
Now, for a graded algebra $A$ and $c=c_1\otimes c_2\in A\otimes A$, let $^\circ c=(-1)^{|c_1||c_2|}c_2c_1$, and given a map $\phi: A\to A\otimes A$, write $^\circ\phi: A\to A: c\mapsto\;\!^\circ(\phi(c))$. Define the [*reduced contraction operator*]{} $\iota$ and the [*reduced Lie derivative*]{} $\mathscr L$ by $$\label{formula:reducedcontraction}
\iota_{\Theta}(-): \Omega^\bullet_{\mathrm{nc}}R \to \Omega^{\bullet-1}_{\mathrm{nc}}R,\;
\omega \mapsto \iota_{\Theta}\omega=\;\!^{\circ}(i_{\Theta}(\omega))$$ and $$\mathscr L_{\Theta}(-): \Omega^\bullet_{\mathrm{nc}}R \to \Omega^{\bullet}_{\mathrm{nc}}R,\;
\omega \mapsto \mathscr L_{\Theta}\omega=\;\!^\circ(L_{\Theta}(\omega))$$ respectively. The following two lemmas are proved in [@CBEG].
1. The reduced contraction defined above only depends on the image of $\omega$ in $\mathrm{DR}_{\mathrm{nc}}^\bullet R $; in other words, $\iota_{\Theta}$ descends to a well-defined map $\mathrm{DR}_{\mathrm{nc}}^\bullet R \to \Omega^{\bullet-1}_{\mathrm{nc}}R$.
2. For $\omega\in\mathrm{DR}^n_{\mathrm{nc}} R $, the map $\Theta\mapsto\iota_{\Theta}\omega$ gives an $R$-bimodule morphism $
\mathbb D\mathrm{er}\; R[1]\to\Omega^{n-1}_{\mathrm{nc}} R
$.
\[Cartanformula\] Let $\iota$ and $\mathscr L$ be as above. Then for any $\Theta\in\mathbb D\mathrm{er}\; R[1]$, we have $$d\circ \iota_{\Theta}+\iota_{\Theta}\circ d=\mathscr L_{\Theta}\quad\mbox{and}\quad
d\circ\mathscr L_{\Theta}=\mathscr L_{\Theta}\circ d.$$
Moreover, from the definitions of $\iota$ and $\mathscr L$, it is clear that both of them respect $\partial$. With these preparations, we now recall the definition of [*bi-symplectic structures*]{} of Crawley-Boevey, Etingof and Ginzburg introduced in [@CBEG] (we here rephrase it for DG algebras):
Suppose $R$ is a DG associative algebra. A closed form $\omega\in\mathrm{DR}^2_{\mathrm{nc}} R $ of total degree $2-n$ is called an [*$n$-shifted bi-symplectic structure*]{} if the following map $$\label{isofrombisymplectic}
\iota_{(-)}\omega: \mathbb{D}\mathrm{er}\; R[1] \to (\Omega^1_{\mathrm{nc}}R[-1])[2-n],\;
\Theta \mapsto \iota_{\Theta}\omega$$ is a quasi-isomorphism of complexes of $R$-bimodules.
The reader may refer to Remark \[rmk:degreeshifting\] for some discussions on the degree shifting in the above definition.
In general, it is difficult to check the existence of a bi-symplectic structure for a given DG algebra. The difficulty lies in the fact there is in general no closed formula for the noncommutative vector fields and 1-forms. However, for free associative algebras, such as the path algebra of a quiver, both of them can be explicitly written down. This leads Crawley-Boevey et. al. [@CBEG] to give an explicit identification of them, and hence to give a bi-symplectic structure on the path algebra of a doubled quiver. In the following two sections, we generalize their construction to the DG case, where Koszul Calabi-Yau algebras naturally appear.
Koszul duality {#sect:Koszulduality}
==============
Koszul duality was originally introduced by Priddy to compute the Hochschild homology and cohomology of associative algebras. Nowadays it plays a more and more important role in the study of noncommutative algebraic geometry. Suppose $A$ is a Koszul algebra, and $A^{{\textup{!`}}}$ its Koszul dual coalgebra. Let $\tilde R$ be the cobar construction of $\tilde A^{{\textup{!`}}}$, where $\tilde A^{{\textup{!`}}}$ is the co-unitalization of $A^{{\textup{!`}}}$. In this section, we give explicit formulas for $\Omega^1_{\mathrm{nc}}\tilde R$ and $\mathbb D\mathrm{er}\;\tilde R$, which are very much related to the Hochschild homology and cohomology of $A$ (see Propositions \[Prop:cotangentcomplex\] and \[Prop:tangentcomplex\] below).
Koszul algebra {#subsect:Koszul}
--------------
Let $W$ be a finite-dimensional vector space over $k$. Denote by $TW$ the free algebra generated by $W$ over $k$. Suppose $Q$ is a subspace of $W\otimes W$, and let $(Q)$ be the two-sided ideal generated by $Q$ in $TW$, then the quotient algebra $A:= TW/(Q)$ is called a [*quadratic algebra*]{}.
Consider the subspace $$U=\bigoplus_{n=0}^\infty U_n:=
\bigoplus_{n=0}^\infty \bigcap_{i+j+2=n}W^{\otimes i}\otimes Q\otimes
W^{\otimes j}$$ of $TW$, then $U$ is a coalgebra whose coproduct is induced from the de-concatenation of the tensor products. The [*Koszul dual coalgebra*]{} of $A$, denoted by $A^{{\textup{!`}}}$, is $$A^{{\textup{!`}}}=\bigoplus_{n=0}^\infty \Sigma^{\otimes n} (U_n).$$ $A^{{\textup{!`}}}$ has a graded coalgebra structure induced from that of $U$ with $$(A^{{\textup{!`}}})_0=k, \quad (A^{{\textup{!`}}})_1=\Sigma W, \quad (A^{{\textup{!`}}})_2=(\Sigma\otimes\Sigma)(Q),\quad\cdots\cdots$$
The [*Koszul dual algebra*]{} of $A$, denoted by $A^!$, is just the linear dual space of $A^{{\textup{!`}}}$, which is then a graded algebra. More precisely, let $W^*=\mathrm{Hom}(W, k)$ be the linear dual space of $W$, and let $Q^\perp$ denote the space of annihilators of $Q$ in $W^*\otimes W^*$. Shift the grading of $W^*$ down by one, denoted by $\Sigma^{-1}W^*$, then $$A^!=T(\Sigma^{-1}W^*)/(\Sigma^{-1}\otimes\Sigma^{-1}\circ Q^{\perp}).$$
Choose a set of basis $\{e_i\}$ for $W$, and let $\{e_i^*\}$ be their duals in $W^*$. Let $\{\xi_i\}$ be the basis in $\Sigma^{-1}W^*$ corresponding to $\{e_i^*\}$, i.e. $\xi_i=\Sigma^{-1}e_i^*$. There is a chain complex associated to $A$, called the [*Koszul complex*]{}: $$\label{Koszul_complex}
\xymatrix{
\cdots\ar[r]^-{b'}&
A\otimes A^{{\textup{!`}}}_{i+1}\ar[r]^-{b'}&
A\otimes A^{{\textup{!`}}}_{i}\ar[r]^-{b'}&
\cdots\ar[r]&
A\otimes A^{{\textup{!`}}}_0\ar[r]^-{b'}& k,
}$$ where for any $r\otimes u\in A\otimes A^{{\textup{!`}}}$, $b'(r\otimes u)=\sum_i re_i\otimes\xi_i\vdash u$. Here $\xi_i\vdash u$ means the interior product (contraction) of $\xi_i$ with $u$, or in other words, the evaluation of $\xi_i$ on the first component of $u$; in the following, we prefer to write it in the form $u\cdot \xi_i$ or simply $u\xi_i$, since the interior product gives a right $A^!$-module structure on $A^{{\textup{!`}}}$. (In the following we will also use the contraction from the right, and in this case we write it in the form $\xi_i u$ for $\xi_i$ and $u$ as above.)
A quadratic algebra $A=TW/(Q)$ is called [*Koszul*]{} if the Koszul chain complex is acyclic.
Applying the graded version of Nakayama Lemma, we have the following (see [@VdB-1 Proposition 3.1] for a proof):
\[KoszulResolutionOfA\] Suppose $A$ is a Koszul algebra. Then the following complex $$\label{Comp:KoszulResolutionOfA}
\xymatrix{
\cdots\ar[r]&A\otimes A^{{\textup{!`}}}_m\otimes A\ar[r]^-b &A\otimes A^{{\textup{!`}}}_{m-1}\otimes A
\ar[r]^-b&
A\otimes A^{{\textup{!`}}}_0\otimes A\cong A\otimes A \ar[r]^-{\mu} &A,
}$$ where $$b(a\otimes c\otimes a')=
\sum_i \Big(ae_i\otimes c\xi_i\otimes a'+(-1)^{m}a\otimes \xi_i c \otimes e_i a'\Big)$$ for $a\otimes c\otimes a' \in A\otimes A^{{\textup{!`}}}_m\otimes A$, and $\mu $ is the multiplication on $A$, gives a resolution of $A$ as an $A^e$-module.
Denote $K(A):=(A\otimes A^{{\textup{!`}}}\otimes A, b)$, then $$\mathrm{Tor}_\bullet^{A^e}(A, A)=\mathrm{H}_\bullet(A\otimes_{A^e}K(A))
\quad\mbox{and}\quad
\mathrm{Ext}^\bullet_{A^e}(A, A)=\mathrm{H}^\bullet(\mathrm{Hom}_{A^e}(K(A), A)),$$ which are also identical to the Hochschild homology $\mathrm{HH}_\bullet(A)$ and cohomology $\mathrm{HH}^\bullet(A)$ of $A$ respectively. This result is due to Priddy:
\[lemma:identityofHochschild\] The complexes $A\otimes_{A^e}K(A)$ and $\mathrm{Hom}_{A^e}(K(A), A)$ are quasi-isomorphic to the Hochschild chain complex $\mathrm{CH}_\bullet(A)$ and the Hochschild cochain complexes $\mathrm{CH}^\bullet(A)$ respectively.
There are explicit formulas for the chain complexes $A\otimes_{A^e}K(A)$ and $\mathrm{Hom}_{A^e}(K(A), A)$: as graded vector spaces, they are the same as $A\otimes A^{{\textup{!`}}}_\bullet$ and $A\otimes A^!_{\bullet}$, while the differentials are given by $$\label{DifferentialInKoszulComplex}
b(a\otimes u)=\sum_{i}
\Big(a e_i \otimes u \xi_i +(-1)^{|a|} e_i a\otimes \xi_i u\Big),$$ and $$\label{DifferentialInComplexComputingHochschildCohomology}
b(a\otimes x)=\sum_i\Big(a e_i \otimes x \xi_i +(-1)^{|a|} e_i a\otimes \xi_i x\Big),$$ respectively. Van den Bergh [@VdB-1 Proposition 3.3] gave a formula for the quasi-isomorphism of these complexes.
Now for a Koszul algebra $A$, view it as a DG algebra concentrated in degree zero with trivial differential. Let $A^{{\textup{!`}}}$ be the Koszul dual coalgebra of $A$, and $\Omega(A^{{\textup{!`}}})$ be the cobar construction of $A^{{\textup{!`}}}$, whose differential is denoted by $\partial$. Then from one can deduce that the natural surjective map $$\label{Koszulresolution}
(\Omega(A^{{\textup{!`}}}),\partial)\to (A, 0),$$ is a quasi-isomorphism of DG algebras ([*c.f.*]{} [@LV Theorem 3.4.4]), which then gives a cofibrant resolution of $A$ in the category of DG associative algebras.
Noncommutative geometry for Koszul algebras
-------------------------------------------
Suppose $A$ is a Koszul algebra. Let $A^{{\textup{!`}}}$ and $A^!$ be its Koszul dual coalgebra and algebra respectively. Let $\tilde A^{{\textup{!`}}}$ be the co-unitalization of $A^{{\textup{!`}}}$; that is, $\tilde A^{{\textup{!`}}}=A^{{\textup{!`}}}\oplus k$ with the coproduct $\tilde \Delta$ being $$\begin{aligned}
\tilde\Delta(a)&=&\Delta(a)+a\otimes 1+1\otimes a,\; \mbox{for}\; a\in A^{{\textup{!`}}}, \\
\tilde\Delta(1)&=&1\otimes 1,\end{aligned}$$ where $\Delta(a)$ is the coproduct of $a$ in $A^{{\textup{!`}}}$. Let $\tilde R=\Omega(\tilde A^{{\textup{!`}}})$ be the cobar construction of $A^{{\textup{!`}}}$. The following is straightforward since $\tilde R$ is a quasi-free DG algebra:
Let $\tilde R=\Omega(\tilde A^{{\textup{!`}}})$. Then $$\Omega^1_{\mathrm{nc}}\tilde R[-1]\cong \tilde R\otimes A^{{\textup{!`}}}\otimes \tilde R \quad
\mbox{and}\quad
\mathbb D\mathrm{er}\;\tilde R[1]\cong \tilde R\otimes A^{!}\otimes \tilde R.$$
In fact, since $\tilde R=\Omega(\tilde A^{{\textup{!`}}})=T(A^{{\textup{!`}}}[1])$ is quasi-free, we have the short exact sequence $$0\longrightarrow \tilde R\otimes A^{{\textup{!`}}}[1]\otimes \tilde R\longrightarrow
\tilde R\otimes \tilde R\longrightarrow \tilde R\longrightarrow 0.$$ It follows that $\Omega^1_{\mathrm{nc}} \tilde R=\tilde R\otimes A^{{\textup{!`}}}[1]\otimes \tilde R$, that is, $\Omega^{1}_{\mathrm{nc}}\tilde R[-1]=\tilde R\otimes A^{{\textup{!`}}}\otimes \tilde R$.
From this, we also see that $$\begin{aligned}
\mathbb D\mathrm{er}\;\tilde R[1]&=&\mathrm{Hom}_{{\tilde R}^e}(\Omega^1_{\mathrm{nc}}\tilde R[-1], \tilde R\otimes\tilde R)\\
&=&\mathrm{Hom}_{{\tilde R}^e}(\tilde R\otimes A^{{\textup{!`}}}\otimes \tilde R, \tilde R\otimes\tilde R)\\
&=&\mathrm{Hom}(A^{{\textup{!`}}},\tilde R\otimes \tilde R)\\
&=&\tilde R\otimes A^!\otimes \tilde R,\end{aligned}$$ where in the last equality, we have identified $A^!\otimes (\tilde R\otimes\tilde R)$ with $\tilde R\otimes A^!\otimes
\tilde R$ via $a\otimes (u\otimes v)\mapsto (-1)^{(|a|+|u|)|v|}v\otimes a\otimes u$. Thus $\mathbb D\mathrm{er}\;\tilde R[1]\cong \tilde R\otimes A^{!}\otimes \tilde R$ follows.
The following proposition was obtained in [@BKR] (see also [@CEEY]), and therefore we will only sketch its proof. Given an associative algebra $A$, denote by $(\mathrm{CC}_\bullet(A),b)$ and $(\mathrm{CH}_\bullet(A), b)$ the Connes cyclic complex and the Hochschild chain complex of $A$ respectively.
\[Prop:cotangentcomplex\] Suppose $A$ is a Koszul algebra, and $\tilde R=\Omega(\tilde A^{{\textup{!`}}})$ with differential $\partial$. Then $$\label{qiscyclicandhoch}
(\bar{\tilde R}_\natural,\partial)\simeq(\mathrm{CC}_\bullet(A)[1],b)
\quad\mbox{and}
\quad(\mathrm{DR}^1_{\mathrm{nc}} \tilde R,\partial)\simeq(\mathrm{CH}_\bullet(A),b)$$ as chain complexes, where $\bar{\tilde R}$ is the augmentation ideal of $\tilde R$.
To prove this proposition, we have to recall the definition of the cyclic homology of coalgebras. Suppose $C$ is a DG coalgebra, let $\Omega(\tilde C)$ be the cobar construction of the co-unitalization of $C$. Let $\bar{\Omega}(\tilde C)$ be the augmentation ideal of $\Omega(\tilde C)$. Then by Quillen [@Quillen §1.3] the cyclic complex of $C$, denoted by $(\mathrm{CC}_\bullet(C), b)$, may take to be complex $(\bar{\Omega}(\tilde C)_{\natural}[-1],\partial)$.
On the one hand, by we have $$\label{quasiisoofcyclic1}
\mathrm{CC}_\bullet(\Omega(A^{{\textup{!`}}}))\simeq \mathrm{CC}_\bullet(A)$$ since quasi-isomorphic DG algebras have quasi-isomorphic cyclic chain complexes. On the other hand, by Jones-McCleary [@JM Theorem 1] (see also [@CYZ Lemma 17] for a proof from the Koszul duality point of view) we have $$\label{quasiisoofcyclic2}
\mathrm{CC}_\bullet(\Omega(A^{{\textup{!`}}}))\simeq\mathrm{CC}_\bullet(A^{{\textup{!`}}}).$$ Combining and we obtain $\bar{\tilde R}_\natural\simeq \mathrm{CC}_\bullet(A)[1]$.
Next, we show $\mathrm{DR}^1_{\mathrm{nc}}\tilde R\simeq\mathrm{CH}_\bullet(A)$. Since $\Omega^1_{\mathrm{nc}}\tilde R[-1]=\tilde R\otimes A^{{\textup{!`}}}\otimes\tilde R$, we have $$\mathrm{DR}^1_{\mathrm{nc}} \tilde R=
(\Omega^1_{\mathrm{nc}}\tilde R[-1])_{\natural}=A^{{\textup{!`}}}\otimes \tilde R=\mathrm{CH}_\bullet(A^{{\textup{!`}}}),$$ where $\mathrm{CH}_\bullet(A^{{\textup{!`}}})$ is the underlying space of the Hochschild chain complex of $A^{{\textup{!`}}}$. By a direct computation we also see that $\partial$ on $(\Omega^1_{\mathrm{nc}}\tilde R)_{\natural}$ coincides with the Hochschild boundary map on $\mathrm{CH}_\bullet(A^{{\textup{!`}}})$.
Again by Koszul duality, the same argument as above yields $
\mathrm{CH}_\bullet(A)\simeq\mathrm{CH}_\bullet(A^{{\textup{!`}}})
$ (see [@CYZ Theorem 15] for a complete proof), which implies $\mathrm{DR}^1_{\mathrm{nc}} \tilde R\simeq\mathrm{CH}_\bullet(A)$.
\[conv:cyclichomology\] In the rest of the paper, as adopted by Berest et. al. in [@BKR], when writing $\mathrm{CC}_\bullet(-)$, we always mean $\mathrm{CC}_\bullet(-)[1]$.
\[Prop:tangentcomplex\] Suppose $A$ is a Koszul algebra and $\tilde R=\Omega(\tilde A^{{\textup{!`}}})$. Then $((\mathbb D\mathrm{er}\;\tilde R[1])_{\natural},\partial)$ is quasi-isomorphic to $(\mathrm{Der}\; \tilde R[1],\partial)$, which is further quasi-isomorphic to the Hochschild cochain complex $\mathrm{CH}^\bullet(A)$.
Observe that as graded vector spaces $$\begin{aligned}
\mathbb D\mathrm{er}\; \tilde R[1]
&=&\mathrm{Hom}_{\tilde R^e}(\Omega^1_{\mathrm{nc}}\tilde R[-1],\tilde R\otimes\tilde R)\\
&=&\mathrm{Hom}_{\tilde R^e}(\tilde R\otimes A^{{\textup{!`}}}\otimes\tilde R,\tilde R\otimes\tilde R)\\
&=&\mathrm{Hom}(A^{{\textup{!`}}},\tilde R\otimes\tilde R)\\
&=&\tilde R\otimes A^{!}\otimes \tilde R,\end{aligned}$$ thus $$(\mathbb D\mathrm{er}\;\tilde R[1])_{\natural}=\mathrm{Der}\;\tilde R[1]
\cong \tilde R\otimes A^{!}
=\bigoplus_n\mathrm{Hom}((A^![-1])^{\otimes n}, A^!)
=\mathrm{CH}^\bullet(A^!).$$ Under this identity, a direct calculation identifies the differential on $\mathrm{Der}\;\tilde R$ with the Hochschild coboundary on $\mathrm{CH}^\bullet(A^!)$.
Finally, by Keller [@Keller Theorem 3.5] (see also Shoikhet [@Shoikhet Theorem 4.2]), for a Koszul algebra $A$, $\mathrm{CH}^\bullet(A)$ is quasi-isomorphic to $\mathrm{CH}^\bullet(A^!)$ as DG Lie algebras, and hence in particular, as chain complexes. Thus combining it with the above quasi-isomorphisms, we have $(\mathbb D\mathrm{er}\; \tilde R[1])_{\natural}\simeq \mathrm{CH}^\bullet(A)$.
Koszul Calabi-Yau algebras {#sect:KoszulCY}
==========================
The notion of Calabi-Yau algebras was introduced by Ginzburg [@Ginzburg] in 2007. Let $A$ be a Koszul Calabi-Yau algebra and $\tilde R=\Omega(\tilde A^{{\textup{!`}}})$ be as before. In this section, we show that the volume form of the noncommutative Poincaré duality of $A$ also gives the shifted bi-symplectic structure on $\tilde R$.
Suppose $A$ is an associative algebra over $k$. Then $A$ is called [*Calabi-Yau of dimension $n$*]{} (or [*$n$-Calabi-Yau*]{}) if
1. $A$ is homologically smooth, that is, $A$, viewed as a (left) $A^e$-module, has a bounded, finitely generated projective resolution, and
2. there is an isomorphism $$\label{cond:CY(ii)}
\eta: \mathrm{RHom}_{A^e}(A, A\otimes A)\cong A[n]$$ in the derived category $\mathbf D(A^e)$ of (left) $A^e$-modules.
In the above definition, the $A^e$-module structure on $A\otimes A$ and $\mathrm{RHom}_{A^e}(A, A\otimes A)$ is completely analogous to the case of $\mathbb D\mathrm{er}\; A$.
Suppose $A$ is a homologically smooth algebra, then Van den Bergh [@VdB0] showed that $$\begin{aligned}
\label{VdBsIso}
\mathrm{HH}_i(A) &=&\mathrm{H}_i(A\otimes^{\mathrm L}_{A^e} A)\nonumber\\
&\cong &
\mathrm{Hom}_{\mathbf D(A^e)}(\mathrm{RHom}_{A^e}(A, A\otimes A),A[i]),
\quad\mbox{for all}\; i\end{aligned}$$ and therefore for $A$ being Calabi-Yau, the isomorphism corresponds to an element in $\mathrm{HH}_n(A)$, which is still denoted by $\eta$ and is called the [*volume class*]{} of $A$.
In general, for an arbitrary Calabi-Yau algebra $A$, it is difficult to find its volume class. However, in the case when $A$ is also Koszul, this turns out to be very easy. Besides that, Koszul Calabi-Yau algebras have some other good features; for example, they form so far the most known and interesting examples of Calabi-Yau algebras, and they are all given by a [*superpotential*]{}.
The volume form
---------------
Now suppose $A$ is Koszul Calabi-Yau of dimension $n$. Then by Proposition \[KoszulResolutionOfA\], we have $$\begin{aligned}
&&\mathrm{Hom}_{\mathbf D(A^e)}(\mathrm{RHom}_{A^e}(A, A\otimes A), A[n])\\
&=&
\mathrm{Hom}_{\mathbf D(A^e)}(\mathrm{Hom}_{A^e}(A\otimes A^{{\textup{!`}}}\otimes A, A\otimes A),
A\otimes A^{{\textup{!`}}}[n]\otimes A)\\
&=&\mathrm{Hom}_{\mathbf D(A^e)}(\mathrm{Hom}_k(A^{{\textup{!`}}}, A\otimes A),
A\otimes A^{{\textup{!`}}}[n]\otimes A)\\
&=&\mathrm{Hom}_k(A^{!}, A^{{\textup{!`}}}[n]).\end{aligned}$$ Hence by combining it with , we see that the volume class of a Koszul $n$-Calabi-Yau algebra $A$ also gives an isomorphism of $A^{!}\cong A^{{\textup{!`}}}[n]$ as vector spaces. This implies that $A^!$ is a [*cyclic*]{} associative algebra of degree $n$. Let us recall its definition first.
Suppose $A$ is a graded associative algebra. It is called [*cyclic*]{} of degree $n$ if it admits a degree $n$, non-degenerate bilinear pairing $\langle-,-\rangle: A\times A\to k[n]$ such that $$\langle a\cdot b,c\rangle=(-1)^{(|a|+|b|)|c|}\langle c\cdot a,b\rangle,\quad\mbox{for all}\;
a,b,c\in A.$$
\[dualitybetweenCYcyclic\] Suppose that $A$ is a Koszul algebra. Then $A$ is $n$-Calabi-Yau if and only if $A^{!}$ is cyclic of degree $n$.
See Van den Bergh [@VdB3]. The interested reader may also refer to [@CYZ Proposition 28] for a simple proof.
In the literature, a cyclic associative algebra is sometimes also called a [*symmetric Frobenius*]{} algebra, or simply a [*symmetric*]{} algebra. If $A$ is a cyclic associative algebra of degree $n$, then its linear dual $A^*=\mathrm{Hom}(A, k)$, which is a coassociative coalgebra, also has a pairing $$\langle-,-\rangle: A^*\times A^* \to k[-n],\;
(\alpha, \beta) \mapsto \langle\alpha^*, \beta^*\rangle,$$ where $\alpha^*$ and $\beta^*$ are the images of $\alpha$ and $\beta$ under the map $A^{*}\stackrel{\cong}{\to} A$ induced by the pairing. The cyclicity condition of the pairing becomes $$\label{cyclicconditionforcoalg}
\sum_{(\beta)}\langle \alpha, \beta^1\rangle\cdot \beta^2
=\sum_{(\alpha)}(-1)^{|\alpha||\beta^1|}\langle \beta, \alpha^2\rangle \cdot \alpha^1,$$ for any $\alpha,\beta\in A^*$ with $\Delta(\alpha)=\sum_{(\alpha)}\alpha^1\otimes \alpha^2$ and $\Delta(\beta)=\sum_{(\beta)}\beta^1\otimes \beta^2$.
The above proposition in fact implies that the volume class of a Koszul Calabi-Yau algebra is represented by a(ny) nonzero top-degree element in $A^{{\textup{!`}}}$. More precisely, we have the following.
Suppose $A$ is a Koszul $n$-Calabi-Yau algebra. Then any nonzero element $\eta\in A^{{\textup{!`}}}_{n}$, viewed as an element in $A\otimes A^{{\textup{!`}}}$ via the embedding $A^{{\textup{!`}}}\cong k\otimes A^{{\textup{!`}}}\subset A\otimes A^{{\textup{!`}}}$ and hence a chain in $(A\otimes A^{{\textup{!`}}}_{\bullet}, b)$, is a cycle and represents the volume class of $A$.
First from the symmetric pairing $$\langle x, y\rangle=(-1)^{|x||y|}\langle y,x\rangle$$ we see that in particular $$\xi_i\eta=(-1)^{n}\eta \xi_i,\quad\mbox{for all}\; \xi_i\in \Sigma^{-1}W^*.$$ This means $\eta$ via the above embedding $A^{{\textup{!`}}}\cong k\otimes A^{{\textup{!`}}}\subset A\otimes A^{{\textup{!`}}}$ is a cycle, namely, $$b(1\otimes \eta)=\sum_i
e_i\otimes \xi_i\eta-(-1)^n e_i\otimes \eta \xi_i=0.$$ Also from the following operation $$\label{capgivesPD}
\begin{array}{ccl}
(A\otimes A^{!}, b)&\longrightarrow& (A\otimes A^{{\textup{!`}}}, b)\\
a\otimes f&\longmapsto& a\otimes (\eta\cap f)
\end{array}$$ it is direct to check that this is an isomorphism of chain complexes and induces an isomorphism on the homology $$\mathrm{H}^\bullet(A\otimes A^!, b)\cong\mathrm{H}_{n-\bullet}(A\otimes A^{{\textup{!`}}}, b).$$ Thus by Proposition \[lemma:identityofHochschild\] we have an isomorphism $$\label{NCPDofVdB}
\mathrm{HH}^\bullet(A)\cong\mathrm{HH}_{n-\bullet}(A).$$ We next need to check that the volume class represented above is identical to the one of Van den Bergh ([@VdB0]).
First, in [@dTdVVdB] de Thanhofer de Volcsey and Van den Bergh showed that the noncommutative Poincaré duality for Calabi-Yau algebras is given by a class $\eta\in\mathrm{HH}_n(A)$ such that $$\begin{array}{ccl}
\mathrm{HH}^\bullet(A)&\longrightarrow&\mathrm{HH}_{n-\bullet}(A)\\
f&\longmapsto&\eta\cap f
\end{array}$$ is an isomorphism, where the “cap product" $\cap$ is given on the Hochschild chain level as follows: $$\begin{array}{ccl}
\bar{\mathrm{CH}}_{m}(A)\times \bar{\mathrm{CH}}^{n}(A)&\stackrel{\cap}
\longrightarrow &\bar{\mathrm{CH}}_{m-n}(A)\\
(\alpha,\quad\quad f )&\longmapsto&
\left\{
\begin{array}{cl}
(a_0f(\bar{a}_1,\ldots,\bar{a}_n),\bar{a}_{n+1},\ldots,\bar{a}_m),&\mbox{if}\; m\geq n\\
0,&\mbox{otherwise,}
\end{array}\right.
\end{array}$$ where $\bar{\mathrm{CH}}_\bullet(A)$ and $\bar{\mathrm{CH}}^\bullet(A)$ are the reduced Hochschild chain and cochain complexes respectively ([*c.f.*]{} Loday [@Loday] for these notions).
Second, on the Koszul complexes we have an analogous cap product given by the following (note that is just a special case) $$\begin{array}{ccl}
(A\otimes A^{{\textup{!`}}})\times (A\otimes A^!)&\stackrel{\cap}\longrightarrow& A\otimes A^{{\textup{!`}}}\\
(a\otimes u, b\otimes f)&\longmapsto& ab\otimes ( uf).
\end{array}$$ It has been shown by Berger et. al. [@BLS] that for Koszul algebras, these two versions of cap product on the homology level, via the isomorphism given in Proposition \[lemma:identityofHochschild\] are the same. Thus the isomorphism is a version of noncommutative Poincaré duality in the sense of Van den Bergh.
Third, by Van den Bergh [@VdB3] the volume class of the noncommutative Poincaré duality, if it exists, is unique up to an inner automorphism of $A$.
Thus by the above three arguments, the noncommutative Poincaré duality of is identical to the one of Van den Bergh given in [@VdB0], possibly up to an inner automorphism of $A$. This completes the proof.
Now consider the coproduct $$\Delta(\eta)=\sum \eta^1_i\otimes\eta^2_i\in A^{{\textup{!`}}}\otimes A^{{\textup{!`}}}.$$ Observe that we have an embedding $$\label{embeddingofAac}
A^{{\textup{!`}}} \cong k\otimes A^{{\textup{!`}}} \otimes k\subset \tilde R\otimes
A^{{\textup{!`}}} \otimes \tilde R=\Omega^1_{\mathrm{nc}}\tilde R[-1],\quad a\mapsto d([a]).$$ Via this embedding, $\Delta(\eta)$ corresponds to an element $$\omega:=\sum d([\eta_i^1]) \otimes d([\eta_i^2])
\in\Omega^{2}_{\mathrm{nc}}\tilde R.$$
\[volumeformasdRcycle\] $\omega$ descends to a degree $n$ cycle in $\mathrm{DR}_{\mathrm{nc}}^2 \tilde R$.
Denote by $[\omega]$ the image of $\omega$ in $\mathrm{DR}^{2}_{\mathrm{nc}}\tilde R$. We show $[\omega]$ is closed with respect to both $d$ and $\partial$. First, $[\omega]$ is automatically $d$-closed. Second, applying $\partial$ to $\omega$, we have $$\begin{aligned}
\partial(\omega)
&=&\sum \partial\circ d ([\eta_i^1]) \otimes d ([\eta_i^2])+(-1)^{|\eta_i^1|}d([\eta_i^1])\otimes
\partial\circ d([\eta_i^2]) \\
&=&-\sum d\circ\partial ([\eta_i^1]) \otimes d ([\eta_i^2])+(-1)^{|\eta_i^1|}d([\eta_i^1])\otimes
d\circ\partial ([\eta_i^2]) \\
&=&-\sum (-1)^{|\eta_i^{11}|}d( [\eta_i^{11}|\eta_i^{12}]) \otimes d ([\eta_i^2])
+(-1)^{|\eta_{i}^1|+|\eta_i^{21}|}d([\eta_i^1])\otimes
d([\eta_i^{21}|\eta_i^{11}])\\
&=&-\sum (-1)^{|\eta_i^{11}|} d([\eta_i^{11}])\cdot [\eta_i^{12}]\otimes d([\eta_i^2])
-[\eta_i^{11}]\cdot d([\eta_i^{12}]) \otimes d([\eta_i^2])\\
&&-\sum(-1)^{|\eta_{i}^1|+|\eta_i^{21}|}
d([\eta_i^1])\otimes d([\eta_i^{21}])\cdot [\eta_i^{22}]-
(-1)^{|\eta_{i}^1|}d([\eta_i^1])\otimes [\eta_i^{21}]\cdot
d([\eta_i^{22}]),\end{aligned}$$ where we write $\Delta (\eta_i^1)=\sum \eta_i^{11}\otimes \eta_i^{12}$ and $\Delta (\eta_i^2)=\sum \eta_i^{21}\otimes \eta_i^{22}$. In the last equality, after descending to $\mathrm{DR}^2_{\mathrm{nc}}\tilde R$, the first and the last summands cancel with each other due to the co-associativity of $A^{{\textup{!`}}}$, while the second and the third summands cancel with each other due to the cyclic condition of the pairing on $A^!$, which is equivalent to . This proves the statement.
\[thm:mainthm1\] Let $A$ be a Koszul Calabi-Yau algebra of dimension $n$. Let $\tilde R=\Omega(\tilde A^{{\textup{!`}}})$. Then $\tilde R$ has a $(2-n)$-shifted bi-symplectic structure.
Since $\Omega^1_{\mathrm{nc}} \tilde R[-1]=\tilde R\otimes A^{{\textup{!`}}}\otimes \tilde R$, we have $$\begin{aligned}
\mathbb D\mathrm{er}\; \tilde R[1]&=&
\mathrm{Hom}_{\tilde R^e}(\Omega^1_{\mathrm{nc}} \tilde R[-1], \tilde R\otimes \tilde R)\\
&=&\mathrm{Hom}_{\tilde R^e}(\tilde R\otimes A^{{\textup{!`}}}\otimes \tilde R, \tilde R\otimes \tilde R)\\
&=&\mathrm{Hom}( A^{{\textup{!`}}}, \tilde R\otimes \tilde R)\\
&=&\mathrm{Hom}(A^{{\textup{!`}}}, k)\otimes(\tilde R\otimes\tilde R).\end{aligned}$$ For any $f\otimes r_1\otimes r_2\in\mathrm{Hom}(A^{{\textup{!`}}}, k)\otimes(\tilde R\otimes\tilde R)$, by its reduced contraction with $\omega$ is $$2\cdot\sum f(\eta_i^1)\cdot r_2\otimes d\eta_i^2\otimes r_1
\in\Omega^1_{\mathrm{nc}}\tilde R[-1].$$ Since $\omega$ is non-degenerate of total degree $n$, we thus have $$\label{iso:tangentandcotangent}
\mathbb D\mathrm{er}\tilde R[1]\cong(\Omega^1_{\mathrm{nc}}\tilde R[-1])[2-n]$$ as $\tilde R$-bimodules. This proves the theorem.
By taking the commutator quotient space of both sides of the above and Propositions \[Prop:cotangentcomplex\] and \[Prop:tangentcomplex\], we once again obtain the noncommutative Poincaré duality: $$\mathrm{HH}^\bullet(A)\cong\mathrm{HH}_{n-\bullet}(A),$$ which coincides with Van den Bergh’s one. In general, suppose $R$ has a (shifted) bi-symplectic structure $\omega$, then Crawley-Boevey et. al. showed in [@CBEG Lemma 2.8.6] that there is a commutative diagram $$\xymatrixcolsep{4pc}
\xymatrix{
\mathbb D\mathrm{er}\; R\ar[d]_{\iota_{(-)}\omega}^{\cong}\ar[r]^{\natural}&
(\mathbb D\mathrm{er}\; R)_{\natural}
\ar[d]^{(\iota_{(-)}\omega)_{\natural}}_{\cong}\\
\Omega^1_{\mathrm{nc}}R\ar[r]^{\natural}&\mathrm{DR}^1_{\mathrm{nc}}R.
}$$
Representation schemes and the shifted symplectic structure {#sect:defofsss}
===========================================================
In this section, we briefly go over the relationship between the shifted bi-symplectic structure of a DG algebra and the shifted symplectic structure on its DG representations.
Representation functors
-----------------------
Let $\mathbf{DGA}$ be the category of associative, unital DG $k$-algebras, and $\mathbf{CDGA}$ its subcategory of DG commutative $k$-algebras. Fix a finite dimensional vector space $V$, and consider the following functor: $$\mathrm{Rep}_V(A): \mathbf{CDGA}\to \mathbf{Sets},\;
B \mapsto \mathrm{Hom}_{\mathbf{DGA}}(A, B\otimes\mathrm{End}\; V).$$ The following result generalizes the result of Bergman [@Bergman] and Cohn [@Cohen] for associative algebras:
\[idoftwosets\] The functor $\mathrm{Rep}_V(A)$ is representable, that is, there exists an object in $\mathbf{CDGA}$, say $A_V$, depending only on $A$ and $V$, such that $$\label{equiv:homsets}
\mathrm{Hom}_{\mathbf{DGA}}(A, B\otimes\mathrm{End}\; V)=
\mathrm{Hom}_{\mathbf{CDGA}}(A_V, B).$$
More precisely, in [@BKR] the authors considered the following two functors: $$\sqrt[V]{-}:\mathbf{DGA}\to\mathbf{DGA},\; A\mapsto(A\ast_k\mathrm{End}\; V)^{\mathrm{End}\; V}.$$ and $$(-)_{\natural\natural}:\mathbf{DGA}\to\mathbf{CDGA},\; A\mapsto A/\langle [A,A]\rangle,$$ where $\langle [A,A]\rangle$ is the ideal of $A$ generated by the commutators. Then the functor $$\label{functor:Rep}
(-)_V:\mathbf{DGA}\to\mathbf{CDGA},\; A\mapsto A_V$$ is given by the composition of the above two functors, namely, $
A_V=(\sqrt[V]{A})_{\natural\natural}$. In the following we shall also write $A_V$ as $\mathrm{Rep}_V(A)$.
In , if we take $B=A_V$, then we have $$\label{equiv:universalmap}
\mathrm{Hom}(A, A_V\otimes\mathrm{End}\; V)=\mathrm{Hom}(A_V, A_V).$$ The identity map on the right hand side corresponds to a map $$\pi_V: A\to A_V\otimes\mathrm{End}\; V$$ on the left hand side, which is usually called the [*universal representation map*]{}.
For a quasi-free algebra $(R, \partial)$ and $V=k^n$, the DG commutative algebra $R_V$ can be described explicitly: Let $\{x^\alpha\}_{\alpha\in I}$ be a set of generators of $R$. Consider a free graded algebra $R'$ on generators $\{x_{ij}^\alpha: 1\le i, j\le n, \alpha\in I\}$, where $|x_{ij}^\alpha|=|x^\alpha|$ for all $i, j$. Form matrices $X^{\alpha}:=(x_{ij}^{\alpha})$, and define the algebra map $$\pi: R\to M_n(R'),\quad x^\alpha\mapsto X^{\alpha},$$ where $M_n(R')$ is the algebra of $n\times n$-matrices with entries in $R'$. Let $$\partial(x_{ij}^{\alpha}):=(\pi(\partial x^{\alpha}))_{ij},$$ and extend it to $R'$ by linearity and the Leibniz rule. We thus obtain a DG algebra $(R', \partial)$, and $R_V$ is $R_{\natural\natural}'$ with the differential induced from $\partial$ (see [@BKR Theorem 2.8] for a proof).
$\mathrm{GL}$-invariants and the trace map
------------------------------------------
Observe that $\mathrm{GL}(V)$ acts on $\mathrm{Rep}_V(A)$ by conjugation. More precisely, then any $g\in\mathrm{GL}(V)$ gives a unique automorphism of $A_V$ which, under the identity , corresponds to the composition $$A\stackrel{\pi_V}\longrightarrow \mathrm{End}(A_V)
\stackrel{\mathrm{Ad}(g)}\longrightarrow \mathrm{End}(A_V).$$ This action is natural in $A$, and hence defines a functor $$\label{functor:GLinvariants}
\mathrm{Rep}_V(-)^{\mathrm{GL}}: \mathbf{DGA}\to\mathbf{CDGA},\; A\mapsto (A_V)^{\mathrm{GL}},$$ where $(-)^{\mathrm{GL}}$ means the $\mathrm{GL}(V)$-invariants.
Now consider the following composite map $$A\stackrel{\pi_V}{\longrightarrow} \mathrm{End}(A_V)
\stackrel{\mathrm{Tr}}\longrightarrow A_V,$$ which is $\mathrm{GL}(V)$-invariant and factors through $A_{\natural}$, we get a map $$\label{tracemap}
\mathrm{Tr}: A_\natural\to\mathrm{Rep}_V(A)^{\mathrm{GL}}.$$ If $A$ is an associative algebra (i.e., a DG algebra concentrated in degree 0), then the famous result of Procesi says that the image of $\mathrm{Tr}$ generates $\mathrm{Rep}_V(A)^{\mathrm{GL}}$; in other words, if we extend $\mathrm{Tr}$ to be a commutative algebra map $$\mathrm{Tr}: \mathbf{\Lambda}^\bullet A_\natural\to
\mathrm{Rep}_V(A)^{\mathrm{GL}},$$ then it is surjective. However, for an arbitrary DG algebra, Berest and Ramadoss showed in [@BR] that this is in general not true; there are some homological obstructions for $\mathrm{Tr}$ to be so.
Van den Bergh’s functor
-----------------------
Let $(R,\partial)$ be a DG algebra. Suppose $(M,\partial_M)$ is a DG $R$-bimodule. Let $\pi: R\to R_V\otimes\mathrm{End}\; V$ be the universal representation of $R$, which means the map on the left hand side of that corresponds to the identity map on the right hand side. Then $\pi$ gives an $R$-bimodule structure on $R_V\otimes\mathrm{End}\; V$. Denote $$\label{functor:VdB}
M_V:=M\otimes_{R^e} (R_V\otimes\mathrm{End}\; V),$$ which is now a DG $R_V$-module. More specifically, let $V=k^n$, then $M_V$ is generated by symbols $m_{ij}$, $1\le i, j\le n$, for each $m\in M$, with the action of $R_V$ given by $$(r\cdot m)_{ij}=\sum_{k}r_{ik}\cdot m_{kj}, \quad
(m\cdot r)_{ij}=\sum_{k} r_{kj}\cdot m_{ik},$$ and with the differential, denoted by $\partial_{M_V}$, given by $$\partial_{M_V}(m_{ij})=(\partial_M m)_{ij},\quad\mbox{for all}\; m\in M.$$ The assignment from the category of DG $R$-bimodules to the category of DG $R_V$-modules $$\mathbf{DGBimod}\; R\to\mathbf{DGMod}\; R_V,\; M\mapsto M_V$$ is a well-defined functor, and is first introduced by Van den Bergh in [@VdB2]. Next, we apply Van den Berg’s idea to the case of noncommutative differential forms and poly-vectors.
Let $(R, \partial)$ be a DG commutative algebra over $k$. Let $$I:=\mathrm{ker}(R\otimes R\stackrel{\mu }
\longrightarrow R)$$ be the kernel of the multiplication map and let $\Omega^1_{\mathrm{com}}R:=I/I^2$, which is the set of [*Kähler differentials*]{} of $R$. Let $$\Omega^p_{\mathrm{com}}R=\mathbf{\Lambda}_R^p(\Omega^1_{\mathrm{com}}R[-1]).$$ Similarly to the DG algebra case, we have the degree $1$ de Rham differential $$d: \Omega^\bullet_{\mathrm{com}}R\to \Omega^{\bullet}_{\mathrm{com}}R,$$ which makes $(\Omega^\bullet_{\mathrm{com}}R, d)$ into a DG cochain algebra. The differential $\partial$ on $R$ also gives a degree $-1$ differential on $\Omega^\bullet_{\mathrm{com}}R$, which also respects the product and commutes with $d$.
The dual space of the cotangent space $
\mathrm{Hom}_R(\Omega^1_{\mathrm{com}} R, R)
$ is called [*the complex of vector fields*]{} of $R$, and is identified with $\mathrm{Der}\; R$.
\[prop:VdBonformsandfields\] Suppose $R$ is a DG algebra. Then
1. $\big(\Omega^1_{\mathrm{nc}}R\big)_V=\Omega^1_{\mathrm{com}}(R_V)$;
2. $\big(\mathbb D\mathrm{er}\; R\big)_V=\mathrm{Der}\; R_V$.
From this proposition, we immediately have:
Suppose $R$ is a quasi-free DG algebra. Then $$\label{identityofformsandvectors}
\big(T_R(\Omega^1_{\mathrm{nc}}R[-1])\big)_V=\Omega^\bullet_{\mathrm{com}}(R_V)\quad
\mbox{and}\quad
\big(T_R(\mathbb D\mathrm{er}\;R[1])\big)_V=\mathbf\Lambda^\bullet (\mathrm{Der}\;R_V[1]).$$
Now applying to , we have the trace maps $$\label{tracemapofquotientspaces}
(T_R(\Omega^1_{\mathrm{nc}}R[-1]))_{\natural}=\mathrm{DR}^\bullet_{\mathrm{nc}}R\to
\Omega^\bullet_{\mathrm{com}}(R_V)^{\mathrm{GL}}$$ and $$(T_R(\mathbb D\mathrm{er}\;R[1]))_{\natural}\to
\mathbf\Lambda^\bullet (\mathrm{Der}\;R_V[1])^{\mathrm{GL}}.$$ When restricting to the first component, we have the trace map $$\label{tracemapofquotientspaces1}
\mathrm{DR}^1_{\mathrm{nc}}R
\to \Omega^1_{\mathrm{com}}(R_V)^{\mathrm{GL}}
\quad\mbox{and}\quad
(\mathbb D\mathrm{er}\; R[1])_{\natural}\to
(\mathrm{Der}\;R_V[1])^{\mathrm{GL}}.$$
The shifted symplectic structure
--------------------------------
The notion of shifted symplectic structure is introduced by Pantev-Toën-Vaquié-Vezzosi in [@PTVV]; see also [@CPTVV; @Melani; @Pridham] for some further studies. In the following we only consider the affine case, which is enough for our purpose.
Suppose $(R,\partial)$ is a DG algebra. For any closed form $\omega\in \Omega^2_{\mathrm{com}}R$ of total degree $2-n$, the [*contraction map*]{} $$\iota_{(-)}\omega: \mathrm{Der}\; R [1]\to (\Omega^1_{\mathrm{com}} R[-1])[2-n],\;
\alpha\mapsto\omega(\alpha, -)[2-n]$$ is a map of $\partial$-complexes.
Suppose $R$ is a DG commutative algebra over $k$. An [*$n$-shifted symplectic structure*]{} on $R$ is a closed form $\omega\in\Omega^2_{\mathrm{com}}R$ of total degree $2-n$ such that the contraction $$\label{mapfromnctangent}
\iota_{(-)}\omega: \mathrm{Der}\;R [1]\to(\Omega_{\mathrm{com}}^1 R[-1])[2-n],$$ is a quasi-isomorphism.
\[rmk:degreeshifting\] In both and we have shifted the degrees on the right hand side of the equations, namely on the (noncommutative) differential 1-forms, up by $n-2$, which looks different from [@PTVV]. However, they are the same in the following sense: and can be alternatively written as $$\mathbb D\mathrm{er}\; R\to\Omega^1_{\mathrm{nc}} R[-n]\quad
\mbox{and}\quad
\mathrm{Der}\; R\to\Omega^1_{\mathrm{com}} R[-n]$$ respectively, which coincides with [@PTVV]. Such a degree shifting guarantees that an $n$-shifted symplectic structure gives an $n$-shifted Poisson structure, whose Poisson bracket has degree $n$.
Suppose $R$ is a DG algebra which admits an $n$-shifted bi-symplectic structure. Then $\mathrm{Rep}_V(R)$ has an $n$-shifted symplectic structure.
Follows from Proposition \[prop:VdBonformsandfields\] and the functoriality of Van den Bergh’s functor. More precisely, it is shown in [@CBEG] that $\mathrm{Tr}(\omega)$ defines a shifted symplectic structure on $\mathrm{Rep}_V(\tilde R)$ and the following diagram $$\label{diag:bisymplectictosymplectic}
\xymatrixcolsep{4pc}
\xymatrix{
\mathbb D\mathrm{er}\; R[1]\ar[r]^{\iota_{(-)}\omega}\ar@{~>}[d]&(\Omega^1_{\mathrm{nc}} R[-1])[n]
\ar@{~>}[d]\\
\mathrm{Der}(\mathrm{Rep}_V( R))[1]\ar[r]^{\iota_{(-)}\mathrm{Tr}(\omega)}&
(\Omega^1_{\mathrm{com}}(\mathrm{Rep}_V( R))[-1])[n]
}$$ is commutative, where the vertical arrows are given by Van den Bergh’s functor.
Combining the above theorem with Theorem \[thm:mainthm1\], we immediately have the following:
Suppose $A$ is a Koszul Calabi-Yau algebra of dimension $n$, and let $\tilde R=\Omega(\tilde A^{{\textup{!`}}})$ as before. Then $\mathrm{Rep}_V(\tilde R)$ has an $(2-n)$-shifted symplectic structure.
Identification of $\mathrm{GL}$-invariant 1-forms and vectors
-------------------------------------------------------------
In this subsection, we show that $\mathrm{Rep}_V(\tilde R)^{\mathrm{GL}}$ is “symplectic". What we mean is the following:
\[conj:Lam\] Let $A$ be a Koszul Calabi-Yau algebra of dimension $n$. Then $$\mathrm{Der}(\mathrm{Rep}_V(\tilde R))^{\mathrm{GL}}[1]
\cong
(\Omega^1_{\mathrm{com}}(\mathrm{Rep}_V(\tilde R))^{\mathrm{GL}}[-1])[n].$$
Observe that in , $\mathrm{Tr}(\omega)$ in fact lies in $\Omega^2_{\mathrm{com}}(\mathrm{Rep}_V(\tilde R))^{\mathrm{GL}}$ (see ). Thus by taking the $\mathrm{GL}(V)$-invariant vector fields and 1-forms, we get the desired isomorphism.
Combining this proposition with Propositions \[Prop:cotangentcomplex\] and \[Prop:tangentcomplex\], we obtain that the trace map gives the following commutative diagram of chain complexes $$\label{diag:tracefromHochschildcomplex}
\xymatrixcolsep{4pc}
\xymatrix{
\mathrm{CH}^\bullet(A)\ar[r]^-{\mathrm{Tr}}\ar[d]^{\cong}
&\mathrm{Der}(\mathrm{Rep}_V(\tilde R))^{\mathrm{GL}}[1]\ar[d]^{\cong}\\
\mathrm{CH}_{n-\bullet}(A)\ar[r]^-{\mathrm{Tr}}
&(\Omega^1_{\mathrm{com}}(\mathrm{Rep}_V(\tilde R))^{\mathrm{GL}}[-1])[n].
}$$ Taking the (co)homology on both sides, we get the the commutative diagram stated in §\[sect:intro\]. As we remarked before, the vertical maps are both isomorphisms, while the two horizontal maps are neither surjective nor injective in general (for see [@BR] more details).
The shifted double Poisson structure {#sect:shiftedPoisson}
====================================
Shifted bi-symplectic structures are intimately related to shifted double Poisson structures. Let us remind the work [@VdB] of Van den Bergh (here we rephrase it in the DG case; see also [@BCER]).
Suppose $R$ is a DG algebra over $k$. A *double bracket* of degree $n$ on $R$ is a DG map ${\mathopen{\{\!\!\{}}
\newcommand{\rdb}{\mathclose{\}\!\!\}}}-,-\rdb: R\times R\to R\otimes R$ of degree $n$ which is a derivation in its second argument and satisfies $$\begin{aligned}
{\mathopen{\{\!\!\{}}
\newcommand{\rdb}{\mathclose{\}\!\!\}}}a,b\rdb&= &-(-1)^{(|a|+n)(|b|+n)}{\mathopen{\{\!\!\{}}
\newcommand{\rdb}{\mathclose{\}\!\!\}}}b,a\rdb^\circ,
\end{aligned}$$ where $(u\otimes v)^\circ =(-1)^{|u||v|}v\otimes u$.
\[Def:DPs\] Suppose that ${\mathopen{\{\!\!\{}}
\newcommand{\rdb}{\mathclose{\}\!\!\}}}-,-\rdb$ is a double bracket of degree $n$ on $R$. For $a,b_1,...,b_n\in R$, let $${\mathopen{\{\!\!\{}}
\newcommand{\rdb}{\mathclose{\}\!\!\}}}a, b_1 \otimes \cdots \otimes b_n\rdb_L \,:=\, {\mathopen{\{\!\!\{}}
\newcommand{\rdb}{\mathclose{\}\!\!\}}}a,b_1 \rdb \otimes b_2 \otimes \cdots \otimes b_n,$$ and for $s$ is a permutation of $\{1,2,\cdots, n\}$, let $$\sigma_s(b_1 \otimes \cdots \otimes b_n):=(-1)^{\sigma(s)}b_{s^{-1}(1)} \otimes \cdots \otimes b_{s^{-1}(n)},$$ where $$\sigma(s)=\displaystyle\sum_{i<j; s^{-1}(j)<s^{-1}(i)}|a_{s^{-1}(i)}||a_{s^{-1}(j)}|.$$ If furthermore $R$ satisfies the following *double Jacobi identity* $$\begin{gathered}
\label{dJ}
{\mathopen{\bigl\{\!\!\bigl\{}}
\newcommand{\rdbg}{\mathclose{\bigr\}\!\!\bigr\}}}a , {\mathopen{\{\!\!\{}}
\newcommand{\rdb}{\mathclose{\}\!\!\}}}b,c \rdb \rdbg_L +(-1)^{(|a|+n)(|b|+|c|)} \sigma_{(123)}{\mathopen{\bigl\{\!\!\bigl\{}}
\newcommand{\rdbg}{\mathclose{\bigr\}\!\!\bigr\}}}b,{\mathopen{\{\!\!\{}}
\newcommand{\rdb}{\mathclose{\}\!\!\}}}c,a\rdb \rdbg_L\\
+ (-1)^{(|c|+n)(|a|+|b|)}\sigma_{(132)}
{\mathopen{\bigl\{\!\!\bigl\{}}
\newcommand{\rdbg}{\mathclose{\bigr\}\!\!\bigr\}}}c,{\mathopen{\{\!\!\{}}
\newcommand{\rdb}{\mathclose{\}\!\!\}}}a,b\rdb \rdbg_L =0,\end{gathered}$$ then $R$ is called a [*double Poisson algebra*]{} of degree $n$ (or *$n$-shifted double Poisson algebra*).
Van den Bergh [@VdB] showed that, if $R$ is equipped with a double Poisson structure, then there is a Poisson structure on the affine scheme $\mathrm{Rep}_V(R)$ of all the representations of $R$ in $V$. Independently and simultaneously, Crawley-Boevey gave in [@CB] the [*weakest*]{} condition for $\mathrm{Rep}_V(R)/\!/{\mathrm{GL}(V)}$ of $R$ to have a Poisson structure. He called such condition the [*$\mathrm H_0$-Poisson structure*]{}, since it involves the zeroth Hochschild/cyclic homology of $R$. It turns out Van den Bergh’s double Poisson structure exactly satisfies this condition, and is so far the most interesting example therein.
From shifted bi-symplectic to shifted double Poisson
----------------------------------------------------
In [@VdB Appendix] Van den Bergh showed that a bi-symplectic structure gives a double Poisson structure. We rephrase it in the Koszul Calabi-Yau case. Suppose $A$ is a Koszul $n$-Calabi-Yau algebra, and let $\tilde R=\Omega(\tilde A^{{\textup{!`}}})$. For any $r\in\tilde R$, since $$\iota_{(-)}\omega: \mathbb D\mathrm{er}\;\tilde R[1]\to(\Omega^1_{\mathrm{nc}}\tilde R[-1])[n]$$ is an [*isomorphism*]{} of chain complexes, there exists an element $H_r\in\mathbb D\mathrm{er}\;\tilde R$ such that $$\iota_{H_r}\omega=d(r).$$ $H_r$ is called the [*bi-Hamiltonian vector field*]{} associated to $r$. Now consider the following bracket $${\mathopen{\{\!\!\{}}
\newcommand{\rdb}{\mathclose{\}\!\!\}}}-,-\rdb : \tilde R\times \tilde R \to \tilde R\otimes\tilde R,\;
(r_1,r_2) \mapsto H_{r_1}(r_2),$$ then we have:
\[prop:shiftedbPontildeR\] ${\mathopen{\{\!\!\{}}
\newcommand{\rdb}{\mathclose{\}\!\!\}}}-,-\rdb $ gives a $(2-n)$-shifted double Poisson structure on $\tilde R$.
By Van den Bergh [@VdB Lemma A.3.3], the $(2-n)$-shifted double Poisson bracket on $\tilde R=\Omega(\tilde A^{{\textup{!`}}})$ is given by the following formula: $$\label{def:doublePoissonbracket}
{\mathopen{\{\!\!\{}}
\newcommand{\rdb}{\mathclose{\}\!\!\}}}x, y\rdb:=\sum_{i=1}^k\sum_{j=1}^\ell
(-1)^{\sigma_{ij}}
(x_i, y_j)\cdot
( y_1\cdots y_{j-1} x_{i+1}\cdots x_k)
\otimes
( x_1\cdots x_{i-1} y_{j+1}\cdots y_{\ell}),$$ for $x=( x_1\cdots x_k)$ and $y=( y_1\cdots y_{\ell})$ in $\tilde R$, where $(-1)^{\sigma_{ij}}$ is the Koszul sign. Here $(x_i, y_j)\in k$ for $x_i, y_j\in A^{{\textup{!`}}}[1]$ is the graded skew-symmetric pairing induced from the pairing on $A^{{\textup{!`}}}$; more precisely, $(x_i, y_j)=(-1)^{n-|x_i|}\langle\Sigma x_i, \Sigma y_j\rangle$.
Formula also appeared in [@BCER Theorem 15] (see also [@CEEY Lemma 4.4]) for $R=\Omega(A^{{\textup{!`}}})$, where we also showed that ${\mathopen{\{\!\!\{}}
\newcommand{\rdb}{\mathclose{\}\!\!\}}}-,-\rdb$ given above commutes with the differential on $R$. The only difference between $\tilde R$ and $R$ is that the generators of $\tilde R$ contain one more element, namely the desuspension of the co-unit of $A^{{\textup{!`}}}$. The sufficient condition for ${\mathopen{\{\!\!\{}}
\newcommand{\rdb}{\mathclose{\}\!\!\}}}-,-\rdb$ commuting with the differentials on $R$ and on $\tilde R$ is the cyclic condition for $A^{{\textup{!`}}}$. Thus the proof of [@BCER Theorem 15] symbolically applies to the above proposition, too.
Suppose $R$ has a double Poisson structure, then Van den Bergh gave an explicit formula for the Poisson structure on $\mathrm{Rep}_V(R)$ (see [@VdB Propositions 7.5.1 and 7.5.2]). For a Koszul Calabi-Yau algebra $A$, the Poisson structure on $\mathrm{Rep}_n(\tilde R)$ is given as follows: suppose $A^{{\textup{!`}}}$ has a set of basis $\{x^\alpha\}_{\alpha\in I}$, and $\Delta(x^{\alpha})=\sum_{(x^{\alpha})} x^{\alpha_1}\otimes x^{\alpha_2}$. Then $\mathrm{Rep}_n(\tilde R)$ is the quasi-free DG commutative algebra generated by $$\label{RepofcobarofKoszulCY}
\big\{x_{ij}^{\alpha}\;\big|\;\alpha\in I, 1\le i,j\le n, |x_{ij}^{\alpha}|=|x^{\alpha}|-1\big\},$$ with $$d(x_{ij}^{\alpha})=\sum_{(x^{\alpha})}(-1)^{|x^{\alpha_1}|}\sum_{k=1}^{n}x_{ik}^{\alpha_1}\cdot x_{kj}^{\alpha_2}.$$ The $(2-n)$-Poisson bracket is given by $$\label{PoissononDRep}
\{x_{ij}^{\alpha}, x_{k\ell}^\beta\}=(-1)^{n-|x^{\alpha}|}\delta_{i\ell}\delta_{jk}\langle x^\alpha, x^\beta\rangle$$ on the generators, which extends to the whole $\mathrm{Rep}_n(\tilde R)$ by the Leibniz rule.
The work of Crawley-Boevey
--------------------------
In [@CB] Crawley-Boevey introduced what he called the [*$\mathrm{H}_0$-Poisson structure*]{}. Let us recall its definition:
Suppose $R$ is an associative algebra. An [*$\mathrm H_0$-Poisson structure*]{} on $R$ is a Lie bracket $$\{-,-\}: R_{\natural}\times R_{\natural}\to R_{\natural}$$ such that the adjoint action $$ad_{\bar u}: R_\natural\to R_\natural, \; \bar v\mapsto \{\bar u,\bar v\}$$ can be lifted to be a derivation $$d_{u}: R\to R,
\quad\mbox{for all}\; u\in R.$$
Crawley-Boevey proved that if $R$ admits an $\mathrm H_0$-Poisson structure, then there is a unique Poisson structure on $\mathrm{Rep}_V(R)^{\mathrm{GL}}$ such that the trace map $$\mathrm{Tr}: R_{\natural}\to \mathrm{Rep}_V(R)^{\mathrm{GL}}$$ is a map of Lie algebras, or in other words, $
\mathrm{Tr}: \mathbf\Lambda^\bullet R_{\natural}\to \mathrm{Rep}_V(R)^{\mathrm{GL}}
$ is a map of Poisson algebras. Here the Poisson bracket on $\mathbf\Lambda^\bullet R_{\natural}$ is the extension of the bracket on $R_{\natural}$ by derivation.
Now suppose $R$ has a double Poisson bracket ${\mathopen{\{\!\!\{}}
\newcommand{\rdb}{\mathclose{\}\!\!\}}}-,-\rdb$. Then $$\{-,-\}: R\times R\to R,\; (u,v)\mapsto\mu \circ {\mathopen{\{\!\!\{}}
\newcommand{\rdb}{\mathclose{\}\!\!\}}}u,v\rdb$$ descends to a well-defined Lie bracket (see [@VdB Lemma 2.4.1]) $$\label{fromdbtoLie}
\{-,-\}: R_\natural\times R_\natural\to R_{\natural},$$ which exactly says that $R$ admits an $\mathrm H_0$-Poisson structure. The restriction of the Poisson structure on $\mathrm{Rep}_V(R)$ gives the one on $\mathrm{Rep}_V(R)^{\mathrm{GL}}$.
This result was later generalized to the DG and derived category in [@BCER]. Going to the Koszul Calabi-Yau algebra case, we have the following.
Suppose $A$ is a Koszul Calabi-Yau algebra, and let $\tilde R=\Omega(\tilde A^{{\textup{!`}}})$ be as above. The trace map $$\label{diag:tracefromcycliccomplex}
\mathrm{Tr}: \tilde R_{\natural}\to\mathrm{Rep}_V(\tilde R)^{\mathrm{GL}}$$ is a map of degree $(2-n)$ DG Lie algebras. In other words, $$\label{tracemapofPoissonalgebras}
\mathrm{Tr}: \mathbf\Lambda^\bullet\tilde R_{\natural}\to
\mathrm{Rep}_V(\tilde R)^{\mathrm{GL}}$$ is a map of $(2-n)$-shifted Poisson algebras.
See [@BCER Corollary 3]. Again we emphasize that in [@BCER] we proved the statement for $R=\Omega(A^{{\textup{!`}}})$, but the same proof symbolically applies to the current case.
Alternatively, one can prove the above corollary by considering the $\mathrm{GL}(V)$-invariant elements in $\mathrm{Rep}_V(\tilde R)$ (see [@BR Theorem 3.1]) and then applying to them directly. We leave it to the interested readers.
Relations with quivers and quiver representations {#subsect:relationswithquivers}
-------------------------------------------------
The shifted bi-symplectic and double Poisson structures on $\tilde R$ generalize the ones of quivers given in [@CBEG; @VdB].
Let $Q$ be a quiver. For simplicity let us assume $Q$ has only one vertex. Let $\bar Q$ be the double of $Q$. Then the path algebra $k\bar Q$ is an associative algebra over $k$; viewing $\bar Q$ as a 1-dimensional CW complex, then $k\bar Q$ is exactly the cobar construction of the coalgebra of the chain complex of $\bar Q$.
Denote the set of the edges of $Q$ by $\{e_i\}$ and their duals by $\{e_i^*\}$, then there is a graded symmetric pairing on $\{e_i\}\cup\{e_i^*\}$ given by $$\langle e_i, e_j\rangle=\langle e_i^*,e_j^*\rangle=0,\quad
\langle e_i, e_j^*\rangle=-\langle e_j^*, e_i\rangle=\delta_{ij}$$ Such pairing is non-degenerate and cyclically invariant, and therefore by results in previous sections, the cobar construction of $\bar Q$, that is, $k\bar Q$, has a $0$-shifted bi-symplectic and double Poisson structure. The bi-symplectic and double Poisson structures are exactly the ones obtained in [@CBEG; @VdB]. The corresponding symplectic structure on $\mathrm{Rep}_V(k\bar Q)$ as well as the Lie bracket on $k\bar Q_{\natural}$ was also previously studied by Ginzburg in [@Ginzburg01]; see also Bocklandt-Le Bruyn [@BLB], where the Lie algebra on $k\bar Q_{\natural}$ is called the [*necklace Lie algebra*]{}.
Now assign the gradings of the edges $\bar Q$ other than one and obtain a DG coalgebra, say $C$. To obtain the shifted bi-symplectic and double Poisson structures on $\Omega(C)$ (respectively $\Omega(\tilde C)$, where $\tilde C$ is the co-unitalization of $C$), then a sufficient condition is that $\bar C=C\backslash k$ (respectively $C$) is cyclic, in other words, the dual space $\mathrm{Hom}(\bar C, k)$ (respectively $\mathrm{Hom}(C, k)$) is a cyclic associative and not necessarily unital algebra.
Quantization {#sect:quantization}
============
In this section, we study the quantization problem. In [@Schedler], Schedler proved that the necklace Lie algebra of a doubled quiver is in fact an involutive Lie bialgebra, and constructed a Hopf algebra which quantizes this Lie bialgebra. He also showed that the Hopf algebra is mapped to the Moyal-Weyl quantization of the quiver representation spaces as associative algebras.
Later in [@GS], he together with Ginzburg constructed a Moyal-Weyl type quantization of the necklace Lie bialgebra, and showed that such quantization is isomorphic to the Hopf algebra constructed in [@Schedler].
The purpose of this section is to generalize their results to the Koszul Calabi-Yau case. The main result is the commutative diagram . Some partial results have been previously obtained in [@CEG].
Quantization of $\mathrm{Rep}_n(\tilde R)$
------------------------------------------
In the bivector form, the shifted Poisson structure on $\mathrm{Rep}_n(\tilde R)$ is given by $$\label{formula:Poissonstructure}
\pi=\sum_{\alpha,\beta\in I}\sum_{i,j}
(-1)^{n-|x^{\alpha}|}
\langle x^{\alpha},x^{\beta}\rangle\frac{\partial}{\partial x^{\alpha}_{ij}}\wedge
\frac{\partial}{\partial x^{\beta}_{ji}}.$$ Observe that $\pi$ is of constant coefficients, and thus we have the [*Moyal-Weyl quantization*]{} of $\mathrm{Rep}_n(\tilde R)$ which is given by $$\label{def:starprodonDRep}
f\star g:=\mu \circ e^{\frac{\hbar}{2}\pi}(f\otimes g)\in \mathrm{Rep}_n(\tilde R)[\hbar],\quad
\mbox{for any}\; f, g\in \mathrm{Rep}_n(\tilde R),$$ where “Mult" is the original multiplication on $\mathrm{Rep}_n(\tilde R)$, and $\hbar$ is a formal parameter of degree $n-2$.
$(\mathrm{Rep}_n(\tilde R)[\hbar],\star)$ is a DG associative algebra over $k[\hbar]$, which quantizes $\mathrm{Rep}_n(\tilde R)$.
We only need to show that the differential $d$ commutes with $\star$, or equivalently, $\partial$ commutes with $\mu \circ \pi^r(-,-)$, for all $r\in\mathbb N$.
In fact, since $\pi^r(f, g)$ is of $r$-th order for both arguments, we only need to check the case when $f$ and $g$ are degree $r$ monomials. In this case, $\mu \circ \pi^n(f, g)$ is a number, whose differential is zero, and hence we need to check $$\label{boundarycommuteswithquantization}
\mu \circ\pi^r(\partial f, g)+\mu \circ\pi^r(f, \partial g)=0.$$ Suppose $f=x_{i_1j_1}^{\alpha_1}\cdots x_{i_rj_r}^{\alpha_r},
g=y_{k_1\ell_1}^{\beta_1}\cdots y_{k_r\ell_r}^{\beta_r}$. Then up to sign, $\mu \circ\pi^r(\partial f, g)$ and $\mu \circ\pi^r(f, \partial g)$ both contain a common scalar factor which is obtained by applying $\pi^{r-1}$ to $x_{i_1j_1}^{\alpha_1}\cdots\widehat
{x_{i_pj_p}^{\alpha_p}}\cdots x_{i_rj_r}^{\alpha_r}$ and $y_{k_1\ell_1}^{\beta_1}\cdots \widehat{y_{k_q\ell_q}^{\beta_q}}\cdots y_{k_r\ell_r}^{\beta_r}$ where $\widehat{\;\;\;\;}$ means the corresponding component is omitted. The rest factors are just $\pi(d({x_{i_pj_p}^{\alpha_p}}), y_{k_q\ell_q}^{\beta_q})$ and $\pi({x_{i_pj_p}^{\alpha_p}}, d(y_{k_q\ell_q}^{\beta_q}))$ respectively. This means, to prove it is sufficient to show $
\pi(f, g)
$ commutes with the boundary, which is already done. This proves the statement.
Quantization of $\tilde R_\natural$
-----------------------------------
In this subsection we study the quantization of the Lie bialgebra on $\tilde R_{\natural}$. Let us start with several definitions.
Suppose that $(L,\{-,-\})$ is a graded Lie algebra with the bracket $\{-,-\}$ having degree $m$ and $(L,\delta)$ is a graded Lie coalgebra with the cobracket $\delta$ having degree $n$. The triple $(L,\{-,-\},\delta)$ is called a [*Lie bialgebra*]{} of degree $(m,n)$ if the following Drinfeld compatibility (also called the cocycle condition) holds: for all $a, b\in L$, $$\label{Drinfeld com}
\delta(\{a,b\})=(ad_a\otimes id+id\otimes ad_a)\delta (b)+(ad_b\otimes id+id\otimes ad_b)\delta (a),$$ where $ad_a(b)=\{a,b\}$ is the adjoint action. If furthermore, $\{-,-\}\circ\delta(g)\equiv 0$ for any $g\in L$, then the Lie bialgebra is called [*involutive*]{}.
In the above definition, if $L$ is equipped with a differential which commutes with both the Lie bracket and Lie cobracket, then it is called a [*DG Lie bialgebra*]{}.
Now let $A$ be a Koszul Calabi-Yau algebra. Recall that the DG Lie bracket on $\tilde R_{\natural}$ is given by the following formula ([*c.f.*]{} and ): $$\{ x, y\}:=\sum_{i=1}^k\sum_{j=1}^\ell
(-1)^{\sigma_{ij}}
( x_i, y_j)\cdot\mathrm{pr}
( y_1\cdots y_{j-1} x_{i+1}\cdots x_kx_1\cdots x_{i-1} y_{j+1}\cdots y_{\ell}),$$ for $x, y\in\tilde R_{\natural}$ represented by $( x_1\cdots x_k)$ and $y=( y_1\cdots y_{\ell})$ in $\tilde R$, where $\mathrm{pr}(-)$ means the projection of $\tilde R$ to $\tilde R_{\natural}$. Now define $$\delta:\tilde R_{\natural}\longrightarrow \tilde R_{\natural}\otimes
\tilde R_{\natural}$$ by $$\begin{gathered}
\delta( x_1 x_2 \cdots x_n)\\
:=\sum_{i,j:\, i<j} (-1)^{\sigma_{ij}}(x_i, x_j)\cdot\mathrm{pr}
( x_1 \cdots x_{i-1} x_{j+1} \cdots x_n)\otimes\mathrm{pr} ( x_{i+1} \cdots x_{j-1})\\
-\sum_{i,j:\, i<j} (-1)^{\sigma_{ij}'}(x_i, x_j)\cdot\mathrm{pr} ( x_{i+1} \cdots x_{j-1})
\otimes \mathrm{pr}( x_1 \cdots x_{i-1} x_{j+1} \cdots x_n).\end{gathered}$$
$(\tilde R_{\natural}, \{-,-\},\delta)$ forms an involutive DG Lie bialgebra of degree $(2-d, 2-d)$.
This Theorem, together with the following Theorem \[thm:existenceofquantization\], is proved in [@CEG]. They are directly inspired by [@Schedler]; what is new there is that the Lie bracket and cobracket thus defined are compatible with the differential. As we remarked in the proof of Proposition \[prop:shiftedbPontildeR\], the cyclic condition guarantees that all constructions respect it. We thus omit the proof and refer the interested reader to [@CEG] for more details.
Suppose $(L, \{-,-\},\delta)$ is a DG Lie bialgebra of degree $(m, n)$ over the field $k$. A [*quantization*]{} of $L$ is a Hopf algebra $(A,\star,\Delta)$, flat over $k[\hbar,h]$, where $\hbar$ and $h$ are formal parameters of degree $-m$ and $-n$ respectively, together with a surjective map $$\Psi: A\to \mathbf\Lambda^\bullet L$$ such that for all $x, y\in A$, $$\label{def:quantization}
\Psi\Big(\frac{x\star y-y\star x}{\hbar}\Big)=\{\Psi(x),\Psi(y)\},\quad
\Psi\Big(\frac{\Delta(x)-\Delta^{\mathrm{op}}(x)}{h}\Big)=\delta(\Psi(x)),$$ where $\Delta^{\mathrm{op}}(-)$ means the opposite coproduct, and $\delta: \mathbf\Lambda^\bullet L\to \mathbf\Lambda^\bullet L\otimes \mathbf\Lambda^\bullet L$ is the cobracket induced by $\delta: L\to L\otimes L$.
Let us remind that the cobracket on $ \mathbf\Lambda^\bullet L$ is defined as follows: for any Lie coalgebra $(L,\delta)$, its graded symmetric product $ \mathbf\Lambda^\bullet L$ admits a co-Poisson algebra structure which is induced by the Lie coalgebra structure $\delta$ via the formula: $$\delta\circ \mu=(\mu\otimes \mu)\circ(1\otimes\tau\otimes 1)\circ (\delta\otimes\Delta+\Delta\otimes\delta),$$ where $\mu$ is the product of $ \mathbf\Lambda^\bullet L$, $\tau:a\otimes b\mapsto(-1)^{|a||b|} b\otimes a$ is the switching operator, and $\Delta$ is the coproduct of $ \mathbf\Lambda^\bullet L$: $$\Delta(x^I)=\sum_{I_1\cup I_2=I}x^{I_1}\otimes x^{I_2}.$$
In the rest of this subsection, we study the quantization of $\tilde R_{\natural}$. Recall that $\tilde R=T(A^{{\textup{!`}}}[1])$ and $A^{{\textup{!`}}}$ is a cyclic coalgebra.
Now let $LA^{{\textup{!`}}}=A^{{\textup{!`}}}\otimes k[\nu,\nu^{-1}]$, where $\nu$ is a formal parameter of degree $0$. An element $a\otimes \nu^r$ in $LA^{{\textup{!`}}}$ is denoted by $(a, r)$. Let $T(LA^{{\textup{!`}}}[1])$ be the free tensor algebra of the desuspension of $LA^{{\textup{!`}}}$, where its elements are written in the form $
[(a_{1}, r_{1})|\cdots|(a_{p}, r_{p})]
$, for $(a_1, r_1),\cdots, (a_p, r_p)\in LA^{{\textup{!`}}}$. Let $LH$ be the commutator quotient space of $T(LA^{{\textup{!`}}}[1])$, namely $LH=T(LA^{{\textup{!`}}}[1])_{\natural}$. To avoid complicated notations, in the following we write elements in $LH$ again in the form $
[(a_{1}, r_{1})|\cdots|(a_{p}, r_{p})]
$; in other words we omit the symbol $\mathrm{pr}(-)$.
Let $SLH$ be the graded symmetric algebra generated by $LH$, whose product is denoted by $\bullet$. Define a differential $\partial$ on $SLH$ given by the following formula: $$\begin{gathered}
\label{dofSLH}
\partial\big([(a_{1,1}, r_{1,1})|\cdots|(a_{1, p_1}, r_{1, p_1})]
\bullet\cdots\bullet[(a_{s,1}, r_{s,1})|\cdots|(a_{s, p_s}, r_{s, p_s})]\big)\\
:=\sum_{i=1}^s\sum_{j=1}^{p_s}\sum_{(a_{i,j})}(-1)^{\sigma_{ij}}
[(a_{1,1},\tilde r_{1,1})|\cdots|(a_{1, p_1}, \tilde r_{1, p_1})]\bullet\cdots\quad\quad\quad\quad\\
\bullet[(a_{i,1}, \tilde r_{i,1})|\cdots|
(a_{i,j}',r_{i,j})|(a_{i,j}'', 1+r_{i,j})|\cdots
(a_{i, p_i}, \tilde r_{i, p_i})]\bullet\cdots,\end{gathered}$$ where $(-1)^{\sigma_{ij}}$ is the Koszul sign, $a_{i,j}'$ and $a_{i,j}''$ come from $\Delta(a_{i,j})=\sum_{(a_{i,j})}a_{i,j}'\otimes a_{i,j}''$, and for $(i',j')\ne (i,j)$, $$\tilde r_{i',j'}=\left\{
\begin{array}{ll}
r_{i',j'},&\mbox{if}\; r_{i',j'}\le r_{i,j}\\
1+r_{i',j'},&\mbox{if}\; r_{i',j'}>r_{i,j}.
\end{array}
\right.$$ The coassociativity of $A^{{\textup{!`}}}$ implies that $b^2=0$.
Now let $\hbar, h$ be formal parameters both of degree $n-2$, and let $SLH[\hbar,h]$ be the $k[\hbar,h]$-module genenrated by $SLH$, on which $\partial$ extends $k[\hbar,h]$-linearly. Let $\widetilde{SLH}$ be the subcomplex of $SLH[\hbar,h]$ spanned by $$\label{elementsinH}
[(a_{1,1}, r_{1,1})|\cdots|(a_{1, p_1}, r_{1, p_1})]
\bullet\cdots\bullet[(a_{s,1}, r_{s,1})|\cdots|(a_{s, p_s}, r_{s, p_s})]$$ where $r_{i,j}$ are all distinct.
Consider the quotient space of $\widetilde{SLH}$ by identifying $$[(a_{1,1}, r_{1,1})|\cdots|(a_{1, p_1}, r_{1, p_1})]
\bullet\cdots\bullet[(a_{s,1}, r_{s,1})|\cdots|(a_{s, p_s}, r_{s, p_s})]$$ with $$[(a_{1,1}, r_{1,1}')|\cdots|(a_{1, p_1}, r_{1, p_1}')]
\bullet\cdots\bullet[(a_{s,1}, r_{s,1}')|\cdots|(a_{s, p_s}, r_{s, p_s}')]$$ under the condition that $r_{i,j}<r_{i',j'}$ if and only if $r_{i,j}'<r_{i',j'}'$. Denote this quotient space by $\tilde A$. Pick an element in $\tilde A$, suppose it is represented by $[(a_{1,1}, r_{1,1})|\cdots|(a_{1, p_1}, r_{1, p_1})]
\bullet\cdots\bullet[(a_{s,1}, r_{s,1})|\cdots|(a_{s, p_s}, r_{s, p_s})]$, without loss of generality, we may assume all $r_{i,j}$ are even, then the image of represents an element in $\tilde A$. This means that $\tilde A$ with the differential induced by $b$ is a chain complex.
Now let $\tilde B$ be the submodule of $\tilde A$ generated by elements of the following form:
1. $X- X_{i,j,i',j'}'-\hbar \cdot X_{i,j,i',j'}''$, where $i\ne i'$, $r_{i,j}<r_{i',j'}$, and there does not exist $(i'',j'')$ with $r_{i,j}<r_{i'',j''}<r_{i,j'}$, and
2. $X- X_{i,j,i,j'}'- h\cdot X_{i,j,i,j'}''$, where $r_{i,j}<r_{i,j'}$, and there does not exist $(i'',j'')$ with $r_{i,j}<r_{i'',j''}<h_{i,j'}$,
where $X_{i,j,i',j'}'$ and $X_{i,j,i',j'}''$ are given as follows: if $i\ne i'$, then $X_{i,j,i',j'}'$ is the same as $X$ except that $r_{i,j}$ and $r_{i',j'}$ are interchanged, while $X_{i,j,i',j'}''$ replaces the factors $[(a_{i,1}, r_{i,1})|\cdots|(a_{i,p_i}, r_{i, p_i})]$ and $[(a_{i',1}, r_{i',1})|\cdots|(a_{i',p_{i'}}, r_{i', p_{i'}})]$ by $$(-1)^{\sigma_{iji'j'}}(a_{i,j}, a_{i',j'})
[(a_{i,j+1}, r_{i,j+1})|\cdots|(a_{i,j-1}, r_{i,j-1})|(a_{i',j'+1}, r_{i',j'+1})|\cdots|(a_{i',j'-1}, r_{i',j'-1})];$$ similarly, $X_{i,j,i,j'}'$ is the same as $X$ but with $r_{i,j}$ and $r_{i,j'}$ interchanged, while $X_{i,j,i,j'}''$ replaces the factor with the following factor: $$(-1)^{\sigma_{ijij'}}( a_{i,j}, a_{i,j'})
[(a_{i,j'+1}, r_{i,j'+1})|\cdots|(a_{i,j-1}, r_{i,j-1})]\bullet[(a_{i,j+1}, h_{i,j+1})|\cdots|(a_{i,j'-1}, r_{i,j'-1})].$$ It is proved in [@CEG Lemma 14] that $\tilde B$ is a subcomplex of $\tilde A$. Let $H=\tilde A/\tilde B$.
\[thm:existenceofquantization\] There is a DG Hopf algebra structure on $H$ over $k[\hbar,h]$, which quantizes $\tilde R_{\natural}$.
The product on $H$, denoted by $\star$, is easy to describe (as we will only use it): for two elements in $H$, say $X$ and $Y$, suppose they are both represented by elements in the form , raise those $r_{i,j}$ in $Y$ such that they are all greater than those in $X$, then the product of $X$ and $Y$, $X\star Y$, is represented by $X\bullet Y$.
Lifting the trace map
---------------------
In this subsection we relate the quantization of $\tilde R_{\natural}$ with the one of $\mathrm{Rep}_V(\tilde R)$. We set $\hbar=h$.
Extending the trace map by $k[\hbar]$-linearity, we obtain a $k[\hbar]$-linear map $$\widetilde{\mathrm{Tr}}: \mathbf\Lambda^\bullet\tilde R_{\natural}[\hbar] \to
\mathrm{Rep}_V(\tilde R)[\hbar].$$ Clearly $\widetilde{\mathrm{Tr}}$ commutes with the differential. We have the following.
\[thm:traceofquantization\] For a Koszul Calabi-Yau algebra $A$, the following map $$\label{tracemapofquantizedPoisson}
\widetilde{\mathrm{Tr}}:(\mathbf\Lambda^\bullet\tilde R_{\natural}[\hbar], \star )\to
(\mathrm{Rep}_V(\tilde R)[\hbar], \star).$$ is a map of DG algebras over $k[\hbar]$.
Since $\widetilde{\mathrm{Tr}}$ commutes with the differential on both sides, we only need to show it is a graded algebra map; however, this is already done in Schedler [@Schedler §3.4].
More precisely, for a quiver $Q$, if we denote its double by $\bar Q$ then what Schedler constructed in [@Schedler] is the following:
1. a Lie bialgebra structure, called the necklace Lie bialgebra, on the commutator quotient space $(k\bar Q)_\natural$;
2. a Hopf algebra $H$ over $k[\hbar]$, completely analogous to the construction in previous subsection, quantizing the necklace Lie bialgebra $(k\bar Q)_{\natural}$;
3. an algebra map from this Hopf algebra to the Weyl algebra of differential operators on $\mathrm{Rep}_V(kQ)$.
Note that $\mathrm{Rep}_V(k\bar Q)$ is the cotangent space of $\mathrm{Rep}_V(k Q)$, the algebra of differential operators is exactly the Moyal-Weyl quantization of $\mathrm{Rep}_V(k\bar Q)$.
As we remarked in §\[subsect:relationswithquivers\], if we assume the the number of vertices of $Q$ is one, and the edges of $Q$ are graded, then $\tilde R$ in the current paper is exactly the path algebra $k\bar Q$. Therefore Schedler’s construction and proof hold in our case.
Later Ginzburg and Schedler showed in [@GS] that the quantization constructed above is also of Moyal-Weyl type.
In summary, we obtain the following commutative diagram $$\label{Diag:quantization}
\xymatrixcolsep{4pc}
\xymatrix{
(\mathbf\Lambda^\bullet\tilde R_{\natural}[\hbar], \star )
\ar[r]^{\widetilde{\mathrm{Tr}}}\ar@{~>}[d]_{\textup{quantization}}
&(\mathrm{Rep}_V(\tilde R)[\hbar], \star)\ar@{~>}[d]^{\textup{quantization}}\\
\mathbf\Lambda^\bullet\tilde R_{\natural}
\ar[r]^{\mathrm{Tr}}
&
\mathrm{Rep}_V(\tilde R).
}$$ Recall that the images of $\mathrm{Tr}$ in fact lie in $\mathrm{Rep}_V(\tilde R)^{\mathrm{GL}}$, and the restriction of the Poisson structure on $\mathrm{Rep}_V(\tilde R)$ gives the Poisson structure on $\mathrm{Rep}_V(\tilde R)^{\mathrm{GL}}$. From we thus obtain the following commutative diagram $$\label{Diag:quantizationviatrace}
\xymatrixcolsep{4pc}
\xymatrix{
(\mathbf\Lambda^\bullet\tilde R_{\natural}[\hbar], \star )
\ar[r]^{\widetilde{\mathrm{Tr}}}\ar@{~>}[d]_{\textup{quantization}}
&(\mathrm{Rep}_V(\tilde R)^{\mathrm{GL}}[\hbar], \star)\ar@{~>}[d]^{\textup{quantization}}\\
\mathbf\Lambda^\bullet\tilde R_{\natural}
\ar[r]^{\mathrm{Tr}}
&
\mathrm{Rep}_V(\tilde R)^{\mathrm{GL}}.
}$$ Recall that $\bar{\tilde R}_{\natural}\simeq\mathrm{CC}_\bullet(A)$ (see Proposition \[Prop:cotangentcomplex\] and Convention \[conv:cyclichomology\]) and observe that the bracket of any element in $\tilde R_\natural$ with unit of $\tilde R$ vanishes, we have an embedding $\mathrm{HC}_\bullet(A)\hookrightarrow\mathrm{H}_\bullet(\tilde R_{\natural})$ as graded vector spaces, where $\mathrm{HC}_\bullet(A)$ is the cyclic homology of $A$. By pulling the Lie bialgebra structure on $\mathrm{H}_\bullet(\tilde R_{\natural})$, $\mathrm{HC}_\bullet(A)$ thus has a Lie bialgebra structure of degree $(2-n, 2-n)$. Thus taking the homology on both sides of and combining the above argument we in fact obtain a commutative diagram $$\xymatrixcolsep{4pc}
\xymatrix{
(\mathbf\Lambda^\bullet\mathrm{HC}_\bullet(A)[\hbar], \star )
\ar[r]^{\widetilde{\mathrm{Tr}}}\ar@{~>}[d]_{\textup{quantization}}
&\big(\mathrm H_\bullet(\mathrm{Rep}_V(\tilde R))^{\mathrm{GL}}[\hbar], \star\big)\ar@{~>}[d]^{\textup{quantization}}\\
\mathbf\Lambda^\bullet\mathrm{HC}_\bullet(A)
\ar[r]^{\mathrm{Tr}}
&
\mathrm H_\bullet(\mathrm{Rep}_V(\tilde R))^{\mathrm{GL}},
}$$ where on the right hand side, we have used the fact that $\mathrm{H}_\bullet(\mathrm{Rep}_V(\tilde R)^{\mathrm{GL}})\cong
\mathrm H_\bullet(\mathrm{Rep}_V(\tilde R))^{\mathrm{GL}}$ (see [@BKR Theorem 2.6(b)] for a proof).
Derived representation schemes {#sect:DRep}
==============================
In this section, we briefly discuss the results in previous sections with the derived representation schemes, introduced by Berest, Khachatryan and Ramadoss. The interested reader may refer to [@BCER; @BFR; @BKR; @BR] for more details.
Derived representation schemes {#derived-representation-schemes}
------------------------------
In algebraic geometry, there is an equivalence of categories between affine schemes and commutative algebras, and many geometric structures on affine schemes have an algebraic description, and vice versa. However, for associative algebras, this correspondence does not exist. In 1998, Kontsevich and Rosenberg [@KR] proposed a heuristic principle to study the [*non-commutative geometry*]{} on associative, not-necessarily commutative, algebras, which is roughly stated as follows: for an associative algebra over a field $k$, say $A$, any non-commutative geometric structure, such as non-commutative Poisson, non-commutative symplectic, etc., should induce its classical counterpart in a natural way on its representation scheme $\mathrm{Rep}_V(A)$, for all $k$-vector space $V$.
During the past decade, much progress has been made in the study of non-commutative geometry under the guidance of the Kontsevich-Rosenberg Principle (see [@CBEG; @Ginzburg01; @VdB2]). However, for a general algebra $A$, $\mathrm{Rep}_V(A)$ is very singular. In 2011 Berest et. al. [@BKR] suggested that one should instead consider the [*derived*]{} representation schemes of $A$. In this case, one replaces $A$ with its cofibrant resolution, say $QA$, in the category of differential graded algebras, and then considers the DG representation schemes of $QA$, which are then smooth in the DG sense. The cofibrant resolution of a DG algebra is not unique, but unique up to homotopy. Correspondingly, $\mathrm{Rep}_V(QA)$ is also unique up to homotopy. By modulo such ambiguity, the DG representation scheme of $QA$ in $V$ in the homotopy category of DG commutative algebras, denoted by $\mathrm{DRep}_V(A)$ and called the [*derived representation scheme*]{} of $A$ in $V$, is a very good object that we can successfully apply the Kontsevich-Rosenberg principle. Formally, their result is stated as follows.
For any two objects $A, B\in\mathbf{DGA}$ and any $f\in\mathrm{Hom}(A, B)$, let $QA$ and $QB$ be any cofibrant replacement of $A$ and $B$ respectively, and $Qf\in\mathrm{Hom}(QA,QB)$ be the corresponding cofibrant lifting of $f$. The functor has a total left derived functor $$\mathbf L(-)_V: \mathbf{Ho}(\mathbf{DGA})\to\mathbf{Ho}(\mathbf{CDGA}),\quad A\mapsto (QA)_V,\; f\mapsto (Qf)_V.$$
According to [@BKR], $\mathbf L(A)_V$ is called the [*derived representation scheme*]{} (or [*DRep*]{} for short) of $A$ in $V$, and is also denoted by $\mathrm{DRep}_V(A)$; its homology $\mathrm{H}_\bullet(\mathrm{DRep}_V(A))$ is called the [*representation homology*]{} of $A$, and is sometimes denoted by $\mathrm H_\bullet(A, V)$.
Suppose $A$ is a Koszul algebra. Then $A$ has an explicit cofibrant resolution $\Omega(A^{{\textup{!`}}})\stackrel{\sim}{\twoheadrightarrow} A$. Thus $\mathrm{DRep}_n(A)$ is explicit in the Koszul case: suppose $\bar A^{{\textup{!`}}}$ has a set of basis $\{x^\alpha\}_{\alpha\in I}$, and the reduced coproduct $\bar\Delta(x^{\alpha})=\sum_{(x^{\alpha})} x^{\alpha_1}\otimes x^{\alpha_2}$. Then $\mathrm{DRep}_n(A)$ is the quasi-free DG commutative algebra generated by $$\label{DRepofKoszulCY}
\big\{x_{ij}^{\alpha}\;\big|\;\alpha\in I, 1\le i,j\le n, |x_{ij}^{\alpha}|=|x^{\alpha}|-1\big\},$$ with $$\partial(x_{ij}^{\alpha})=\sum_{(x^{\alpha})}(-1)^{|x^{\alpha_1}|}\sum_{k=1}^{n}x_{ik}^{\alpha_1}\cdot x_{kj}^{\alpha_2}.$$
Comparison of $\mathrm{DRep}_n(A)$ and $\mathrm{Rep}_n(\tilde R)$
-----------------------------------------------------------------
Suppose $A$ is a Kosuzl algebra. Then one obtains $R=\Omega A^{{\textup{!`}}}$ from $\tilde R=\Omega\tilde A^{{\textup{!`}}}$ by identifying the counit of $A^{{\textup{!`}}}$ with zero. Therefore, if we identify the coordinates of $\mathrm{Rep}_V(\tilde R)$ corresponding the co-unit of $A^{{\textup{!`}}}$ with zero, we obtain $\mathrm{Rep}_V(R)$, which is isomorphic to $\mathrm{DRep}_V(A)$ in the homotopy category of DG algebras.
We here give two remarks regarding the role of the co-unit of $A^{{\textup{!`}}}$. On the one hand, since $R=\Omega(A^{{\textup{!`}}})\simeq A$ is a cofibrant resolution of $A$, $\mathrm{Rep}_V(R)$ is a derived functor of $A$ rather than $\mathrm{Rep}_V(\tilde R)$. $\mathrm{Rep}_V(\tilde R)$ is hardly a derived functor since $\tilde R$ is acyclic.
On the other hand, considering $\mathrm{Rep}_V(\tilde R)$ has the advantage when considering the $\mathrm{GL}$-invariants; namely the $\mathrm{GL}(V)$-invariant functions, the cotangent vector space and the tangent vector space on $\mathrm{Rep}_V(\tilde R)$ is directly related to the cyclic complex, the Hochschild chain and cochain complexes of $A$ via the trace maps (see and ). Sometimes, keeping the unit (respectively co-unit) of an algebra (respectively coalgebra) is important in the study of Noncommutative Geometry; see for example the early works of Connes and Quillen in this field.
Now suppose $A$ is also Calabi-Yau. In [@BCER; @CEEY], we showed that $R$ has a shifted Poisson structure, while in the current paper we deduced the shifted Poisson structure on $\tilde R$. We may understand them in this way: geometrically, $\mathrm{DRep}_V(A)$ can be understood as a DG subscheme of $\mathrm{Rep}_V(\tilde R)$. It is direct to see that the Poisson bracket on $\mathrm{Rep}_V(\tilde R)$ restricts to a Poisson bracket on $\mathrm{Rep}_V(R)$ studied in [@BCER] and [@CEEY]; in other words, $\mathrm{Rep}_V(R)$ is a shifted Poisson subscheme of $\mathrm{Rep}_V(\tilde R)$.
Example: The Sklyanin algebras {#sect:example}
==============================
In this section, we give two examples of Koszul Calabi-Yau algebras, namely the 3- and 4-dimensional Sklyanin algebras, and study the corresponding shifted bi-symplectic structure with some detail. In general the geometry of the representations of Sklyanin algebras are complicated, however, both the derived representation schemes $\mathrm{DRep}_V(A)$ and the representation schemes $\mathrm{Rep}_V(\tilde R)$ are easy to describe; here $A$ is a Sklyanin algebra and $\tilde R=\Omega(\tilde A^{{\textup{!`}}})$ as before.
Let $a,b,c\in k$ satisfying the following two conditions:
1. $[a:b:c]\in\mathbb{P}^2_k\backslash D$, where $$D=\{[1:0:0], [0:1:0], [0:0:1]\}\cup \{[a:b:c]|a^3=b^3=c^3=1\};$$
2. $abc\ne 0$ and $(3abc)^3\ne (a^3+b^3+c^3)^3$.
The 3-dimensional Sklyanin algebra $A=A(a,b,c)$ is the graded $k$-algebra with generators $x, y, z$ of degree one, and relations $$\begin{array}{c}
cx^2+bzy+ayz=0,\\
azx+cy^2+bxz=0,\\
byx+axy+cz^2=0.
\end{array}$$ Smith showed in [@Smith Example 10.1] that $A$ is Koszul, whose dual algebra $A^!$ is generated by $\xi_1,\xi_2,\xi_3$ with relations $$\begin{array}{lll}
c\xi_2\xi_3-b\xi_3\xi_2,& b\xi_1^2-a\xi_2\xi_3, & c\xi_3\xi_1-b\xi_1\xi_3,\\
b\xi_2^2-a\xi_3\xi_1, & c\xi_1\xi_2-b\xi_2\xi_1,& b\xi_3^2-a\xi_1\xi_2.
\end{array}$$ In degree 3, we have relations $$\begin{array}{l}
\xi_i^3=\displaystyle\frac{a}{b}\xi_1\xi_2\xi_3
=\displaystyle\frac{a}{b}\xi_3\xi_1\xi_2
=\displaystyle\frac{a}{b}\xi_2\xi_3\xi_1
=\displaystyle\frac{a}{c}\xi_2\xi_1\xi_3
=\displaystyle\frac{a}{c}\xi_3\xi_2\xi_1
=\displaystyle\frac{a}{c}\xi_1\xi_3\xi_2,\quad i=1,2,3,\\
\xi_i\xi_j^2=\xi_j^2\xi_i=0,\; i,j=1,2, 3\;\mbox{and}\; i\ne j.
\end{array}$$ From these relations, we obtain that $A^{{\textup{!`}}}$ is isomorphic to a graded coalgebra spanned by $$\label{generatorsofdualof3Skl}
\begin{array}{cl}
A_0^{{\textup{!`}}}: & {\mathbf 1}\\
A_1^{{\textup{!`}}}: & \xi_1^*,\xi_2^*,\xi_3^*\\
A_2^{{\textup{!`}}}: &\xi_1^*\xi_1^*,\xi_2^*\xi_2^*,\xi_3^*\xi_3^*\\
A_3^{{\textup{!`}}}:&\xi_3^*\xi_2^*\xi_1^*
\end{array}$$ with the coproducts given by $$\begin{aligned}
\Delta(\mathbf 1)&=&\mathbf 1\otimes\mathbf 1,\nonumber\\
\Delta(\xi_i)&=&\mathbf 1\otimes \xi_i+\otimes \xi_i\otimes \mathbf 1,\quad i=1,2,3,\nonumber\\
\Delta(\xi_1^*\xi_1^*)&=&\mathbf 1\otimes \xi_1^*\xi_1^* +\xi_1^*\otimes \xi_1^*
+\frac{a}{b}\xi_2^*\otimes\xi_3^*+\frac{a}{c}\xi_3^*\otimes\xi_2^*
+\xi_1^*\xi_1^*\otimes\mathbf 1,
\nonumber\\
\Delta(\xi_2^*\xi_2^*)&=&\mathbf 1\otimes \xi_2^*\xi_2^*+\xi_2^*\otimes\xi_2^*
+\frac{a}{c}\xi_1^*\otimes\xi_3^*
+\frac{a}{b}\xi_3^*\otimes\xi_1^*
+\xi_2^*\xi_2^*\otimes\mathbf 1,
\nonumber \\
\Delta(\xi_3^*\xi_3^*)&=& \mathbf 1\otimes \xi_3^*\xi_3^*+\xi_3^*\otimes\xi_3^*
+\frac{a}{b}\xi_1^*\otimes \xi_2^*
+\frac{a}{c}\xi_2^*\otimes \xi_1^*
+\xi_3^*\xi_3^*\otimes\mathbf 1,\nonumber\\
\Delta(\xi_3^*\xi_2^*\xi_1^*)
&=&{\mathbf 1}\otimes\xi_3^*\xi_2^*\xi_1^*
+\frac{b}{a}\xi_1^*\otimes\xi_1^*\xi_1^*
+\frac{b}{a}\xi_2^*\otimes\xi_2^*\xi_2^*
+\frac{b}{a}\xi_3^*\otimes\xi_3^*\xi_3^*
\nonumber\\
&&+\frac{b}{a}\xi_1^*\xi_1^*\otimes\xi_1^*
+\frac{b}{a}\xi_2^*\xi_2^*\otimes\xi_2^*
+\frac{b}{a}\xi_3^*\xi_3^*\otimes\xi_3^*
+\xi_3^*\xi_2^*\xi_1^*\otimes {\mathbf 1}.\label{coprodof3Skl}\end{aligned}$$ The pairing on $A^!$ given by is cyclic, and therefore $A$ is 3-Calabi-Yau.
Now, $\tilde R$ is the DG algebra freely generated by those elements in with degree shifted down by one. From we see that the $(-1)$-shifted bi-symplectic structure $\omega\in\mathrm{DR}^{2}_{\mathrm{nc}}\tilde R$ is given by $$\begin{gathered}
\omega
=d({\mathbf 1})\otimes d(\xi_3^*\xi_2^*\xi_1^*)
+\frac{b}{a}d(\xi_1^*)\otimes d(\xi_1^*\xi_1^*)
+\frac{b}{a}d(\xi_2^*)\otimes d(\xi_2^*\xi_2^*)
+\frac{b}{a}d(\xi_3^*)\otimes d(\xi_3^*\xi_3^*)
\\
+\frac{b}{a}d(\xi_1^*\xi_1^*)\otimes d(\xi_1^*)
+\frac{b}{a}d(\xi_2^*\xi_2^*)\otimes d(\xi_2^*)
+\frac{b}{a}d(\xi_3^*\xi_3^*)\otimes d(\xi_3^*)
+d(\xi_3^*\xi_2^*\xi_1^*)\otimes d({\mathbf 1}).\end{gathered}$$ Note that here (as well as in the next example) $d({\mathbf 1})\ne 0$.
Let $\alpha, \beta,\gamma\in k$ such that $$\alpha+\beta+\gamma+\alpha\beta\gamma=0,\quad \{\alpha,\beta,\gamma\}\cap\{0,\pm 1\}=\emptyset.$$ The 4-dimensional Sklyanin algebra $A=A(\alpha,\beta,\gamma)$ is the graded $k$-algebra with generators $x_0,x_1,x_2,x_3$ of degree one, and relations $f_i=0$, where $$\begin{aligned}
f_1=x_0x_1-x_1x_0-\alpha(x_2x_3+x_3x_2),&& f_2=x_0x_1+x_1x_0-(x_2x_3-x_3x_2),\\
f_3=x_0x_2-x_2x_0-\beta(x_3x_1+x_1x_3),&& f_4=x_0x_2+x_2x_0-(x_3x_1-x_1x_3),\\
f_5=x_0x_3-x_3x_0-\gamma(x_1x_2+x_2x_1),&& f_6=x_0x_3+x_3x_0-(x_1x_2-x_2x_1).\end{aligned}$$ As proved by Smith and Stafford ([@SS Propositions 4.3-4.9]), $A$ is Koszul, whose Koszul dual algebra $A^!$ is generated by $\xi_0,\xi_1,\xi_2,\xi_3$ with the following relations: $$\begin{aligned}
&& \xi_0^2=\xi_1^2=\xi_2^2=\xi_3^2=0,\nonumber\\
&& 2\xi_2\xi_3+(\alpha+1)\xi_0\xi_1-(\alpha-1)\xi_1\xi_0=0,\nonumber\\
&& 2\xi_3\xi_2+(\alpha-1)\xi_0\xi_1-(\alpha+1)\xi_1\xi_0=0,\nonumber\\
&& 2\xi_3\xi_1+(\beta+1)\xi_0\xi_2-(\beta-1)\xi_2\xi_0=0,\nonumber\\
&& 2\xi_1\xi_3+(\beta-1)\xi_0\xi_2-(\beta+1)\xi_2\xi_0=0,\nonumber\\
&& 2\xi_1\xi_2+(\gamma+1)\xi_0\xi_3-(\gamma-1)\xi_3\xi_0=0,\nonumber\\
&& 2\xi_2\xi_1+(\gamma-1)\xi_0\xi_3-(\gamma+1)\xi_3\xi_0=0.\nonumber\end{aligned}$$ Smith and Stafford also showed that $A^!$ admits a non-degenerate symmetric pairing, and hence $A$ is 4-Calabi-Yau. In particular, in degree 4, we have the following identities: $$\begin{aligned}
&&\xi_0\xi_i\xi_0\xi_i=-\xi_i\xi_0\xi_i\xi_0\quad
\mbox{for}\; 1\le i\le j, \quad
\xi_0\xi_i\xi_0\xi_j=0\quad\mbox{for}\; i\ne j,\\
&&\xi_0\xi_2\xi_0\xi_2=\frac{1+\gamma}{1-\beta}\frac{1+\alpha}{1-\gamma}\xi_0\xi_1\xi_0\xi_1,\\
&&\xi_0\xi_3\xi_0\xi_3=\frac{1+\alpha}{1-\gamma}\xi_0\xi_1\xi_0\xi_1.\end{aligned}$$ From the above two groups of identities, we obtain that the Koszul dual coalgebra $A^{{\textup{!`}}}$ is isomorphic to a coalgebra spanned by $$\begin{array}{cl}
A_0^{{\textup{!`}}}: & {\mathbf 1}\\
A_1^{{\textup{!`}}}: &\xi_0^*,\xi_1^*,\xi_2^*,\xi_3^*\\
A_2^{{\textup{!`}}}: &\xi_1^*\xi_0^*,\xi_2^*\xi_0^*,\xi_3^*\xi_0^*,\xi_0^*\xi_1^*,\xi_0^*\xi_2^*,\xi_0^*\xi_3^*\\
A_3^{{\textup{!`}}}:&\xi_0^*\xi_1^*\xi_0^*,\xi_0^*\xi_2^*\xi_0^*,\xi_0^*\xi_3^*\xi_0^*,\xi_1^*\xi_0^*\xi_1^*\\
A_4^{{\textup{!`}}}:&\xi_1^*\xi_0^*\xi_1^*\xi_0^*,
\end{array}$$ with the coproduct of $\xi_1^*\xi_0^*\xi_1^*\xi_0^*$ given by $$\begin{gathered}
\Delta(\xi_1^*\xi_0^*\xi_1^*\xi_0^*)
={\mathbf 1}\otimes \xi_1^*\xi_0^*\xi_1^*\xi_0^*
+\xi_0^*\otimes \xi_1^*\xi_0^*\xi_1^*
-\xi_1^*\otimes \xi_0^*\xi_1^*\xi_0^*\\
-\frac{1-\beta}{1+\gamma}\frac{1-\gamma}{1+\alpha}\xi_2^*\otimes \xi_0^*\xi_2^*\xi_0^*
-\frac{1-\gamma}{1+\alpha}\xi_3^*\otimes \xi_0^*\xi_3^*\xi_0^*
-\xi_1^*\xi_0^*\otimes \xi_1^*\xi_0^* \\
-\frac{1-\beta}{1+\gamma}\frac{1-\gamma}{1+\alpha}\xi_2^*\xi_0^*\otimes \xi_2^*\xi_0^*
-\frac{1-\gamma}{1+\alpha}\xi_3^*\xi_0^*\otimes \xi_3^*\xi_0^*
+\xi_0^*\xi_1^*\otimes \xi_0^*\xi_1^* \\
+\frac{1-\beta}{1+\gamma}\frac{1-\gamma}{1+\alpha}\xi_0^*\xi_2^*\otimes \xi_0^*\xi_2^*
+\frac{1-\gamma}{1+\alpha}\xi_0^*\xi_3^*\otimes \xi_0^*\xi_3^*
-\xi_1^*\xi_0^*\xi_1^*\otimes \xi_0^*\\
+\xi_0^*\xi_1^*\xi_0^*\otimes \xi_1^*
+\frac{1-\beta}{1+\gamma}\frac{1-\gamma}{1+\alpha} \xi_0^*\xi_2^*\xi_0^*\otimes \xi_2^*
+ \frac{1-\gamma}{1+\alpha}\xi_0^*\xi_3^*\xi_0^*\otimes \xi_3^*
+\xi_1^*\xi_0^*\xi_1^*\xi_0^*\otimes {\mathbf 1}.\end{gathered}$$ Thus the $(-2)$-shifted bi-symplectic structure $\omega\in\mathrm{DR}^{2}_{\mathrm{nc}}\tilde R$ is given by $$\begin{gathered}
d({\mathbf 1})\otimes d(\xi_1^*\xi_0^*\xi_1^*\xi_0^*)
+d(\xi_0^*)\otimes d(\xi_1^*\xi_0^*\xi_1^*)
-d(\xi_1^*)\otimes d(\xi_0^*\xi_1^*\xi_0^*)\\
-\frac{1-\beta}{1+\gamma}\frac{1-\gamma}{1+\alpha}d(\xi_2^*)\otimes d(\xi_0^*\xi_2^*\xi_0^*)
-\frac{1-\gamma}{1+\alpha}d(\xi_3^*)\otimes d(\xi_0^*\xi_3^*\xi_0^*)
-d(\xi_1^*\xi_0^*)\otimes d(\xi_1^*\xi_0^* )\\
-\frac{1-\beta}{1+\gamma}\frac{1-\gamma}{1+\alpha}d(\xi_2^*\xi_0^*)\otimes d(\xi_2^*\xi_0^* )
-\frac{1-\gamma}{1+\alpha}d(\xi_3^*\xi_0^*)\otimes d(\xi_3^*\xi_0^* )
+d(\xi_0^*\xi_1^*)\otimes d(\xi_0^*\xi_1^*) \\
+\frac{1-\beta}{1+\gamma}\frac{1-\gamma}{1+\alpha} d(\xi_0^*\xi_2^*)\otimes d(\xi_0^*\xi_2^* )
+\frac{1-\gamma}{1+\alpha}d(\xi_0^*\xi_3^*)\otimes d(\xi_0^*\xi_3^* )
-d(\xi_1^*\xi_0^*\xi_1^*)\otimes d(\xi_0^*)\\
+ d(\xi_0^*\xi_1^*\xi_0^*)\otimes d(\xi_1^*)
+ \frac{1-\beta}{1+\gamma}\frac{1-\gamma}{1+\alpha}d(\xi_0^*\xi_2^*\xi_0^*)\otimes d(\xi_2^*)
+\frac{1-\gamma}{1+\alpha} d(\xi_0^*\xi_3^*\xi_0^*)\otimes d(\xi_3^*)
\\
+d(\xi_1^*\xi_0^*\xi_1^*\xi_0^*)\otimes d({\mathbf 1}).\end{gathered}$$
This work is partially supported by NSFC No. 11671281. We also thank R. Nest and Song Yang for helpful comments during the preparation of the work.
[100]{}
Y. Berest, X. Chen, F. Eshmatov and A. Ramadoss, [Noncommutative Poisson structures, derived representation schemes and Calabi-Yau algebras]{}. Contemp. Math. [**583**]{} (2012), 219–246.
Y. Berest, G. Felder and A. Ramadoss, [Derived representation schemes and noncommutative geometry]{}. Contemp. Math. [**607**]{} (2014), 113–162.
Y. Berest, G. Khachatryan and A. Ramadoss, [Derived representation schemes and cyclic homology]{}. Adv. Math. [**245**]{} (2013), 625-689.
Y. Berest and A. Ramadoss, [Stable representation homology and Koszul duality]{}. J. Reine Angew. Math. [**715**]{} (2016), 143–187.
R. Berger, T. Lambre and A. Solotar, Koszul calculus, Glasgow Math. J. 2017, 1-39. doi:10.1017/S0017089517000167.
G.M. Bergman, Coproducts and some universal ring constructions. Trans. Amer. Math. Soc. [**200**]{} (1974), 33–88.
R. Bocklandt and L. Le Bruyn, Necklace Lie algebras and noncommutative symplectic geometry. Math. Z. [**240**]{} (2002), no. 1, 141–167.
D. Calaque, T. Pantev, B. Toën and M. Vaquié, Shifted Poisson structures and deformation quantization. J. Topol. [**10**]{} (2017), no. 2, 483–584.
X. Chen, A. Eshmatov, F. Eshmatov and S. Yang, The derived noncommutative Poisson bracket on Koszul Calabi-Yau algebras. J. Noncommut. Geom. [**11**]{} (2017), no. 1, 111–160.
X. Chen, F. Eshmatov and W.L. Gan, Quantization of the Lie bialgebra of string topology, Communications in Mathematical Physics [**301**]{}:1(2011), 37–53.
X. Chen, S. Yang and G. Zhou, Batalin-Vilkovisky algebras and the noncommutative Poincaré duality of Koszul Calabi-Yau algebras. J. Pure Appl. Algebra [**220**]{} (2016), no. 7, 2500–2532.
P.M. Cohn, The affine scheme of a general ring. Applications of sheaves (Proc. Res. Sympos. Appl. Sheaf Theory to Logic, Algebra and Anal., Univ. Durham, Durham, 1977), pp. 197–211, Lecture Notes in Math. [**753**]{}, Springer, Berlin, 1979.
W. Crawley-Boevey, Poisson structures on moduli spaces of representations. J. Algebra [**325**]{} (2011), 205–215.
W. Crawley-Boevey, P. Etingof and V. Ginzburg, Noncommutative geometry and quiver algebras. Adv. Math. [**209**]{} (2007), no. 1, 274–336.
L. de Thanhoffer de Völcsey and M. Van den Bergh, Calabi-Yau Deformations and Negative Cyclic Homology, to appear in Journal of Noncommutative Geometry, arXiv:1201.1520.
V. Ginzburg, Noncommutative symplectic geometry, quiver varieties, and operads. Math. Res. Lett. [**8**]{} (2001), no. 3, 377–400.
V. Ginzburg, Calabi-Yau algebras, arXiv:math/0612139.
V. Ginzburg and T. Schedler, Moyal quantization and stable homology of necklace Lie algebras. Mosc. Math. J. [**6**]{} (2006), no. 3, 431–459, 587.
J.D.S. Jones and J. McCleary, Hochschild homology, cyclic homology, and the cobar construction, Adams Memorial Symposium on Algebraic Topology, Cambridge University Press (1992) 53–65.
B. Keller, Derived invariance of higher structures on the Hochschild complex, available at https://webusers.imj-prg.fr/$~\sim$bernhard.keller/publ/index.html.
B. Keller, Calabi-Yau triangulated categories. Trends in representation theory of algebras and related topics, 467–489, EMS Ser. Congr. Rep., Eur. Math. Soc., Zürich, 2008.
M. Kontsevich and A. Rosenberg, Noncommutative smooth spaces. The Gelfand Mathematical Seminars, 1996–1999, 85–108, Gelfand Math. Sem., Birkhäuser Boston, Boston, MA, 2000.
Y. T. Lam, Calabi-Yau categories and quivers with superpotential, Ph.D. Thesis 2015, available at http://people.maths.ox.ac.uk/joyce/theses/theses.html.
J.-L. Loday, Cyclic homology, 2nd edition, Grundl. Math. Wiss. [**301**]{}, Springer-Verlag, Berlin, 1998.
J.-L. Loday, D. Quillen, Cyclic homology and the Lie algebra homology of matrices, Comment. Math. Helv. [**59**]{} (1984) 565–591.
J.-L. Loday and B. Vallette, Algebraic operads. Grundlehren der Mathematischen Wissenschaften [**346**]{}. Springer, Heidelberg, 2012.
V. Melani, Poisson bivectors and Poisson brackets on affine derived stacks. Adv. Math. [**288**]{} (2016), 1097–1120.
T. Pantev, B. Toën, M. Vaquié and G. Vezzosi, Shifted symplectic structures. Publ. Math. Inst. Hautes Études Sci. [**117**]{} (2013), 271–328.
S.B. Priddy, Koszul resolutions. Trans. Amer. Math. Soc. [**152**]{}, 1970, 39–60.
Pridham, Shifted Poisson and symplectic structures on derived N-stacks. J. Topol. [**10**]{} (2017), no. 1, 178–210.
D. Quillen, Algebra cochains and cyclic cohomology. Inst. Hautes Études Sci. Publ. Math. [**68**]{} (1989), 139–174.
T. Schedler, A Hopf algebra quantizing a necklace Lie algebra canonically associated to a quiver. Int. Math. Res. Not. 2005, no. [**12**]{}, 725–760.
B. Shoikhet, Koszul duality in deformation quantization and Tamarkin’s approach to Kontsevich formality. Adv. Math. [**224**]{} (2010), no. 3, 731–771.
S.P. Smith, [Some finite dimensional algebras related to elliptic curves]{}, in [Representation Theory of Algebras and Related Topics]{} (Mexico City, 1994), CMS Conf. Proc. 19, Amer. Math. Soc., Providence, 1996, 315–348.
S.P. Smith and J.T. Stafford, [Regularity of the four dimensional Sklyanin algebra]{}. Compositio Math. [**83**]{} (1992), 259–289.
M. Van den Bergh, Noncommutative homology of some three-dimensional quantum spaces, K-Theory [**8**]{} (1994) 213–230.
M. Van den Bergh, A relation between Hochschildt homology and cohomology for Georenstein rings, Proc. Amer. Math. Soc. [**126**]{} (1998) 1345-1348, and Erratum, Proc. Amer. Math. Soc. 130 (2002) 2809–2810.
M. Van den Bergh, Double Poisson algebras. Trans. Amer. Math. Soc. [**360**]{} (2008), no. 11, 5711–5769.
M. Van den Bergh, Non-commutative quasi-Hamiltonian spaces. Poisson geometry in mathematics and physics, 273–299, Contemp. Math. [**450**]{}, Amer. Math. Soc., Providence, RI, 2008.
M. Van den Bergh, Calabi-Yau algebras and superpotentials. Selecta Math. (N.S.) [**21**]{} (2015), no. 2, 555–603.
[^1]: Remind that for a graded $R$-bimodule $M$, $\Sigma M$ is a graded $R$-bimodule with $a\cdot \Sigma m\cdot b=(-1)^{|a|}\Sigma (amb)$, for homogeneous $a, b\in R$ and $m\in M$.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The so-called [*Modern Theory of Polarization*]{}, which rigorously defines the spontaneous polarization of a periodic solid and provides a route for its computation in electronic structure codes through the Berry phase, is introduced in a simple qualitative discussion.'
address: 'Materials Theory, ETH Zurich, Wolfgang-Pauli-Strasse 27, 8093 Zürich, Switzerland'
author:
- 'Nicola A. Spaldin'
bibliography:
- '/Users/nspaldin/papers/Nicola.bib'
title: 'A Beginner’s Guide to the Modern Theory of Polarization'
---
polarization, Berry phase, electronic structure calculation
Introduction
============
The concept of electric dipole moment is central in the theory of electrostatics, particularly in describing the response of systems to applied electric fields. For finite systems such as molecules it poses no conceptual or practical problems. In the ionic limit the dipole moment, $d$, of a collection of charges, $q_i$, at positions $\mathbf{r}_i$ is defined as $$d = \Sigma_i q_i \mathbf{r}_i \quad ;$$ for the case of a continuous charge density, $e n(\mathbf{r})$ (where $e$ is the electronic charge and $n(\mathbf{r})$ is the number density) this expression is straightforwardly extended to $$d = \int e n(\mathbf{r}) \mathbf{r} d\mathbf{r} \quad .$$ Provided that the molecule or cluster carries no net charge these expressions are well defined, can be straightforwardly evaluated and yield results – for example for the direction of the dipole moment – that are consistent with our intuitive understanding.
Things apparently start to turn to custard, however, when we try to extend this simple reasoning to bulk solids. The usual way to define intrinsic quantities in macroscopic systems is to introduce the property per unit volume or mass. For example the magnetization is the magnetic moment per unit volume, and the bulk analogue to the electric dipole moment, the electric polarization, should be represented by the electric dipole moment per unit volume. The relevant quantity is then evaluated within a small repeat unit – the unit cell – of the solid, and normalized with the volume of the chosen unit cell. The problem with this simple method in the case of electric polarization can be understood in the simple one-dimensional cartoon of Figure \[1Dchain\]: Without performing any calculations, we can see that the two equally valid unit cells shown with dashed lines have completely [*opposite*]{} orientations of the polarization!
This difficulty led to tremendous confusion in the field, with discussions as fundamental as whether the polarization (and related quantities such as the piezoelectric response) could be considered as intrinsic properties in bulk solids, or are in fact determined by details of the surface termination. Thankfully the confusion was resolved around 20 years ago with the introduction of the so-called [*Modern theory of polarization*]{}. This very elegant theory showed that [*changes*]{} in polarization are in fact rigorously defined, can be calculated quantum mechanically using electronic structure methods, and correspond to experimentally measurable observables.
The purpose of this article is to introduce in the simplest possible terms the apparent difficulties associated with defining polarization in bulk solids, and the solutions provided by the modern theory. It is motivated by my having explained these concepts repeatedly to many and diverse students ranging from experimentalists with a casual interest in understanding obscure theory papers to beginning hard-core theoretical solid-state physicists and quantum chemists. This article in no way intends to substitute for the elegant early papers on the topic, nor the subsequent detailed and rigorous review papers which are referenced throughout. Indeed I hope that this informal introduction provides sufficient background for the reader to tackle these excellent articles without intimidation.
Bulk periodicity, the polarization lattice and the polarization quantum
=======================================================================
We begin by reconciling the different values for polarization obtained for the different choices of unit cells in Fig. \[1Dchain\] by introducing a formal concept that at first sight is even more confusing – that is the [*multi-valuedness*]{} of the bulk polarization. We will show, however, that a multi-valued polarization is a natural consequence of the periodicity in a bulk solid, and hopefully that it is actually not so frightening. We will see, in fact that [*changes*]{} in polarization – which are the quantities that are anyway measured in experiments – are single valued and well defined, and we can once again sleep without anxiety.
We take the simplest possible example of a one-dimensional chain of singly charged alternating anions and cations – the closest real-life analogue would be rock-salt structure sodium chloride in just one direction. Look at Figure \[1Dchain\] which shows such a chain with the atoms spaced a distance $a/2$ apart so that the lattice constant is $a$. The first thing to notice is that all of the ions are centers of inversion symmetry: If I sit on any ion and look to the left, then to the right I see no difference. So by definition this lattice is non-polar.
![One-dimensional chain of alternating anions and cations, spaced a distance $a/2$ apart, where $a$ is the lattice constant. The dashed lines indicate two representative unit cells which are used in the text for calculation of the polarization.[]{data-label="1Dchain"}](1Dchain_small){width="1.0\columnwidth"}
Now let’s work out the polarization by calculating the dipole moment [*per unit length*]{} (the definition in three dimensions is dipole moment per unit volume) using in turn the two unit cells shown as the dashed rectangles to compute the local dipole moment. First, the cell on the left. Taking the left edge of the shell as the origin ($x=0$), we have an ion with charge -1 at position $a/4$, and an ion with charge +1 at position $3a/4$. So the polarization, or dipole moment per unit length is $$\begin{aligned}
p & = & \frac{1}{a} \sum_i q_i x_i \nonumber\\
& = & \frac{1}{a} \left(-1 \times \frac{a}{4} +1 \times \frac{3a}{4} \right) \nonumber\\
& = & \frac{1}{a} \frac{2a}{4} \nonumber\\
& = & \frac{1}{2} \end{aligned}$$ in units of $|e|$ per unit length. Immediately we have an apparent problem: Using this method, our non-polar chain has a non-zero polarization.
I am afraid that things will get worse before they get better. Next, let’s do the same exercise using the right-most unit cell. Again taking the left edge of the unit cell as the origin, this time there is a positively charged ion at position $a/4$, and a negatively charged ion at $3a/4$. So $$\begin{aligned}
p & = & \frac{1}{a} \left(+1 \times \frac{a}{4} -1 \times \frac{3a}{4} \right) \nonumber\\
& = & \frac{1}{a} \times -\frac{2a}{4} \nonumber\\
& = & -\frac{1}{2} \quad .\end{aligned}$$ Again a non-zero value, and this time [*different*]{} from the value we obtained using the other, equally valid unit cell, by an amount $a$.
So what is going on here, and how can we connect it to physical reality? Well, if we were to repeat this exercise with many choices of unit cell (convince yourself by choosing a couple of arbitrary unit cells and giving it a try!), we would obtain many values of polarization, with each value differing from the original value by an integer. We call this collection of polarization values the [*polarization lattice*]{}. In this case it is $..., -5/2, -3/2, -1/2, 1/2, 3/2, 5/2 ...$. Notice that the lattice of polarization values is symmetric about the origin. In fact this is the signature of a non-polar structure: The polarization lattice may or may not contain zero as one of its elements, but it must be centrosymmetric around zero.
Now what is the significance of the spacing (in this case 1) between the allowed values? Well, imagine removing an electron from one of the anions in the lattice (leaving a neutral atom) and moving it by one unit cell to put it on the next anion to the right. Because of the periodic boundary conditions of the infinite lattice, the next anion simultaneously has it’s electron removed and moved one unit cell to the right, and so it is able to accept the incoming electron and appear unchanged at the end of the process. There has been no change in the physics of the system resulting from the relocation of the electrons by one unit cell to the right. But what has happened to the polarization? Well, in each unit cell a charge of $-1$ has moved a distance $a$, changing the dipole moment by $-a$ and the polarization by -1. We can clearly perform this thought experiment any number of times, and in either direction, changing the polarization by any integer without changing the physical system! We call the value of polarization resulting from moving one electron by one unit cell the [*polarization quantum*]{}, $P_q$. In one dimension it is equal to the lattice constant divided by the length of the unit cell, which is just an integer (in units of the electronic charge per unit length). Going back to the polarization lattice of our non-polar chain, we see that it’s polarization values correspond to half-polarization quanta. In fact all non-polar systems have polarization lattices of either $0 \pm nP_q$ or $\frac{P_q}{2} \pm nP_q$.
![Schematic of the Sawyer-Tower method of measuring ferroelectric polarization. The material on the left is polarized in the up direction and its surface charge is screened by electrons in the upper electrode (grey) and holes in the lower electrode. When the polarization is switched (right), electrons and holes flow through the external circuit to screen the new opposite surface charges, and are counted by comparing the voltage across the material with that across a reference capacitor.[]{data-label="Sawyer-Tower"}](Sawyer-Tower_small){width="1.0\columnwidth"}
If this all seems too esoteric, please bear with me for one more paragraph by which time I hope things should start making sense. First, let’s think about how we measure electrical polarization, and what a reported measured polarization really means. Look at Figure \[Sawyer-Tower\] – this is a cartoon of a standard way of measuring the electrical polarization using a so-called Sawyer-Tower circuit. On the left the material has become polarized in the up direction as a result of the cation sub-lattice displacing upwards relative to the anion sub-lattice. This could happen, for example during a ferroelectric phase transition with an external electric field applied in the up direction (the light colored cations with the dashed-line bonds indicate their positions in the high-symmetry, paraelectric structure). Electrons accumulate at the upper electrode, and holes (or a depletion of electrons) at the lower electrode in order to screen the surface charge resulting from the ionic displacements. In fact, on each electrode, the accumulated charge per unit area is exactly equal to the polarization of the sample. So if we could measure the amount of charge accumulation we would have a direct measure of the polarization. But how can we do this? Well, next, imagine reversing the orientation of the polarization – for example by applying an external electric field in the down direction – to reach the configuration on the right. Now electrons accumulate at the lower electrode and holes at the upper electrode to achieve the screening. They achieve this by flowing through the external circuit connecting the two electrodes, where they can be counted by comparing the voltage across the series reference capacitor then using $Q=CV$! The amount of charge per unit area of electrode that flows during the transition is equal to the change in polarization between the up- and down-polarized states; the “absolute” value of polarization which is reported is half of this number.
Now, bearing in mind that what is measured in an experiment is a [*change*]{} in polarization, let’s go back to our cartoon one-dimensional model and make some sense out of this multi-valuedness business. In the upper part of Figure \[1Dchain\_polar\] we reproduce the non-polar one-dimensional chain of Fig. \[1Dchain\], and below it we show a similar chain in which the cations have been displaced by a distance $d$ relative to the anions in the manner of a ferroelectric distortion to create a polar system. Let’s repeat our earlier exercise of calculating the polarization using the two unit cells shown as the dashed rectangles.
![The upper panel reproduces the one-dimensional chain of alternating anions and cations of Fig. \[1Dchain\]. In the lower panel, the cations are displaced to the right by a distance $d$ relative to the anions, with the vertical dotted lines indicating their original positions.[]{data-label="1Dchain_polar"}](1Dchain_polar_small){width="1.0\columnwidth"}
In the left hand case, $$\begin{aligned}
p = \frac{d}{a} & = & \frac{1}{a} \sum_i q_i x_i \nonumber\\
& = & \frac{1}{a} \left(-1 \times \frac{a}{4} +1 \times (\frac{3a}{4} +d) \right) \nonumber\\
& = & \frac{1}{2} + \frac{d}{a} \end{aligned}$$ and in the right hand case, $$\begin{aligned}
p & = & \frac{1}{a} \left(+1 \times (\frac{a}{4} + d) -1 \times \frac{3a}{4} \right) \nonumber\\
& = & -\frac{1}{2} + \frac{d}{a} \quad .\end{aligned}$$ Again the two answers are different, but this time that doesn’t worry us, because we recognize that they differ by exactly one polarization quantum. Next comes the key point: Let’s calculate the [*change*]{} in polarization between the polar and non-polar chains using each unit cell as our basis. First for the cell on the left, $$\delta p = (\frac{1}{2} + \frac{d}{a} ) -\frac{1}{2} = \frac{d}{a}$$ and for the cell on the right $$\delta p = -(-\frac{1}{2} + \frac{d}{a} ) - (-\frac{1}{2}) = \frac{d}{a} \quad.$$ In both cases the change in polarization between polar and non-polar chains is the same. In fact this would have been the case whatever unit cell we had chosen to make the calculation. So, while the absolute value of polarization in a bulk, periodic system, is multivalued, the change in polarization – which remember is the quantity that can be measured in an experiment – is single valued and well defined. Phew.
Just to really drive the point home, in Fig. \[p\_of\_d\] we plot the polarization of the ideal one-dimensional ionic chain as a function of the displacement of the cations (as a fraction of the lattice constant) from their non-polar positions. As we calculated earlier, for zero displacement the polarization lattice is centrosymmetric and consists of all half-integer values (black circles). As the displacement increases, the polarization increases linearly and by the same amount along each branch of the polarization lattice (labeled by $n = -1, 0, 1$ etc.) The branches are always separated from each other by the same amount, the polarization quantum, which is equal to $1$ in this case. The dashed lines on the $n=1$ branch show that for a displacement of 0.25$a$, the polarization increases from 0.5 to 0.75, and so the change in polarization is 0.25. If the displacement is increased artificially to 0.5 – that is half of the unit cell – the ions end up on top of each other; in our thought experiment this causes the polarization to jump between branches of the polarization lattice, although in practice we would have achived nuclear fusion which would likely dominate the physics.
![Polarization as a function of the displacement, $d$, of the cations in the 1D chain of Figure \[1Dchain\_polar\]. The polarization lattice is zero-centered, and the branches are separated by the polarization quantum. Notice that the branches of the lattice run exactly parallel to each other, so that differences in polarization along each branch for the same displacement are identical.[]{data-label="p_of_d"}](p_of_d_small){width="0.7\columnwidth"}
Extension to three dimensions
-----------------------------
The one-dimensional example that we chose here for simplicity is not entirely without physical relevance; for example ferroelectric or polar polymers closely resemble one-dimensional chains. Such an application is discussed in Ref. [@Kudin/Car/Resta_2:2007], along with an excellent analysis of the development of an infinite chain from a finite one. In most cases, however, we are interested in three-dimensional systems, and fortunately the extension to three dimensions is straightforward: A polarization lattice can now be defined along all three lattice vector directions, with the polarization quantum equal to $$P_i = \frac{1}{\Omega} e R_i$$ Here $e$ is the electronic charge, $R_i$ is the $i$th lattice vector, and $\Omega$ is the unit cell volume. Note that in non-magnetic systems, the polarization quantum is usually multiplied by an additional factor of two because the up- and down-spin electrons are equivalent, and shifting an up-spin electron by a lattice vector also shifts the corresponding down-spin electron. Polarization values along cartesian coordinates, for example, can then be readily obtained using the appropriate linear transformation.
Wannier representation and Berry Phase {#section_WR_and_BP}
======================================
In the previous section we discussed the multivaluedness of the polarization in a bulk periodic solid and reconciled it with what can be measured experimentally for the simple example of an array of ions. Of course in a real solid, there is (thankfully) more chemistry to take care of. In this section we will first explain how this chemistry can be incorporated rather simply by extending the ionic model through the method of Wannier functions. (A similar approach is followed in Ref. [@Kudin/Car/Resta_1:2007], where an algorithm is developed that is particularly suited to localized-basis quantum chemistry codes.) Once we are comfortable with this conceptually we will move on to the real meat of the modern theory of polarization – the Berry phase method.
Remember that the Wannier function, $w_n(\mathbf{r})$, in unit cell $\mathbf{R}$ associated with band $n$ is defined as $$\begin{aligned}
w_n(\mathbf{r}-\mathbf{R}) & = &
\frac{\Omega}{(2\pi)^3} \int_{BZ} d^3 \mathbf{k} e^{-i \mathbf{k.R}} \Psi_{n \mathbf{k}}(\mathbf{r})\nonumber \\
& = &
\frac{\Omega}{(2\pi)^3} \int_{BZ} d^3 \mathbf{k} e^{i \mathbf{k.(r} - \mathbf{R})} u_{n \mathbf{k}}(\mathbf{r})\end{aligned}$$ where $\Psi_{n \mathbf{k}}(\mathbf{r})= e^{i\mathbf{k.r}} u_{n\mathbf{k}}(\mathbf{r})$ are the Bloch functions, written as usual in terms of the cell-periodic part, $u_{n\mathbf{k}}(\mathbf{r})$. Here $\Omega$ is the unit cell volume, and the integral is over the Brillouin zone.
Unlike the Bloch functions which are delocalized in space, the Wannier functions are localized. As a result they are often used in visualization of chemical bonding, as well as for basis sets in electronic structure calculations, where their minimal overlap can lead to favorable scaling with system size. They are relevant here, because their localized nature provides a convenient atomic-like description of the charge density in a solid: While we know in reality that the charge density in a solid is a continuous function, the localized picture will allow us to continue to calculate dipole moments by summing over charges multiplied by positions.
Let’s go back to our 1D chain, and relax the constraint that it is composed of point charge ions to give it some chemistry. If it’s helpful you could think of it as say a chain of Na$^{+}$ cations alternating with Cl$^{-}$ anions. In the following figures we associate pink with Na ions or electrons, and green with Cl ions or electrons. In Figure \[NaCl\_BandStructure\] (left) we show the molecular orbitals that would form between two such ions in an Na-Cl “molecule” – the lower energy, bonding orbital is occupied by two electrons and more localized on the $p$ orbital of the anion, and the higher energy, antibonding orbital is empty and consists primarily of cation $s$ character. The corresponding band structure cartoon is shown to the right; you can derive the dispersion using simple linear-combination-of-atomic-orbitals (LCAO) methods; see for example the book by Cox [@Cox:Book]. In Figure \[1Dchain\_WF\] we show a cartoon of our 1D chain again, but this time we have separated out the charge on the ions (all of which are +1, and which we continue to treat as point charges) from the charge on the electrons which are spread through the system, but piled up more on the anions than the cations. The blobs around the anions illustrate what we might expect the Wannier functions of the occupied band to look like, with each Wannier function containing two electrons. The character of the Wannier function is mostly Cl $p$-like, with a little bit of Na $s$ character, indicated by the slight pink tinge on the edges. Note that if we consider both of the electrons in each Wannier function to be associated with the Cl ion, then the formal charge on the Cl is +1 (the ionic charge) -2 = -1, and that on the Na ion is +1 +0 = +1, and we recover our simple ionic model of Figs. \[1Dchain\] and \[1Dchain\_polar\].
![Left: The molecular orbitals formed in an Na-Cl “molecule”. Right: Band structure of a one-dimensional Na-Cl chain. The valence band is derived from Cl-like molecular orbitals each containing two electrons, and is fully occupied; the Na-like conduction band is empty. In both cases pink represents Na-derived states and green Cl-derived states. []{data-label="NaCl_BandStructure"}](NaCl_BandStructure_small){width="1.0\columnwidth"}
![One-dimensional chain of alternating cations (pink postively charged ion cores) and anions (green positively charged ion cores with their associated negatively charged valence electron cloud). The dimensions and dashed unit cells are as in Figure \[1Dchain\].[]{data-label="1Dchain_WF"}](1Dchain_WF_small){width="1.0\columnwidth"}
How should we now calculate the polarization of the chain? We would like again to reduce our polarization integral again to a sum over localized charges multiplied by their positions. This is straightforward for the ions which we are still treating as point charges. For the electrons, it turns out that this procedure will work too. Since the Wannier functions are localized, we work out the average position of the electrons in the Wannier function, and treat them all as sitting at that point. This “position” of the Wannier function is called the [*Wannier center*]{}, $\overline{\mathbf{r}}_n$. The Wannier center associated with band $n$ is defined to be the expectation value of the position operator $\mathbf{r}$ for Wannier function $w_n(\mathbf{r})$: $$\overline{\mathbf{r}}_n=\int w_n^*(\mathbf{r}) \mathbf{r} w_n(\mathbf{r}) d^3 \mathbf{r}
\label{Eq_Wannier_Center_r}$$ Later we will find it useful to rewrite this expression in terms of the periodic cell functions using the momentum representation of the position operator $\mathbf{r} = -i \frac{\partial}{\partial \mathbf{k}}$: $$\overline{\mathbf{r}}_n=i\frac{\Omega}{(2\pi)^3} \int_{BZ}d^3\mathbf{k}
e^{-i \mathbf{k}.\mathbf{R}} \left< u_{n \mathbf{k}} | \frac{\partial u_{n \mathbf{k}}}
{\partial \mathbf{k}} \right>
\label{Eq_Wannier_Center_k}$$ You can spend your next free Sunday morning showing that Eqns. \[Eq\_Wannier\_Center\_r\] and \[Eq\_Wannier\_Center\_k\] are equivalent, take my word for it, or follow the derivation by Blount in Ref. [@Blount:1962]
With this concept of the Wannier center, the expression for polarization that we used previously for the ionic chain extends simply to a sum over the contribution from the point charge ions, plus a sum over the electronic charges centered at the Wannier centers of each occupied Wannier function, $n$: $$p = \frac{1}{a} \left(\sum_i (q_i x_i)^{ions} + \sum_n^{occ} (q_n \overline{\mathbf{r}}_{n})^{WFs} \right) \\
\label{PofWF}$$ Let’s try it for the case of the left-hand unit cell in our 1D chain. In the non-polar case, we can see by symmetry that the Wannier center is at the same position as the green anion; remember now also that the charge on all of the ions is +1, and that each Wannier function contains two electrons. So the dipole moment per unit length in the left unit cell is $$\begin{aligned}
p & = & \frac{1}{a} \left(+1 \times \frac{a}{4} +1 \times \frac{3a}{4} + -2 \times \frac{a}{4} \right) \nonumber\\
& = & \frac{1}{a} \frac{2a}{4}\nonumber\\
& = & \frac{1}{2} \quad .\end{aligned}$$ The same result as we obtained previously! This is as expected – the allowed values of the polarization lattice for a centrosymmetric structure are dictated by the symmetry of the crystal and the ionic charges, and are not modified by factors such as the details of the chemical bonding within the material.
Now let’s think about the off-centered case, in the lower part of Fig. \[1Dchain\_WF\]. As before, the cations have moved a distance $d$ to the right, but this time the Wannier centers have also moved – by a distance $\Delta$ say – to the left. This occurs as the chemical bond between the near neighbor anion-cation pairs becomes stronger, and develops more cation $s$ character, whereas that between the distant neighbor pairs weakens; you can think of it as a flow of electrons from the anion (which previously had all of the valence electrons) towards the cation in the process of covalent bond formation. Let’s see what this additional covalency does to the polarization: $$\begin{aligned}
p & = & \frac{1}{a} \left(+1 \frac{a}{4} +1 (\frac{3a}{4} + d) +
-2 (\frac{a}{4} - \Delta) \right) \nonumber\\
& = & \frac{1}{a} (\frac{2a}{4} + d + 2\Delta)\nonumber\\
& = & \frac{1}{2} +\frac{d}{a} +2\frac{\Delta}{a}\quad .\end{aligned}$$ Compared to the purely ionic case, the polarization has increased by an amount $\frac{2\Delta}{a}$. This is because, in addition to the positively charged cation moving to the right (along the positive $x$ axis) some negatively charged electron density has moved to the left (along the negative $x$ axis). This results in a larger [*effective*]{} displacement of positive charge along $+x$ and a larger polarization. We’ll return to this picture later when we discuss the concept of the Born effective charge.
Let’s summarize the discussion so far before we go on to formalize it mathematically. Using our Wannier function picture we can continue to write our polarization as the sum over the charges times their positions. We include both the contribution from the postively charged ion cores, and the contribution from the negatively charged valence electrons, and we take the “position” of each valence electron to be its Wannier center. Had we repeated our analysis for the right-hand unit cell, we would have seen that, as in the purely ionic model, the polarization is multi-valued, but the difference in polarization for example between a centrosymmetric and polar structure, is well-defined, and corresponds to the experimentally measurable spontaneous polarization.
Now we will derive the formal mathematical expression for the spontaneous polarization $\delta p$ in the Wannier representation. Since we already have an expression for the Wannier centers this is going to be rather painless. Remember that the spontaneous polarization is the difference in polarization, on the same branch of the polarization lattice, between the final, polarized and initial, unpolarized states. Using $p = \frac{1}{\Omega} \sum_i q_i \mathbf{r}_i$ for the ionic part, and Eqn. \[Eq\_Wannier\_Center\_k\] for the Wannier centers, we obtain: $$\begin{aligned}
\delta p & = & p^f - p^0 \nonumber \\
& = & \frac{1}{\Omega} \sum_i \left[ q_i^f \mathbf{r}_i^f - q_i^0 \mathbf{r}_i^0 \right] \nonumber \\
& - & \frac{2ie}{(2\pi)^3} \sum_n^{occ} \left[ \int_{BZ} d^3 \mathbf{k} e^{-i\mathbf{k.R}}
\left< u_{n \mathbf{k}}^f| \frac{\partial u_{n \mathbf{k}}^f}{\partial \mathbf{k}} \right>
-\left< u_{n \mathbf{k}}^0| \frac{\partial u_{n \mathbf{k}}^0}{\partial \mathbf{k}} \right>
\right]
\label{P_BP}\end{aligned}$$ where $f$ and $0$ indicate the final (polar) and intial (high symmetry) positions/wavefunctions. Since the wavefunctions, at least at the Kohn-Sham level, are a direct output of standard electronic structure codes, Eqn. \[P\_BP\] can be used to evaluate the polarization with only a small extension to a standard density functional theory code. (A rigorous extension to correlated, many-body wavefunctions also exists, see for example Refs. [@Resta:1999] and [@Resta:1998].) Notice of course, that the issues discussed earlier about multivaluedness of the polarization and the polarization lattice persist here, and in taking the difference in Eqn. \[P\_BP\] one must be careful to remain on the same branch of the polarization lattice.
If you are familiar with the concept of the [*Berry phase*]{} [@Berry:1984] and its extension to periodic solids [@Zak:1989] you will recognize the integrals in Eqn. \[P\_BP\] to be the Berry phase developed by the wavefunction $u_{n \mathbf{k}}$ as it evolves along the path $\mathbf{k}$. As a result, the formalism for calculating polarization using this method is often called the [*Berry phase theory of polarization*]{}. Refs. [@King-Smith/Vanderbilt:1993; @King-Smith/Vanderbilt:1994; @Resta:1993; @Resta:1994] are the original papers providing the detailed derivations of the Berry phase formalism, and excellent reviews can be found in Refs. [@Resta:1996; @Martin:Book; @Resta/Vanderbilt:2007]. If you find the Berry phase concept too frightening, however, just stick with the Wannier function ideas, and regard Eqn. \[P\_BP\] as a tool that we’ll see in Section \[nittygritties\] allows for convenient computation.
Subtlety – gauge transformation!
--------------------------------
Those of you who have managed to stay awake and alert to this point might raise an objection: Since the Bloch functions are defined only to within a phase factor, i.e. $$\Psi_{n \mathbf{k}}(\mathbf{r}) \rightarrow e^{i\phi(\mathbf{k})} \Psi_{n\mathbf{k}}(\mathbf{r})$$ without changing any physically meaningful quantities, the Wannier functions are not uniquely defined! As a result, the Wannier centers, which we have just seen are crucial in defining the polarization, are also not uniquely defined. We are saved, however, by the fact that the sum over the Wannier centers in any given unit cell [*is*]{} uniquely defined, and looking again at Eqn. \[PofWF\] we find that this is in fact the quantity that matters in defining the polarization. In practice, special choices of Wannier functions are often made in calculations of the polarization. The so-called [*Maximally-localized Wannier functions*]{} in which the phases of the Bloch functions are chosen so as to minimize the sum of the mean squares of the positional spread [@Marzari/Vanderbilt:1997] are particularly popular.
The concept of Born Effective Charge
====================================
At this stage I think it’s appropriate to formally introduce the [*Born effective charge*]{}, which is a quantity that is very useful conceptually in thinking about ferroelectric polarization. In fact we have already seen the main idea, in Section \[section\_WR\_and\_BP\], where we saw that the polarization resulting from the displacement of an ion could be different from that expected by multiplying its formal charge times its displacement, in the case when the Wannier center(s) move by a different amount than the ion cores. In fact in the example of Fig. \[1Dchain\_WF\], as the positive cations moved to the right, the Wannier centers shifted to the left, resulting in a larger overall polarization than we would have expected from the formal charges alone. We say that the [*effective charges*]{} on the ions – the amount of charge that effectively contributes to the polarization during the displacement – is larger than the formal charge.
This is formalized in the concept of the Born effective charge, $Z^*$, which is defined as the change in polarization divided by the amount that an ion (or rather the periodic sub-lattice of equivalent ions) is displaced: $$Z^*_{ij} = \frac{\Omega}{e}\frac{\delta P_i}{\delta d_j} \quad.$$ The Born effective charge is a tensor: When an ionic sublattice is displaced in direction $i$, there is of course a change in polarization along the displacement direction, but in addition, the polarization in perpendicular directions, $j$, can change. Turning this expression around we can see immediately what we have been discussing qualitatively – that the change in polarization is determined by these [*effective charges*]{} times their displacements, not by the formal charges: $$\partial P_i = \frac{e}{\Omega} Z^*_{ij} \delta d_j$$ The total polarization is then obtained by summing over the contributions from the displacements of all sublattices.
In materials that are ferroelectric, or that are close to a ferroelectric phase transition, the Born effective charges tend to be anomalously large, particularly on the atoms that displace the furthest from their high symmetry to their ferroelectric positions. For example in the prototypical ferroelectric PbTiO$_3$, in which the formal charges are Pb $+2$, Ti $+4$ and O $2-$, the effective charges on the ions that are active during the ferroelectric phase transition are Pb $+3.9$, Ti $+7.1$ and O $-5.8$ [@Ghosez/Michenaud/Gonze:1998]. This is consistent, with the alternative, equivalent definition of the Born effective charge as the force induced on an ion by a uniform small electric field, $E$: $$Z^*_{ij} = -e\frac{\delta F_i}{\delta E_j} \quad.$$ In highly polarizable ferroelectrics, small electric fields generate large forces on the ions, mediated by the anomalously large Born effective charges.
Lastly, I want to emphasize that it is important to distinguish between the Born effective charge, which is a well-defined [*dynamical*]{} and [*measurable*]{} quantity, and the [*formal, static*]{} charge on an ion. The latter quantity, which reflects the number of electrons sitting at a particular ion site, depends on how you “count”, since there is not a unique way of deciding how to apportion the electrons in a chemical bond to one ion or another. While the static charge indeed indicates a measure of the amount of covalency in a compound, it is not a good indicator of ferroelectricity, which is rather indicated by a [*change*]{} in covalency during ionic displacement.
A few tips on getting a Berry Phase calculation to work {#nittygritties}
=======================================================
Finally we describe a few of the tricks and foibles that we have learned through (sometimes) bitter experience are needed to make a Berry phase calculation of the polarization both run and give the correct answer. We try to keep our comments general – for the specifics of a particular code refer to the relevant manual.
The first step is of course to calculate the structure (if required) and self-consistent charge density, as in any standard total energy calculation. Of course the charge density should be well-converged with respect to the energy cutoff and $\mathbf{k}$-point sampling. In addition, if one is interested in systems such as improper ferroelectrics with small polarization values [@Kimura_et_al_Nature:2003; @Malashevich/Vanderbilt:2008], the ionic positions must be obtained with higher-than-usual accuracy. An extra subtlety is to check that the system is insulating at every point in $\mathbf{k}$-space, otherwise the Berry phase is ill-defined. The relaxed ionic positions and self-consistent charge density are then used as an input to the Berry phase calculation.
One then proceeds to calculate one of the Berry phase values on the right-hand side of Eqn. \[P\_BP\], that is $$\sum_n^{occ} \int_{BZ} d^3 \mathbf{k} e^{-i\mathbf{k.R}}
\left< u_{n \mathbf{k}}| \frac{\partial u_{n \mathbf{k}}}{\partial \mathbf{k}} \right>
\quad ,$$ where $u_{n \mathbf{k}}$ is the cell part of the Bloch function for the structure we are considering. First, the matrix elements are calculated by integrating along strings of $k$-points. Since $ \frac{\partial}{\partial \mathbf{k}}$ is a vector derivative the matrix elements should be computed along any three non-collinear directions; usually the lattice vectors are chosen. Then multiple strings in a particular direction are sampled so that an integration over the Brillouin zone can be performed (see Figure \[string\_integration\]). It’s important to check convergence both with respect to the number of $k$-points along a string, and the number of strings used in the sampling, as the requirements can be quite different in each case. Finally the values for all bands $n$ are summed. One subtletly, which is sometimes not well taken care of in codes, concerns the procedure for averaging the Berry phase over the Brillouin zone. This is usually done by taking the sum of the Berry phase values at each $k$-point, weighted by the fractional contribution of the $k$-point. This procedure works well provided that the value from each $k$-point is on the same branch of the polarization lattice. Figure \[kptaverage\] illustrates a not-uncommon problem that can occur with some codes during the averaging procedure. Here by inspection the average Berry phase is clearly close to $\pi$, modulo the phase quantum of $2\pi$. Taking a simple average of the values mapped into the range between $\pm \pi$, however, would result in an incorrect value close to zero. We recommend checking the values of Berry phase obtained for the individual strings if your code performs an automatic averaging procedure!
For a spin-polarized system, the Berry phase calculation is performed for both up- and down-spin electrons separately; the phases are converted into polarization units by multiplying by $-\frac{ie}{2\pi^3}$ and then added to the ionic contribution $\frac{1}{\Omega}\Sigma_i q_i \mathbf{r}_i$, where $q_i$ is the charge of the pseudopotential or ion, to obtain the total polarization of the system along the chosen lattice vector.
![Choice of the $\mathbf{k}$-point grid for a Berry phase calculation of the polarization. Here the polarization is to be calculated along the $z$-direction. The integration to obtain the Berry phase is carried out along 4 strings of $\mathbf{k}$ points centered around $\Gamma$ in the $\mathbf{k}_x - \mathbf{k}_y$ plane, with 6 sampling points along each string in the $\mathbf{k}_z$ direction. The final Berry phase is obtained by averaging the values obtained from each of the four strings. []{data-label="string_integration"}](string_integration_small){width="0.4\columnwidth"}
![The black dots show the values of Berry phase obtained by integration along six $\mathbf{k}$-point strings in the Brillouin zone. Clearly the average value is close to $\pi$, $\pm$ the Berry phase quantum of $2\pi$. Mapping all of the values into the lowest phase branch then taking the simple average would lead to an incorrect result close to zero. []{data-label="kptaverage"}](kptaverage_small){width="0.8\columnwidth"}
The above procedure is repeated for each lattice vector in turn. Be careful to check in the output of your code whether the results are reported with respect to the lattice vectors or in cartesian coordinates!
Remember that the number that you have now calculated is the absolute value of the polarization, and is only defined modulo a polarization quantum. To calculate the spontaneous polarization in a ferroelectric for example, the procedure should be repeated also for a high symmetry, non-polar reference state. The difference between the two values, taken along the same branch of the polarization lattice, is then the spontaneous polarization. Sometimes it is necessary to re-calculate the polarization for a number of structures along the deformation path between the high- and low-symmetry structures in order to know unambiguously which difference to take. For example, Fig. \[BiFeO3\] shows the calculated polarization values for the case of perovskite structure BiFeO$_3$, one of the most well-studied multiferroic materials[@Neaton_et_al:2005]. Notice first that the polarization lattice for the non-polar structure, labeled with 0% distortion, does not contain zero, but is centered around 92.8 $\mu$C cm$^{-2}$, which is half a polarization quantum. It is clear from following the evolution of the polarization with distortion that the correct value for the spontaneous polarization is 187.8 - 92.8 = 95.0 $\mu$C cm$^{-2}$. From a calculation of only the end-points at the $R3c$ and $-R3c$ structures the appropriate difference to take would be unclear, and one might incorrectly assume a value of $\frac{1}{2}(2.3 - (-2.3)) =
2.3 \mu$C cm$^{-2}$.
![Calculated polarization as a function of percentage distortion from the high symmetry non-polar structure (0% distortion) to the ground state $R3c$ structure for perovskite BiFeO$_3$. The black dots are calculated points and the dashed lines are a guide to the eye illustrating the evolution along branches of the polarization lattice. From Ref. [@Neaton_et_al:2005]. []{data-label="BiFeO3"}](BiFeO3_small){width="0.9\columnwidth"}
Finally a hint for calculating Born Effective Charges. Since these are defined as derivatives, in principle the polarization should be calculated for the structure of interest, and then again for an infinitesimally small displacement of each ion in turn. In practice, however, if the diplacement is too small the result from this approach can be noisy. The best plan is to plot polarization as a function of ionic displacement, starting with very small displacement values, and to take the slope of the line in the region beyond the noise but before the non-linear regime.
Last words
==========
I hope that this article has taken away some of the mystique associated with the modern theory of polarization, and motivated you to start making your own calculations of spontaneous polarization and related dielectric properties. For more practical introductory help, I recommend working through the tutorials that accompany many of the electronic structure computational packages. For example the [*Lesson on polarization and finite electric field*]{} provided by the ABINIT code, [www.abinit.org]{}, is particularly helpful. Or even better, attend a hands-on course hosted by one of the public codes where you will have direct access to leading experts in the field. Good luck!
Acknowledgements
================
My thanks to the pioneers of the Modern Theory of Polarization – Raffaele Resta and David Vanderbilt – who helped me to understand their elegant theory, as well as to the many students who have allowed me to impose my explanations upon them and in turn improve my description. The preparation of the manuscript was supported by ETH Zürich.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
We investigate the observational signatures and physical origin of ram-pressure stripping (RPS) in 63 massive galaxy clusters at $z=0.3-0.7$, based on images obtained with the Hubble Space Telescope. Using a training set of a dozen “jellyfish" galaxies identified earlier in the same imaging data, we define morphological criteria to select 211 additional, less obvious cases of RPS. Spectroscopic follow-up observations of 124 candidates so far confirmed 53 as cluster members. For the brightest and most favourably aligned systems we visually derive estimates of the projected direction of motion based on the orientation of apparent compression shocks and debris trails.
Our findings suggest that the onset of these events occurs primarily at large distances from the cluster core ($>400$ kpc), and that the trajectories of the affected galaxies feature high impact parameters. Simple models show that such trajectories are highly improbable for galaxy infall along filaments but common for infall at high velocities, even after observational biases are accounted for, provided the duration of the resulting RPS events is $\lesssim$500 Myr. We thus tentatively conclude that extreme RPS events are preferentially triggered by cluster mergers, an interpretation that is supported by the disturbed dynamical state of many of the host clusters. This hypothesis implies that extreme RPS might occur also near the cores of merging poor clusters or even merging groups of galaxies.
Finally, we present nine additional “jellyfish" galaxies at z$>$0.3 discovered by us, thereby doubling the number of such systems known at intermediate redshift.
author:
- |
Conor McPartland$^1$, Harald Ebeling$^1$, Elke Roediger$^2$ & Kelly Blumenthal$^1$\
$^1$ Institute for Astronomy, University of Hawai’i at Manoa, 2680 Woodlawn Drive, Honolulu, HI, 96822, USA\
$^2$ E.A. Milne Centre for Astrophysics, Department of Physics & Mathematics, University of Hull, Cottinton Road, Hull, HU6 7RX, United Kingdom
bibliography:
- 'jellies\_paper.bib'
date: Draft version
title: 'Jellyfish: The origin and distribution of extreme ram-pressure stripping events in massive galaxy clusters'
---
galaxies: evolution - galaxies: clusters: intracluster medium - galaxies: structure
INTRODUCTION
============
Evidence of accelerated galaxy evolution in galaxy clusters has been presented as early as 1980, the most well known examples being the increased occurrence of ellipticals in dense environments [i.e., the morphology-density relation; @dressler_1980] and the higher fraction of blue galaxies in clusters at higher redshift [i.e., the Butcher-Oemler effect, @butcher_1984]. The physical mechanisms responsible for these effects are, however, still very much debated. A variety of processes have been proposed in the literature, ranging form slow-acting gravitational interactions such as galaxy-galaxy harassment [@moore_1996] to potentially extremely rapid galaxy transformations brought about by interactions with the gaseous intracluster medium (ICM).
The latter process, ram-pressure stripping (RPS) is expected to be especially efficient in massive galaxy clusters, as the pressure imparted on a galaxy is directly proportional to the local gas density of the ICM and to the square of the galaxy’s velocity with respect to the ICM [@gunn_infall_1972]. The resulting removal of the galaxy’s interstellar medium (ISM) occurs in the direction of motion of the galaxy relative to the ICM, generating a trail of star-forming regions in the galaxy’s wake. For fortuitous viewing angles, this trail, or at least the associated deformation of the galactic disk, is accessible to observation, thus creating a rare opportunity to constrain the motion of galaxies in the plane of the sky. Observations of RPS events thus constitute a valuable complement to spectroscopic radial-velocity surveys and permit a detailed investigation of the kinematics and spatial evolution of galaxies in the dense cluster environment.
The physics and observational signature of RPS have been the subject of extensive numerical simulations which predict that gradual stripping should be pervasive even in low-mass clusters [@vollmer_2001]. Indeed RPS events have been studied in great detail in the Virgo [@chung_2007; @vollmer_influence_2012; @abramson_caught_2011] and Coma clusters [@smith_2010; @yagi_dozen_2010], as well as in other nearby systems, such as the Shapley Concentration [@merluzzi_2013] or Abell 3627 [@sun_2007; @fumagalli_2014]. As expected, these events are relatively modest though, with observations showing atomic hydrogen to be displaced and only partially removed [@scott_2010], while the denser, more centrally located molecular gas is found to be essentially unperturbed [@boselli_1997; @vollmer_2001]. By contrast, in the most massive clusters the environment encountered by infalling galaxies can lead to their entire gas reservoir being stripped in a single pass through the cluster core [e.g. @takeda_1984; @abadi_1999; @kapferer_2009; @steinhauser_2012]. Observational evidence of extreme ram-pressure stripping is, however, sparse, due to their reliance on favourable circumstances, such as suitable infall trajectory, gas mass, galaxy orientation, and high ICM density. Considering the small number and relatively low masses of nearby clusters (except for Coma), these conditions are unlikely to be met in the local Universe.
The extreme environment that is a prerequisite for extreme RPS is, however, routinely encountered by galaxies falling into massive clusters where galaxy peculiar velocities in excess of 1000 km s^-1^ are common and the ICM particle density easily exceeds 10^-3^ cm^-3^. Since massive clusters are rare, larger volumes have to be searched to efficiently probe such truly high-density environments. Although their numbers are still small, striking examples of extreme RPS events have been discovered in *Hubble Space Telescope* (*HST*) images of moderately distant ($z\gtrsim0.2$) massive clusters [@owen_2006; @cortese_2007; @owers_2012] and, most recently, in X-ray selected massive clusters at $z{>}0.3$ [@ESE see Fig. 1]. Importantly, these clusters are not only intrinsically more massive, they are also dynamically less evolved and more likely to be undergoing mergers than systems in the local Universe [@mann_2012], a critical requirement if extreme RPS events are triggered by merger-driven shocks, as suggested by @owers_2012. Increasing the size of the still small sample of RPS examples clearly constitutes a crucial step toward a meaningful statistical investigation of the physics of accelerated galaxy evolution.
In this paper, we aim to compile a statistically significant sample of galaxies that might be undergoing RPS in very massive clusters. We then use this sample to establish which galaxy trajectories are most conducive to creating extreme RPS, and thereby elucidate whether the most dramatic RPS events are triggered by massive cluster mergers [@owers_2012], rather than during regular infall of galaxies from the field or along filaments. In order to compile the required sample, we develop morphological criteria to select RPS candidates from archival *HST* imaging data for a well defined sample of massive clusters at $z>0.3$, and compare the spatial and dynamical distribution of the selected objects with expectations from numerical and theoretical models.
This paper is structured as follows: in Section \[sec:data\] we introduce the cluster sample and present an overview of the observations and data-reduction procedures; in Section \[sec:morph\] we discuss our morphological criteria for the identification of galaxies experiencing ram-pressure stripping and present the sample of RPS candidates; in Section \[sec:models\] we present the a simple model of clustre infall which we use to interpret our data; in Section 5 we present our results for the spatial distribution and dynamical properties of RPS events in massive clusters; and in Section 6 we draw conclusions about the origin, trajectories, and physics of extreme RPS. We present a summary of our work in Section 7.
Throughout this paper, we assume a concordance $\Lambda$CDM cosmology with $\Omega_M$ = 0.3, $\Omega_\Lambda$ = 0.7, H$_0$ = 70 km s^-1^ Mpc^-1^. As the clusters in our sample span a range of redshifts of $0.3<z<0.7$, the metric scale of our images varies from 4.45 to 7.15 kpc arcsec^-1^.
{width="30.00000%"} {width="30.00000%"} {width="30.00000%"}
{width="30.00000%"} {width="30.00000%"} {width="30.00000%"}
{width="30.00000%"} {width="30.00000%"} {width="30.00000%"}
{width="30.00000%"} {width="30.00000%"} {width="30.00000%"}
{width="35.00000%"} {width="35.00000%"} {width="35.00000%"}
Data used in this study {#sec:data}
=======================
The MACS sample
---------------
Our cluster sample is drawn from a master list of clusters identified in the course of the Massive Cluster Survey [MACS; @ebeling_2001; @ebeling_2007; @ebeling_2010; @mann_2012], designed to provide a large, statistically complete sample of X-ray luminous ($L_X \gtrsim 5\times 10^{44}$ erg s^-1^, 0.1-2.4 keV) and moderately distant ($z \gtrsim 0.3$) galaxy clusters. Covering over 22,000 sq.deg., the MACS sample comprises the majority of massive galaxy clusters in the observable Universe, making it ideally suited for our investigation. At redshifts $z\gtrsim0.3$, the sub-kiloparsec angular resolution needed to identify the characteristic morphological traits of RPS events can only be achieved with the Advanced Camera for Surveys (ACS) aboard *HST*. We thus limit our sample to MACS clusters with archival *HST*/ACS images as described in more detail in the following section.
Imaging data
------------
As our primary observational diagnostics revolve around morphological features traced by star-forming regions, we limit our study to MACS clusters that have been observed in the *HST*/ACS F606W band. The F606W filter is well suited as it corresponds roughly to the B band in the cluster rest frame and has been used in a large number of *HST* observations of MACS clusters. We further require clusters in our sample to also have imaging data in the ACS F814W passband, as the resulting F814W–F606W colours provide a straightforward means to discriminate against the population of passively evolving cluster ellipticals.
Of the entire MACS sample, 44 clusters were successfully observed in both the ACS F606W and F814W passbands as part of the *HST* SNAPshot programmes GO-10491, -10875, -12166, and -12884 (PI: Ebeling). These programmes use short exposures (1200 seconds for F606W and 1440 seconds for F814W) designed to reveal bright strong-lensing features and provide constraints on the physical nature of galaxy-galaxy and galaxy-gas interactions in cluster cores. Fundamental properties of this subset of the MACS cluster sample are presented and discussed by Ebeling & Repp (in preparation). Supplementing these SNAPshots, we also include data from observations of 17 additional MACS clusters obtained by the Cluster Lensing and Supernova Survey with Hubble [CLASH; @CLASH], an *HST* Multi-Cycle Treasury Program employing 16 filters from the UV to the NIR, including F606W and F814W. Exposure times for the CLASH observations are nominally one and two orbits for all ACS filters, but vary substantially between cluster fields around median exposure times of 4060 and 8480 seconds for the F606W and F814W passbands, respectively (see Table \[tab:snaps\_obs\] & \[tab:clash\_obs\] for a summary of the observations).
In total, our sample thus comprises 63 MACS clusters. At the redshifts relevant to our study, the field of view of the ACS Wide Field Channel ($202^{\prime\prime}\times202^{\prime\prime}$) covers an inscribed circle of radius between 450 and 720 kpc and thus samples primarily the cluster core region. Charge-transfer-efficiency corrected images in the two passbands were registered using the astrometric solution of the F606W image as a reference, and source catalogs were created using SExtractor [@sextractor] in dual-image mode, with F606W chosen as the detection band. We removed stars as well as cosmic rays and other artefacts as objects falling on or below the star lines in both magnitude-$\mu_\mathrm{max}$ and magnitude-$r_{20\%, \mathrm{ell}}$ space[^1]. After removing spurious detections, we have a 5$\sigma$ 90% completeness limiting magnitude of 24.9 in F606W (here and in the following the magnitudes quoted are measured within the Petrosian radius).
As the quantitative morphological indicators we employ to identify RPS candidates (see Section \[sec:morph\]) require signal-to-noise ratios of $\langle$S/N$\rangle{>}5$ per pixel, we limit our galaxy sample to objects with $m_{\rm F606W}{<}24$, which leaves a total of 15,875 galaxies (11,550 in the SNAPshot data and 4,325 in the CLASH data). We note that, due to the high density of objects in cluster cores and the presence of objects of complex morphology, some of the objects in our master catalogue may in fact be blends of several objects, whereas others have suffered fragmentation, i.e., were broken up into multiple sources.
To mitigate the effect of fragmentation in our master catalog, we enforce strict deblending criteria (DEBLEND\_NTHRESH=16, DEBLEND\_MINCONT=0.2). Due to the relatively shallow depth ($\sim$1200 s) of the imaging data, the faint extraplanar tails that characterize jellyfish galaxies often fall below our detection limit. For the quantitative selection criteria (see Section 3.1), we, therefore, focus on identifying robust morphological features (disturbances) in the high signal to noise regions of galaxies. However, note that the presence of optical tails is a requirement for an object to be classified as a compelling jellyfish candidate during our visual screeening process.
As for the completeness of the sample of candidates presented here, it is almost certain that modest cases of RPS (in particular when occuring in low mass galaxies) will have been missed due the lack of pronounced morphological features, whereas essentially all the brightest objects would have been easily identified by eye. We note however that regarldless of brighness, objects moving close to our line of sight are likely to be missed as the tell-tale debris trails will be obscured by the the much brighter disks of the galaxies. We discuss this bias in detail in Section \[sec:bias\] and \[sec:bias2\].
Spectroscopic data
------------------
The sample of RPS candidates compiled in this work using morphological selection is expected to be heavily contaminated by galaxies that are in fact not members of the respective MACS cluster and / or whose morphology is irregular for reasons other than RPS (see Section \[sec:morph\] for details). In order to eliminate interlopers, we have embarked on a comprehensive spectroscopic survey of our RPS candidate sample, aimed at (a) excluding fore- and background galaxies from our sample of RPS candidates, and (b) obtaining peculiar radial velocities of those systems that are cluster members. We refer to a forthcoming paper (Blumenthal et al., in preparation) for a more extensive report on these efforts, including a description of the data-reduction procedure. We note here though that all spectroscopic observations were conducted with the DEIMOS spectrograph on the Keck-II 10m-telescope on Maunakea, using multi-object spectroscopy with slits of 1mm width, the 600 l/mm Zerodur grating set to a central wavelength of 6300Å, the GG455 blocking filter, and exposure times ranging from 3$\times$10 to 3$\times$15 minutes. For almost all targeted galaxies, redshifts were measured from emission lines detected in these spectra, yielding a precision of approximately 0.0002 in redshift or 60 km s$^{-1}$ in radial velocity.
GALAXY MORPHOLOGY {#sec:morph}
=================
A recent study by @ESE [hereafter ESE] presented six textbook examples of “jellyfish" galaxies (thought to be extreme RPS events[^2]) discovered in *HST* imaging data for 36 of the 63 clusters used in this work. These objects were visually identified, having to meet the following criteria: (1) a strongly disturbed morphology in optical images indicative of unilateral external forces; (2) a pronounced brightness and colour gradient suggesting extensive triggered star formation; (3) compelling evidence of a debris trail. Furthermore, the direction of motion implied by each of these features had to be consistent. We expand the ESE sample by six additional, unpublished, jellyfish candidates, identified by the same authors, that satisfy at least two of these criteria[^3], and use the resulting superset of 12 objects (shown in Fig \[fig:ese\]) as a training set for the identification of additional, less obvious candidate objects.
For each of the galaxies in our catalogue we compute several non-parametric galaxy morphology statistics defined previously in the literature: concentration ($C$) and asymmetry ($A$) [@bershady_2000; @CAS], Gini coefficient ($G$) and $M_{20}$ [@abraham_2003; @lotz_2004]. While these statistics were originally designed to identify the morphological features of galaxy mergers, we find that they can be applied more widely to characterise and select objects featuring disturbed morphologies. In addition to the aforementioned four statistics, we introduce two “skeletal decomposition" parameters ($Sk_{0-1}$ and $Sk_{1-2}$; see Appendix \[sec:skel\]).
We compute values for each of these indicators using the ellipticities, position angles, and locations provided by SExtractor but note that the precise location of the centre of each object is iteratively refined through minimisation procedures, as described in @lotz_2004. Acknowledging the difficulty of cleanly separating galaxies in crowded cluster cores, we resort to using SExtractor’s segmentation maps to identify the pixels belonging to a given galaxy rather than relying on an isophotal definition of a galaxy’s extent. We stress that, as a result, the morphological quantities measured here should not be directly compared to those from other work.
![Venn diagram of the sets of galaxies selected by each of the morphological criteria shown in the three panels of Fig. \[fig:cuts\]. Although each type of cut selects a similar number of galaxies (represented by the area of each circle), the modest overlap between these sets makes the final selection, achieved by requiring all criteria to be met, much more restrictive. \[fig:venn\]](jelly-venn.pdf){width="30.00000%"}
Selection criteria and visual screening
---------------------------------------
The fact that the extended ESE sample (Fig. \[fig:ese\]) contains some of the most extreme examples of jellyfish galaxies known to date (i.e., the brightest and most morphologically disturbed) makes it well suited as a training set for an iterative, semi-automated search for additional RPS candidates. To this end, we examine the location of the training-set members in $C$–$A$, Gini–$M_{\rm 20}$, and $Sk_{0-1}$–$Sk_{1-2}$ space, and define cuts in these parameter spaces that preserve the training set but eliminate the vast majority of other galaxies. The physical rationale behind these cuts is to discard extremely diffuse objects (achieved by a cut in $C$), almost perfectly symmetric sources (cut in $A$), morphologically undisturbed disk and elliptical galaxies (cut in $G$-$M_{20}$), and, finally, objects with little substructure (cuts in $Sk_{0-1}$ and $Sk_{1-2}$).
We apply an initial set of morphological criteria (cuts in $C$–$A$ and Gini–$M_{20}$) to galaxies detected in the 10 cluster fields from which the extended ESE sample originates. The $\sim$650 candidate objects thus selected are then visually scrutinised independently by two of us (CM and HE) and classified according to their plausibility as RPS events. We attempt to reduce the subjectivity of this procedure by reviewing jointly, in a second iteration, all objects classified either as compelling jellyfish galaxies or as plausible candidates by one of the inspectors and assigning a consensus classification. From the resulting set of potential RPS events we select the most compelling candidates, add them to our original training set, and re-evaluate our initial morphological constraints. Cuts in colour-magnitude space were considered too during this process but ultimately dismissed as largely redundant with the aforementioned morphological cuts, which already remove the majority cluster ellipticals and faint blue objects. The full set of morphological criteria (now also including cuts in $Sk_{0-1}$–$Sk_{1-2}$) are then applied to the remaining clusters, and the resulting subset is once again visually screened. Fig. \[fig:cuts\] shows the distribution of all galaxies in various projections of our multi-dimensional morphology parameter space, as well as the applied selection criteria. Members of the extended training set and of our final sample of RPS candidates are highlighted. Although the three sets of selection criteria shown in Fig. \[fig:cuts\] all select approximately the same fraction of galaxies (30-40%), their doing so largely non-redundantly leads to a much more restrictive selection of merely 8% (1263 galaxies) when all criteria are combined (Fig. \[fig:venn\]).
It is evident from Fig. \[fig:cuts\] that the adopted selection criteria, although highly efficient in eliminating regular disk galaxies and ellipticals, still select mostly galaxies that, although morphologically disturbed, are not necessarily undergoing RPS. In fact less than 20% of the automatically selected systems are classified as RPS candidates in our visual screening process. The disturbed sources rejected after visual inspection can largely be assigned to one of the following classes: strong-gravitational-lensing features (including both cluster-galaxy and galaxy-galaxy lensing events), foreground irregular galaxies, close pairs of ellipticals, unclassifiable clumpy emission in low signal-to-noise areas, and artefacts due to source confusion in crowded regions. We also note that, while colour information was not directly included in our selection procedure, the availability of images in both the F606W and F814W passbands proved essential in our visual classification to distinguish between the morphological disturbances caused by RPS and irregular extinction due to dust (see Fig. \[fig:need\_color\]).
![Importance of colour information for our visual inspections. Viewed solely in the F606W passband (left) this object could be considered a (remotely) plausible RPS candidate. A false-colour image including data in the F814W filter (right) strongly suggests a slightly disturbed dusty disk galaxy. \[fig:need\_color\]](need_color_gray.pdf "fig:"){width="24.00000%"} ![Importance of colour information for our visual inspections. Viewed solely in the F606W passband (left) this object could be considered a (remotely) plausible RPS candidate. A false-colour image including data in the F814W filter (right) strongly suggests a slightly disturbed dusty disk galaxy. \[fig:need\_color\]](need_color_rgb.pdf "fig:"){width="24.00000%"}
RPS-candidate sample
--------------------
Name $\alpha$ \[J2000\] $\delta$ \[J2000\] $m_{\rm F606W}$ $m_{\rm F814W}$ $r_{\rm BCG}$ \[kpc\] Incidence \[deg.\] $z$
---------------- -------------------- -------------------- ----------------- ----------------- ----------------------- -------------------- --------
MACSJ0257-JFG1 02 57 41.4 $-$22 09 53 18.75 18.22 166 10 0.3241
MACSJ0451-JFG1 04 51 57.3 $+$00 06 53 19.66 19.29 298 50 0.4362
MACSJ0712-JFG1 07 12 18.9 $+$59 32 06 19.10 18.39 87 107 0.3430
MACSJ0947-JFG1 09 47 23.1 $+$76 22 52 19.81 19.69 210 34 0.3417
MACSJ1258-JFG1 12 57 59.6 $+$47 02 46 19.10 18.70 133 45 0.3424
MACSJ1752-JFG1 17 51 56.1 $+$44 40 20 20.13 19.61 370 120 0.3739
MACSJ0035-JFG1 00 35 27.3 $-$20 16 18 19.49 19.02 182 103 0.3597
MACSJ0257-JFG2 02 57 43.5 $-$22 08 38 19.92 19.44 243 130 0.3297
MACSJ0429-JFG1 04 29 33.3 $-$02 53 02 20.97 20.64 203 113 0.4000
MACSJ0429-JFG1 04 29 40.4 $-$02 53 18 20.75 20.36 334 40 0.4049
MACSJ0916-JFG1 09 16 12.9 $-$00 25 01 20.43 19.97 334 81 0.3300
MACSJ1142-JFG1 11 42 37.0 $+$58 31 48 20.25 19.62 549 87 0.3267
MACSJ1720-JFG1 17 20 13.6 $+$35 37 17 20.05 19.52 309 30 0.3832
MACSJ1752-JFG1 17 52 06.3 $+$44 40 05 20.25 20.06 747 86 0.3527
RXJ2248-JFG1 22 48 40.2 $-$44 30 50 20.66 20.18 335 64 0.3515
The process described in the previous section yielded 223 possible ram-pressure stripping events (including the training set). We consider 15 of these to be classical jellyfish galaxies (yellow symbols in Fig. \[fig:cuts\]); an additional 115 objects show characteristic features of RPS (albeit less extreme), and 93 are at least plausible candidates. While we cannot rule out that physical processes other than RPS (e.g., minor mergers or tidal interactions) contribute to, or in fact cause, the observed morphology of our candidates, such alternative scenarios are likely to be relevant mainly for the fainter galaxies in our sample for which the most compelling sign of RPS (evidence of a debris trail) cannot be discerned in the shallow imaging data in hand.
As a complement to the first six “jellyfish” galaxies discovered in MACS clusters by @ESE, we show in Fig. \[fig:jellies\] a second sample of nine compelling jellyfish galaxies; fundamental properties of these systems are further described in Section \[sec:arrows\] and listed in Table \[tab:jellies\].
{width="30.00000%"} {width="30.00000%"} {width="30.00000%"}
{width="30.00000%"} {width="30.00000%"} {width="30.00000%"}
{width="30.00000%"} {width="30.00000%"} {width="30.00000%"}
### Observational biases {#sec:bias}
Impressive as the list of 223 RPS candidates may appear, we caution again that most of these galaxies may not even be cluster members, and that, for those that are, the cause of the observed morphological features need not be RPS. In addition, our list is almost certainly incomplete. Two primary observational biases are to blame: (a) our inability to reliably discriminate against non-RPS events solely from morphological data (leading to contamination by non-cluster galaxies) and (b) our inability to identify RPS events in galaxies moving close to our line of sight (leading to incompleteness regarding true RPS events in our target clusters).
First results from a comprehensive spectroscopic survey of all candidates (Blumenthal et al., in preparation) indeed indicate that more than half of the objects we selected are in fact fore- or background galaxies. The hazards of morphological selection alone are underlined not just by this high percentage of projection effects, but also by the elimination of three members of our extended training set (see bottom row Fig. \[fig:ese\]): the edge-on disk with a stellar tail in MACSJ1236.9+6311 is in the foreground of the cluster, while the dramatically distorted face-on spiral galaxy near the core of MACSJ1652.3+5534 was found to be a background object gravitationally lensed by the massive MACS cluster. The bright blue face-on spiral in MACSJ1731.6+2252, finally, turned out to be a member of a foreground group of galaxies. Although the removal of these three objects from our training set has no effect on our selection criteria, as can be seen from Fig. \[fig:cuts\] in which these galaxies are marked by red circles, the misidentification of galaxies we considered “textbook” cases of RPS serves as a warning about the robustness of morphological selection and underlines the need for spectroscopic follow-up observations.
The impact of the second observational bias cannot trivially be quantified by means of additional observations. Galaxies moving close to our line of sight lack the tell-tale debris trail and bow-shock morphology readily apparent for RPS proceeding in the plane of the sky (see Fig. \[fig:numsim\_models\]) and are thus likely to be missed. We attempt to account for the resulting systematic incompleteness when modelling galaxy trajectories in Section \[sec:models\].
Direction of motion and location within the cluster {#sec:arrows}
---------------------------------------------------
Since one of the goals of our study is to distinguish between the different geometric and kinematic scenarios associated with “stream-fed” infall along filaments, and cluster mergers, we focus on two key properties of cluster galaxies: the angle of incidence of their trajectory with respect to the gravitational centre of the cluster and the distance from the cluster centre. To observationally constrain the former, we consult the results of hydrodynamical modeling of RPS [e.g., @roediger_2006; @kronberger_2008; @roediger_2014] for insights regarding the correlation between the morphological disturbances caused by RPS and the galaxy’s direction of motion. Figure \[fig:numsim\_models\] shows model predictions for the distribution of gas and newly formed stars in galaxies undergoing RPS while moving face-on through the ICM. As expected, identifying the direction of motion becomes challenging when a galaxy moves through the ICM along our line of sight or is observed early in the stripping process.
![Distribution of gas (white) and newly formed stars (turquoise) for a simulated RPS event involving a spiral galaxy moving face-on through the ICM. A comparison with Fig. \[fig:ese\] shows that our morphological selection is, unsurprisingly, most sensitive to features typical of mature RPS events in galaxies viewed edge-on. [Reproduced from @kronberger_2008]. \[fig:numsim\_models\]](kronberger08-fig6.pdf){width="47.00000%"}
We attempt to assign projected directions of motion visually according to the following prescriptions: (1) if tails are discernible, the velocity vector is assumed to be parallel to the tail; (2) edge-on disks showing significant curvature are assigned velocity vectors oriented perpendicular to said curvature and originating at its apex; (3) if extended regions of star formation appear to be present, the velocity vector is placed perpendicular to the dominant elongation of said regions; (4) if none of the previously mentioned indicators are present (or if they are contradictory), we attempt to make the best physically motivated estimate. To avoid systematic biases, galaxies are inspected using small thumbnail images covering only the region immediately surrounding the galaxy with no indication of the direction to the cluster centre. In recognition of the subjective nature of our visual measurements [especially for galaxies moving partly or largely along our line of sight, @roediger_2006], the process is performed independently by three reviewers to derive an approximate grade for the robustness of each estimated direction of motion. Figure \[fig:arrows\] shows examples of objects falling into each of our quality grades with uncertainty increasing top to bottom and left to right. We then define the angle of incidence as the angle between the apparent velocity vector and the position vector with respect to the cluster centre (taken to be the location of the brightest cluster galaxy, BCG), i.e., the angular deviation from a purely radial infall trajectory (note again that all of these quantities are defined and observed in projection).
A second galaxy property that is critical to our efforts to deduce trajectories is location within the cluster. For RPS candidates lacking radial-velocity measurements, we are unable to assess whether an object is located in front or behind the cluster centre (defined by the redshift of the BCG), let alone further constrain its physical distance to the latter along our line of sight. Projected distances, however, measured in the plane of the sky and relative to the location of the BCG, are trivially obtained for comparison with the distribution expected for different geometries of galaxy infall.
![Examples of RPS candidate events illustrating our process to estimate direction of motion and the associated error. The arrows are the same as in Fig. \[fig:jellies\]. \[fig:arrows\]](macsj0451_all_id_318.pdf "fig:"){width="23.00000%"} ![Examples of RPS candidate events illustrating our process to estimate direction of motion and the associated error. The arrows are the same as in Fig. \[fig:jellies\]. \[fig:arrows\]](macsj0429_all_id_298.pdf "fig:"){width="23.00000%"}
![Examples of RPS candidate events illustrating our process to estimate direction of motion and the associated error. The arrows are the same as in Fig. \[fig:jellies\]. \[fig:arrows\]](macsj0416_all_id_107.pdf "fig:"){width="23.00000%"} ![Examples of RPS candidate events illustrating our process to estimate direction of motion and the associated error. The arrows are the same as in Fig. \[fig:jellies\]. \[fig:arrows\]](macsj0647_all_id_342.pdf "fig:"){width="23.00000%"}
A simple model of galaxy trajectories {#sec:models}
=====================================
In order to understand which kind of galaxy trajectories are most compatible with the observed distributions of (projected) incidence angle and cluster-centric distance, we compare our observations with the results of a simple theoretical model. To this end, we calculate orbits in a canonical cluster representative of the MACS clusters in our sample and use simple prescriptions, described below, to predict the projected radii and incidence angles at which extreme RPS events are most likely to occur.
As an infalling galaxy approaches the dense cluster core, the ICM exerts an increasing ram-pressure, $p_\mathrm{ram} = \rho_\mathrm{ICM}v_\mathrm{gal}^2$, where $\rho_\mathrm{ICM}$ is the ICM mass density and $v_\mathrm{gal}$ is the relative velocity between the galaxy and ICM [@gunn_infall_1972 hereafter GG]. By comparing $p_\mathrm{ram}$ to the gravitational restoring force per unit area on the gas within the galaxy, $$f_\mathrm{grav}(R) = \Sigma_\mathrm{gas}(R)\frac{\partial\Phi}{\partial Z}(R),$$ we find the critical radius where $p_\mathrm{ram} = f_\mathrm{grav}(R_\mathrm{strip})$ [@roediger_2007]. Here $\Sigma_\mathrm{gas}$, $\Phi$, and $Z$ are the ISM mass surface density, the gravitational potential of the galaxy, and its scale height, respectively. Beyond $R_\mathrm{strip}$, the galaxy potential is not strong enough to retain the gas and stripping sets in. @vollmer_2001 give an analytic estimate for the GG criterion which determines the stripping radius: $$\Sigma_\mathrm{gas} v_\mathrm{rot}^2 R_\mathrm{strip}^{-1} = p_\mathrm{ram},$$ where $v_\mathrm{rot}$ is the rotation speed of the galaxy. Although, in reality, the onset of RPS is likely to be a highly non-linear process, the simple GG criterion has proven suitable for global characterisations of RPS in in-depth numerical simulations [e.g., @roediger_2007; @kronberger_2008].
Galaxy properties
-----------------
Since our simple model aims only to predict the distribution of RPS events along galaxy orbits, but not the detailed properties of such events, we model all galaxies in our simulation as thin disks with radius $R_\mathrm{gal} = 15$ kpc and gas surface density $\Sigma_\mathrm{gas}$ = $10^{21}$ atoms per cm$^{2}$ moving face-on through the ICM.
To account for galaxy-to-galaxy variation in $f_\mathrm{grav}$, we also run all models for a range of galaxy masses, parametrized by the rotational velocity $v_\mathrm{rot}$ (see Eq. 2). The explored range of $v_\mathrm{rot}$ from 150 to 350 km s$^{-1}$ corresponds to dynamical masses, within 15 kpc, of $8\times10^{10}$, $2\times10^{11}$, and $4\times10^{11}$ M$_\odot$. The adopted range of rotational velocities covers a spectrum of masses from sub- to super-Milky Way sized objects.
Cluster properties
------------------
We describe the gas and total mass distribution within the cluster using a spherical $\beta$-model [@cavaliere_1976] $$\rho = \rho_0\left[1 + \left(\frac{r}{r_0}\right)^2\right]^{-\frac{3}{2}\beta },$$ where $\rho_0$ is the central mass (or gas) density, $\beta$ and $r_0$ are the power-law index and core radius, respectively, and $r$ is the cluster-centric radius. We adopt a total mass of $1.3\times10^{15} M_\odot$ [the average weak-lensing mass, $M(r<1.5$ Mpc$)$, of MACS clusters at $z>0.3$ according to @applegate_2014]. As the majority ($\sim$2/3) of the clusters in our sample do not show dramatic large-scale substructure, we assume that our model cluster is largely relaxed, featuring gas and total mass distributions that share a common centre, core radius $r_0$ and power-law slope $\beta$. We adopt $r_0$ = 180 kpc and $\beta$ = 0.59, the median of the values from the spatial X-ray analysis of @mantz_2010. Assuming a gas fraction $f_{\rm gas}=0.074$ [@mantz_2014] and the model parameters above, our canonical cluster has a central particle density $n_0$ of $2.29\times10^{-3}$ cm^-3^.
Galaxy trajectories
-------------------
The orbits of test particles falling into our model cluster are computed for a wide range of initial orbital parameters that encompass expectations for infall along connected filaments and from cluster mergers. Orbit calculations begin at the end of a filament which is assumed to be at a distance of 2.5 Mpc from the cluster core ($\approx R_\mathrm{vir}$). In Fig. \[fig:vectors\], we show a schematic of the quantities that characterise orbits in our model: the speed of a galaxy in the direction of the filament axis $v_\parallel$, the transverse velocity perpendicular to the filament flow $v_\perp$, and the impact parameter $b$.
Radial profiles of filaments in cosmological simulations show a well defined edge at a radius of 1.0–2.0 $h^{-1}$ Mpc ($\sim$1.4–2.8 Mpc in our assumed cosmology) beyond which the matter density essentially vanishes [@colberg_2005]. We therefore model filaments as cylinders of constant density with radius $b_\mathrm{max}$. We populate these filaments with $3\times10^4$ galaxies using Monte Carlo sampling designed to provide constant density within $b_\mathrm{max}$ and a normal distribution in $v_\perp$ to account for the velocity dispersion of galaxies within the filament.
In the following, we consider three infall scenarios that differ primarily in the approach velocity of galaxies at the cluster’s virial radius: 1) *stream-fed* infall along filaments; 2) a *slow merger*; and 3) a *fast merger*. Table \[tab:models\] lists the model parameters that characterise each of these scenarios. For each infall scenario, we fix the initial velocity $v_\parallel$ at one value for all orbits. For the stream-fed model, we choose $b_{\rm max}$=1.5 Mpc and $v_\parallel$ = 200 km s$^{-1}$, the average filament radius and the average velocity of matter at the cluster-filament interface, respectively [@colberg_2005], as well a velocity dispersion characteristic of group environments ($\sim$100 km s$^{-1}$). The slow and fast merger models are characterised by initial velocities of 1000 and 3000 km s$^{-1}$, respectively, and a velocity dispersion of 1000 km s$^{-1}$ and $b_\mathrm{max}$ = 2.5 Mpc for either merger scenario. Since we know neither the number and orientation of connected filaments for our cluster sample, nor the orientation of a putative merger axis, we place filaments/merging clusters at $10^3$ positions, sampled isotropically on a 2.5 Mpc sphere. In total, this results in $3\times10^6$ orbits per scenario which are each followed for 5 Gyr ($\sim t_\mathrm{cross}$) in time steps of 5 Myr.
![Schematic diagram of the quantities that characterise the initial conditions and orbits of galaxies in our infall models: the maximal impact parameter $b_\mathrm{max}$, the initial velocities $v_\parallel$ and $v_\perp$, the cluster-centric radius $\hat{r}$, and the inclination angle $i$. \[fig:vectors\]](model_quantities.pdf){width="50.00000%"}
Model $v_\parallel$ \[km s$^{-1}$\] $\sigma_v$ \[km s$^{-1}$\] $b_\mathrm{max}$ \[Mpc\]
------------- ------------------------------- ---------------------------- --------------------------
Stream-fed 200 100 1.5
Slow Merger 1000 1000 2.5
Fast Merger 3000 1000 2.5
: Model Parameters\[tab:models\]
See Fig. \[fig:vectors\] for a schematic illustration of $v_\parallel$ and $b_\mathrm{max}$; $\sigma_v$ indicates the velocity dispersion of infalling galaxies.
Defining the start of the RPS event as the time step in which the GG criterion is first satisfied, we explore a range of RPS event durations, from 50 Myr to 1 Gyr, during which the resulting event is assumed to remain observationally detectable. This choice is motivated by numerical simulations: @roediger_2014 find that the signature RPS morphology should be observable in galaxies overrun by an ICM shock for between $\sim$several 10 Myr to a few 100 Myr. Slightly longer durations are quoted by @kronberger_2008 for a scenario similar to our stream-fed infall model (see also Fig. \[fig:numsim\_models\]).
For comparison with our observational results, segments of the orbits corresponding to an RPS event (under our definition) are projected onto the plane of the sky, thus providing the projected angle of incidence (the projected angle between the galaxy’s velocity and position vectors), $i$, and the projected radius from the cluster centre. We then tabulate the amount of time spent in bins of projected radius and inclination angle to construct simulated probability distributions for each scenario.
Accounting for observational bias {#sec:bias2}
---------------------------------
As mentioned in Section \[sec:bias\] and illustrated in Fig. \[fig:numsim\_models\], RPS events in galaxies moving along or close to our line of sight are likely to be missed, as, for this particular geometry, the pronounced morphological features that our selection process is build upon are obscured by the galaxy being stripped. We examine the importance of this observational bias by imposing on our modeling results that all RPS events are undetectable that occur in galaxies moving along an axis that is inclined to our line of sight by 0, 15, 30, or 45 degrees. As detailed in the following section, even the most severe implementation of this line-of-sight bias results in only modest changes in the model predictions, suggesting that the effect does not significantly affect the conclusions drawn from our comparison with the data.
{width="\textwidth"}
{width="\textwidth"}
{width="\textwidth"}
Results
=======
In order to reduce contamination by interlopers (fore- or background galaxies), we restrict our analysis to the subset of candidate RPS events with measured redshifts within $\pm$4000 km s$^{-1}$ of the redshift of the host cluster; the 53 objects (of 124 with measured redshifts) meeting this criterion are hereafter referred to as the “spectroscopic sample". Of these, we select a subset of the 15 galaxies exhibiting the most compelling “jellyfish" morphology comprised of the six systems presented by ESE and the nine shown in Fig. \[fig:jellies\] (“jellyfish sample"). We further restrict the comparison between data and model predictions to a projected radius of 415 kpc from the cluster core, which leaves 23 and 11 galaxies in the spectroscopic and jellyfish samples respectively. This radial cutoff minimises systematic incompleteness introduced at larger cluster-centric radii, which are covered only by images of the most distant clusters in our sample.
Fig. \[fig:incid\_cum\_dist\] shows the cumulative distributions of the incidence angle for our two RPS subsamples plotted against predictions from our infall model, with (bottom row) and without (top row) correction for the bias discussed in Section \[sec:bias\] and \[sec:bias2\]. The bottom The left, center, and right columns of Fig. \[fig:incid\_cum\_dist\] show predictions for the stream-fed, slow-merger, and fast-merger models, respectively (see Table \[tab:models\] for the parameters characterising these models).
Visual comparison suggests that the observations are best matched by the model predictions for the slow-merger scenario, provided that the duration of the stripping process is less than a Gyr[^4]. Contrary to the traditional picture of RPS being driven purely by infall from the low-density field, preferably along filaments, we find poor agreement between the data and stream-fed models which over-predict events at extreme incidence angles (at $\lesssim$40$^\circ$ for almost all combinations of model parameters explored by us, and at $\gtrsim$140$^\circ$ for low-mass galaxies experiencing long RPS events). In this scenario, the motion of galaxies is dominated by the cluster potential, which leads to a preferential alignment of trajectories toward the cluster core (at least in our projected view) and thus a highly anisotropic distribution of incidence angles.
Fig. \[fig:rad\_cum\_dist\] shows the cumulative distributions of the number of RPS events within a given projected cluster-centric radius. To provide more natural, equal-area sampling, we bin the data in equal steps of $r_\mathrm{proj}^2$; a uniform areal distribution thus appears as a straight line from zero to one. We find that both of the cluster merger models predict a nearly uniform areal distribution of events in agreement with our observations. Stream-fed models with the most massive galaxies and/or the longest event timescales predict an excess at small projected radii which is not supported by our data. Note, that this comparison also effectively rules out the stream-fed model with a Milky-Way sized galaxy and 300 Myr timescale that at least marginally matched the observed distribution of incidence angles and is shown as the green dashed line in Fig. \[fig:incid\_cum\_dist\].
A more quantitative assessment of the significance of the discrepancies between the observed and predicted distributions can be obtained with Kolmogorov-Smirnov (KS) tests. In Fig. \[fig:ks\_scores\], we show KS probabilities for the null hypothesis that the observed distributions are drawn from the same parent population as the predictions of a given model. Correcting all models for the aforementioned line-of-sight bias (Sections \[sec:bias\] and \[sec:bias2\]) does not change our conclusions significantly. For simplicity, we therefore ignore the bias due to motion along the line of sight in the KS tests. To maximize the number of objects in the comparison, we show results for the spectroscopic sample only. However, considering the smaller jellyfish subsample does not significantly alter our conclusions.
Consistent with our qualitative assessments above, we find no agreement with the observed distribution of incidence angles for any model assuming infall along filaments, although the distribution of projected radii does not rule out such models (at least not for low-mass galaxies, see bottom panel of Fig. \[fig:ks\_scores\]). By contrast, practically all of the models for the two merger scenarios provide an acceptable (or good) description of the data, with the exception of those involving the most massive galaxies, for which models assuming long RPS durations of $\tau_\mathrm{event}\gtrsim$ 300 Myr are ruled out at more than $2\sigma$ confidence.
Conclusions
===========
Since our models are intrinsically three dimensional, the comparisons presented above, although involving solely parameters measured in projection, allow us to distinguish between distinctly different three-dimensional scenarios.
In the merger scenarios, RPS events are triggered in fast-moving galaxies near the outskirts of the cluster and, due to the relatively short duration of $\sim$500 Myr required by our incidence angle data (see top row of Fig. \[fig:ks\_scores\]), remain confined to a shell well outside a (three-dimensional) cluster-centric radius of 400 kpc. On the other hand, the projected radius data favour event durations longer than $\sim$100 Myr to explain the uniform areal distribution (Fig. \[fig:rad\_cum\_dist\]). The RPS candidates detected by us are thus the projection of the essentially uniform distribution of much more distant RPS events in the fore- and background segments of this shell. In principle, galaxies of all masses may contribute to the observed RPS distribution; however, the majority are likely to be systems of low to intermediate mass, since models for extremely massive galaxies generally require finely tuned, short RPS lifetimes of about 100 Myr approximately to match the observations (red lines in Fig. \[fig:ks\_scores\]).
By contrast, galaxies falling into the cluster along filaments do so at much lower peculiar velocities and thus require higher ICM densities for the GG criterion to be met; as a result, RPS events are triggered only much closer to the cluster core. To match the observed, broad distribution of incidence angles, these galaxies need time to enter our field of view from all sides, which mandates that the associated RPS events remain observable for 300 Myr or longer (Fig. \[fig:incid\_cum\_dist\]). Such long life-times, however, lead in turn to an excess in the number of events close to the cluster core that is not observed (Fig. \[fig:rad\_cum\_dist\]).
We therefore tentatively conclude that extreme RPS events in massive clusters are generally short-lived ($\lesssim$500 Myr) and triggered far from the cluster core, likely driven by cluster mergers. Interestingly this preference of our analysis for RPS events being most readily observed in galaxies moving at high speed through an only modestly dense ICM suggests that textbook cases of “jellyfish galaxies" might also be observed near the cores of less massive clusters (or even groups of galaxies, see also @poggianti_2015) provided a cluster or group merger event ensures sufficiently high peculiar initial velocity. Note also that, while our data disfavour infall along filaments as the primary trigger, they do not rule out a contribution from such a scenario. Wide-field imaging surveys that are able to detect RPS events out to the virial radius are needed to determine the relative contributions of stream-fed infall and cluster mergers.
Summary
=======
We have conducted a systematic search for galaxies experiencing ram-pressure stripping (RPS) in 63 MACS clusters at $z{=}$0.3–0.7. Using quantitative morphological parameters for $\sim$16,000 galaxies detected in *Hubble Space Telescope* images of these systems we identify 211 potential cases of RPS that complement a training set of 12 “jellyfish" galaxies used to define our selection criteria. Where possible, the direction of motion in the plane of the sky is estimated for these systems based on morphological indicators such as the curvature and orientation of the apparent galaxy-ICM interface region or a visible debris trail. Several systematic biases are inherent to our approach: (a) the classification of galaxies according to their likelihood of undergoing RPS is partly based on visual inspection and thus to some extent subjective, (b) the small field of field of view our observations prevents us from sampling the galaxy population in the outer regions of our cluster targets (except in projection) where RPS events might be initially triggered, and (c) our selection process is fundamentally unable to robustly identify RPS events in galaxies moving along, or close to, our line of sight.
We attempt to address the first of these biases by obtaining spectroscopic redshifts of all our RPS candidates. While the resulting spectra do not immediately confirm or refute an RPS event, they allow us to establish whether or not a morphologically selected candidate is in fact a cluster member and whether its spectral characteristics are consistent with ongoing or recent star formation. So far, 53 of 124 systems targeted in spectroscopic follow-up observations were confirmed as cluster members. A detailed analysis of these galaxies’ spectral properties will be presented in a forthcoming paper (Blumenthal et al., in preparation).
The remaining two observational biases mentioned above can be accounted for by three-dimensional modelling of the trajectories and environment of galaxies falling into a massive cluster. Specifically, we compare the distributions of the observed projected incidence angle and distance from the BCG with predictions from simple models of galaxy orbits in a MACS-like cluster. We investigate two scenarios: accretion of galaxies from an attached filament, and a cluster merger event.
We find significantly better agreement for the merger scenario, provided the duration of RPS events is $\lesssim$500 Myr. We thus tentative conclude that extreme ram-pressure stripping events is primarily triggered in massive cluster mergers (rather than by infall alone) where relative velocities between galaxies and the ICM are large enough to initiate RPS far from the cluster core ($\gg$ 400 kpc). Although our study is, by design, limited to relatively massive clusters, we note that this result implies that extreme RPS events may also occur in mergers of poorer clusters and even groups of galaxies, where the required ingredients (high peculiar velocity and moderately high ICM density) are both met by galaxies close to core passage. We also find that galaxies of mass similar to, or less than, our Milky Way are likely to dominate the set of observable RPS events in massive clusters, although more massive galaxies may contribute too at a lower level. Although models assuming infall along a filament were found to yield predictions that are largely in conflict with our data, both processes (accretion along filaments and via cluster mergers) can be expected to contribute. The extent to which the two mechanisms are responsible for the observed population of RPS events in our sample is difficult to quantify but could be tested by imaging surveys that probe the distribution of RPS events to larger cluster-centric radii.
In-depth studies of the X-ray properties of RPS host clusters along with spectroscopic investigations of the star-formation rates and histories of the candidates identified in this study will be critical to test our conclusions and allow a quantitative comparison of observational diagnostics with predictions of numerical models of ram-pressure stripping.
Acknowledgements {#acknowledgements .unnumbered}
================
We thank the anonymous referee for their helpful comments, questions, and suggestions on revising the manuscript. CM thanks J. Lotz for providing the galaxy morphology source code which was adapted for this work. HE gratefully acknowledges financial support from STScI grants GO-10495, -10875, -12166, and -12884. This research made use of Astropy, a community-developed core Python package for Astronomy (Astropy Collaboration, 2013).
Skeletal Decomposition Parameters {#sec:skel}
=================================
The morphological indicators discussed in Section 3 were generally defined to identify characteristic morphological traits of galaxy mergers [e.g. @lotz_2011]. We introduce a new metric based on the concept of the morphological skeleton [@maragos_1986] to both quantify the amount of substructure in a galaxy while concurrently identifying arm/tail-like structures. Conceived in the context of mathematical morphology [see @serra1988image] and originally introduced as a means for binary image compression, the morphological skeleton (or medial axis transform) reduces a shape to a line that maintains the topological structure of the full image, thus allowing exact reconstruction.
We here generalise the definition of the morphological skeleton to images with non-binary, continuous greyscale pixel values. However, we must be cautious as noise in relatively short exposures used in this survey ($\sim$1200 sec.) can manifest as small scale substructure in the skeleton if applied naively. To reduce this erroneous signal from noise, we smooth the image using a Gaussian kernel before determining the skeleton and then prune the result to remove any disconnected segments. We define the result of this process as $Sk_i$. We perform skeletal decompositions under three smoothing scales corresponding to the Petrosian radius $r_p$, the half light radius $r_{50\%}$, and the 10% light radius $r_{10\%}$ which define $Sk_0$, $Sk_1$, and $Sk_2$. Note that due to the cleaning process we apply here, exact reconstruction of the original image is not possible.
To further reduce erroneous signal due to residual noise, we define $Sk_{x+y}$ (where $y=x+1$) as comprising all pixels in the higher-order skeleton (i.e. under a smaller smoothing kernel) connected to that of the lower-order skeleton (larger smoothing kernel). To generate a common reference point and to avoid bias due to image size and lower order structure, we then subtract the length of the lower-order skeleton from $Sk_{x+y}$ and normalise by the length of the lower-order skeleton (e.g. $[|Sk_{0+1}|-|Sk_0|]/|Sk_0|$) defining a final numerical measure $Sk_{x-y}$ which quantifies the excess in substructure under smoothing scale $y$ with respect to $x$ (see Fig. \[fig:Sk\]).
A simple way to understand this qualitatively is to consider a case where $Sk_{0-1}$ or $Sk_{0-1}$ is equal to zero. This would imply that image smoothed on a finer scale (smaller kernel) does not reveal any more substructure or that the galaxy’s light profile is essentially smooth below the upper smoothing scale. However, as a full interpretation of the meaning and reliability of these indicators is beyond of the scope of this paper, we here characterise $Sk_{0-1}$ only to be a measure of bending in the galaxy or the deviation from a symmetric object (somewhat correlated with asymmetry), while $Sk_{1-2}$ quantifies the amount of clumpy substructure connected to the brighter regions of the galaxy.
![An example of the results of the greyscale skeletonization process which we use to define the skeletal decomposition parameters $Sk_{0-1}$ and $Sk_{1-2}$. \[fig:Sk\]](sk_params.pdf){width="50.00000%"}
Name $\alpha$ \[J2000\] $\delta$ \[J2000\] $t_{exp}$ \[s\] GO Prop. ID
------------------- -------------------- -------------------- ----------------- -------------
EMACSJ1057.5+5759 10:57:31.680 +57:59:33.72 1200 12884
MACSJ0032.1+1808 00:32:11.344 +18:07:49.37 1200 12166
MACSJ0035.4-2015 00:35:26.957 -20:15:50.66 1200 10491
MACSJ0140.0-0555 01:40:01.626 -05:55:06.71 1200 10491
MACSJ0152.5-2852 01:52:35.361 -28:53:39.88 1200 10491
MACSJ0257.6-2209 02:57:40.596 -22:09:27.80 1200 10875
MACSJ0308.9+2645 03:08:56.839 +26:45:43.91 1200 12166
MACSJ0451.9+0006 04:51:55.443 +00:06:11.66 1200 10491
MACSJ0521.4-2754 05:21:25.808 -27:55:06.91 1200 10491
MACSJ0547.0-3904 05:47:01.796 -39:04:13.24 1200 12166
MACSJ0553.4-3342 05:53:23.850 -33:42:42.21 2092 12362
MACSJ0712.3+5931 07:12:21.985 +59:32:24.82 1200 10491
MACSJ0845.4+0327 08:45:28.224 +03:27:28.46 1200 10491
MACSJ0916.1-0023 09:16:12.344 -00:23:47.00 1200 10491
MACSJ0947.2+7623 09:47:10.744 +76:23:21.62 1200 10491
MACSJ0949.8+1708 09:49:52.655 +17:07:06.38 1200 10491
MACSJ1006.9+3200 10:06:55.632 +32:01:33.91 1200 10491
MACSJ1115.2+5320 11:15:15.968 +53:19:47.47 1200 10491
MACSJ1124.5+4351 11:24:29.365 +43:51:32.97 1200 12166
MACSJ1133.2+5008 11:33:14.109 +50:08:29.50 1200 10491
MACSJ1142.4+5831 11:42:26.434 +58:32:01.30 1200 12166
MACSJ1226.8+2153C 12:26:41.421 +21:53:07.58 1200 12166
MACSJ1236.9+6311 12:36:59.868 +63:11:02.26 1200 10491
MACSJ1258.0+4702 12:58:02.708 +47:02:42.87 1200 10491
MACSJ1319.9+7003 13:20:09.685 +70:04:28.16 1200 10491
MACSJ1354.6+7715 13:54:31.253 +77:15:08.71 1200 10491
MACSJ1447.4+0827 14:47:26.289 +08:28:37.08 1200 12166
MACSJ1452.9+5802 14:52:57.957 +58:02:43.28 1200 12166
MACSJ1526.7+1647 15:26:42.342 +16:47:48.83 1200 12166
MACSJ1621.3+3810 16:21:23.928 +38:10:16.28 1200 12166
MACSJ1644.9+0139 16:45:01.729 +01:40:09.83 1200 12166
MACSJ1652.3+5534 16:52:19.726 +55:34:46.63 1200 10491
MACSJ1731.6+2252 17:31:39.268 +22:52:05.09 1200 12166
MACSJ1738.1+6006 17:38:05.383 +60:06:14.92 1200 12166
MACSJ1752.0+4440 17:51:57.961 +44:39:45.45 1200 12166
MACSJ1806.8+2931 18:06:51.898 +29:30:23.03 1200 12166
MACSJ2050.7+0123 20:50:42.381 +01:23:24.69 1200 12166
MACSJ2051.1+0215 20:51:10.058 +02:16:00.72 1200 12166
MACSJ2135.2-0102 21:35:12.822 -01:02:51.52 1200 10491
MACSJ2241.8+1732 22:41:56.386 +17:32:47.33 1200 12166
SMACSJ0234.7-5831 02:34:43.512 -58:31:16.51 1200 12166
SMACSJ0549.3-6205 05:49:18.358 -62:05:07.88 1200 12166
SMACSJ0600.2-4353 06:00:12.915 -43:53:19.33 1200 12166
SMACSJ0723.3-7327 07:23:18.709 -73:27:06.01 1200 12166
SMACSJ2031.8-4036 20:31:46.993 -40:37:03.68 1200 12166
SMACSJ2131.1-4019 21:31:05.693 -40:19:12.22 1200 12166
Name $\alpha$ \[J2000\] $\delta$ \[J2000\] $t_{exp}$ \[s\] GO Prop. ID
----------------------- -------------------- -------------------- ----------------- -------------
MACSJ0329-0211 03:29:41.560 -02:11:46.10 4104 12452
MACSJ0416-2403 04:16:08.380 -24:04:20.79 4036 12459
MACSJ0429-0253 04:29:36.049 -02:53:06.10 3938 12788
MACSJ0647+7015 06:47:50.269 +70:14:54.99 4128 12101
MACSJ0717.5+3745-POS5 07:17:32.629 +37:44:59.70 7920 10420
MACSJ0744+3927 07:44:52.819 +39:27:26.89 4128 12067
MACSJ1115+0129 11:15:51.900 +01:29:55.10 3870 12453
MACSJ1149+2223 11:49:34.704 +22:24:04.75 4128 12068
MACSJJ1206.2-0847 12:06:12.055 -08:47:59.44 6608 10491
MACSJ1311-0310 13:11:01.800 -03:10:39.79 4158 12789
RXJ1347-1145 13:47:32.110 -11:45:11.36 3878 12104
MACS1423+2404 14:23:47.88 +24:04:42.49 4240 12790
RXJ1532+3021 15:32:53.779 +30:20:59.39 4060 12454
MACSJ1720+3536 17:20:16.780 +35:36:26.49 4040 12455
MACSJ1931-2635 19:31:49.62 -26:34:32.90 3850 12456
MACSJ2129-0741 21:29:26.059 -07:41:28.79 3728 12100
RXJ2248-4431 22:48:43.960 -44:31:51.30 3976 12458
[^1]: Here, $\mu_\mathrm{max}$ and $r_{20\%, \mathrm{ell}}$ are the peak surface brightness and the elliptical radius encircling 20% of the total flux, respectively.
[^2]: Although the observed morphology of these objects does not prove the occurrence of RPS, in-depth follow-up studies of galaxies sharing the same striking features unambiguously confirmed RPS to be at work (@sun_2010; @sivanandam_2010; @cortes_2015).
[^3]: Note that the inferred direction of motion for two candidates (leftmost two in the bottom row of Fig. \[fig:ese\]) is largely aligned with our line of sight.
[^4]: Note that, in the top panel of Fig. \[fig:incid\_cum\_dist\], all of the solid lines, as well as the red dashed line, fall on top of each other and are thus indistinguishable by eye.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'HR 8799 is a four planet system that also hosts a debris disk. By numerically integrating both planets and a planetesimal disk, we find interactions between an exterior planetesimal disk and the planets can influence the lifetime of the system. We first consider resonant planetary configurations that remained stable for at least 7 Myrs sans debris disk. An exterior debris disk with only $\sim1\%$ the mass of the outermost planet (approximately a Neptune mass) was sufficiently large enough to pull the system out of resonance after 2 to 6 Myrs. Secondly, we consider configurations which are unstable in less than a few hundred thousand years. We find that these can be stabilized by a debris disk with a mass of more than $\sim10\%$ that of the outermost planet. Our two sets of simulations suggest that estimates of the long term stability of a planetary system should take into account the role of the debris disk.'
author:
- |
Alexander Moore & Alice C. Quillen\
Department of Physics and Astronomy, University of Rochester, Rochester, NY 14627, USA\
title: Effects of a planetesimal debris disk on stability scenarios for the extrasolar planetary system HR 8799
---
Introduction
============
Photometric surveys conducted with Kepler space telescope [@borucki11] in addition to radial velocity surveys such as those conducted by @wright09 have indicated that multiple planet systems are common. Numerical integrations can be used to determine if these systems are stable and estimate the time to first orbit crossing event or collision [@gozdziewski09; @reidemeister09; @fabrycky10; @marshall10; @marois10]. Any configuration that has a short lifetime is considered less likely. Consequently, integrations can be used to place constraints on both the orbital elements and masses of the planets. These numerical investigations often neglect planetesimals. However, extrasolar planetary systems can harbor planetesimal debris disks [@su09; @moro-martin10].
HR 8799 is a $1.5 \pm 0.3 M_{sol}$ A5V star found $39.4 \pm 1.0 pc$ from Earth [@marois08]. It has at least four planets which have been directly imaged [@marois10]. Mass estimates for the planets, even taking into account HR 8799’s young age of $60^{+100}_{-30}$ Myr determined by various techniques [@marois08], or $30^{+20}_{-10}$ as part of the Columba association [@marois10], have values of $10 \pm 3$ $M_{Jupiter}$ for HR 8799c, d, and e and $7^{+4}_{-2}$ $M_{J}$ for HR 8799b for an assumption of a 60 Myr age, and masses of $7^{+3}_{-2}$ $M_{J}$ for HR 8799c, d, e and $5 M_{J}$ for an assumption of a 30 Myr age. Projected separations for the planets of HR 8799e, d, c and b from the star are observed to be 14.5, 24, 38 and 68 AU, respectively [@marois10]. However, this measurement lacks a long baseline helpful for constraining the planets’ positions. Astrometric measurements with a much longer baseline taken with *Hubble Space Telescope* in 1998 for the outer three planets HR 8799b, c and d confirm reported values for HR 8799b and add new observations for HR 8799c and d [@soummer11].
Dynamical studies of HR 8799 indicate that it is likely in a 4:2:1 dual mean motion resonance (MMR). In other words, the inner two planets are in a 2:1 mean motion resonance while the outer pair are also in a 2:1 mean motion resonance. This architecture is required to explain how the system has remained stable over its observed age [@gozdziewski09; @fabrycky10; @reidemeister09; @marshall10; @marois10]. Possible orbital configurations determined by numerical simulations are summarized by @moro-martin10.
HR 8799 also has an inner debris disk ranging from 6-15 AU, an outer debris disk which is thought to extend from 90 to 300 AU and a dusty halo out to $\sim1000$ AU [@su09]. The specific details of this model will be discussed further in subsection 1.2.
Planetesimal debris disks can influence the long term stability of a planetary system. Within the context of the ‘Nice’ model [@tsiganis05], planets migrate due to interactions with planetesimals, and instability occurs when two planets cross a strong mean motion resonance. @thommes08 considered systems put in resonance by a gas disk, a possible scenario explaining HR 8799’s current configuration. However, after the gas disk had been depleted, they found that planetary interactions with the remnant planetesimal disk tended to remove these systems from resonances and induce dynamical instability.
It is in this context that we examine the role of the debris disk in affecting the stability of the HR 8799 system. First, we investigate if it is possible to delay the onset of instability for an initially highly unstable orbital configuration. Then we consider a configuration with a long lifetime and determine whether a debris disk can cause instability.
Resonant Structure of HR 8799
-----------------------------
As discussed in the introduction, initial solutions tested by @gozdziewski09 [@reidemeister09; @fabrycky10] had suggested that HR 8799b, c and d were most likely in 4:2:1 dual mean motion resonances. Nearly all orbital configurations that remain stable during a reverse integration for the estimated age of the system which also agree with the observed orbital elements and nominal masses for the planets minimally called for HR 8799c and d to be in a 2:1 MMR. Furthermore, the dual mean motion 4:2:1 resonance followed by the inner pair of HR 8799c and d in a 2:1 MMR allow for the largest possible range of masses that still produce simulations stable over the lifetime of the system. Alternate solutions included a 2:1 MMR among the outer pair HR 8799b and c as well as a few finely tuned solutions. In a later analysis, @marshall10 found that when the three planet configuration is placed in the 4:2:1 MMR with low planet eccentricities, HR 8799 could survive the observed age of the system and potentially longer. Re-reduction of Hubble space telescope astrometric data of HR 8799 by @soummer11, originally taken in 1998, robustly supports the previous 4d:2c:1b dual MMR hypothesis with results which have only small departures from the exact integer period ratios in previously published data.
This previous dynamical work precedes the discovery of the innermost planet HR 8799e by @marois10. However, simulations by @marois10 have indicated that the extra planet only places more restrictions on the possible system architectures. In order for those simulations to remain stable over the minimum estimated lifetime of the HR 8799 system, the planets e, d and c are required to be in a 4:2:1 MMR along with masses on the minimum end of the estimated values. Note that the 4:2:1 dual mean motion resonance is still the most likely configuration for this system to be in if long term stability is desired - the difference being which planets are in the mean motion resonance, their projected masses due to dynamical considerations, and the maximum age of the system which remains stable given these orbital elements.
Distribution and Total Mass of HR 8799’s Debris Disk
----------------------------------------------------
The total mass and distribution of the debris found in the disk affects the dynamics, migration rate, extent of migration, and the smoothness of migration. However, the total mass of the inner and outer debris disk along with the dust halo is not well known for HR 8799. Estimating the total mass in debris within a disk is difficult to compute accurately for both observational and theoretical reasons. Solid (non-gaseous) matter, which emits continuously in the $\mu m$ to mm wavelengths produces nearly all of the radiation [@beckwith96]. However, while the total cross section of the dust particles is many orders of magnitude larger than that of larger objects (m to km sized asteroids, comets, and planetesimals), it is these later objects which comprise a majority of the total mass of the debris. These objects are not bright enough to individually detect in the wavelengths that they emit. But estimating the total mass of the debris is important to determining the dynamics.
The halo and disk dust has been modelled by @su09. The halo mass is estimated to be $1.9 \times
10^{-2} M_{\oplus}$ with a radius of up to 1000 AU. Estimating the dust mass of the other disk components is more difficult - particularly for the cold outer disk because the inner and outer edges are not well known. In the case of the inner edge, it was first thought to be at 90 AU from temperature estimates while the outer radius had been modelled at 300 AU to account for all the observational constraints, giving a total dust mass of $1.2 \times 10^{-1} M_{\oplus}$. The inner disk is quite warm at $\sim150 K$, allowing for a very low dust mass estimate of $1.1 \times 10^{-6} M_{\oplus}$. A brief summary of the modelled parameters that were used including assumed surface densities, inner and outer radii, minimum and maximum grain size, as well as total mass can be found in @su09 in their Table 3.
Recent submillimeter observations have added some additional constraints. @hughes11 examined HR 8799 and its debris disk with the Submillimeter Array at $880 \mu m$ - a wavelength which is suitable for examining larger dust grains. Low signal-to-noise prevented a full multi-parameter modelling of the dust at this wavelength but a combination of the SMA data and spectral energy distributions (SEDs) seem to rule out a narrow ring of dust and favor a broad outer ring starting with an inner edge at $\sim 150$ AU.
Scaling up from these modelled dust masses to a total debris disk mass can be difficult. Small dust grains are typically evacuated in debris disks by a combination of Poynting-Robertson drag and radiation blow-out (depending on grain size) on very short timescales relative to the age of an average debris disk. To constantly replenish the dust in a system that is many millions of years old, a collisional cascade is required [@safronov69; @dohnanyi69; @williams94; @tanaka96; @kenyon99; @kenyon01; @kenyon02].
It is assumed that the differential number distribution of the particles in mass is a power law which takes the form $$dn(m) = Am^{-\alpha}dm$$ or simply $$dn(a) = Ca^{2-3\alpha}da$$ in radius. The steady-state solution depends on an $\alpha$ index of $\frac{11}{6}$. This parameter has been estimated empirically and numerically in the previously mentioned literature and, assuming that $\alpha$ is determined only by the mass-dependence of the collision rate and that the model is self-similar, can be shown to be that value analytically [@tanaka96].
To estimate the total mass of the disk, we integrate the differential number distribution times the mass of these objects at a constant density from the minimum to maximum sizes in the cascade $$M_{T} = \int _{a_{min}} ^{a_{max}} M(a) dn(a).$$ The constant in our number distribution can be set by using the model of the outer debris disk by @su09. This model includes the minimum and maximum grain sizes that were used in order to create a synthetic SED which was matched to the observed data. With the dust mass estimate from that model, and a minimum and maximum grain size of $10\mu m$ and $1000\mu m$ respectively, the constant $C$ can be computed.
To find the total mass of the disk we repeat the previous calculation only with a new $a_{max}$ that corresponds to the largest radius objects in our collisional cascade. The total mass that this integrand will yield is entirely set by the upper bound because there are many orders of magnitude between the minimum and maximum object sizes. Unfortunately, determining $a_{max}$ is difficult.
Assuming the cascade operates for the system age @quillen07 estimated $a_{max}$, assuming an alpha parameter of 11/6, that used only observable properties of the debris disk. As per their equation 16, $$\begin{split}
a_{top}& \approx 5.4 \rm{km} \left(\frac{\lambda}{10 \mu m}\right) \left(\frac{M_{\star}}{M_{\odot}}\right)^{8/3}
\left(\frac{r}{100 \rm{AU}}\right)^{-14/3}\\
& \quad \times \left(\frac{Q ^{\star} _{D}}{2 \times 10^6 \rm{erg} \cdot \rm{g^{-1}}}\right)^{-5/3}
\left(\frac{t_{age}}{10^7 \rm{yr}} \right)^2 \left(\frac{h}{0.02}\right)^{10/3}\\
& \quad \times \left[\frac{\bar{\tau}(\lambda)}{10^{-2}}\right]^2 \left(\frac{f_{\tau}}{4}\right)^{-2}
\end{split}$$ where $h$, the disk aspect ratio, and $\bar{\tau}$, the normal disk opacity at wavelength $\lambda$, are our disk observables. Other parameters include $M_{\star}$, $\lambda$, r, $Q ^{\star} _{D}$, $t_{age}$, and $f_{\tau}$, which correspond to the stellar mass, observation wavelength, the radii at which there is a break in the surface brightness profile, specific energy, age, and an uncertainty factor.
However, HR 8799 has no constraints on the scale height because the disk is not resolved. We therefore adopt the $\beta$ Pictoris estimate for $h$ by @quillen07. $\beta$ Pictoris is also an A star which has a similar age and mass to HR 8799 as well as an extended debris disk. Estimates for the dust mass around $\beta$ Pictoris have been found to be $7.8M_{Moon}$, or about $0.096M_{\oplus}$ [@holland98]. This is similar to the $0.12M_{\oplus}$ suggested by @su09 for HR 8799. @su09 also notes that the amount of excess emission in the HR 8799 disk is similar to that of other debris disks around A stars (See @su06’s study of the evolution of debris disks around A stars.) See table 1 by @quillen07 and references therein for mass, age, and other parameters for $\beta$ Pictoris.
We note that our biggest uncertainty is therefore in $h$ in the above formula, which goes as $\frac{10}{3}$. This makes even small errors in estimates of $h$ have a large impact on $a_{top}$. However, the goal is not to determine the exact mass of the disk, only to determine what a reasonable range is.
To find $a_{top}$ we must also compute the opacity at a specific wavelength. The opacity can be measured by modifying our total dust mass integral. The opacity at a specific wavelength is a measure of the fractional area covered by particles of radius $a$ (i.e. the opacity depends on the number of particles per unit area times the cross sectional area). In this way we can relate the number of particles of a specific radius to the opacity and total disk surface area. Secondly, it is possible to equate the differential number distribution to $N_{a}$ by $$\frac{dN(a)}{dln(a)} \equiv N(a) ,$$ [@quillen07]. Thirdly, we make use of a relation which describes how the opacity scales with the radius of the object, $$\tau_{(a)} = \tau_{d} \left( \frac{a}{a_d} \right) ^{3-q}$$ where $q = 3.5$ is equivalent to $\alpha = \frac{11}{6}$ from above and $a_d$ and $\tau _d$ are the radius and opacity at a specific radius/wavelength [@quillen07]. Substituting the above three relationships into our mass computation yields an integral that is dependent only on the radius of the dust. Using $1000 \mu m$ dust particles as the largest grains we find that the opacity at the specific dust particle radius is $$\tau _d \sim \frac{M_d}{A}\frac{1}{\rho a_d}$$ where $M_d$ is the total mass of the dust, $A$ is the total disk area, $\rho$ is the density of the dust grains, and $a_d$ is the radius of the largest dust grain in the model. Using dust masses from @su09 as well as a disk area computed from the inner and outer edges and a maximum grain size from the same, along with a density appropriate for dust grains $\rho = 1.5-3.0 \frac{g}{cm^3}$, we compute a value for the opacity of $\tau_d \sim 10^{-4}$.
With these computed disk observables along with the other known parameters for HR 87999 we find an $a_{top}$ value of $a_{top} \sim 1 km$. This $a_{top}$ yields a disk mass estimate of $M_{disk} \sim 150 M_{\oplus}$ or about one half of a Jupiter mass. Large debris disks are often attributed masses in the region of $50-100M_{\oplus}$.
While this estimate would place the disk mass usable in our simulations at of order a Jupiter mass or less, there are a number of caveats to our estimates. We note that our estimate for $a_{top}$ is low compared to other disks by @quillen07 which place the radius at values anywhere from a few to a few hundred times larger. Using an estimate for $a_{top}$ in line with those for $\beta$ Pictoris and similar disks would yield a significantly more massive disk. Also, we recognize that the masses of the planets are extreme, with four planets of ten Jupiter masses each. It is not inconceivable that the disk may be significantly more massive than those observed previously. We also note that planetesimals with a radius of $a_{top}$ are the largest objects that contribute to the collisional cascade. Significant mass could be found in other, more massive objects that would have collision times too long to contribute to dust production. Finally, we note that the major phenomenon discussed in the paper occur at masses of one Jupiter mass or less in the debris disk. It is not required for us to include much of the more massive disks, however, given the difficulty of estimating the mass and the general peculiarity and size of HR 8799, we have included fairly generous debris disk masses.
A second more generic estimate of the largest sized objects that can be grown was found by @cuzzi93. Using a numerical model which uses the Reynold’s averaged Navier-Stokes equations with both turbulence and full viscosity, they found that it was possible to create 10-100km sized planetesimals. This would place our disk mass estimate of order $\sim M_J$.
Finally, we can use our own solar system as a reference point. The ‘Nice’ model required approximately 30-50 $M_{\oplus}$ to explain the outward migration and eventual configuration of our own system [@tsiganis05]. This corresponds to a few tenths of a Jupiter mass of debris.
Total debris disk masses used in our simulations varied, but had measurable effects at a Neptune mass, or about 17 $M_{\oplus}$. This puts those simulated disk masses in a similar range to that of the ‘Nice’ model. Larger masses than the ‘Nice’ model are justified by way of the discussion above. Given the uncertainties in measuring $a_{top}$, it is very difficult to determine if the largest total masses can be used in our simulations.
The total particle number used in our simulations, 1024, is also comparable to the 1000-5000 particles used in the ‘Nice’ model simulations. Above we noted that the diameter of objects at the top end of the collisional distribution is anywhere between one to several hundred km. An $a_{top}$ value similar to $\beta$ Pictoris of 180km is about 6.5 times smaller than the diameter of Pluto. $a_{top}$ values near 1 km would be approximately one thousand times smaller. This suggests that our simulated disks which are made up of at least a Neptune mass in debris (those that are massive enough to have measurable effects on the lifetime of the system) would have planetesimals that are larger than those that are predicted by the collisional cascade. Therefore, while our planetesimal masses are close to those used in the ‘Nice’ model, they could be more massive than those in HR 8799’s actual disk. This may result in more stochastic interactions between the planetesimals and planets. However, the planetesimal masses suggested by the collisional cascade are a distribution which could reasonably involve both smaller or larger planetesimals then those which we have suggested.
Integrator
==========
All simulations were run with the software package *QYMSYM*[^1]. *QYMSYM* is a GPU-accelerated hybrid second order symplectic integrator which permits close encounters similar to the *Mercury* software package developed by @chambers99. Like *Mercury*, *QYMSYM* uses a second order symplectic integrator to advance the positions of all particles in a simulation in the typical manner described by @duncan98 and @levison94. Exploratory work on symplectic maps for N-body integrators was elucidated by @wisdom91. Additional analytical work on the formulations of the symplectic integrator can be found by @saha92 and by @yoshida90 [@yoshida93]. Unlike these first symplectic integrators but similar to *Mercury*, *QYMSYM* flags any particle from an integration when it is deemed that it has a close approach to another massive particle during a time step. These are then integrated separately with an adaptive step-size conventional integrator. The criterion for closest approach is decided based on the Hill radii of the interacting objects. *QYMSYM* uses the $4^{th}$ order Hermite integration scheme detailed in @makino92 rather than the Bulirsch-Stoer algorithm used in *Mercury*. See [@moore11] for more details on the *QYMSYM* integrator.
We note that due to the nature of our CUDA based code, it is not possible to run less than one block worth of particles, even in the case where we wish a massless debris disk. Furthermore, due to memory latencies, low particle counts are not run efficiently on the GPU. Currently, we do not have the ability to disable the $O(N^2)$ kick computation or a way to offload low particle counts to the CPU. This is something that could be rectified in future revisions of the code. Additionally, because of the compilation process, our code is restricted to operating on machines which have GPU’s attached.
These two caveats have an effect on the number and size of the parameter space that we are able to explore.
The inefficiency of the GPU at low particle count and single block restriction has the impact of reducing the length for which we can integrate our debris-less test case. We do not want to run multiple integrators. Two hybrid-symplectic integrators will not get identical answers unless the collision integration method is the same. Rather than use another hybrid-symplectic integrator (like *Mercury*) to run debris-less test cases, we continue to use our own code. However, this does make it difficult to run simulations with as many orbital periods as those possible with integrators which can be run with very low particle counts.
The GPU requirement also makes it more difficult to test a wide parameter space. A node must be equipped with a GPU in order to be able to run our code. The GPU clusters that we have access to are much smaller than the CPU clusters that are available. This limits the number of possible trials.
Last, we note that our integrator is truly $O(N^2)$ - all particles feel the effects of all others. This is unlike many integrators available in the field which typically do not compute planetesimal-planetesimal effects. This makes our simulations more precise, but does increase the arithmetic intensity.
All simulations were run on either dual or quad core Intel Core2 Duo architecture CPU’s with either GT200 of GF100 architecture NVIDIA GPU’s, both of which are CUDA capable double precision architectures. The code can easily be optimized to run on any Linux kernel 2.6+ distribution with appropriate GPU hardware.
Accuracy, integration parameters and eccentricity corrections
-------------------------------------------------------------
Timestep sizes were set to 0.016 (out of a possible 2$\pi$ orbit) corresponding to 42 days in simulations including HR 8799e and 93 days in simulations without. We used smoothing lengths which correspond to radii at least a few times larger than the size of the planets assuming a planet and planetesimal density of the order of $1 g/cm^3$ and disregard any simulation during which the energy conservation (given by $\Delta E/E$) dips below $1.0 \times 10^{-4}$. Additionally, we are only interested in the time to the onset of first instability, and do not need to concern ourselves with larger energy errors that may occur after a planet-planet interaction or ejection.
In a few simulations, very high eccentricities were detected for planetesimals which had been ejected from the system during a close approach. Because our integrator explicitly conserves angular momentum through the use of f and g functions when solving Kepler’s equations (see @moore11 for more details), these large eccentricities and their corresponding high velocities produced a drift in the location of the center of momentum. This is most likely due to the way a hybrid symplectic integrator typically updates the particles positions. If a particle has its position updated such that it is now in close proximity to another particle, it would feel a large force. Typically this force is removed by reverse integration of the Hamiltonian and the particle is then moved to a more accurate integration regime (the Hermite integrator in our case). However, if by chance that particle is updated to be very close to another, the assumptions made in the perturbation theory used to split the Hamiltonian into components, namely that Keplerian motion is dominant over interparticle forces, no longer applies. This means that the particle will not be correctly reverse integrated. This problem is common to many symplectic integrators.
Once a planetesimal has been ejected with a very high eccentricity this phenomenon is more likely to repeat. The high eccentricity means that its angle of impact will be nearly perpendicular to the motion of the other orbiting objects exterior to its current position. This newly large velocity relative to the size of the collision detection criteria makes it more likely for a particle to be updated from outside the collision detection criteria to be in close proximity to the colliding particle without being previously removed.
Including a larger collision detection criterion or reducing the timestep size of the simulation could resolve this problem, but at greatly increased simulation time. Decreasing timestep size goes directly with total simulation time. Due to the high velocities achieved by the particles, it would need to be decreased significantly. Increasing the collision detection radius, which is some factor of the Hill radius, will force more particles to be offloaded to the significantly slower Hermite integration routines. Not only is the Hermite integrator $O(N^2)$ (and therefore suffers from non-linear increases in simulation time based on increasing particle count), but the number of particles included in the Hermite integrator will increase by a factor that goes with the collisional volume - even small increases in the collision detection criteria can lead to significant increases in particle counts. Alternatively, increasing the smoothing length can partially alleviate this problem but has the negative effect of smoothing out many of the important planetesimal/planetesimal dynamics.
To correct for this occasional error without greatly increasing simulation time, we wrote an additional check that could remove planetesimals which had eccentricities above an arbitrary threshold or those that collided with the star. Additionally, particles which are no longer bound and have a high enough semi-major axis are removed. Energy error checks are common in comparable literature (see appendix of @raymond11 for an example) as is planetesimal removal due to ejection.
This check will have no effect on simulations in which planetesimals are not ejected or ejected with realistic velocities and eccentricities that are still close enough to interact or collide with another particle. Only simulations which have particles with extremely large eccentricities will be quantitatively altered. As we would expect, in simulations where these particles are removed, it appears that the dynamics are qualitatively identical, only without the aforementioned drift in center of momentum. We also note that the center of mass is conserved to a high degree of accuracy and we maintain a satisfactory level of energy conservation throughout either type of integration. Conservation of center of mass is typically on the order of one part in $10^{12}$ or better. We therefore present the results of the corrected and uncorrected simulations together - using uncorrected simulations when no drift is detected in their outputs and corrected versions elsewhere.
The above timestep choice and numerical energy integration error are comparable to those reported by @fabrycky10.
Simulation configurations of HR 8799
------------------------------------
We set up two initial configurations to test in our simulations detailed in each respective section. Generally, given the importance of the 4:2:1 MMR in previous work, our simulations begin with this planetary configuration and have an additional debris disk or alternatively use a somewhat modified version of this planetary configuration. Minimally, the simulations have the innermost two planets in a 2:1 MMR.
Three planet (b,c,d) plus debris disk simulations were based off of the stable configurations discussed by [@fabrycky10], using planet masses of 10, 10 and 7 $M_{Jup}$ for planets d, c and b. Four planet (b,c,d,e) plus debris disk simulations are based off the predicted orbital elements found by [@marois10] and use masses of 7, 7, 7 and 5 $M_{Jup}$ for planets e, d, c and b. Our debris disk is simply a uniform distribution of 1024 equal mass particles. Depending on total disk mass desired, the planetesimal mass was modified accordingly.
This later set of initial positions and masses reflects the most current observational results by @marois10 by including the fourth planet and the corresponding reduced masses required from the dynamical simulations that were run. The three planet configuration, while no longer representative of those more recent observations and simulations, is both practical and illuminating for several reasons.
First, we note that the planet which is not included in the three-planet configuration is the most recently discovered innermost object. Unless its presence destabilizes the system, we expect its effect on the migration rate of the outermost planet to be small in those simulations due to the proximity and mass of the debris disk to the outermost planet. Additionally, we only use the three planet configuration for simulations in which the system is arbitrarily placed in an unstable configuration from the beginning. The four planet configuration has much smaller regions of stability. While the addition of the fourth planet would make it even easier for us to create an initially unstable configuration which shares orbital elements similar to those that are observed, it makes it significantly more unlikely for a system to move from an unstable region to a stable region via orbital migration. Last, we recall that the recent work by [@soummer11] which reduced Hubble Space Telescope observations suggests best orbital fits for HR 8799b,c,d that coincide with the 1:2:4 mean motion resonance. This data agrees with previous work that fits only the three planets, but does not agree with dynamical simulations which included all four planets [@marois10]. @marois10 found that the innermost three planets, HR 8799c,d,e were the most likely planets to share mean motion resonances with b excluded. There is as of yet no consensus on the possible orbital elements or dynamical structure of HR 8799 other than the most long-lived systems having the innermost three planets in a 4:2:1 mean motion resonance or at least having the innermost two planets in a 2:1 mean motion resonance. Therefore we assume that the 2:1 mean motion is the primary initial requirement for our simulated systems.
Because we are measuring the effects of planetary migration on system stability to determine if all unstable configurations remain so, it is useful to run the three planet system - it has larger regions of stability for the planets to migrate into. Due to the computational intensity of the simulations and their corresponding time requirements, we were only able to run on the order of hundreds of simulations over a few months rather than tens or hundreds of thousands. Given this limitation, it proved difficult to migrate large numbers of four planet configurations into regions of stability. This lower simulation number and corresponding low number of stabilized configurations would introduce a fine tuning problem to our analysis so we do not discuss those in greater detail, instead focusing on the three planet configurations for the stabilization through migration scenarios. This fine tuning issue could potentially be resolved via more stable initial conditions which would require significantly less planet migration to move from regions of instability to regions of stability. This issue is discussed in further detail in our results section.
In the three planet configurations the inner disk edge has a semi-major axis of $a_{min}$ = 2.5 and an outer disk edge of $a_{max}$ = 7.0, corresponding to separations of 60 and 170 AU. In the four planet configurations the inner disk edge has a semi-major axis of $a_{min}$ = 6.14 and an outer disk edge of $a_{max}$ = 20.69, corresponding to separations of 90 and 300 AU. This second set matches the estimated outer debris disk’s observed inner and outer edges.
The inner disk edge with reduced semi-major axis for the three planet configuration was used to encourage rapid planet-planetesimal crossings and therefore rapid migration. As mentioned previously, we forced our three planet configuration to become unstable on very short timescales. Because of this, we require rapid migration in order for any outcome other than planet-planet scattering to be a possibility. The inclusion of a debris disk at a position more coincident with that which is observed would be possible with either faster or more GPUs. This is because configurations that are stable on longer timescales - which allow for reduced amounts of planet migration to be necessary in order to stabilize an unstable configuration - could be used. We could additionally begin to experiment with the inclusion of the fourth planet as well as keep the outermost planet at its observed position rather than moving it in to encourage the system to become unstable. We found both the removal of the innermost planet and movement of the inner disk edge were helpful in finding these post-migration stable configurations.
Other simulation parameters include the Hill factor for encounter detection and the K function (see @moore11), both of which are set to 2.0. The K function is an arbitrary function which weights certain part of the Hamiltonian to be integrated in order to allow it to be broken up into evolutions operators which are otherwise not possible. When the distance between two particles is large, the value of K goes to zero while when the distance between two particles is small, the value of K goes to 1 (or vice versa). When K or (1-K) are multiplied into the respective separated Hamiltonians described in [@moore11], it prevents either from becoming too large and breaking the perturbation theory used to separate them. K is a function of Hill radii, but was created by trial and error.
The distribution of planetesimal semi-major axes is flat with probability independent of a within $a_{min}$ and $a_{max}$. The initial eccentricity and inclination distributions were chosen using Rayleigh distributions with the mean eccentricity $\overline{e}$ equivalent to twice the mean value of the inclination $\overline{i}$ and $\overline{i}$ = 0.01. The initial orbital angles (mean anomalies, longitudes of pericenter and longitudes of the ascending node) were randomly chosen.
A group of 150 simulations with three planets were run, 58 using the eccentricity correction while 92 without. Disk mass was varied from $1.0 \times 10^{-30}$ to $10$ $M_{Jup}$, although we only present results from a disk mass of $1.0 \times 10^{-3} M_{Jup}$ and larger here. 18 simulations with four planets were run, half using the eccentricity-correction while the other half without. These simulations differed only in the initial conditions of the planetesimals.
Simulations of initially unstable planetary orbital architectures
=================================================================
Does the addition of a relatively massive planetesimal disk allow for an increase in stability timescale? To answer this question we created an unstable initial configuration for both the three and four planet configurations of HR 8799. In either case, this was done by starting with actual observed orbital elements found by @fabrycky10 and @marois10 and reducing the outermost planet’s semi-major axis in small increments until the system was unstable on the time scale of thousands of orbits or less. In the three planet simulations, this gives an initial configuration with all three planets having orbital elements the same as those reported by @fabrycky10 except that the outermost planet’s semi-major axis was reduced by $14\%$. For the four planet simulations, the planets’ orbital configuration was identical to those found by @marois10 but the outermost planet has a semi-major axis reduced by $8\%$. This arbitrary and largely unstable configuration was chosen to both reduce simulation time as well as allow for pronounced effects by migration. Due to the previously mentioned limitation in the number of total simulations it is possible to run and the extremely limited regions of stability available to the four planet configuration, we opted to run far more three planet simulations than four planet systems and present only that data.
Simulations both with and without eccentricity corrections are presented simultaneously and see no significant difference between them.
Results
-------
In figure \[fig:stabilizing\] we plot the difference between the stability time for each simulation and that of a system lacking a debris disk. The stability time is the time to first planet-planet encounter. The time difference is plotted as a function of disk mass. Simulations are run for a maximum of 5 Myrs. If no encounters had taken place in that time, they are plotted as upper limits in the figure.
In figure \[fig:3p-migration\] we plot the evolution in time of the semi-major axis of the three planets for a sample simulation. In this case the total disk mass simulated was $1.6 M_{Jup}$.
{width="150mm"}
![In this figure we show the evolution in semi-major axis of an example three planet configuration over time. a) Shows the evolution in time of the semi-major axis of the three planets over the entire 5 Myr simulation. b) Shows the evolution in time of the semi-major axis of the three planets over the first Myr. In b) we also include error bars that indicate the maximum and minimum separation (given by a(1+e) and a(1-e) respectively) of the planets given their eccentricity. This second image more clearly shows the migration of the outermost planet planet from its initial position at 58.6 AU to approximately 65 AU over the first few hundred thousand years. In this example the total disk mass was $1.6 M_{Jup}$ and the configuration remained stable over the entire 5 Myr simulation.[]{data-label="fig:3p-migration"}](3p.eps "fig:"){width="85mm"} ![In this figure we show the evolution in semi-major axis of an example three planet configuration over time. a) Shows the evolution in time of the semi-major axis of the three planets over the entire 5 Myr simulation. b) Shows the evolution in time of the semi-major axis of the three planets over the first Myr. In b) we also include error bars that indicate the maximum and minimum separation (given by a(1+e) and a(1-e) respectively) of the planets given their eccentricity. This second image more clearly shows the migration of the outermost planet planet from its initial position at 58.6 AU to approximately 65 AU over the first few hundred thousand years. In this example the total disk mass was $1.6 M_{Jup}$ and the configuration remained stable over the entire 5 Myr simulation.[]{data-label="fig:3p-migration"}](3p_migration.eps "fig:"){width="85mm"}
Figure \[fig:stabilizing\] shows that at low disk masses the difference in lifetime is near 0. Only massive disks with masses of ten percent or more of the outermost planet can remain stable until the end of the simulation. Given enough mass, migration via planetesimal scattering can take a tightly packed configuration of HR8799 b, c, d and e, or in the case of figures \[fig:stabilizing\] and \[fig:3p-migration\], simply b, c and d, and migrate the outermost planet into a more stable configuration. The total migration for the outermost planet in system configurations which remained stable over the 5 Myr simulation ranged from 5 to 15 AU in semi-major axis depending on total disk mass. This migration occurred over a few hundred thousand years.
If stable regions are small compared to unstable regions, migration is unlikely to put an initially unstable configuration into a stable region. Only large migrations could significantly increase the stability time. This is consistent with what we see; rapid migration caused by a massive disk is able to significantly affect the lifetime of our simulations. Migration is expected to reduce planetary eccentricities and this could increase stability [@lissauer93; @fernandez96]. We searched for signs that the outermost planet’s eccentricity decreased during migration. In figure \[fig:3p-migration\] we see oscillations in eccentricity of the outermost planet were large throughout the simulation due to the proximity and high masses of the inner planets. We therefore attribute the increased stability time to the wider separation rather than any eccentricity damping.
As a planet migrates outwards it can cross mean motion resonances with other planets. For simulations with short lifetimes, we searched for evidence that resonant crossing caused instability. We examined any relevant $1^{st}$, $2^{nd}$, $3^{rd}$ and $4^{th}$ order resonance possible between the outer planet and innermost planet as well as the outer planet and the middle planet. We saw no large semi-major axis or eccentricity jumps caused by resonant crossing so they are unlikely to be the cause of instability in the simulations with shorter lifetimes. Resonances near the outer planet’s position are primarily $2^{nd}$ or $3^{rd}$ order and cause smaller eccentricity jumps due to the weaker dependence on mass [@quillen06].
As discussed in our section on simulation parameters, our choice of unstable configurations was done in part to reduce the amount of simulation time. With an initial planetary configuration so far from a region of stability, a large amount of migration is required in order to move the planet to a more stable region of phase space. Because of this, only a massive debris disk can migrate the outermost planet sufficiently far to increase the lifetime.
Another concern with this set of simulations is that with a planetesimal count of 1024 and a total disk mass of at least a Jupiter mass, interactions between planetesimals and planets would cause stochastic migration [@zhou02]. We would have expected such large planetesimal masses to cause instability rather than increase the lifetime of the system. Therefore, we have no reason to suspect that stochastic behavior is a problem. Because the typical disk mass required to stabilize the system is unrealistically large there may be no need to resolve the disk more accurately. As mentioned previously, an approximate upper limit for debris mass is thought to be on the order of $10^{-3} M_{\odot}$, or approximately one Jupiter mass. For HR 8799, this is effectively the minimum amount of mass required for any observable effects in our simulations.
Simulations of initially stable planetary orbital architectures
===============================================================
In this second set of simulations we examine the impact of a planetesimal debris disk on the stability time of a configuration of planets which was stable for about 150k orbits (or around 7 Myr) sans debris disk. In these simulations we begin with all four planets situated in an orbital configuration based off the current estimated positions by @marois10. This four planet configuration is generally extremely unstable unless situated in a 4e:2d:1c MMR. As before, smoothing lengths are set to sizes a few times the radii of the planet to prevent the unrealistic forces possible if two particles form a very tight binary or collide within what would have been considered an approximate planetary radius. Energy error is typically even lower in these simulations with a $\Delta E / E$ at $1.0 \times 10^{-5}$ or $1.0 \times 10^{-6}$ due to reduced planetesimal masses. We stopped simulations after their first planet-planet encounter regardless of the eventual fate of the interacting planets. Stability time in the histogram is the time to first planet-planet encounter.
We began by simulating the four planets with an initial orbital configuration comparable to that by @marois10 with massless debris and plotted the resonant arguments. We see constrained resonant angles which indicate the presence of a 2:1 MMR between the inner two planets, shown in figure \[fig:angles\]. While @marois10 did not directly show the resonant argument to illustrate that HR 8799d and e were in resonance, our plot is similar to Figure 10 by @fabrycky10 although for HR 8799e and d rather than d and c. The integrator confirms that this planetary configuration is in resonance as suggested by @marois10.
Matching the exact age found by @marois10 is difficult despite the fact that there is no longer a requirement of planetesimals with mass. This is due to the way the integrator operates. Specifically, it is not possible to integrate less than a certain number of particles, depending on compile time and hardware restrictions. For more details on the exact nature of the code, we refer the reader to @moore11. However, while the control simulation is not as long lived as those found by @marois10, it is still stable on long time scales ($10^5$ orbits).
![Here we show the resonant angle for 2:1 mean motion resonance for HR 8799d and e as function of time. a) For a simulation without a debris disk. b) For a simulation with a debris disk that has a mass of 1/100th the outer planet. The initial conditions for the planets were stable for 7 Myrs. In b) the extreme values of $\phi$ slowly begin to increase over the length of the simulation. This effect is not present in a). The resonant island is so small that even a disk mass of $1\%$ the outer planet is capable of effecting the systems dynamics. The resonant angle plotted is given by $\phi _e = 2 \lambda _d - \lambda _e - \varpi _e$.[]{data-label="fig:angles"}](res1221_1.eps "fig:"){width="85mm"} ![Here we show the resonant angle for 2:1 mean motion resonance for HR 8799d and e as function of time. a) For a simulation without a debris disk. b) For a simulation with a debris disk that has a mass of 1/100th the outer planet. The initial conditions for the planets were stable for 7 Myrs. In b) the extreme values of $\phi$ slowly begin to increase over the length of the simulation. This effect is not present in a). The resonant island is so small that even a disk mass of $1\%$ the outer planet is capable of effecting the systems dynamics. The resonant angle plotted is given by $\phi _e = 2 \lambda _d - \lambda _e - \varpi _e$.[]{data-label="fig:angles"}](res1221_1_disk.eps "fig:"){width="85mm"}
This simulation is then compared to 18 simulations with Neptune mass disks. These 18 simulations are identical in terms of planetesimal mass, total disk mass, initial outer and inner disk edge, and initial planetary orbital configuration. However, the planetesimals initial orbital configurations have been randomized 18 different ways.
The histogram shown in figure \[fig:histogram\] shows the distribution in lifetimes of the eighteen simulations each with a Neptune mass disk. Figure \[fig:4p\] shows the evolution in time of the semi-major axis of all four planets of a sample simulation. The error bars indicate the minimum and maximum separations.
![Here we show a histogram of time to the onset of instability, which we define as planet interaction or ejection. The disk mass is about one Neptune in mass. The planets are initially in a configuration that is stable for approximately 7 Myrs. We see that the distribution of lifetimes is very broad and many of the simulations had lifetimes less than this. The reduction in lifetimes was as large as $\sim6$ Myrs.[]{data-label="fig:histogram"}](histogram.eps){width="85mm"}
![In this figure we show the evolution in time of the semi-major axis of an example 4 planet configuration. This particular simulation became unstable in a little over 3 Myr. We include error bars that indicate the maximum and minimum separation (given by a(1+e) and a(1-e) respectively) of the planets given their eccentricity. In all four-planet simulations the total disk mass was equivalent to Neptune.[]{data-label="fig:4p"}](4p.eps){width="85mm"}
Results
-------
In figure \[fig:histogram\] we see that in 16 of 18 simulations the stability time has decreased by at least 2 Myrs from the initial ‘stable’ configuration which had its first planet interactions after 7 Myrs. We see a broad distribution of lifetimes. Even though a Neptune mass disk is only about $1\%$ the mass of the outer most planet it has an effect on the stability time. With a Neptune mass disk each simulated planetesimal has a mass of about 8 Plutos. Debris disk models estimate the top of the collisional cascade to contain similar (but slightly smaller) size bodies. For example, @wyatt02 argued for a distribution of planetesimals in the Fomalhaut system with a range of sizes between 4 and 1000km. As discussed previously, this second value roughly coincides with Pluto’s diameter within a factor of a few of what we used in our simulations. Stochastic perturbations are expected in real systems; however they will be somewhat smaller than those present in our simulation do to our factor of 8 difference in mass of planetesimals. Figure 3 by @gozdziewski09 demonstrates why a planetesimal disk with a total mass around that of Neptune can have such a pronounced effect on the stability time. In this plot, dark regions indicate highly likely configurations of eccentricity and semi-major axis for planet d, whereas yellow regions indicate strongly chaotic systems. Here we see that semi-major axis changes of less than 0.1 AU can move planet d from a stable region to a strongly chaotic one. Similarly, minor changes in eccentricity can also have a large effect. @fabrycky10’s Figure 7 shows a similar behavior. Note that a change in current separation of planet d by 0.05 in the figure (corresponding to semi-major axis change of 1.22 AU) can decrease the stability time by as much as four orders of magnitude. A three order of magnitude change in stability time is possible with a semi-major axis increase of approximately half an AU. In either example the regions of stability have sharp edges between stable and chaotic solutions.
We would expect a disk mass of $1\%$ the mass of a planet to be able to migrate a planet roughly $1\%$ of the semi-major axis if a majority of its angular momentum is transferred to the planet via scattering. If this were the case than a Neptune mass disk would be able to migrate the outer planet (which has a mass of $5 M_{Jup}$ at 68 AU) approximately 0.5 AU. This is more than enough to vary the lifetime by many orders of magnitude. Therefore we looked for the number of planetesimals which had become orbit crossing by the time each simulation had become unstable. In all 16 simulations, at least $15\%$ of the disk mass was orbit crossing. This corresponds to a minimum of 2.5 $M_{\oplus}$. By starting a simulation near the edges of a region of stability, we see that even lower mass debris disks can affect the stability time.
Pulling HR 8799’s planets out of resonance
------------------------------------------
We continuously monitor the planets’ orbital elements throughout all simulations to see if they are in resonance with one another. We do this by noting whether or not the resonant angle librates around 0 or $\pi$. We looked for $1^{st}$, $2^{nd}$, $3^{rd}$ and $4^{th}$ order mean motion resonances that could exist near each of the planets’ semi-major axis at varying times. We did not observe any strong two body resonances beyond the 2:1 MMR between e and d, a much weaker 2:1 MMR between d and c, and no clear resonant angles for b. We attribute this to the overwhelmingly large gravitational perturbations caused by the massive planets as well as the fact that most of the nearby resonances for the planets were of higher order. This prevented behavior similar to the ‘Nice’ model during which migration causes resonance crossing and corresponding eccentricity increases. Finally, we see that in all simulations when the planets e and d are pulled out of the 2:1 MMR the system rapidly becomes unstable.
In figure \[fig:angles\]a we see that the resonant angle for the 2:1 MMR between HR 8799e and d over the 7 Myr simulation remains largely unaffected until the end of the simulation. However, in figure \[fig:angles\]b we note that the extreme values of $\phi$ begin to increase over time, leading to e and d moving out of resonance and the rapid disruption of the system. In this particular example, the system’s lifetime is decreased by about 50$\%$. It is possible the debris disk is responsible for pulling HR 8799 out of resonance and so, reducing its lifetime.
Conclusion
==========
In this paper we discuss how a planetesimal debris disk can effect the stability of the multiple planet system HR 8799. Two questions are considered; whether it is possible to destabilize a stable planetary configuration with a planetesimal debris disk and conversely, whether it is possible to stabilize an unstable configuration with a planetesimal debris disk. We examine both three planet (b,c,d) and four planet configurations (b,c,d,e).
In three planet configurations which were unstable on short timescales without a debris disk, only massive debris disks ($10\%$ the mass of the outermost planet) could cause increases in stability time. In four planet configurations which were stable over long timescales without a debris disk, debris disks of only a Neptune in mass ($1\%$ the outer planet) can cause large decreases in lifetimes. We attribute this sensitivity to the small size of HR 8799’s resonant region of stability.
While the amount of disk mass required for system stabilization in our simulations is unrealistic, it is only required to be this large due to the initial conditions. In order to run a large number of simulations, a highly unstable configuration that rapidly had planetary encounters was required. Given an initial configuration which is just outside a region of stability, it would be possible to cause sufficient migration to allow an unstable configuration to become stable with a more reasonable disk mass.
Similarly, while the disk properties required for system disruption in our simulations are reasonable in both total mass and planetesimal size, the mass distribution for the planetesimals is not a true mass distribution. In order to keep total particle number down we restrict all planetesimals to have an identical 8 Pluto sized mass. We do not believe that this had a large effect on the distribution of stability times of the system when a disk is included because a realistic mass distribution could encourage even more stochastic migration if there are even a few planetesimals of greater mass.
These results suggest that HR 8799 may be destined for eventual planet scattering. The very high masses of the planets causes regions of stability to be relatively small and makes it possible for low mass debris disks to pull the system out of its mean motion resonances and induce instability. We also see that migration is much more likely to make the system unstable than migrate the system to a region of stability. It may be the case that the debris disk has already been responsible for removing the system from a maximally stable region and is currently near the boundary of instability. All of these possibilities tend to induce instability rather than stability and suggest that HR 8799 may be headed towards instability, possibly sooner than would otherwise be predicted.
In this study we have used initial conditions consistent with observed positions of the planets when attempting to determine whether a stable orbital configuration could be made unstable by a debris disk. As we have shown here, a low mass disk can cause evolution of the planetary configuration. Consequently, in the past the planets could have been in a different configuration. By exploring different initial conditions future investigations could explore scenarios for the past evolution of HR 8799 that would be consistent with its current configuration. For example, if the planetesimal disk is causing the system to move out of resonance, in the past it may have been in a more stable region of this resonance, instead of on its boundary. Increasing simulation length and particle count could also increase the resolution of our simulations.
Alternatively, it appears that similar simulations could be used to put upper limits on planetesimal debris disk mass if it is assumed that the debris disk has a negligible dynamical effect. In the case of our simulations of HR 8799, planetesimal disk masses would need to be much smaller than one Neptune in mass.
The stability timescale for an HR 8799-like system is very sensitive to small alterations in planetary orbital elements. This sensitivity implies that the effects of planetesimal debris disks may not be negligible for systems with such small regions of stability. However, dynamical simulations of HR 8799 sans planetesimal disk were used to determine the resonant structure of the system and constrain observations while not including all the required dynamics. The assumption of long term maximum stability may be erroneous when debris disks could allow for migration of planets from currently observed stable regions to unstable configurations or vice versa. In general, it would appear that using dynamical models which do not include the effects of planetesimals to constrain observations may be unwise when a system is known to have such limited regions of stability.
Additionally, the long term stability of systems with lower mass disks could still be important. In these systems the planets will typically be smaller and the regions of stability would be larger. However, reduced planet mass would increase the migration rate.
Beckwith S. V. W., Sargent A. I., 1996, Nature, 383, 139
Borucki W. J., Koch D. G., 2011, IAU Symposium, 276, 34
Chambers J. E., 1999, MNRAS, 304, 793
Cuzzi J. N., Dobrovolskis A. R., Champney J. M., 1993, Icarus, 106, 102
Dohnanyi J. S., 1969, J. Geophys. Res., 74, 2531
Duncan M. J., Levison H. F., Lee M. H., 1998, AJ, 116, 2067
Fabrycky D.C., Murray-Clay R.A., 2010, ApJ, 710, 1408
Fernández J. A., Ip W.-H., 1996, P&SS, 44, 431
Goździewski K., Migaszewski C., 2009, MNRAS, 397, 16
Holland W. S., Greaves J. S., Zuckerman B., Webb R. A., McCarthy C., Coulson I. M., Walther D. M., Dent W. R. F., Gear W. K., Robson I., 1998, Nature, 392, 788
Hughes A. M., Wilner D. J., Andrews S. M., Williams J. P., Su K. Y. L., Murray-Clay R. A., Qi C., 2011, ApJ, 740, 38
Ida S., Bryden G., Lin D. N. C., Tanaka H., 2000, ApJ, 534, 428
Kenyon S. J., 2002, PASP, 114, 265
Kenyon S. J., Bromley B. C., 2001, AJ, 121, 538
Kenyon S. J., Bromley B. C., 2002, AJ, 123, 1757
Kenyon S. J., Luu J. X., 1999, AJ, 118, 1101
Kirsh D. R., Duncan M., Brasser R., Levison H. F., 2009, Icarus, 199, 197
Levison H. F., Duncan M. J., 1994, Icarus, 108, 18
Lissauer J. J., Stewart G. R., 1993, in Levy E. H., Lunine J. I., eds, Protostars and Planets III. Univ. Arizona Press, Tuscon, p. 1061
Makino J., Aarseth S. J., 1992, PASJ, 44, 141
Marois C., Macintosh B., Barman T., Zuckerman B., Song I., Patience J., Lafrenière D., Doyon R., 2008, Science, 322, 1348
Marois C., Zuckerman B., Konopacky Q. M., Macintosh B., Barman T., 2010, Nature, 468, 1080
Marshall J., Horner J., Carter A., 2010, ASB, 9, 259
Moore A., Quillen A., 2011, NewAST, 16, 445
Morbidelli A., 2001, Ann. Rev. Earth and Pl. Sci., 30, 89 Moro-Martín A., Malhotra R., Bryden G., Rieke G. H., Su K. Y. L., Beichman C. A., Lawler S. M., 2010, ApJ, 717, 1123
Quillen A. C., 2006, MNRAS, 365, 1367
Quillen A. C., Morbidelli A., Moore A., 2007, MNRAS, 380, 1642
Raymond S. N., Armitage P. J., Moro-Martín A., Booth M., Wyatt M. C., Armstrong J. C., Mandell A. M., Selsis F., West A. A., 2011, A&A, 530A, 62R
Reidemeister M., Krivov A. V., Schmidt T. O. B., Fiedler S., Müller S., Löhne T., Neuhäuser R., 2009, A&A, 503, 247
Safronov V. S., Zvjagina E. V., 1969, Icarus, 10, 109
Saha P., Tremaine S., 1992, AJ, 104, 1633
Soummer R., Hagan J. B., Pueyo L., Thormann A., Rajan A., Marois C., 2011, ApJ, 741, 55
Su K. Y. L., Rieke G. H., Stansberry J. A., Bryden G., Stapelfeldt K. R., Trilling D. E., Muzerolle J., Beichman C. A., Moro-Martín A., Hines D. C., Werner M. W, 2006, ApJ, 653, 675
Su K. Y. L., Rieke G. H., Stapelfeldt K. R., Malhotra R., Bryden G., Smith P. S., Misselt K. A., Moro-Martín A., Williams J. P., 2009, ApJ, 705, 314
Tanaka H., Inaba S., Nakazawa K., 1996, Icarus, 123, 450
Thommes E. W., Bryden G., Wu Y., Rasio F. A., 2008, ApJ, 675, 1538
Tsiganis K., Gomes R., Morbidelli A., Levison H. F., 2005, Nature, 435, 459
Williams D. R., Wetherill G. W., 1994, Icarus, 107, 117
Wisdom J., Holman M., 1991, AJ, 102, 1528
Wright J. T., Upadhyay S., Marcy G. W., Fischer D. A., Ford E. B., Johnson J. A., 2009, ApJ, 693, 1084
Wyatt M. C., Dent W. R. F., 2002, MNRAS, 334, 589
Yoshida H., 1990, Phys. Lett. A, 150, 262
Yoshida H., 1993, CeMDA, 56, 27
Zhou L., Sun Y., Zhou J., Zheng J., Valtonen M., 2002, MNRAS, 336, 520
[^1]: See author’s Web site for source code.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: |
An identifying code of a graph is a dominating set which uniquely determines all the vertices by their neighborhood within the code. Whereas graphs with large minimum degree have small domination number, this is not the case for the identifying code number (the size of a smallest identifying code), which indeed is not even a monotone parameter with respect to graph inclusion.
We show that every graph $G$ with $n$ vertices, maximum degree $\Delta=\omega(1)$ and minimum degree $\delta\geq c\log{\Delta}$, for some constant $c>0$, contains a large spanning subgraph which admits an identifying code with size $O\left(\frac{n\log{\Delta}}{\delta}\right)$. In particular, if $\delta=\Theta(n)$, then $G$ has a dense spanning subgraph with identifying code $O\left(\log n\right)$, namely, of asymptotically optimal size. The subgraph we build is created using a probabilistic approach, and we use an interplay of various random methods to analyze it. Moreover we show that the result is essentially best possible, both in terms of the number of deleted edges and the size of the identifying code.
author:
- 'Florent Foucaud, Guillem Perarnau and Oriol Serra'
title: Random subgraphs make identification affordable
---
Introduction
============
Consider any graph parameter that is not monotone with respect to graph inclusion. Given a graph $G$, a natural problem in this context is to study the minimum value of this parameter over all spanning subgraphs of $G$. In particular, how many edge deletions are sufficient in order to obtain from $G$ a graph with near-optimal value of the parameter? Herein, we use random methods to study this question with respect to the identifying code number of a graph, a well-studied non-monotone parameter. An identifying code of graph $G$ is a set $C$ of vertices which is a dominating set, and such that the closed neighborhood within $C$ of each vertex $v$ uniquely determines $v$.
Identifying codes were introduced in 1998 in [@KCL98] and have been studied extensively in the literature since then. We refer to [@biblio] for an on-line bibliography. One of the interests of this notion lies in their applications to the location of threats in facilities [@UTS04] and error-detection in computer networks [@KCL98]. One can also mention applications to routing [@LTCS07], to bio-informatics [@HKSZ06] and to measuring the first-order logical complexity of graphs [@KPSV04]. Let us also mention that identifying codes are special cases of the more general notion of *test covers* of hypergraphs, see e.g. [@DHHHLRS03; @MS85] (test covers are also the implicit object of Bondy’s celebrated theorem on *induced subsets* [@B72]).
Let $G$ be a simple, undirected and finite graph. The *open neighborhood* of a vertex $v$ in $G$ is the set of vertices in $V(G)$ that are adjacent to it, and will be denoted $N_G(v)$. The *closed neighborhood* of a vertex $v$ in $G$ is defined as $N_G[v]=N_G(v)\cup \{v\}$. The degree of a vertex $u\in V(G)$, is defined as $d(v)= |N_G(v)|$. Similarly, we define, for a set $S\subseteq V(G)$, $N_G(S)=\bigcup_{v\in S}N_G(v)$ and $N_G[S]=\bigcup_{v\in S}N_G[v]$. If two distinct vertices $u,v$ are such that $N[u]=N[v]$, they are called *twins*. The symmetric difference between two sets $A$ and $B$ is denoted by $A{\oplus}B$. Given a graph $G$ and a subset $C$ of vertices of $G$, $C$ is called a *dominating set* if each vertex of $V(G)\setminus C$ has at least one neighbor in $C$. The set $C$ is called a *separating set* of $G$ if for each pair $u,v$ of vertices of $G$, $N[u]\cap C\neq
N[v]\cap C$ (equivalently, $(N[u]{\oplus}N[v])\cap C\neq\emptyset$). If $x\in N[u]$, we say that $x$ *dominates* $u$. If $x\in N[u]{\oplus}N[v]$, we say that $x$ *separates* $u,v$.
A subset of vertices of a graph $G$ which is both a dominating set and a separating set is called an *identifying code* of $G$.
The following observation gives an equivalent condition for a set to be an identifying code, and follows from the fact that for two vertices $u,v$ at distance at least 3 from each other, $N[u]{\oplus}N[v]=N[u]\cup N[v]$.
\[obs:dist2\] For a graph $G$ and a set ${C}\subseteq V(G)$, if ${C}$ is dominating and $N[u]\cap {C}\neq N[v]\cap {C}$ for each pair of vertices $u,v$ at distance at most two from each other, then $C$ is an identifying code of the graph.
The minimum size of a dominating set of graph $G$, its *domination number*, is denoted by $\gamma(G)$. Similarly, the minimum size of an identifying code of $G$, ${\gamma^{\text{\tiny{ID}}}}(G)$, is the *identifying code number* of $G$. It is known that for any twin-free graph $G$ on $n$ vertices having at least one edge, we have: $$\lceil\log_2(n+1)\rceil\le{\gamma^{\text{\tiny{ID}}}}(G)\leq n-1.$$ The lower bound was proved in [@KCL98] and the upper bound, in [@B01; @GM07]. Both bounds are tight and all graphs reaching these two bounds have been classified (see [@M06] for the lower bound and [@FGKNPV10] for the upper bound). Other papers studying bounds and extremal graphs for identifying codes are e.g. [@CHL07; @FKKR10; @FP12].
In view of the above lower bound, we say that an identifying code $C$ of $G$ is *asymptotically optimal* if $$|C| = O(\log{n})\;.$$
The problem we will address in this paper is to deal with graphs that have a large identifying code number, or are not even identifiable. Our approach will consist in slightly modifying such a graph in order to decrease its identifying code number and obtain an asymptotically optimal identifying code, unless its domination number prevents us from doing so.
One of the reasons for a graph to have a large identifying code number is that it has a large domination number (this one being a monotone parameter under edge deletion). For instance, we need roughly $n/3$ vertices to dominate all the vertices in a path of order $n$. When this is the case, we cannot expect to decrease much the size of a minimum identifying code by deleting edges from $G$, as the deletion of edges cannot decrease the domination number.
However, there are many graphs with small domination number where the identifying code number is very large [@FGKNPV10; @FP12]. Typically, this phenomenon appears in graphs having a specific, “rigid”, structure. Supporting this intuition, Frieze, Martin, Moncel, Ruszinkó and Smyth [@FMMRS07] have shown that the random graph $G(n,p)$ with $p\in (0,1)$, admits an asymptotically optimal identifying code. In particular, they prove in [@FMMRS07] that $$\begin{aligned}
{\gamma^{\text{\tiny{ID}}}}(G(n,p))=(1+o(1)) \frac{2\log{n}}{\log{(1/q)}}\;,\end{aligned}$$ where $q=p^2+(1-p)^2$. This suggests that the lack of structure in dense graphs implies the existence of a small identifying code.
**Our results and structure of the paper.** In Section \[sec:main\], we prove our main result by selecting at random a small set of edges that can be deleted to “add some randomness” in the graph,
\[thm:stronger\] For any graph $G$ on $n$ vertices ($n$ large enough) with maximum degree $\Delta=\omega(1)$ and minimum degree $\delta\geq 66\log{\Delta}$, there exists a subset of edges $F\subset E(G)$ of size $$|F|\leq 83n\log{\Delta}\;,$$ such that $${\gamma^{\text{\tiny{ID}}}}(G\setminus F)\leq 134\frac{n\log{\Delta}}{\delta}\;.$$
Observe that when $\delta=\Theta(n)$, this result is similar to the one in [@FMMRS07].
We then show in Section \[sec:tight\] that our result is asymptotically best possible in terms of both the number of deleted edges and of the size of the final identifying code for any graph with $\Delta={\mbox{Poly}}(\delta)$. For smaller values of the minimum degree, we prove that our result is almost optimal. We also show that the two assumptions $\Delta=\omega(1)$ and $\delta\geq c\log{\Delta}$ for some constant $c$ are necessary.
We present some consequences of our result in Section \[sec:consequences\]. When considering the case of adding edges to the graph, we get analogous (symmetric) results, showing that every graph is a large spanning subgraph of some graph that admits a small identifying code. This result also turns out to be tight. We also describe an application to the closely related topic of *watching systems*.
The paper concludes with some final remarks and open problems.
**Our methods.** To show our results, we use the technique of defining a suitable random spanning subgraph of $G$: we first randomly choose a code $C$, and then we randomly delete edges among the edges containing vertices of $C$. We then analyze the construction by applying concentration inequalities and the use of the local lemma.
A similar approach has been used in the literature when considering *random subgraphs of a graph*: for any graph $G$, consider the graph $G_p$ to be the subgraph of $G$ obtained by keeping *each* edge from $E(G)$ independently with probability $p$. The behavior of random subgraphs of graphs $G$ with minimum degree $\delta$, inspired by applications in epidemiology or social networks, has been widely studied [@ch2007; @fk2012; @kls2012; @ks2013]. A well-known instance of this problem is the classical Erdős-Rényi random graph $G(n,p)$ where $G=K_n$, the complete graph of order $n$. In most of the cases, it was shown that many similarities exist between $G_p$ and the random graph $G(\delta,p)$. The connectivity of a random subgraph of a graph, where every edge has a different probability of being deleted, has been studied in [@A1993]. Our random subgraph model is adapted to the analysis of identifying codes, and can be seen as a weighted version of $G_p$.
Main theorem {#sec:main}
============
In this section, we prove Theorem \[thm:stronger\]. We will need some tools and lemmas.
Important tools and lemmas
--------------------------
In our proofs, we will repeatedly use the Chernoff inequality for the sum of independent bounded random variables:
\[lem:chernoff2\] Let $X_1,\dots,X_{N}$ be independent Bernoulli random variable with probability $p_i$ and define $X=\sum_{i=1}^{N} X_i$. Then , for all ${\varepsilon}>0$, $$\Pr(|X- {\mathbb{E}}(X)|\geq {\varepsilon}{\mathbb{E}}(X))< 2e^{-c_{{\varepsilon}}{\mathbb{E}}(X)}\;,$$ where $$c_{{\varepsilon}}= \min \left\{(1+{\varepsilon})\log(1+{\varepsilon})-{\varepsilon}, \frac{{\varepsilon}^2}{2}\right\}\;.$$
In what follows, for any set of vertices $B\subseteq V(G)$ and any $v\in V(G)$, we let $N_G^B(v)= N_G(v)\cap B$ be the set of neighbors of $v$ in $B$. Analogously, $N_G^B[v]= N_G[v]\cap B$. We denote by $d_B(v)=|N_G^B (v)|$, the degree of $v$ within set $B$.
\[def:G(B,f)\] Given a graph $G$ and $B\subseteq V(G)$, a function $f: V(G)
\rightarrow\mathbb{R}^+\cup \{0\}$ is said to be $(G,B)$–[ *bounded*]{} if for each vertex $u$, $f(u)\leq d_B(u)$ and for each pair $u,v$ of vertices with $d_B(u)\geq d_B(v)$, $f(u)/d_B(u)\leq
f(v)/d_B(v)$. Given a $(G,B)$–bounded function $f$, we define the random spanning subgraph $G(B,f)$ of $G$ as follows:
- $G(B,f)$ contains all edges of the subgraph $G[V(G)\setminus B]$ induced by $V(G)\setminus B$, and
- each edge $uv$ incident with $B$ is independently chosen to be in $G(B,f)$ with probability $1-p_{uv}$, where $$p_{uv}=\frac{1}{4}\left(\frac{f(u)}{d_B(u)} + \frac{f(v)}{d_B(v)}\right)\;.$$ Observe that, since $f(u)\leq d_B(u)$ for each vertex $u\in V(G)$, we have $p_{uv}\leq 1/2$.
The next lemma gives an exponential upper bound on the probability that two vertices of $G(B,f)$ are not separated by $B$. This lemma is a crucial one in our main proof.
\[lem:symmetric\] Let $G$ be a graph, $B\subseteq V(G)$, and $f$ a $(G,B)$–bounded function. In the random subgraph $G(B,f)$, for every pair $u,v$ of distinct vertices with $d_B(u)\geq d_B(v)$, we have $$\begin{aligned}
\Pr\left(N_{G(B,f)}^B[u]=N_{G(B,f)}^B[v]\right)& \leq
e^{-3 f(u)/16 }\;.\end{aligned}$$
Consider the following partition of $S=N^B_G[u]\cup N^B_G[v]$ into three parts: $S_1$, the vertices of $B$ dominating $u$ but not $v$; $S_2$, the vertices of $B$ dominating $v$ but not $u$; and $S_3$, the vertices of $B$ dominating both $u$ and $v$.
Let $D$ be the random variable which gives the size of the symmetric difference of $N_{G(B,f)}^B[u]$ and $N_{G(B,f)}^B[v]$. The statement of the lemma is equivalent to $\Pr(D=0) < e^{-3f(u)/16}$.
The random variable $D= |N_{G(B,f)}^B[u]\oplus N_{G(B,f)}^B[v]|$ can be written as the sum of independent Bernoulli variables $$D=\sum_{w\in S} D_w\;,$$ where $D_w=1$ if and only if $w$ dominates precisely one of the two vertices $u$ or $v$ in $G(B,f)$. Therefore, fro any $w\notin\{u,v\}$, $$\Pr (D_w=1)=\left\{ \begin{array}{ll} 1-p_{uw} & w\in S_1\\ 1-p_{vw} & w\in
S_2\\p_{uw}(1-p_{vw})+p_{vw}(1-p_{uw})&w\in S_{3}\end{array}\right.$$
Since we want to bound from above the probability that $D=0$, we can always assume that $u,v\notin
N_{G(B,f)}^B[u]{\oplus}N_{G(B,f)}^B[v]$. Recall that $d_B(u)\geq d_B(v)$. By the definition of a $(G,B)$–bounded function, we have that $p_{uw}\leq p_{vw}$ for each $w\in S_3$. Since $x(1-x)$ has a unique maximum at $x=1/2$ and $p_{uw},
p_{vw}\le 1/2$, we also have: $$\begin{aligned}
\label{eq:3}
p_{vw}(1-p_{uw})\geq p_{uw}(1-p_{uw}) \geq \frac{f(u)}{4d_B(u)}\left( 1-
\frac{f(u)}{4d_B(u)}\right)=g(u)\;,\end{aligned}$$ for each $w\in S_3$.
For $w\in S$, denote by $q_w$ the parameter of the Bernoulli random variable $D_w$. Then, $$\begin{aligned}
\label{eq:1}
{\mathbb{E}}(D) &\geq\sum_{w\in N^B_G(u)} q_w \nonumber \\
& = \sum_{w\in S_1} q_w+ \sum_{w\in S_3} q_w \nonumber \\
& = \sum_{w\in S_1} (1-p_{uw})+ \sum_{w\in S_3} \left(p_{uw}(1-p_{vw})+p_{vw}(1-p_{uw})\right)
\nonumber \\
&\geq \sum_{w\in S_1} p_{uw}(1-p_{uw})+ \sum_{w\in S_3} p_{uw}(1-p_{uw}) \nonumber \\
& \geq g(u) d_B(u) \nonumber\\
&= \frac{f(u)}{4}\left(1-\frac{f(u)}{4d_B(u)}\right) \nonumber\\
&\geq \frac{3}{16} f(u)\;.\end{aligned}$$
Finally, we have that $$\begin{aligned}
\Pr(D=0)&= \prod_{w\in S} (1-q_{w})\leq e^{-\sum_{w\in S} q_w}
= e^{-{\mathbb{E}}(D)}
\leq e^{-3 f(u)/16 }\;,\end{aligned}$$ and the lemma follows.
In the proof of our main result, we will use the following version of the Lovász local lemma, which can be found in e.g. [@AS00 Corollary $5.1.2$] (the lower bound on $\Pr(\bigcap_{i=1}^M\overline{E_i})$ can be derived from the general local lemma, see [@AS00 Lemma $5.1.1$], by setting $x_i=e\cdot p_{LL}$).
\[lem:LL\] Let $\mathcal{E}=\left\{E_1,\ldots,E_M\right\}$ be a set of (typically “bad”) events such that each $E_i$ is mutually independent of $\mathcal{E}\setminus(\mathcal{D}_i\cup\left\{E_i\right\})$ for some $\mathcal{D}_i\subseteq \mathcal E$. Let $d_{LL}=|\mathcal D_i|$, and suppose that there exists a real $0<p_{LL}<1$ such that, for each $1\le i\le
M$,
- $\Pr(E_i)\leq p_{LL}$, and
- $e\cdot p_{LL}\cdot (d_{LL}+1)\leq 1$.
Then $\Pr(\bigcap_{i=1}^M\overline{E_i})\ge(1-e\cdot p_{LL})^M>0$.
Proof of the main result
------------------------
We are now ready to prove the main theorem.
The proof is structured in the following steps:
1. We select a set $C$ at random, where each vertex is selected independently with probability $p$. Using the Chernoff inequality, we estimate the probability of the event $A_C$ that $C$ is small enough for our purposes. From $C$, we construct the spanning subgraph $G(C,f)$ of $G$ as given in Definition \[def:G(B,f)\], for some suitable function $f$.
2. We use the local lemma (Lemma \[lem:LL\]) and Lemma \[lem:symmetric\] to bound from below the probability that the following events (whose conjunction we call $A_{LL}$) hold jointly: (i) in $G(C,f)$, each pair of vertices that are at distance at most 2 from each other are separated by $C$; and (ii) for each such pair and each member of this pair in $G$, its degree within $C$ in $G$ is close to its expected value $d(v)p$. We show that with nonzero probability, $A_C$ and $A_{LL}$ hold jointly.
3. We find a dominating set $D$ of $G$ with $|D|=O(|C|)$; by Observation \[obs:dist2\], if $A_{LL}$ holds, then $C\cup D$ is an identifying code.
4. Finally, we show that, subject to $A_C$ and $A_{LL}$, the expected number of deleted edges is as small as desired.
**Step 1. Constructing $C$ and $G(C,f)$**
Let $C\subseteq V(G)$ be a subset of vertices, where each vertex $v$ in $G$ is chosen to be in $C$ independently with probability $$p=\frac{66\log{\Delta}}{\delta}\;.$$ Observe that $p\leq 1$ since $\delta\geq 66\log{\Delta}$.
Consider the random variable $|C|$ and recall that ${\mathbb{E}}(|C|)=n
p$.
Define $A_C$ to be the event that $$\begin{aligned}
\label{eqn:einstein}
|C|\leq 2np =\frac{132n\log{\Delta}}{\delta}.\tag{$A_C$}
\end{aligned}$$ Since the choices of the elements in $C$ are done independently, by setting ${\varepsilon}=1$ in Lemma \[lem:chernoff2\], notice that $c_{\varepsilon}>1/3$, we have $$\begin{aligned}
\label{eqn:A_C}
\Pr(\overline{A_C}) < e^{-\frac{22n\log{\Delta}}{\delta}}\;.
\end{aligned}$$
We let $$f(u)=\min\left(66\log\Delta,d_C(u)\right).$$
Observe that $f$ is $(G,C)$–bounded. We construct $G(C,f)$ as the random spanning subgraph of $G$ given in Definition \[def:G(B,f)\], where each edge $uv$ incident to a vertex of $C$ is deleted with probability $p_{uv}$. **Step 2. Applying the local lemma**
Let $u,v$ be a pair of vertices at distance at most 2 in $G$. We define the following events:
- $A_{uv}$ is the event that there exists a vertex $w\in\{u,v\}$ such that the degree of $w$ within $C$ is deviating from its expected value $d(w)
p$ by half, i.e. $|d_C(w)-d(w) p| \geq \frac{d(w) p}{2}$;
- $B_{uv}$ is the event that $N_{G(C,f)}^C[u]=N_{G(C,f)}^C[v]$;
- $E_{uv}$ is the event that $A_{uv}$ or $B_{uv}$ occurs;
- $A_{LL}$ is the event that no event $E_{uv}$ occurs.
In order to apply the Local Lemma, we wish to upper bound the probability of $E_{uv}$. We have: $$\begin{aligned}
\Pr(E_{uv})&\leq \Pr(A_{uv})+\Pr(B_{uv})\\
&=\Pr(A_{uv})
+\Pr(B_{uv}|A_{uv})\cdot\Pr(A_{uv})+\Pr(B_{uv}|\overline{A_{uv}})\cdot\Pr(\overline{A_{ uv}})\;.\end{aligned}$$
Let us upper bound $\Pr(A_{uv})$. We use Lemma \[lem:chernoff2\] with ${\varepsilon}=1/2$. Observe that $c_{{\varepsilon}}>\frac{1}{10}$, and thus $$\begin{aligned}
\Pr(A_{uv})& <\Pr\left(|d_C(u)-d(u) p|\geq \frac{d(u) p}{2}\right)+\Pr\left(|d_C(v)-d(v) p|\geq \frac{d(v) p}{2}\right)\\
&\leq 2e^{-\frac{1}{10} d(u)p} +2e^{-\frac{1}{10} d(v)p}\\
&= 2e^{-\frac{66d(u)\log\Delta}{10\delta}} +2e^{-\frac{66d(v)\log\Delta}{10\delta}}\\
& \leq 4e^{-\frac{33\log\Delta}{5}}\\
& \leq 4\Delta^{-\frac{33}{5}}\;.\end{aligned}$$ Next, we give an upper bound for $\Pr(B_{uv}|\overline{A_{uv}})$. For such a purpose, we apply Lemma \[lem:symmetric\] with $B=C$ and $f(u)=\min(66\log\Delta,d_C(u))$. Observe that $f$ is $(G,C)$–bounded. Since $A_{uv}$ does not hold, we know that $d_C(u)$ and $d_C(v)$ are large enough, i.e. for $w\in\{u,v\}$, $d_C(w)\geq\frac{d(w) p}{2}\geq\frac{\delta
p}{2}=33\log\Delta$; thus $f(u),f(v)\geq 33\log\Delta$. We have: $$\begin{aligned}
\Pr(B_{uv}|\overline{A_{uv}})\leq e^{-\frac{3\cdot 33\log\Delta}{16}}\leq\Delta^{-\frac{99}{16}}\;.\end{aligned}$$ The probability that the event $E_{uv}$ holds is $$\begin{aligned}
\Pr(E_{uv})\leq &
\Pr(A_{uv})
+\Pr(B_{uv}|A_{uv})\cdot\Pr(A_{uv})+\Pr(B_{uv}|\overline{A_{uv}})\cdot\Pr(\overline{A_{uv } })\\
& \leq 4\Delta^{-\frac{33}{5}}+1\cdot 4\Delta^{-\frac{33}{5}} + \Delta^{-\frac{99}{16}} \cdot 1\\
& \leq 2\Delta^{-\frac{99}{16}}=p_{LL}\;,\end{aligned}$$ where we used $\Delta=\omega(1)$.
We now note that each event $E_{uv}$ is mutually independent of all but at most $2\Delta^6$ events $E_{u'v'}$. Indeed, $E_{uv}$ depends on the random variables determining the existence of the edges incident to $u$ and $v$. This is given by probabilities $p_{uw}$ and $p_{vw}$ that depend on $d_C(w)$, where $w$ is at distance at most one from either $u$ or $v$. Thus, $E_{uv}$ depends only on the vertices at distance at most two from either $u$ or $v$ belonging to $C$. In other words, $E_{uv}$ and $E_{u'v'}$ are mutually independent unless there exist a vertex $w$ at distance at most two from both pairs; in other words, $d(\{u,v\},\{u',v'\})\leq 4$. Hence, there are at most $2\Delta^4$ choices for the vertex among $\{u',v'\}$ that is closest from $\{u,v\}$ (say $u'$), and at most $\Delta^2$ additional choices for $v'$, since $d(u',v')\leq 2$.
Therefore, we can apply Lemma \[lem:LL\] if $$e\cdot 2\Delta^{-\frac{99}{16}}\cdot (2\Delta^6+1) \leq 1\;,$$ which holds since $\Delta=\omega(1)$.
Now, by Lemma \[lem:LL\] and since there are at most $\frac{n\Delta^2}{2}$ events $E_{uv}$ (one for each pair of vertices at distance at most 2 from each other) and $p_{LL}=2\Delta^{-\frac{99}{16}}$, $$\begin{aligned}
\label{eqn:A_LL}
\Pr(A_{LL})&\geq (1-e\cdot p_{LL})^M \geq e^{-2e\cdot p_{LL}M}\geq
e^{-2en\Delta^{2-\frac{99}{16}}}\;,\end{aligned}$$ where we have used $(1-x)= e^{-x (1-O(x))}\geq e^{-2x}$, if $x=o(1)$.
**Step 3. Revealing the identifying code**
Let us lower bound the probability that both $A_C$ and $A_{LL}$ hold, by using Inequalities \[eqn:A\_C\] and \[eqn:A\_LL\]: $$\begin{aligned}
\Pr(A_C\cap A_{LL})&\geq\Pr(A_{LL})-\Pr(\overline{A_C})\\
& \geq e^{-2en\Delta^{2-\frac{99}{16}}}-e^{-\frac{22n\log{\Delta}}{\delta}}\;,\end{aligned}$$ which is strictly positive if $$\frac{22\log\Delta}{\delta} > 2e\Delta^{2-\frac{99}{16}}\;,$$ which holds since $n$ is large (and hence $\Delta=\omega(1)$ is large too), and $\delta\leq \Delta$.
Hence, there exists a set $C$ of size $132\frac{n\log{\Delta}}{\delta}$ such that all vertices at distance 2 from each other are separated by $C$, and such that the degree in $C$ of all vertices is large.
In order to build an identifying code, we must also make sure that all vertices are dominated. It is well-known that for any graph $G$, $\gamma(G) \leq (1+o(1))\frac{n\log{\delta}}{\delta}$ (see e.g. [@AS00 Theorem $1.2.2$]). Hence, we select a dominating set $D$ of $G$ with size $(1+o(1))\frac{n\log{\delta}}{\delta}$. Then, by Observation \[obs:dist2\], $C\cup D$ is an identifying code of size at most $$(132+1+o(1))\frac{n\log{\Delta}}{\delta}\leq
134\frac{n\log{\Delta}}{\delta}.$$
**Step 4. Estimating the number of deleted edges**
Let $Y=|E(G)\setminus E(G(C,f))|$ be the number of edges we have deleted from $G$ to obtain $G(C,f)$. Recall that each edge $uv\in E(G)$ is deleted independently from $G$ with probability $$p_{uv}=\frac{1}{4}\left(\frac{f(u)}{d_C(u)}+\frac{f(v)}{d_C(v)}\right)\;,$$ if one of its endpoints is in $C$.
Since $\Pr(A_C\cap A_{LL})> 0$, there is a small identifying code of $G$ obtained by deleting at most ${\mathbb{E}}(Y|A_C\cap A_{LL})$ edges. We next give an upper bound for ${\mathbb{E}}(Y|A_C\cap A_{LL})$. If both $A_C$ and $A_{LL}$ hold, then $$p_{uv}\leq \frac{1}{4}\left(\frac{66\log{\Delta}}{d_C(u)}+\frac{66\log{\Delta}}{d_C(v)}\right)\;.$$ The expected number of deleted edges is $${\mathbb{E}}(Y|A_C\cap A_{LL}) =\sum_{\substack{uv\in E(G)\\\left(\{u,v\}\cap C\right)\neq\emptyset}} p_{uv}\;.$$ Observe that in order to estimate this quantity, we can split the two additive terms in each $p_{uv}$: for every $u\notin C$, we sum all the terms $\frac{66\log\Delta}{4d_C(u)}$ for all $v\in C$ being neighbors of $u$; for every $u\in C$, we sum all the terms $\frac{66\log\Delta}{4d_C(u)}$ for all $v\in V(G)$ being neighbors of $u$. $$\begin{aligned}
{\mathbb{E}}(Y|A_C\cap A_{LL}) &\leq \frac{1}{4}\left(\sum_{u\notin C} \sum_{v\in N^C_G(u)}
\frac{66\log\Delta}{d_C(u)}+ \sum_{u\in C}
\sum_{v\in N_G(u)} \frac{66\log\Delta}{d_C(u)}\right) \\
&\leq\frac{1}{4}\left(\sum_{u\notin C} d_C(u)\frac{66\log\Delta}{d_C(u)}+\sum_{u\in C}d(u)
\frac{66\log\Delta}{d_C(u)}\right)\\
&\leq\frac{1}{4}\left( |V(G)\setminus C|\cdot 66\log\Delta + \sum_{u\in C}
2\frac{66\log\Delta}{p}\right)\\
&\leq\frac{1}{4}\left(n\cdot 66\log\Delta + 2|C|\delta\right)\\
&\leq \frac{66 n\log\Delta+ 264 n\log\Delta}{4}\\
& \leq 83n\log{\Delta}\;,\end{aligned}$$ where we used the fact (implied by $A_{LL}$) that for any vertex $v$, $\frac{d(v)p}{2}\leq d_C(v)$ at the second line, and that $A_C$ implies $|C|\leq 132\frac{n\log{\Delta}}{\delta}$ at the fifth line.
Summarizing, we have shown the existence of a small identifying code in a spanning subgraph of $G$ obtained by deleting at most ${\mathbb{E}}(Y|A_C\cap
A_{LL})$ edges from $G$, which completes the proof.
Asymptotic optimality of Theorem \[thm:stronger\] {#sec:tight}
=================================================
In this section, we discuss the optimality of Theorem \[thm:stronger\], first with respect to the size of the constructed code and the number of deleted edges, and then with respect to the hypothesis $\Delta=\omega(1)$ and $\delta\geq 66\log\Delta$.
On the size of the code and the number of deleted edges
-------------------------------------------------------
Charon, Honkala, Hudry and Lobstein showed that deleting an edge from $G$ can decrease by at most $2$ the identifying code number of a graph [@CHHL13]. That is, for any graph $G$ and any edge $uv \in
E(G)$, $${\gamma^{\text{\tiny{ID}}}}(G)\leq {\gamma^{\text{\tiny{ID}}}}(G\setminus uv)+2\;.$$
This directly implies that for every graph with linear identifying code number, one needs to delete a subset $F$ of at least $\Omega(n)$ edges, to get a graph with ${\gamma^{\text{\tiny{ID}}}}(G\setminus F)=o(n)$.
We will show that, indeed, one needs to delete at least $\Omega(n\log{n})$ edges from the complete graph to get a graph with an asymptotically optimal identifying code. Using this, we will derive a family of graphs with arbitrary minimum degree $\delta$, that asymptotically attains the bound of Theorem \[thm:stronger\], both in number of edges and size of the minimum code, when $\Delta={\mbox{Poly}}(\delta)$.
First of all, we prove that every graph with an asymptotically optimal identifying code cannot contain too few edges.
\[lem:sparse\] For any $M'\geq 0$, there exists a constant $c_0>0$ such that any graph $G$ with ${\gamma^{\text{\tiny{ID}}}}(G)\leq M'\log{n}$ contains at least $c_0 n\log{n}$ edges.
Set $\alpha_0$ as the smallest positive root of $$\begin{aligned}
\label{eq:cond_alpha}
f(\alpha) =\alpha \log{\left(\frac{M'+\alpha}{\alpha} e\right)} - 1/2\;.\end{aligned}$$ Note that $f(\alpha)$ is well-defined since $\lim_{\alpha\to 0}
f(\alpha) = -1/2$ and $f(1)=\log(M'+1)+1/2>0$.
Suppose by contradiction that there exists a graph $G$ containing less than $c_0 n\log{n}$ edges, with $c_0=\alpha_0 /4$, that admits an identifying code $C$ of size at most $M'\log{n}$. Let $U$ be the subset of vertices of degree at least $\alpha_0 \log{n}$. Notice that $$|U|\leq \frac{2|E(G)|}{\alpha_0 \log{n}} \leq \frac{2c_0}{\alpha_0}n = \frac{n}{2}\;.$$
Since $|C|\leq M'\log{n}$ and any $v\in V(G)\setminus U$ has degree smaller than $\alpha_0 \log{n}$, the number of possible nonempty sets $N_G[v]\cap C$, is smaller than $$\begin{aligned}
\sum_{i=1}^{\alpha_0 \log{n}} \binom{|C|}{i} &\leq \binom{M'\log{n}+\alpha_0\log{n}}{\alpha_0
\log{n}} \\
&\leq \left(\frac{(M'+\alpha_0)e}{\alpha_0}\right)^{\alpha_0 \log{n}} \\
&= n^{\alpha_0 \log{\left(\frac{M'+\alpha_0}{\alpha_0} e\right)}} \\
&= \sqrt{n}\;.\end{aligned}$$ where we have used that $\binom{a}{b}\leq \left(\frac{ae}{b}\right)^b$ for the second inequality and the fact that $\alpha_0$ is a root of for the last one.
Since $|V(G)\setminus U|\geq n/2$ there must be at least two vertices $v_1,v_2\in V(G)\setminus U$ such that $N_G[v_1]\cap C=N_G[v_2]\cap C$, and thus $C$ cannot be an identifying code, a contradiction.
The following lemma relates the identifying code number of a graph $G$ to the one of its complement $\overline{G}$.
\[lem:complement\] Let $G$ be a twin-free graph. If $\overline{G}$ is twin-free, then $${\gamma^{\text{\tiny{ID}}}}(\overline{G})\leq 2 {\gamma^{\text{\tiny{ID}}}}(G)\;.$$
Let $C_0$ be a minimum identifying code of $G$. We will show that there exists a set $C_1$ of size at most ${\gamma^{\text{\tiny{ID}}}}(G)-1$ and a special vertex $v$, such that $C=C_0\cup C_1\cup \{v\}$ is an identifying code of $\overline{G}$.
For the sake of simplicity, we define the following relation. Two vertices $u, v\in V(G)$ are in relation with each other if and only if $N_{G}(u)\cap C_0= N_{G}(v)\cap C_0$ and $u\not\sim v$ (i.e. considering $C_0$ in $G$, $u,v$ are separated by one of $u,v$). This will be denoted as $u\equiv_G v$. It can be checked that this relation is an equivalence relation.
Every pair of distinct vertices $u\not\equiv_G v$ is separated by $C_0$ in $\overline{G}$.
By the definition of $\equiv_G$, either $N_{G}(u) \cap
C_0 \neq N_{G}(v)\cap C_0$ or $u\sim v$.
If $N_{G}(u) \cap C_0 \neq N_{G}(v)\cap C_0$, there exists $w\in C_0$ (and $w\notin\{u,v\}$) such that $w\in N_{G}(u){\oplus}N_{G}(v)$. Then, $w\in N_{\overline{G}}(u){\oplus}N_{\overline{G}}(v)$, hence $w$ still separates $u,v$ in $\overline{G}$.
If $N_{G}(u) \cap C_0 = N_{G}(v)\cap C_0$, then $u\sim v$. If at least one of them belongs to $C_0$, then this vertex separates $u,v$ in $\overline{G}$. Otherwise, $u, v\notin C_0$ and we have $N_{G}(u)
\cap C_0=N_{G}[u] \cap C_0$ and $N_{G}[v]\cap C_0 = N_{G}(v)\cap
C_0$. Hence $N_{G}[u] \cap C_0=N_{G}[v] \cap C_0$. But then $C_0$ does not separate $u,v$ in $G$, a contradiction.
In particular, this implies that any vertex in an equivalence class of size one is separated by $C_0$ from all other vertices in $\overline{G}$.
If $u\equiv_G v$ and both $u,v\notin C_0$, then $u=v$.
Since $u,v\notin C_0$, $N_{G}[u] \cap C_0=N_{G}(u) \cap C_0$ and $N_{G}[v] \cap C_0=N_{G}(v) \cap C_0$. Using that they are equivalent, we have that $N_{G}[u] \cap C_0 =N_{G}[v] \cap
C_0$. Since $C_0$ is an identifying code of $G$, we must have $u=v$.
Let $U=\{u_1,\dots, u_s\}$ be an equivalence class of $\equiv_G$. Then all the pairs in $U$ can be separated in $\overline{G}$ by using $s-1$ vertices.
We will prove the claim by induction. For $s=2$ it is clearly true: since $\overline{G}$ is twin-free, we can select $w\in N_{\overline{G}}[u_1]{\oplus}N_{\overline{G}}[u_2]$, and $w$ separates $u$ and $v$ in $\overline{G}$.
For any $s>2$, consider the vertices $u_1,u_2 \in U$ and let $w\in
N_{\overline{G}}[u_1]{\oplus}N_{\overline{G}}[u_2]$. Since $U$ forms a clique in $\overline{G}$, $w\notin U$. Then $w$ splits the set $U$ into $U_1$, the set of vertices of $U$ adjacent to $w$ in $\overline{G}$, and $U_2$, the set of vertices in $U$ non-adjacent to $w$ in $\overline{G}$. Let $|U_1|=s_1$ and $|U_2|=s_2$; by construction, $s_1,s_2<s$.
Now, the pairs of vertices of $U$ with one vertex from $U_1$ and one vertex from $U_2$ are separated by $w$. By induction, the pairs of vertices in $U_1$ can be separated using $s_1-1$ vertices and the ones in $U_2$ using $s_2-1$. Thus we need at most $(s_1-1)+(s_2-1) +1 = s-1$ vertices to separate all the pairs of vertices in $U$.
From the previous claims, it is straightforward to deduce that there is a set $C_1$ of size at most $|C_0|-1$ vertices that separates all the pairs in $\overline{G}$ that are not separated by $C_0$.
Eventually, there might be a unique vertex $v$ such that $N_{\overline{G}}[v]\cap (C_0\cup C_1)= \emptyset$ (if there were two such vertices, they would not be separated by $C_0\cup C_1$, a contradiction). Hence, $C=C_0\cup C_1 \cup \{v\}$ is an identifying code of $\overline{G}$ of size at most $2|C_0|=2{\gamma^{\text{\tiny{ID}}}}(G)$.
\[prop:optimal\] For any $M\geq 0$, there exists a constant $c>0$ such that for any set of edges $F\subset E(K_n)$ satisfying ${\gamma^{\text{\tiny{ID}}}}(K_n\setminus F)\leq
M\log{n}$, $|F|\geq c n\log{n}$.
Set $M'=M/2$ and let $c=c_0$ be the constant given by Lemma \[lem:sparse\] for this $M'$. Suppose that there exists a set $F$ of edges, $|F|< c n\log{n}$ such that $G=K_n\setminus F$ satisfies ${\gamma^{\text{\tiny{ID}}}}(G)\leq M\log{n}$. By Lemma \[lem:complement\], the graph $\overline{G}$ admits an identifying code of size at most $2M\log{n}= M'
\log{n}$. By Lemma \[lem:sparse\], we get a contradiction.
Using the former proposition, for any $\delta$ we can provide an example of a graph with minimum degree $\delta$ for which the result of Theorem \[thm:stronger\] is asymptotically tight when assuming that $\Delta={\mbox{Poly}}(\delta)$.
For any $\delta>0$, consider the graph $H_\delta$ to be the disjoint union of cliques of order $\delta+1$. We may assume that $\delta+1$ divides $n$ for the sake of simplicity. Denote by $H_\delta^{(1)},\dots,
H_\delta^{(s)}$, $s=\frac{n}{\delta+1}$, the cliques composing $H_\delta$.
Since $H_\delta^{(i)}$ is a connected component, an asymptotically optimal identifying code for $H_\delta$ must also be asymptotically optimal for each $H_\delta^{(i)}$. By Proposition \[prop:optimal\], we must delete at least $\Omega(\delta\log{\delta})$ edges from $H_\delta^{(i)}$ to get an identifying code of size $O(\log{\delta})$.
Thus, one must delete at least $\Omega(s \delta\log{\delta}) =
\Omega(n\log{\delta})$ edges from $H_\delta$ to get an optimal identifying code.
\[cor:disjointcliques\] For any $\delta= \omega(1)$ and any $M\geq 0$, there exists a constant $c>0$ such that for any set of edges $F\subset E(H_\delta)$ satisfying ${\gamma^{\text{\tiny{ID}}}}(H_\delta\setminus F)\leq M\frac{n\log{\delta}}{\delta}$, we have $|F|\geq
c n\log{\delta}$.
We remark that a connected counterexample can also be constructed from $H_\delta$ by connecting its cliques using few edges, without affecting the above result.
Corollary \[cor:disjointcliques\] implies that Theorem \[thm:stronger\] is asymptotically tight when $\Delta={\mbox{Poly}}(\delta)$, since in that case $\log{\Delta}=O(\log{\delta})$. However, when $\delta$ is sub-polynomial with respect to $\Delta$, we do not know if Theorem \[thm:stronger\] is asymptotically tight.
On the hypothesis
-----------------
We conclude this section by discussing the necessity of the hypothesis $\Delta=\omega(1)$ and $\delta\geq 66 \log{\Delta}$ in Theorem \[thm:stronger\].
First note that, if $\Delta$ is bounded by a constant, we need at least $\tfrac{n}{\Delta+1}=\Theta(n)$ vertices to dominate $G$. Thus, no code of size smaller than $\Theta(n)$ can be obtained by deleting edges of the graph.
On the other hand, the condition $\delta\geq 66
\log{\Delta}$ in Theorem \[thm:stronger\], is also necessary (up to a constant factor) as can be deduced from the following proposition.
For arbitrarily large values of $\Delta$, there exists a graph $G$ with maximum degree $\Delta$ and minimum degree $\delta=\frac{\log_2{\Delta}}{2}$ such that, for any spanning subgraph $H\subseteq G$, $${\gamma^{\text{\tiny{ID}}}}(H) = (1-o(1))n\;.$$
Consider the bipartite complete graph $G=K_{r,s}$ where $s=2^{2r}$. Denote by $V_1$ the stable set of size $r$ and by $V_2$ the stable set of size $s$. Observe that $\delta= r= \frac{\log_2{s}}{2} = \frac{\log_2{\Delta}}{2}$.
For any given twin-free spanning subgraph $H\subseteq G$, let $C\subseteq V(G)$ be an identifying code of $H$. Let us show that most of the vertices in $V_2$ must be in $C$. Let $S\subseteq V_2$ be the subset of vertices in $V_2$ that are not in the code. Thus, for any $u\in S$, $N_C[u]=
N_C(u)$. Observe that $N_C(u)\subseteq V_1$, and hence, there are at most $2^{r}$ possible candidates for such $N_C(u)$. Since $C$ is dominating and separating all the pairs in $S$, all the subsets $N_C(u)$ must be non empty and different, which implies, $|S|<2^{r}$. Hence, we have $$|C|\geq |V_2\setminus S|\geq 2^{2r}-2^r = (1-o(1))2^{2r} =(1-o(1))n\;.$$
Consequences of our results {#sec:consequences}
===========================
We now describe consequences of our results on the case when we want to *add* edges to a graph to decrease its identifying code number, and to the notion of watching systems.
Adding edges {#sec:adding}
------------
In the previous sections, we have studied how much can the identifying code number decrease when we delete few edges from the original graph. In this section, we discuss the symmetric question of how much can the addition of edges help to decrease this parameter.
The question of how much can a parameter decrease when deleting/adding edges has been already studied for some monotone parameters. However, if the parameter is monotone, only one of either deleting or adding, can help to decrease it. One of the interesting facts of studying the identifying code number is that, since it is a non-monotone parameter, we can have similar results for both procedures.
As before, let $G$ be a graph with maximum degree $\Delta$ and minimum degree $\delta$. We aim to find a set of edges $F$ with $F \cap E(G)=\emptyset$ such that ${\gamma^{\text{\tiny{ID}}}}(G\cup F)$ is small. This set $F$ will be provided by applying Theorem \[thm:stronger\] to the graph $\overline{G}$, that has maximum degree $\Delta(\overline{G})=n-1- \delta$ and minimum degree $\delta(\overline{G})
=n-1-\Delta$. Thus, it will have size $$|F|= O\left(n\log{\Delta(\overline{G})}\right),$$ and $${\gamma^{\text{\tiny{ID}}}}( \overline{G}\setminus F) =
O\left(\frac{n\log{\Delta(\overline{G})}}{\delta(\overline{G})}\right)\;.$$
Since $\overline{G}\setminus F= \overline{G\cup F}$, we have the following corollary of Theorem \[thm:stronger\] and Lemma \[lem:complement\].
\[cor:adding\] For any graph $G$ on $n$ vertices with minimum degree $\delta= n-\omega(1)$ and maximum degree $\Delta$ such that $n-\Delta\geq 66\log{(n-\delta)}$, there exists a set of edges $F$ with $F
\cap E(G)=\emptyset$ of size $$|F|= O\left(n\log{(n-\delta)}\right)\;,$$ such that $${\gamma^{\text{\tiny{ID}}}}(G\cup F) = O\left(\frac{n\log{n}}{n-\Delta}\right)\;.$$
This result is also asymptotically tight. Otherwise, by using again Lemma \[lem:complement\], we could translate our case to the case of deleting edges and we would get a contradiction with the optimality of Theorem \[thm:stronger\].
Watching systems {#sec:watching}
----------------
The result of Theorem \[thm:stronger\] has a direct application for *watching systems*, which are a generalization of identifying codes [@ACHL10; @ACHL12]. In a watching system, we can place on each vertex $v$ a set of *watchers*. To each watcher $w$ placed on $v$, we assign a nonempty subset $Z(w)\subseteq N[v]$, its *watching zone*. We now ask each vertex to belong to a unique and nonempty set of watching zones; the minimum number of watchers that need to be placed on the vertices of $G$ to obtain a watching system is the *watching number* $w(G)$ of $G$.
It is clear from the definition that $\gamma(G)\leq w(G)\leq{\gamma^{\text{\tiny{ID}}}}(G)$, since the vertices of any identifying code form a watching system (where the watching zones are the closed neighborhoods). In fact, even the following holds:
\[obs:watch\] For any twin-free graph $G$, $w(G)\leq\min\{{\gamma^{\text{\tiny{ID}}}}(H), \mbox{ where $H$ is a spanning
subgraph of $G$}\}$. Indeed, consider the spanning subgraph $H_0$ of $G$ with smallest identifying code number, and define the watching system to be the vertices of an optimal identifying code of $H_0$, with the watching zones being the closed neighborhoods in $H_0$.
In [@ACHL10 Theorems $2$ and $3$], the authors propose the following upper bound for graphs with given maximum degree:
\[thm:watchDelta\] Let $G$ be a graph with maximum degree $\Delta$, then $$\lceil \log_2(n+1) \rceil \leq w(G) \leq \gamma(G) \lceil\log_2(\Delta+2)\rceil\;.$$
Note that for any values of parameters $\gamma$ and $\Delta$, the upper bound from the above theorem is tight for the graph consisting of $\gamma$ disjoint copies of a star on $\Delta+1$ vertices.
It is well-known (see e.g. [@AS00 Theorem $1.2.2$]) that the domination number of a graph with minimum degree $\delta$ satisfies $$\gamma(G) \leq (1+o(1))\frac{n\log{\delta}}{\delta}\;.$$ This bound is sharp and, in particular, the “typical” $\delta$-regular graph is an asymptotically tight example. Indeed, for such a “typical” graph $G$, the upper bound of Theorem \[thm:watchDelta\] gives $$\begin{aligned}
\label{eq:2}
w(G)\leq \gamma(G) \lceil\log{\Delta+2}\rceil =
\Omega\left(\frac{n\log^2{\delta}}{\delta}\right)\;.\end{aligned}$$ By Observation \[obs:watch\], a direct corollary of Theorem \[thm:stronger\] is the following:
\[cor:WS\] For any graph $G$ on $n$ vertices with minimum degree $\delta\geq 66\log\Delta$ and maximum degree $\Delta=\omega(1)$, we have: $$w(G) \leq 134\frac{n\log{\Delta}}{\delta}\;.$$
Note that this bound improves Theorem \[thm:watchDelta\] when the maximum degree is $\Delta={\mbox{Poly}}(\delta)$.
Concluding remarks and open questions
=====================================
**1.** The kind of results we provide in this paper can be connected to the notion of resilience. Given a graph property ${\mathcal{P}}$, the *global resilience* of $G$ with respect to ${\mathcal{P}}$ is the minimum number of edges one has to delete to obtain a graph not satisfying ${\mathcal{P}}$. The resilience of monotone properties is well studied, in particular, in the context of random graphs [@SV08].
Our result can be interpreted in terms of the resilience of the following (non-monotone) property ${\mathcal{P}}$: “$G$ has a large identifying code number in terms of its degree parameters, $\delta$ and $\Delta$”. For any graph $G$ satisfying the hypothesis $\Delta=\omega(1)$ and $\delta\geq 66\log{\Delta}$, Theorem \[thm:stronger\] can be stated as: the resilience of $G$ with respect to ${\mathcal{P}}$ is $O(n\log{\Delta})$. Moreover, Corollary \[cor:disjointcliques\] shows that there are graphs that attain this value of the resilience.
**2.** In Theorem \[thm:stronger\], we show the existence of a small identifying code for a large spanning subgraph of $G$. However, our proof is not constructive and, besides, the probability that such pair exists is exponentially small, due to the use of the local lemma. The algorithmic version of the local lemma proposed by Moser and Tardos, allows to explicitly find a configuration that avoids all the bad events $E_{uv}$, when these events are determined by a finite set of mutually independent random variables. Unfortunately, this is not the case here, since $E_{uv}$ depends on the random variables determining the existence of certain edges close to $uv$. These random variables are not independent because of the definition of $p_{uv}$.
On the other hand, if we do not want to argue in terms of the maximum degree $\Delta$, one can show that by deleting a set of $O(n\log{n})$ random edges we have an identifying code of size $O\left(\frac{n\log{n}}{\delta}\right)$ with probability $1-o(1)$. In such a case, the proof provides a randomized algorithm which constructs the desired code for almost all subgraphs.
**3.** Note that a notion similar to identifying codes, *locating-dominating sets*, was also extensively studied in the literature (see e.g. [@biblio] for many references). A set $C$ of vertices of $G$ is a locating-dominating set if $C$ is a dominating set which separates all pairs of vertices in $V(G)\setminus C$. It follows that any identifying code is a locating-dominating set, hence Theorem \[thm:stronger\] also holds for this notion. In fact, the proof of Corollary \[cor:disjointcliques\] can be adapted for this case too.
**4.** As further research, it would be very interesting to close the gap between the result in Theorem \[thm:stronger\] and the lower bound given by the example in Corollary \[cor:disjointcliques\]. Motivated by this example, we ask the following question:
Is it true that for any graph $G$ with minimum degree $\delta$, there exists a subset of edges $F\subset E(G)$ of size $$|F|= O\left(n\log{\delta}\right)\;,$$ such that $${\gamma^{\text{\tiny{ID}}}}(G\setminus F)=O\left(\frac{n\log{\delta}}{\delta}\right)\;?$$
It seems to us that the techniques used in this paper will not provide an answer to the previous question. The main obstacle is the use of the local lemma, which forces us to take into account the role of the maximum degree of $G$.
[0]{}
N. Alon. A note on network reliability. *Discrete probability and algorithms ([M]{}inneapolis, [MN]{}, 1993)*, *IMA Volumes in Mathematics and its Applications* 72:11–14, 1995.
N. Alon and J. H. Spencer. *The probabilistic method*, 3rd edition, Wiley-Interscience, 2008.
D. Auger, I. Charon, O. Hudry and A. Lobstein. Watching systems in graphs: an extension of identifying codes. To appear in *Discrete Applied Mathematics*.
D. Auger, I. Charon, O. Hudry and A. Lobstein. Maximum size of a minimum watching system and the graphs achieving the bound. To appear in *Discrete Applied Mathematics*.
L. Babai. On the complexity of canonical labeling of strongly regular graphs. *SIAM Journal of Computing* 9(1):212–216, 1980.
N. Bertrand. *Codes identifiants et codes localisateurs-dominateurs sur certains graphes*. Master thesis, ENST, France, June 2001.
J. A. Bondy. Induced subsets. *Journal of Combinatorial Theory, Series B* 12(2):201–202, 1972.
I. Charon, I. Honkala, O. Hudry and A. Lobstein. Minimum sizes of identifying codes in graphs differing by one vertex. *Cryptography and Communications* 5(2):1–18, 2013.
I. Charon, O. Hudry and A. Lobstein. Extremal cardinalities for identifying and locating-dominating codes in graphs. *Discrete Mathematics* 307(3-5):356–366, 2007.
F. Chung and P. Horn. The spectral gap of a random subgraph of a graph. *Internet Mathematics* 4(2-3):225–244, 2007.
K. M. J. De Bontridder, B. V. Halldórsson, M. M. Halldórsson, C. A. J. Hurkens, J. K. Lenstra, R. Ravi and L. Stougie. Approximation algorithms for the test cover problem. *Mathematical Programming Series B* 98:477–491, 2003.
F. Foucaud, E. Guerrini, M. Kovše, R. Naserasr, A. Parreau and P. Valicov. Extremal graphs for the identifying code problem. *European Journal of Combinatorics* 32(4):628–638, 2011.
F. Foucaud, R. Klasing, A. Kosowski and A. Raspaud. On the size of identifying codes in triangle-free graphs. *Discrete Applied Mathematics* 160(10-11):1532–1546, 2012.
F. Foucaud and G. Perarnau. Bounds for identifying codes in terms of degree parameters. *The Electronic Journal of Combinatorics* 19:P32, 2012.
A. Frieze and M. Krivelevich. On the non-planarity of a random subgraph. ArXiv e-prints, 2012.
A. Frieze, R. Martin, J. Moncel, M. Ruszinkó and C. Smyth. Codes identifying sets of vertices in random networks. *Discrete Mathematics* 307(9-10):1094–1107, 2007.
S. Gravier and J. Moncel. On graphs having a $V\setminus\{x\}$ set as an identifying code. *Discrete Mathematics* 307(3-5):432–434, 2007.
T. W. Haynes, D. J. Knisley, E. Seier and Y. Zou. A quantitative analysis of secondary RNA structure using domination based parameters on trees. *BMC Bioinformatics* 7:108, 2006.
M. G. Karpovsky, K. Chakrabarty, and L. B. Levitin. On a new class of codes for identifying vertices in graphs. *IEEE Transactions on Information Theory*, 44:599-611, 1998.
J. H. Kim, O. Pikhurko, J. Spencer and O. Verbitsky. How complex are random graphs in First Order logic? *Random Structures and Algorithms* 26(1-2):119–145, 2005.
M. Krivelevich, C. Lee and B. Sudakov. Long paths and cycles in random subgraphs of graphs with large minimum degree. ArXiv e-prints, 2012.
M. Krivelevich and B. Sudakov. The phase transition in random graphs — a simple proof. ArXiv e-prints, 2012.
M. Laifenfeld, A. Trachtenberg, R. Cohen and D. Starobinski. Joint monitoring and routing in wireless sensor networks using robust identifying codes. *Proc. IEEE Broadnets 2007*, pp. 197–206, 2007.
A. Lobstein. Watching systems, identifying, locating-dominating and discriminating codes in graphs: a bibliography. <http://www.infres.enst.fr/~lobstein/debutBIBidetlocdom.pdf>
J. Moncel. On graphs on $n$ vertices having an identifying code of cardinality $\log_2(n + 1)$. *Discrete Applied Mathematics*, 154(14):2032–2039, 2006.
B. M. E. Moret and H. D. Shapiro. On minimizing a set of tests. *SIAM Journal of Scientifical and Statistical Computation* 6(4):983–1003, 1985.
B. Sudakov and and V. H. Vu. Local resilience of graphs, *Random Structures Algorithms* 33(4):409–433, 2008.
R. Ungrangsi, A. Trachtenberg and D. Starobinski. An implementation of indoor location detection systems based on identifying codes. *Proc. Intelligence in Communication Systems, INTELLCOMM 2004*, *Lecture Notes in Computer Science* 3283:175–189, 2004.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'The singularity space-time metric obtained by Krori and Barua[@Krori1975] satisfies the physical requirements of a realistic star. Consequently, we explore the possibility of applying the Krori and Barua model to describe ultra-compact objects like strange stars. For it to become a viable model for strange stars, bounds on the model parameters have been obtained. Consequences of a mathematical description to model strange stars have been analyzed.'
author:
- '[Farook Rahaman$^{\ast}$ [^1], Ranjan Sharma$^{\dag}$[^2], Saibal Ray$^{\ddag}$ [^3], Raju Maulick$^{\S}$[^4] and Indrani Karar$^{\clubsuit}$[^5]]{}'
- '$^{\ast}$[Department of Mathematics, Jadavpur University, Kolkata - 700032, India ]{}'
- '$^{\dag}$ [Department of Physics, P. D. Women’s College, Jalpaiguri 735101, India]{}'
- '$^{\ddag}$[Department of Physics, Government College of Engineering & Ceramic Technology, Kolkata 700 010, West Bengal, India ]{}'
- '$^{\S}$[Department of Mathematics, Jadavpur University, Kolkata 700 032, West Bengal, India]{}'
- '$^{\clubsuit}$[Department of Mathematics, Saroj Mohan Institute of Technology, Guptipara, West Bengal, India]{}'
date: 'Received: date / Accepted: date'
title: 'Strange stars in Krori-Barua space-time'
---
Introduction {#sec:1}
============
In relativistic astrophysics, understanding the nature and exact composition of a specific class of compact stars which are more compact than ordinary neutron stars, has become a field of active research in recent years. A neutron star is the final stage of a gravitationally collapsed star which, after exhausting all its thermo-nuclear fuel, gets stabilized by degenerate neutron pressure. Soon after the discovery of the particle ‘neutron’ by Chadwick, the existence of neutron star was predicted. Later on, the concept got observational support with the discovery of pulsars[@Hewish]. With the progress in our understanding of particle interactions at high energy, theoretical modelling of neutron stars have improved considerably over the last few decades[@Shapiro]. However, the nature of particle interactions beyond nuclear density is still poorly understood. The conjecture that quark matter might be the true ground state of hadrons[@Witten; @Farhi], has led to the discussions of an entirely new class of stellar bodies composed of deconfined $u$, $d$ and $s$ quarks, called ‘strange stars’. It is interesting to note that a strange matter equation of state (EOS) seems to explain the observed compactness of many astrophysical objects like Her X-1, 4U 1820-30, SAX J 1808.4-3658, 4U 1728-34, PSR 0943+10 and RX J185635-3754, whose estimated compactness, otherwise, cannot be explained in terms of a neutron star EOS[@Alcock; @Haensel; @Weber; @Garcia; @Rodrigues; @Bordbar]. Though many more exotic phases may exist at the interior of such class of stars, in this paper, we shall restrict our discussions to the strange matter EOS only. We shall choose a simple EOS for strange quark matter based on the MIT bag model where the quark confinement is assumed to be caused by a universal pressure $B_g$, called the bag constant[@Chodos]. The bag model essentially describes the confinement mechanism of quarks inside hadrons. By imposing the condition that the energy per baryon of strange matter be less than that of the nucleon ($939~$MeV), Farhi and Jaffe[@Farhi] have shown that for a stable strange quark matter the bag constant should be $B_g \sim 60~$MeV/fm$^{-3}$.
Once the strange matter EOS is known, one can employ numerical techniques to get an estimate of the gross features of a strange star by integrating the Tolman-Oppenheimer-Volkoff(TOV) equations. From the general relativistic view point, question is, what would be the relevant back ground space-time to model such class of compact stars? In a recent work, Avellar and Horvath[@Avellar] have considered a wide set of exact and approximate solutions to model strange stars. The objective of the current investigation is to look for a physically viable analytical model which can describe such class of compact stars. To this end, we explore the possibility of applying the Krori and Barua[@Krori1975] (henceforth KB) metric to describe the interior space-time of a strange star. The KB model is singularity free and has earlier been shown to be useful to describe realistic stars[@Junevicus1976]. Ivanov[@Ivanov2002] has shown that the KB model satisfies the necessary energy conditions of a realistic star. Varela [*et al*]{}[@Varela2010], for a Einstein-Maxwell system, have used the KB model to describe a self-gravitating, charged, anisotropic fluid sphere satisfying a linear and/or non-linear EOS. The key observations made by them are as follows: (i) spheres with vanishing net charge contain fluid elements with unbounded proper charge density located at the fluid-vacuum interface; (ii) inward-directed fluid forces caused by pressure anisotropy may allow equilibrium configurations with larger net charges and electric field intensities than those found in studies of charged isotropic fluids; (iii) possible applications of the model to describe charged strange quark stars, dark matter distributions and massive charged particles. In a separate paper, Farook [*et al*]{}[@Rahaman2010] have used the KB model to analyze an anisotropic, charged, static, spherically symmetric fluid source. It has been shown that the inclusion of a tangential pressure-like variable admits a non-linear, Chaplygin-type EOS. Interestingly, the two approaches coincide for an EOS of the form $p= H\rho$, where $\rho$ is the energy density, $p$ the pressure and $H$ is a model parameter describing the stiffness of the EOS. The results obtained by Varela [*et al*]{} may be regained by Farook [*et al*]{}’s[@Rahaman2010] approach. They too predicted a possible extrapolation of the investigation to the case of astrophysical bodies, in particular, for a quark star of radius $\sim 8~$ km.
The present investigation is a follow up of the earlier works done by Varela [*et al*]{}[@Varela2010] and Rahaman [*et al*]{}[@Rahaman2010]. In the present work, we shall incorporate the bag model EOS in the KB model and study the subsequent stellar configurations. It is to be noted here that once we specify the EOS, we can integrate the TOV equations to derive the gross features of a stellar configuration. On the other hand, if one of the metric functions is assumed a priori, one can determine the subsequent EOS of the material composition of the star. However, if both the metric functions as well as the EOS are provided, it becomes an over determined system. To overcome the situation, we include two additional input parameters into the system. Note that the KB model was originally developed for a charged neutral, isotropic and spherically symmetric object in static equilibrium. We assume here that the composition of the strange star is anisotropic in nature coupled with high electric field. These assumptions are justified due to the following reasons: Strange stars are extremely dense objects and at very high density, it is expected that the pressure should be anisotropic in nature[@Bowers]. The electric field at the surface of a strange star has also been reported to be very high[@Usov]. Implications of these additional parameters on the physical behaviour of the strange star will be discussed in the following sections.
The paper has been organized as follows: In Sect. \[sec:2\], we have written the basic field equations. In different sub-sections of Sect. \[sec:3\], we have derived bounds on the model parameters based on various physical requirements. In Sect. \[sec:4\], we have discussed implications of applying the KB model for the description of strange stars. In different sub-sections of Sect. \[sec:5\], we have analyzed various features of the model including stability of the resultant configurations. Finally, some concluding remarks have been made in Sect. \[sec:6\].
Basic Equations {#sec:2}
===============
We assume that the interior space-time of a ‘strange star’ is well described by the Krori and Barua[@Krori1975] metric given by $$ds^2 = -e^{\nu(r)}dt^2 + e^{\lambda(r)}dr^2 +r^2(d\theta^2 +sin^2\theta d\phi^2),\label{eq1}$$ where, $\lambda(r)=Ar^2$ and $\nu(r) = Br^2 + C$. In Eq. (\[eq1\]), $A$, $B$ and $C$ are arbitrary constants which will be fixed on the ground of various physical requirements.
For a static charged fluid source with density $\rho(r)$, radial pressure $p_r(r)$, tangential pressure $p_t(r)$, proper charge density $\sigma(r)$ and electric field $E(r)$, the Einstein-Maxwell(EM) equations take the form (we employ the geometrized units $G = c = 1$) $$\begin{aligned}
8\pi\rho+E^2 &=& e^{-\lambda}\left(\frac{\lambda^\prime}{r}-\frac{1}{r^2}\right)+\frac{1}{r^2},\label{eq2}\\
8\pi p_r-E^2 &=& e^{-\lambda}\left(\frac{\nu^\prime}{r}+\frac{1}{r^2}\right)-\frac{1}{r^2},\label{eq3}\\
8\pi p_t+E^2 &=& \frac{e^{-\lambda}}{2}\left(\nu^{\prime\prime}+\frac{\nu^{\prime 2}}{2}+\frac{\nu^{\prime}
-{\lambda^{\prime}}}{r}-\frac{\nu^{\prime}\lambda^{\prime}}{2}\right), \label{eq4}\\
E(r) &=& \frac{1}{r^2}\int_0^r 4\pi r^2 \sigma
e^{\frac{\lambda}{2}}dr = \frac{q(r)}{r^2}, \label{eq5}\end{aligned}$$ where $q(r)$ is the total charge within a sphere of radius $r$.
Following the MIT bag model, we take the simple form of the strange matter EOS $$p_r = \frac{1}{3} (\rho -4B_g), \label{eq6}$$ where, $B_g$ is the bag constant. With the choice of the above EOS, we have a system of five independent equations with five unknown parameters namely, $\rho$, $p_r$, $p_t$, $E(r)$ and $\sigma(r)$. Substituting the metric potentials given by $\lambda(r)=A r^2$, $\nu(r) = B r^2 + C$, and their derivatives in Eqs. (\[eq2\])-(\[eq6\]), we obtain $$\begin{aligned}
\rho &=& \frac{3}{ 16 \pi}(A+B)e^{-Ar^2} + B_g,\label{eq7}\\
p_r &=& \frac{ 1}{ 16 \pi}(A+B)e^{-Ar^2} - B_g,\label{eq8}\\
p_t &=& \frac{ 1}{ 8\pi}\left[ \left(\frac{7}{2}B-
\frac{3}{2}A +B^2r^2-ABr^2+\frac{1}{r^2} \right)e^{-Ar^2}\right.\nonumber \\
& &\left. - \frac{1}{r^2}\right] + B_g,\label{eq9}\\
E^2 &=& \frac{ 1}{2}\left(A-3B-\frac{2}{r^2}\right)e^{-Ar^2} + \frac{1}{r^2}- 8 \pi B_g.\label{eq10}\end{aligned}$$ The charge density is obtained as $$\begin{aligned}
\sigma &=& \frac{ e^{-\frac{Ar^2}{2}}}{2\pi r }\psi +
\frac{ Ae^{-\frac{3Ar^2}{2}}}{8\pi r\psi}\left[2-(A -3B)r^2
\right] \nonumber \\
&& + \frac{e^{-\frac{Ar^2}{2}}}{4\pi r^3 \psi}[
e^{-Ar^2}-1],\label{eq11}\end{aligned}$$ where $$\psi = \sqrt{~\left[ \frac{ 1}{
2}\left(A-3B-\frac{2}{r^2}\right)e^{-Ar^2} + \frac{1}{r^2}- 8 \pi
B_g\right]}.$$ The charge within a sphere of radius $r$ turns out to be $$q = r^2\sqrt{\left[\frac{ 1}{
2}\left(A-3B-\frac{2}{r^2}\right)e^{-Ar^2} + \frac{1}{r^2}- 8 \pi
B_g\right]}.\label{eq12}$$ The anisotropic stress is obtained as $$\begin{aligned}
\Delta &=& p_t-p_r = 2B_g -\frac{1}{8\pi r^2} \nonumber \\
&& +\frac{ 1}{8\pi}\left(3B - 2A + B^2 r^2 -A B r^2 +\frac{1}{r^2}
\right)e^{-Ar^2}.\label{eq13}\end{aligned}$$
Bounds on the model parameters {#sec:3}
==============================
One of the advantages of using the KB metric is that there there is no singularity in its metric functions. Proper bounds, however, should be imposed on the conatants appearing in the metric functions so that all the physically significant parameters remain well behaved at all interior points of the star.
Regularity at the centre ($r=0$): {#sec:3.1}
---------------------------------
From Eq. (\[eq7\]), we obtain the central density in the form $$\rho_0 = \rho(r=0) = \frac{3}{ 16 \pi}(A+B) + B_g.\label{eq14}$$ For regularity of the electric field, it must vanish at the centre, i.e., $$E^2(r=0) = \frac{3}{ 2}(A-B)- 8\pi B_g = 0,\label{eq15}$$ which yields $$B_g = \frac{3}{16\pi}(A-B).\label{eq16}$$ Substituting the value of $B_g$ in Eq. (\[eq14\]), we note that the parameter $A$ corresponds to the central density given by $$A = \frac{8\pi \rho_0}{3}.\label{eq17}$$ Eq. (\[eq17\]) implies that $A$ is finite and positive. From Eq. (\[eq16\]), it then follows that for a positive value of the Bag constant, we must have $A > B$.
The two pressures and density should be decreasing functions of $r$. In our model, radial variation $p_r$ is obtained as $$\frac{d p_r}{dr} = -\frac{ 1}{ 8 \pi}(A+B)rAe^{-Ar^2}.\label{eq18}$$ Obviously, at $r=0$, $\frac{d p_r}{dr}$ = 0. Now, $\frac{d^2 p_r}{dr^2} < 0$, if the condition $- A(A+B) < 0$ is satisfied. Since $A$ is positive and $A > B$, this implies that $B > 0$. In Sect. \[sec:5.2\], we have shown that if the strong energy condition has to be satisfied then $B >0$. Therefore, if we assume that the matter within the sphere satisfies the strong energy condition so that $B>0$, then $p_r$ will decrease radially outward. Similarly, it can be shown that $p_t$ also decreases radially outward. The radial variation of the energy density is obtained as $$\frac{d \rho}{dr} = -\frac{ 3}{ 8\pi}(A+B)rAe^{-Ar^2},\label{eq19}$$ which also shows that at $r=0$, $\frac{d \rho}{dr} $ = 0 and $\frac{d^2\rho}{dr^2} = -\frac{ 3}{ 8 \pi}(A+B) < 0$. Thus, in this set up, a necessary and sufficient condition for regular behaviour of the physical parameters will be $A > B > 0$.
Regularity at the boundary ($r=R$): {#sec:3.2}
-----------------------------------
The radius $R$ of the star can be obtained by utilizing the condition that the radial pressure should vanish at the surface i.e., $$p_r ( r=R) = \frac{1}{ 16 \pi}(A+B)e^{-AR^2}- B_g = 0.\label{eq20}$$ This yields $$R = \sqrt{\frac{1}{A}ln\left[\frac{A+B}{16\pi B_g}\right]}.\label{eq21}$$ Since all the parameters on the right hand side of Eq. (\[eq21\]) have positive values as discussed in Sect. \[sec:3.1\], the radius of the star is finite and positive.
The exterior space-time of the star will be described by the Reissner-Nordström metric[@Reissner1916; @Nordstrom1918] given by $$\begin{aligned}
ds^{2} &=& -\left(1 - \frac{2M}{r} + \frac {Q^2}{r^2}\right)dt^2 + \left(1 - \frac{2M}{r} + \frac {Q^2}{r^2}\right)^{-1}dr^2
\nonumber\\
&& + r^2(d\theta^2+\sin^2\theta d\phi^2), \label{eq22}\end{aligned}$$ where, $Q$ is the total charge enclosed within the boundary $r=R$. Continuity of the metric coefficients $g_{tt}$, $g_{rr}$ and $\frac{\partial g_{tt}}{\partial r}$ across the boundary surface $r= R$ between the interior and the exterior regions of the star yields the following results: $$\begin{aligned}
1 - \frac{2M}{R} + \frac {Q^2}{R^2} &=& e^{BR^2+C},\label{eq23}\\
1 - \frac{2M}{R} + \frac {Q^2}{R^2} &=& e^{-AR^2},\label{eq24}\\
\frac{M}{R^2} - \frac {Q^2}{R^3} &=& B Re^{BR^2+C}.\label{eq25}\end{aligned}$$ Eqs. (\[eq23\]) - (\[eq25\]) determine the values of the constants $A$, $B$ and $C$ in terms of the total mass $M$, radius $R$ and charge $Q$. By solving the above set of equations, we get $$\begin{aligned}
A &=& - \frac{1}{R^2} \ln \left[ 1 - \frac{2M}{R} + \frac {Q^2}{R^2}
\right], \label{eq26}\\
B &=& \frac{1}{R^2} \left[\frac{M}{R} - \frac {Q^2}{R^2}\right] \left[1 - \frac{2M}{R} + \frac {Q^2}{R^2}
\right]^{-1},\label{eq27}\\
C &=& \ln \left[ 1 - \frac{2M}{R} +
\frac {Q^2}{R^2} \right]- \frac{ \frac{M}{R} - \frac {Q^2}{R^2}}{
\left[ 1 - \frac{2M}{R} + \frac {Q^2}{R^2} \right]}. \label{eq28}\end{aligned}$$ Note that the values of the parameters $M$, $R$ and $Q$ should be such that the condition $A > B > 0$ is satisfied. Moreover, consistency of the above equations puts the following constraint on the system: $$\begin{aligned}
\left[\frac{M}{R} - \frac {Q^2}{R^2}\right] \left[1 - \frac{2M}{R}
+ \frac {Q^2}{R^2} \right]^{-1} \nonumber \\
+\left[\frac{2+ \frac{2M}{R} -
\frac {Q^2}{R^2}}{4 - \frac{2M}{R} + \frac {Q^2}{R^2}}\right]
\times \ln \left[1 - \frac{2M}{R} + \frac
{Q^2}{R^2}\right] = 0.\label{eq29}\end{aligned}$$ Eq. (\[eq29\]) is useful to get an estimate of the charge to radius ($Q/R$) ratio for a given compactness ($M/R$) of the star. Based on logarithmic principle, another condition that must be fulfilled is that the total charge $Q^2 < 2R M$. Therefore, physical values of the parameters like the mass, size and charge can not be fixed arbitrarily in this construction. In Sect. \[sec:4\], we have demonstrated that it is possible to choose numerical values of masses and radii, consistent with the above constraints. These physical values indicate that the corresponding configurations are more likely to be strange stars rather than neutron stars.
Estimation of physical values {#sec:4}
=============================
In this section, we assume the mass and radius of a star consistent with the bounds discussed in the previous section so that the compactness the star is greater than that of a neutron star. This will, in turn, help us to get an estimate of the physically relevant parameters like the energy density, pressure and the bag constant. We have considered compact stars of different compactification factors and calculated the corresponding constants. The results have been shown in Table \[tab:1\] & \[tab:2\]. For example, for star of mass $1.4~M_{\odot}$ and radius $R=6.88~$km, the values of the constants are obtained as $A=0.017977861$, $B=0.013506968$, $B_g = 0.0002669721$ in units of km$^{-2}$ and $Q^2/R^2 =0.027$ (Case II). Plugging in $G$ and $c$ in the relevant equations, the values turn out to be $\rho_0 = 2.895\times 10^{15}$ gm cm$^{-3}$, $\rho_R = 1.443\times 10^{15}$ gm cm$^{-3}$, $p_r(r=0) = p_t(r=0) = 4.361\times 10^{35}$ dyne cm$^{-2}$ and the bag constant $B_g = 202.275$ MeV fm$^{-3}$. Numerical values for other cases have been shown in Table \[tab:3\]. Note that each case satisfies the condition $A > B > 0$. Making use of the constraint Eq. (\[eq29\]), we also note that for $Q^2/R^2 = 0.004$, the minimum value of $M/R = 0.25$. Therefore, for a star of mass $1.4~M_{\odot}$, the corresponds maximum radius turns out to be $R=8.26~$km. To illustrate the behaviour of physical parameters at the interior of the star, we have considered case II and plotted the variations of the energy density and the two pressures in Fi. \[fig1\] - \[fig3\], respectively.
The bag constant, in this framework, increases with the increase of the compactification factor, i.e., the bag constant is density dependent. A more dense star requires a greater bag constant. Similar observations may be found in Ref. [@Bordbar2], where a density dependent bag constant has been employed to model magnetized strange quark stars. The pioneering works of Farhi and Jaffe[@Farhi] showed that for a stable strange matter distribution, the bag constant should be around $\sim 60~$Mev/[fm]{}$^3$. Our results show that with the inclusion of anisotropy and electric field, the bag constant turns out to much more than its representative value. However, we would like to point out here that in Ref. [@Farhi], the calculation was made for a $\beta$-equilibrium strange quark matter satisfying the baryon number conservation principle where the charge neutrality condition was employed. The window of stability was parametrized by three factors namely, the bag constant, the mass of the quark particles and the QCD coupling constant. What happens when the charge neutrality condition is not imposed is not obvious from the analysis. For a relatively massive strange quark mass there could be an accumulation of net positive charge within the system. Probably, in the presence of charge, to counter the repulsive force generated due to the electric field, the bag pressure increases. The issue, however, is a matter of further investigation. What we have shown here is that if one wishes to use a mathematically consistent and physically reasonable analytic solution to model strange stars, the bag constant does not remain constant. Rather it becomes a free paramter which depends on the compactness of the star.
Case $M$ ($M_{\odot}$) $R$ (km) $\frac{M}{R}$ $\frac{Q^2}{R^2}$
------ ------------------- ---------- --------------- -------------------
I 1.4 8.26 0.25 0.004
II 1.4 6.88 0.30 0.027
III 1.4 5.90 0.35 0.061
IV 1.4 5.16 0.40 0.105
: Values of $Q^2/R^2$ for different choices of the compactification factor $M/R$.[]{data-label="tab:1"}
Case $A$ (km$^{-2}$) $B$ (km$^{-2}$) $B_g$ (km$^{-2}$)
------ ----------------- ----------------- -------------------
I 0.0102 0.0073 0.0001732
II 0.01798 0.01351 0.000267
III 0.0292 0.0231 0.0003643
IV 0.044 0.037 0.000418
: Values of the model parameters $A$ and $B$ as well as the bag constant $B_g$ for different choices of compactification factor $M/R$. Data obtained in Case II have been utilized to plot the figures.[]{data-label="tab:2"}
------ ----------------------- ----------------------- ----------------------- ----------------- --
Case $\rho (r=0)$ $\rho (r=R)$ $p_r(r=0)$ $B_g$
(gm cm$^{-3}$) (gm cm$^{-3}$) (dyne cm$^{-2}$) (MeV fm$^{-3}$)
I $1.643\times 10^{15}$ $9.362\times 10^{14}$ $2.123\times 10^{35}$ 131.204
II $2.895\times 10^{15}$ $1.443\times 10^{15}$ $4.361\times 10^{35}$ 202.275
III $4.703\times 10^{15}$ $2.015\times 10^{15}$ $8.204\times 10^{35}$ 275.98
IV $7.087\times 10^{15}$ $2.58\times 10^{15}$ $1.448\times 10^{35}$ 316.699
------ ----------------------- ----------------------- ----------------------- ----------------- --
: Energy density, pressure and bag constant for different cases shown in Table \[tab:1\] & \[tab:2\].[]{data-label="tab:3"}
![The energy density ($\rho$) plotted against $r$. \[fig1\]](fig-1.eps)
![Radial ($p_r$) and transverse ($p_t$) pressures plotted against $r$. \[fig2\]](fig-2.eps)
![The electric field ($E^2$) plotted against $r$. \[fig3\]](fig-3.eps)
Some salient features of the model {#sec:5}
==================================
Generalized TOV equations {#sec:5.1}
-------------------------
The generalized Tolman-Oppenheimer-Volkoff(TOV) equations, in this set up, gets the form $$-\frac{M_G\left(\rho+p_r\right)}{r^2}e^{\frac{\lambda-\nu}{2}}-\frac{dp_r}{dr}+\sigma
\frac{q
}{r^2}e^{\frac{\lambda}{2}}+\frac{2}{r}\left(p_t-p_r\right) = 0,\label{eq30}$$ where $M_G$ is the effective gravitational mass given by $$M_G(r)=\frac{1}{2}r^2e^{\frac{\nu-\lambda}{2}}\nu^{\prime} = B r^3 e^{\frac{1}{2}[(B-A)r^2 -C]}.\label{eq31}$$ Eq. (\[eq30\]) describes the equilibrium condition for a charged anisotropic fluid subject to gravitational ($F_g$), hydrostatic ($F_h$), electric ($F_e$) and anisotropic stress ($F_a$) so that $$F_g+ F_h+ F_e+ F_a=0, \label{eq32}$$ where, $$\begin{aligned}
F_g &=& -B r\left(\rho+p_r\right) = - \frac{Br}{4\pi} (A+B)e^{-Ar^2}, \label{eq33}\\
F_h &=& -\frac{dp_r}{dr} = \frac{ 1}{ 8 \pi}(A+B)rAe^{-Ar^2},\label{eq34}\\
F_e &=& \sigma E e^{\frac{A r^2}{2}} = \frac{1}{2\pi r}\left[\frac{e^{-A r^2/2}}{2} \left(A-3B-\frac{2}{r^2}\right)\right.\nonumber\\
&&\left. +\frac{1}{r^2} -8\pi B_g\right] + \frac{A e^{-A r^2}}{8\pi r}\left(2-(A-3B)r^2\right) \nonumber\\
&&+\frac{1}{4\pi r^2}\left(e^{-A r^2}-1\right),\label{eq35}\\
F_a &=& \frac{2}{r}\left(p_t-p_r\right) = \frac{2}{r}\left[\frac{1}{8\pi}\left[\left(3B -2A +B^2
r^2 -AB r^2 \right.\right.\right. \nonumber\\
&& \left.\left.\left. +\frac{1}{r^2} \right)e^{-Ar^2} - \frac{1}{r^2}
\right]+2B_g \right]. \label{eq36}\end{aligned}$$ In Fig. \[fig4\], the profiles of these force terms at the interior of the star for the case II have been shown. The plots indicate that an equilibrium stage can be achieved under the combined effects of gravitational, electric, hydrostatic and anisotropic stresses.
![Contributions of different force terms acting on fluid elements in static equilibrium. \[fig4\]](fig-4.eps)
Energy conditions {#sec:5.2}
-----------------
The anisotropic charged fluid sphere composed of strange matter will satisfy the null energy condition (NEC), weak energy condition (WEC) and strong energy condition (SEC) if the following inequalities hold simultaneously at all points within the star: $$\begin{aligned}
\rho + \frac{ E^2}{8\pi}\geq 0, \label{eq37}\\
\rho + p_r\geq 0, \label{eq38}\\
\rho + p_t + \frac{E^2}{4\pi}\geq 0, \label{eq39}\\
\rho + p_r + 2p_t+\frac{E^2}{4\pi}\geq 0.\label{ec40}\end{aligned}$$ Employing these energy conditions at the centre ($r=0$), we get the following bounds on the constants $A$ and $B$:\
(i) NEC: $\rho + \frac{ E^2}{8\pi}\geq 0 \Rightarrow A \geq 0 $.\
(ii) WEC: $\rho+p_r\geq 0 \Rightarrow A +B \geq 0,~~~\rho+p_t+\frac{E^2}{4\pi}\geq 0 \Rightarrow A +B \geq 0$\
(iii) SEC: $ \rho+p_r+ 2p_t+\frac{E^2}{4\pi}\geq 0 \Rightarrow B \geq 0$.\
Since the central density is given by $\rho_0 = \frac{3A}{8\pi}$, we must have $A > 0$, i.e., condition (i) is satisfied. The weak and strong energy conditions (ii) and (iii) will then be satisfied if $ B \geq 0$. With the set of values discussed Sect. \[sec:4\], we have shown in Fig. \[fig5\] that the energy conditions are satisfied simultaneously within the sphere.
![Different energy conditions plotted against $r$ for case II.[]{data-label="fig5"}](fig-5.eps)
Stability {#sec:5.3}
---------
To examine stability of the resultant configuration, we employ the technique based on Herrera’s[@Herrera1992] cracking (or overturning) concept. Physical acceptability conditions for a fluid distribution include the condition of causality. It suggests that the squares of the radial and tangential sound speeds should be within the limit $[0,1]$. Herrera’s [@Herrera1992] cracking (or overturning) concept implies that the region for which the radial speed of sound is greater than that of transverse speed is a potentially stable region. It also suggests that, for ‘no cracking’ to occur, the difference of the two sound speeds, i.e., $v_{st}^2 - v_{sr}^2$ should retain the same sign everywhere within the matter distribution. In our model, we have $$\begin{aligned}
v_{sr}^2 &=& \frac{dp_r}{d\rho} = \frac{1}{3},\label{eq41}\\
v_{st}^2 &=& \frac{dp_t}{d\rho} = \frac{\alpha +\beta}{-3(A+B)rAe^{-Ar^2}},\label{eq42}\end{aligned}$$ where, $$\begin{aligned}
\alpha &=& e^{-Ar^2}\left( 2B^2r -2ABr - \frac{2}{r^3}\right)+\frac{1}{r^2},\nonumber \\
\beta &=& -2Are^{-Ar^2}\left(\frac{7}{2}B- \frac{3}{2}A +B^2r^2-ABr^2 +\frac{1}{r^2}\right).\nonumber\end{aligned}$$ For causality condition to be satisfied we must have $$0 < \frac{\alpha+\beta}{-3(A+B)rAe^{-Ar^2}} < 1. \label{eq43}$$ For the assumed set of values, we note that $0\leq v_{sr}^2 \leq 1$ and $0\leq v_{st}^2 \leq 1$ as shown in Fig. \[fig6\]. Following the works of Herrera[@Herrera1992] and Andréasson[@Andreasson1992], we note that the configuration will be stable provided $\mid v_{st}^2 - v_{sr}^2 \mid \leq 1 $. Fig. \[fig7\]-\[fig8\] show that the above condition is satisfied for the assumed configuration implying stability of the configuration.
![$\mid v_{st}^2 -
v_{sr}^2 \mid $ plotted against $r$. []{data-label="fig6"}](fig-6.eps)
![$\mid v_{st}^2 -
v_{sr}^2 \mid $ plotted against $r$. []{data-label="fig7"}](fig-7.eps)
![$ v_{st}^2 - v_{sr}^2$ plotted against $r$. []{data-label="fig8"}](fig-8.eps)
Effective mass-radius relation {#sec:5.4}
------------------------------
For a static spherically symmetric perfect fluid star, Buchdahl[@Buchdahl1959] derived an upper limit for maximum allowed mass to radius ratio as $\frac{2M}{R} < \frac{8}{9}$ which was later generalized by Mak [*et al*]{}[@Mak2001] for a charged sphere. In this model, the effective gravitational mass has the form $$M_{eff} = 4\pi\int^{R}_{0}\left(\rho+\frac{E^2}{8\pi}\right)r^2 dr =
\frac{1}{2}R\left( 1-e^{-AR^2}\right).\label{eq44}$$ In Fig. \[fig9\], we have shown the variation of mass against radius. We have also plotted $\frac{M _{eff}}{R}$ against $R$ in Fig. \[fig10\] which shows that the ratio $\frac{M _{eff}}{R}$ is an increasing function of the radial parameter. We note that the constraint on the maximum allowed mass-radius ratio in this case is similar to that of an isotropic fluid sphere, i.e., $\frac{M}{R} < \frac{4}{9}$, as obtained by Buchdahl[@Buchdahl1959]. Defining the compactification factor as $$u = \frac{ M_{eff}(R)} {R}= \frac{1}{2}\left(1-e^{-AR^2} \right),\label{eq45}$$ the surface red-shift ($Z_s$) corresponding to the above compactness ($u$) is obtained as $$Z_s= ( 1-2 u)^{-\frac{1}{2}} - 1=
e^{\frac{AR^2}{2}}-1.\label{eq46}$$ The maximum surface redshift, in this set up, for a strange star of mass $1.4~M_{\odot}$ and radius $6.88~$km turns out to be $Z_s = .5303334$.
A lower bound on the mass to radius ratio for a charged sphere has been reported by Andréasson[@Andreasson1992] which has the form $$\sqrt{M} < \frac{\sqrt{R}}{3} + \sqrt{\frac{ R }{9} + \frac{Q^2}{3R}}.\label{eq47}$$ This inequality is applicable to stellar objects satisfying the inequality $p_r+ 2p_t \leq \rho$. In Fig. \[fig12\], we have plotted $p_r+ 2p_t - \rho$ against $r$ which shows that at all interior points, the above condition is satisfied, i.e., Andréasson’s inequality holds in our model.
![$M_{eff}$ plotted against $R$. \[fig9\]](fig-9.eps)
![$\frac{M _{eff}}{R}$ plotted against $R$. \[fig10\]](fig-10.eps)
![The redshift function $Z_s$ plotted against $R$. \[fig11\]](fig-11.eps)
![$2p_t+p_r - \rho$ plotted against $r$. \[fig12\]](fig-12.eps)
Discussions {#sec:6}
===========
We have explored the relevance of KB model[@Krori1975] in the modelling of strange quark stars where the strange matter EOS based on the MIT bag model has been assumed. The inclusion of the EOS does not make the system over determined since the matter distribution in the set up has been assumed to anisotropic in nature together with high electric field. We have shown that a self-consistent mathematical model can generate physical values which are consistent with a strange star. The bag constant, however, in this framework turns out to be above the range specified for a stable strange quark matter, i.e., ($60-80~$Mev/fm$^3$)[@Farhi; @Weber; @Alcock]. However, experimental results from CERN-SPS and RHIC show that a wide range of values of the bag constant is possible for a density-dependent bag model[@Burgio]. Perhaps, to compensate the extra repulsion in the presence of electric field, the bag value increases in our set up. The impacts of anisotropy and high electric field on the bag model is a matter of further investigation and will be taken up elsewhere.
FR, RS and SR gratefully acknowledge support from the Inter-University Centre for Astronomy and Astrophysics (IUCAA), Pune, India, under the visiting research associateship programme, under which a part of this work was carried out. FR is also thankful to PURSE, DST and UGC, Govt. of India, under Research Award, for providing financial support.
K. D. Krori and J. Barua, (1975) [*J. Phys. A.: Math. Gen.*]{} [**8**]{}, 508. A. Hewish, S. J. Bell, J. D. H. Pilkington, P. F. Scott and R. A. Collins (1968) [*Nature*]{} [**217**]{}, 709. S. L. Shapiro and S. A. Teukolosky, [*Black Holes, White Dwarfs and Neutron Stars: The Physics of Compact Objects*]{}, (Wiley, New York, 1983). E. Witten, (1984) [*Phys. Rev. D*]{} [**30**]{}, 272. E. Farhi and R. L. Jaffe, (1984) [*Phys. Rev. D*]{} [**30**]{}, 2379. C. Alcock, E. Farhi and A. Olinto, (1986) [*Astrophys. J.*]{} [**310**]{}, 261. P. Haensel, J. L. Zdunik and R. Schaeffer, (1986) [*Astron. Astrophys.*]{} [**160**]{}, 121. F. Weber, (2005) [*Prog. Part. Nucl. Phys.*]{} [**54**]{}, 193. M. A. Perez-Garcia, J. Silk and J. R. Stone, (2010) [*Phys. Rev. Lett.*]{} [**105**]{}, 141101. H. Rodrigues, S. B. Duarte and J. C. T. de Oliveira, (2011) [*Astrophys. J.*]{} [**730**]{}, 31. G. H. Bordbar and A. R. Peivand, (2011) [*Research in Astron. Astrophys.*]{} [**11**]{}, 851. M. G. B. de Avellar and J. E. Horvath, (2010) [*Int. J. Mod. Phys. D*]{} [**D19**]{}, 1937. A. Chodos, R. L. Jaffe, K. Johnson, C. B. Thorn and V. F. Weisskopf, (1974) [*Phys. Rev. D*]{} [**9**]{}, 3471. R. L. Bowers and E. P. T. Liang, (1974) [*Astrophys. J.*]{} [**188**]{}, 657. V. V. Usov, (2004) [*Phys. Rev. D*]{} [**70**]{}, 067301. G. J. G. Junevicus, (1976) [*J. Phys. A.: Math. Gen.*]{} [**9**]{}, 2069. B. V. Ivanov, (2002) [*Phys. Rev. D*]{} [**65**]{}, 104001. V. Varela, F. Rahaman, S. Ray, K. Chakraborty and M. Kalam, (2010) [*Phys. Rev. D*]{} [**82**]{}, 044052. F. Rahaman, S. Ray, A. K. Jafry and K. Chakraborty, (2010) [*Phys. Rev. D*]{} [**82**]{} 104055. H. Reissner, (1916) [*Ann. Phys., Lpz.*]{} [**50**]{}, 106. G. Nordstr[ö]{}m, (1918) [*Roc. K. Ned. Akad. Wet.*]{} [**20**]{}, 1238. G. H. Bordbar, H. Bahri and F. Kayanikhoo, (2012) [*hep-ph/1204.0325v1*]{}. L. Herrera, (1992) [*Phys. Lett. A*]{} [**165**]{} 206. H. A. Buchdahl, (1959) [*Phys. Rev.*]{} [**116**]{}, 1027. M. K. Mak, P. N. Dobson and T. Harko, (2001) [*Europhys. Lett.*]{} [**55**]{}, 310. H. Andréasson, (2009) [*Commun. Math. Phys.*]{} [**288**]{}, 715. G. F. Burgio, M. Baldo, P. K. Sahu and H. -J. Schulze, (2002) [*Phys. Rev. C*]{} [**66**]{}, 025802.
[^1]: [email protected]
[^2]: [email protected]
[^3]: [email protected]
[^4]: [email protected]
[^5]: [email protected]
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We consider the late-time tails of spherical waves propagating on even-dimensional Minkowski spacetime under the influence of a long range radial potential. We show that in six and higher even dimensions there exist exceptional potentials for which the tail has an anomalously small amplitude and fast decay. Along the way we clarify and amend some confounding arguments and statements in the literature of the subject.'
author:
- Piotr Bizoń
- Tadeusz Chmaj
- Andrzej Rostworowski
title: Anomalously small wave tails in higher dimensions
---
Introduction
============
It is well known that sharp propagation of free waves along light cones in even-dimensional flat spacetimes, known as Huygens’ property, is blurred by the presence of a potential. Physically, the spreading of waves inside the light cone is caused by the backscattering off the potential. If the potential falls off exponentially or faster at spatial infinity, then the backscattered waves decay exponentially in time, while the long range potentials with an algebraic fall-off give rise to tails which decay polynomially in $1/t$. The precise description of these tails is an important issue in scattering theory. There are two main approaches to this problem in the literature. On the one hand, there are mathematical results in the form of various decay estimates. These results are rigorous, however they rarely give optimal decay rates inside the light cone and provide very poor information about the amplitudes of tails. The notable exception is the work of Strauss and Tsutaya [@st] (recently strengthened by Szpak [@sz]) where the optimal pointwise decay estimate for the tail was proved in four dimensions. Unfortunately, to the best of our knowledge, there is no analogous result in higher dimensions.
On the other hand, there are non-rigorous results in the physics literature based on perturbation theory. The most complete work in this category was done by Ching *et al.* [@ching] who derived first-order approximations of the tails for radial potentials. Although these results were originally formulated for partial waves in four dimensions, they can be easily translated to spherical waves in higher dimensions. Ching *et al.* noticed that there are exceptional potentials for which the first-order tail vanishes, however they did not pursue their analysis to the second order, apart from giving some dimensional arguments. The main purpose of this paper is to analyze the tails for such exceptional potentials in more detail.
One of the physical motivations behind our work stems from the fact that this kind of potentials arise in the study of linearized perturbations of higher even-dimensional Schwarzschild black holes. The behavior of tails on the Schwarzschild background in well known in four dimensions (see [@p], [@l], [@ching], [@gpp], [@b], [@dr]), but not in higher even dimensions (despite statements to the contrary in the literature [@car]). Although our analysis is restricted to the flat background, it sheds some light on the problem of tails on the black hole background because the properties of tails are to some extent independent of what happens in the central region.
The rest of the paper is organized as follows. In section 2 we construct the iterative scheme for the perturbation expansion of a spherically symmetric solution of the linear wave equation with a potential. This scheme is applied in section 3 to derive the first and second-order approximations of the tails for radial potentials which fall off as pure inverse-power at infinity. In section 4 we discuss the modifications caused by subleading terms in the potential. Section 5 contains numerical evidence confirming the analytic formulae from sections 3 and 4. Finally, in section 6 we give a heuristic argument to predict the behavior of tails outside Schwarzschild black holes in higher even dimensions. Technical details of most calculations are given in the appendix.
Throughout the paper we use the succinct notation and summation technics from the excellent book by Graham *et al.* [@gkp]. In particular, we shall frequently use the following abbreviations $$\begin{aligned}
x^{\underline{0}} := 1, &\qquad& x^{\underline{k}} := x \cdot (x-1)
\cdot \dots \cdot (x-(k-1)), \quad k>0\,,
\\
x^{\overline{0}} := 1, &\qquad& x^{\overline{k}} := x \cdot (x+1)
\cdot \dots \cdot (x+(k-1)), \quad k>0\,.\end{aligned}$$
Iterative scheme
================
We consider the wave equation with a potential in even-dimensional Minkowski spacetime $R^{d+1}$ $$\label{eqm}
\partial_t^2 \phi -\Delta \phi + \lambda V \phi =0\,.$$ The prefactor $\lambda$ is introduced for convenience - throughout the paper we assume that $\lambda$ is small which allows us to use it as the perturbation parameter. The precise assumptions about the fall-off of the potential will be formulated below. We restrict attention to spherical symmetry, i.e., we assume that $\phi=\phi(t,r)$ and $V=V(r)$. Then, equation (\[eqm\]) becomes $$\label{eqs}
\mathcal{L} \phi + \lambda V(r) \phi =0\,,\qquad \mathcal{L}
:=\partial_t^2 -\partial_r^2 -\frac{d-1}{r} \partial_r\,.$$ We are interested in the late-time behavior of $\phi(t,r)$ for smooth compactly supported (or exponentially localized) initial data. $$\label{id}
\phi(0,r)=f(r),\qquad \partial_t \phi(0,r)=g(r)\,.$$ To determine the asymptotic behavior of solutions we define the perturbative expansion (Born series) $$\label{pert}
\phi=\sum_{n=0} \lambda^n \phi_n\,,$$ where $\phi_0$ satisfies initial data (\[id\]) and all $\phi_n$ with $n>0$ have zero initial data. Substituting this expansion into equation (\[eqs\]) we get the iterative scheme $$\label{scheme}
\mathcal{L} \phi_n = -V \phi_{n-1}\,, \qquad \phi_{-1}=0\,,$$ which can be solved recursively. The zeroth-order solution is given by the general regular solution of the free radial wave equation which is a superposition of outgoing and ingoing waves [@k] $$\label{f0}
\phi_0(t,r)=\phi_0^{ret}(t,r)+\phi_0^{adv}(t,r)\,,$$ where $$\label{f1}
\phi_0^{ret}(t,r)= \frac{1}{r^{l+1}}\,\sum_{k=0}^{l} \frac {(2l-k)!} {k!(l-k)!} \frac
{a^{(k)}(u)}{(v-u)^{l-k}}\,, \qquad \phi_0^{adv}(t,r) =
\frac{1}{r^{l+1}}\,\sum_{k=0} ^{l} (-1)^{k+1} \frac {(2l-k)!}
{k!(l-k)!} \frac {a^{(k)}(v)}{(v-u)^{l-k}}\,,$$ and $u=t-r$, $v=t+r$ are the retarded and advanced times, respectively. Here and in the following, instead of $d$, we use the index $l$ defined by $d=2l+3$ (remember that we consider only *odd* space dimensions $d$). Note that for compactly supported initial data the generating function $a(x)$ can be chosen to have compact support as well (this condition determines $a(x)$ uniquely).
To solve equation (\[scheme\]) for the higher-order perturbations we use the Duhamel representation for the solution of the inhomogeneous equation $\mathcal{L} \phi = N(t,r)$ with zero initial data $$\label{a2}
\phi(t,r)= \frac{1}{2 r^{l+1}}
\int\limits_{0}^{t} d\tau \int\limits_{|t-r-\tau|}^{t+r-\tau} \rho^{l+1} P_l(\mu)
N(\tau,\rho) d\rho\,,$$ where $P_l(\mu)$ are Legendre polynomials of degree $l$ and $\mu=(r^2+\rho^2-(t-\tau)^2)/2r\rho$ (note that $-1\leq \mu \leq 1$ within the integration range). This formula can be readily obtained by integrating out the angular variables in the standard formula $\phi=G^{ret} * N$ where $G^{ret}(t,x)=(2\pi^{l+1})^{-1}\Theta(t)
\delta^{(l)}(t^2-|x|^2)$ is the retarded Green’s function of the wave operator in $d+1$ dimensions (see, for example, [@ls]). It is convenient to express (\[a2\]) in terms of null coordinates $\eta=\tau-\rho$ and $\xi=\tau+\rho$ $$\label{duh1}
\phi(t,r)= \frac{1}{2^{l+3} r^{l+1}}
\int\limits_{|t-r|}^{t+r} d\xi \int\limits_{-\xi}^{t-r} (\xi-\eta)^{l+1} P_l(\mu)
N(\eta,\xi) d\eta\,,$$ where now $ \mu=(r^2+(\xi-t)(t-\eta))/r(\xi-\eta)$. Using this representation we can rewrite the iterative scheme (\[scheme\]) in the integral form $$\label{iter}
\phi_{n}(t,r)= -\frac{1}{2^{l+3} r^{l+1}}
\int\limits_{|t-r|}^{t+r} d\xi \int\limits_{-\xi}^{t-r}
(\xi-\eta)^{l+1} P_l(\mu) V(\rho(\eta,\xi))\phi_{n-1}(\eta,\xi)
d\eta\,.$$ This “master” equation will be applied below to evaluate the first two iterates for a special class of potentials. It is natural to expect that for sufficiently small $\lambda$ these iterates provide good approximations of the true solution.
Pure inverse-power potentials at infinity
=========================================
In this section we consider the simple case (below referred to as type I) when the potential is *exactly* $V(r)= r^{-\alpha}$ for $r$ greater than some $r_0>0$. We assume that $\alpha>2$. The modifications caused by subleading corrections to the pure inverse-power decay of the potential will be discussed in section 4.
Generic case
------------
We wish to evaluate the first iterate $\phi_1(t,r)$ near timelike infinity, i.e, for $r=const$ and $t\rightarrow \infty$. Thanks to the fact that $\phi_0(\eta,\xi)$ has compact support we may interchange the order of integration in (\[iter\]) and drop the advanced part of $\phi_0(\eta,\xi)$ to obtain $$\label{iter1}
\phi_{1}(t,r) = -\frac{2^{\alpha}}{2^{l+3} r^{l+1}}
\int\limits_{-\infty}^{\infty} d\eta \int\limits_{t-r}^{t+r} (\xi-\eta)^{l+1-\alpha}
P_l(\mu) \phi_0^{ret}(\eta,\xi)
d\xi\,,$$ where we have substituted $V= 2^{\alpha} (\xi-\eta)^{-\alpha}$. Plugging (\[f1\]) into (\[iter1\]), after a long calculation (see appendix A for the technical details), we get $$\begin{aligned}
\phi_1(t,r) &=& - 2^{\alpha+3l-1} \left( \frac {\alpha-3} {2}
\right)^{\underline{l}} \left( \frac {\alpha} {2}
\right)^{\overline{l}} \int \limits_{-\infty}^{+\infty} d\eta \,
a(\eta) \, \frac {(t-\eta)^{\alpha-2}} {\left[ (t-\eta)^2 - r^2 \right]^{\alpha-1+l}}
\nonumber\\
&\times& \sum_{0 \leq n \leq \lfloor (\alpha-2) / 2 \rfloor} (-1)^{n} \frac {2^{2n}(l+n)!} {n! (2l+2n+1)!}
\left( - \frac {\alpha-2} {2} - l - 1 \right)^{\underline{n}} \left( \frac {\alpha-1} {2} - l - 1 \right)^{\underline{n}}
\nonumber\\
&\times& \sum_{n \leq m \leq n+l} (-1)^{m} \left( \begin{array}{c} l \\ m-n \end{array} \right)
\frac {\left( - \frac {\alpha} {2} + 1 \right)^{\overline{m}}} {\left( \frac {\alpha} {2} \right)^{\overline{m}}} \left( \frac {r} {t-\eta} \right)^{2m}.
\label{tailf1}\end{aligned}$$ Asymptotic expansion of (\[tailf1\]) near timelike infinity yields the following first-order approximation of the tail $$\label{tail1}
\phi(t,r) \approx \lambda \phi_1(t,r) = \lambda \, \frac {C(l,\alpha)} {t^{\alpha+2l}} \left[ A + (\alpha+2l)\frac{B}{t} \, + \mathcal{O} \left( \frac{1}{t^2}
\right) \right] \, ,$$ where $$\label{B}
C(l,\alpha) = - \frac {2^{\alpha+2l-1}}{(2l+1)!!} \left( \frac {\alpha-3} {2} \right)^{\underline{l}}
\left( \frac {\alpha} {2} \right)^{\overline{l}}\,,$$ and $$\label{ab}
A=\int \limits_{-\infty}^{+\infty} a(\eta)\, d\eta\,,\qquad B=\int \limits_{-\infty}^{+\infty}
a(\eta)\,\eta\,d\eta\,.$$ In general $A\neq 0$ and the tail decays as $t^{-\alpha-2l}$, however there are nongeneric initial data for which $A=0$ and then the tail decays as $t^{-\alpha-2l-1}$; in particular this happens for time symmetric initial data for which $a(x)$ is an odd function. 0.2cm *Remark 1.* It is easy to check that if the function $\phi(t,r)$ satisfies equation (\[eqs\]), then the function $\psi=r^{l+1}\phi$ satisfies the radial wave equation for the $l$th multipole $$\label{a3}
(\partial_t^2-\partial_r^2+l(l+1)/r^2)\psi + \lambda V(r)\psi =0\,.$$ The late-time tails for this equation were studied by Ching *et al.* [@ching] who derived the formula equivalent to (\[tail1\]) via the Fourier transform methods.
Exceptional case
----------------
It follows from (\[tail1\]) that if $\alpha$ is an odd integer satisfying $3\leq \alpha \leq 2l+1$, then $\phi_1(t,r)$ vanishes identically due to factor $\left( \frac {\alpha-3} {2} \right)^{\underline{l}}$ in (\[B\]) and there is no (polynomial) tail whatsoever in the first order. Thus, in order to compute the tail in this exceptional case we need to go the second order of the perturbation expansion.
Using (\[iter\]) and proceeding as above we get the second iterate $$\label{iter2}
\phi_{2}(t,r) = -\frac{2^{\alpha}}{2^{l+3} r^{l+1}}
\int\limits_{-\infty}^{\infty} d\eta \int\limits_{t-r}^{t+r}
(\xi-\eta)^{l+1-\alpha} P_l(\mu) \phi_1^{ret}(\eta,\xi)
d\xi\,,$$ where $\phi_1^{ret}$ is the outgoing solution of the inhomogeneous equation $$\label{NH}
\mathcal{L} \phi_1=-V \phi_0\,.$$ In general $\phi_1$ is a sum of the solution of the homogeneous equation and the particular solution of the inhomogeneous equation. The homogeneous part has the form (\[f1\]) (with a different generating function than $a$, but still compactly supported), thus for the same reason as above it gives no contribution to the tail. The particular solution of the inhomogeneous equation (\[NH\]) reads $$\label{f1nul} \phi_l^{NH} = \frac {1}{2 (\alpha-1) r^{\alpha +l}}
\sum_{q=0}^{l-\alpha/2+1/2} (l-\alpha/2+1/2)^{\underline{q}} \;
\frac {2^q \left(\alpha/2 \right)^{\overline{q}}}
{\alpha^{\overline{q}}} \frac {\phi_{l-1-q}^{H}} {r^q} \, ,$$ where $\phi_{l-1-q}^{H}$ denotes the solution of the homogeneous equation with $d=2(l-1-q)+3$ and the same generating function $a$ as in $\phi_0$ (see (\[f1\])). The formula (\[f1nul\]) can be easily derived by the method of undetermined coefficients (we emphasize that this formula is valid *only* for odd $\alpha$ satisfying $3\leq \alpha \leq 2l+1$). Substituting (\[f1nul\]) into (\[iter2\]), after a long calculation (see appendix A for the technical details), we obtain the following asymptotic behavior near timelike infinity $$\label{tail2}
\phi(t,r) \approx \lambda^2 \, \phi_2(t,r) = \lambda^2 \frac {D(l,\alpha)} {t^{2(\alpha+l-1)}}
\left[ A + 2(\alpha+l-1)\frac{B}{t} + \mathcal{O} \left( \frac{1}{t^2} \right) \right]\,,$$ where the coefficients $A$ and $B$ are defined in (\[ab\]) and $$\label{c2}
D(l,\alpha)= \frac {2^{2(\alpha+l-2)}}{(2l+1)!!} \cdot \frac{(2\alpha-3)}{2(\alpha-1)}
\left( \alpha - \frac {5} {2} \right)^{\underline{l-1}} \left( \alpha - 2 + l
\right)^{\underline{l-1}}\,
F\left( \left.
\begin{array}{c} -l+\alpha/2-1/2,\, \alpha/2,\, 2\alpha-2,\, 1 \\ \alpha,\, \alpha,\,
\alpha - l-1/2 \end{array} \right| 1
\right)\,.$$ Here $F$ stands for the generalized hypergeometric function $$\label{F}
F\left( \left.
\begin{array}{c}
a_1,\, \dots,\, a_m
\\
b_1,\, \dots,\, b_n
\end{array}
\right| z \right) = \sum _{k \geq 0} \frac {a_1^{\overline{k}},\,
\dots,\, a_m^{\overline{k}}}
{b_1^{\overline{k}},\, \dots,\, b_n^{\overline{k}}}
\frac{z^k}{k!}\,.$$
[|c||\*[4]{}[c|]{}]{} & & & &\
$1$ & $4$ & & &\
$2$ & $-8/5$ & $2240/3$ & &\
$3$ & $96/35$ & $1792$ & $2523136/5$ &\
$4$ & $-64/7$ & $-17920/9$ & $16580608/5$ & $4638965760/7$\
We remark that the behavior $\mathcal{O}(\lambda^2)\,
t^{-2(l+\alpha-1)}$ of the tail (\[tail2\]) was conjectured before by Ching *et al.* [@ching] on the basis of dimensional analysis.
General polynomially decaying potentials
========================================
In this section we analyze how the presence of subleading corrections to the pure inverse-power asymptotic behavior of the potential affects the results obtained in section 3. We restrict ourselves to the most interesting and common case (below referred to as type II) when near infinity $$\label{case2}
V(r) = \frac{1}{r^{\alpha}} \left(1+\frac{\beta}{r^{\gamma}}\right)
+o\left(\frac{1}{r^{\alpha+\gamma}}\right),\,\qquad
\gamma>0\,.$$ If $C(\alpha,l)\neq 0$, then the dominant behavior of the tail is of course the same as in (\[tail1\]): $$\label{gtail1}
\phi(t,r) \sim \lambda \, A \, C(l,\alpha)\,
t^{-(\alpha+2l)}\,.$$ However, in the exceptional case, when $C(\alpha,l)=0$, the situation is more delicate. As we showed above, in this case there is the second-order contribution to the tail given by (\[tail2\]) $$\label{gtail2}
\phi_2(t,r) \sim A \, D(l,\alpha)\,
t^{-2(\alpha+l-1)}\,.$$ In contrast to the type I case where the first-order tail vanishes identically, in the type II case the subleading term in the potential produces the first-order contribution which is given by (\[tail1\]) with $\alpha$ replaced by $\alpha+\gamma$: $$\label{tailf1p}
\phi_1(t,r) \sim \beta \, A\, C(l,\alpha+\gamma)\,
t^{-(\alpha+\gamma+2l)}\,,$$ assuming that $\alpha+\gamma$ is not an odd integer $\leq
d-2$ (otherwise one has to repeat the analysis for the next subleading term in the potential).
Now, comparing the decay rates in (\[gtail2\]) and (\[tailf1p\]) we conclude that the leading asymptotics of the tail is given by the first-order term $\lambda \phi_1(t,r)$ if $\gamma\leq \alpha-2$ (we call it subtype IIa), but otherwise, *i.e.* for $\gamma
> \alpha-2$ (subtype IIb), the second-order term $\lambda^2 \phi_2(t,r)$ is dominant for $t\rightarrow \infty$. 0.2cm *Remark 2.* In the context of equation (\[a3\]) a formula analogous to (\[tailf1p\]) was obtained by Hod who studied tails in the presence of subleading terms in the potential (see subgroup IIIb in [@hod]). However, Hod’s analysis, restricted to the first-order approximation, was inconclusive because, as we just have shown, without the second-order formula (\[gtail2\]) one is not in position to make assertions about the dominant behavior of the tail. 0.2cm
Numerics
========
In order to verify the above analytic predictions we solved numerically the initial value problem (4-5) for various potentials and initial data. Our numerical algorithm is based on the method of lines with finite differencing in space and explicit fourth-order accurate Runge-Kutta time stepping. As was pointed out in [@ching], a reliable numerical computation of tails requires high-order finite-difference schemes, since otherwise the ghost potentials generated by discretization errors produce artificial tails which might mask the genuine behavior. The minimal order of spatial finite-difference operators depends on the fall-off of the potential – for the cases presented below the fourth-order accuracy was sufficient, but for the faster decaying potentials a higher-order accuracy is needed. To eliminate high-frequency numerical instabilities we added a small amount of Kreiss-Oliger artificial dissipation All computations were performed using quadruple precision which was essential in suppressing round-off errors at late times.
The numerical results presented here were produced for initial data of the form $$\label{idn}
\phi(0,r) = \exp(-r^2),\qquad
\partial_t \phi(0,r)= \exp(-r^2)\,.$$ As follows from (\[f1\]) the generating function for these data is $$\label{a}
a(x)=2^{-(l+2)} (1-2x) \exp(-x^2)\,,\quad \mbox{hence}\quad
A=\int_{-\infty}^{+\infty} a(x) dx=\sqrt{\pi}/2^{l+2}\,.$$ We considered the following potentials
[V(r)=]{} & )\
(1+) & ,
for various values of $\alpha$ and $\gamma$. The regularizing factor $\tanh(r)$ introduces exponentially decaying corrections to the pure inverse-power behavior at infinity but such corrections do not affect the polynomial tails. The numerical verification of the formulae (\[tail1\]), (\[tailf1p\]), and (\[tail2\]) is shown in tables II and III. The observed decay rates agree perfectly with analytic predictions, while small errors in the amplitudes are due to (neglected) higher-order terms in the perturbation expansion.
-- ----------- --------- ---------- ---------- ---------- --------- ----------
Theory Numerics Theory Numerics Theory Numerics
Exponent 2.5 2.499 3.01 3.009 4 4.00002
Amplitude -0.1253 -0.0881 -0.1785 -0.1518 -0.3545 -0.3320
Exponent 4.5 4.501 5.01 5.0101 6 5.9999
Amplitude 0.0261 0.0235 -0.00089 -0.00085 -0.2363 -0.2318
Exponent 6.5 6.501 7.01 7.01 8 7.9999
Amplitude -0.0294 -0.0276 0.00089 0.00087 0.1418 0.1404
-- ----------- --------- ---------- ---------- ---------- --------- ----------
: The generic case: numerical verification of the analytic formula (\[tail1\]) for the potential (31a) ($\lambda=0.1$) and initial data (\[idn\]). Comparing the second column of this table (corresponding to $\alpha=3.01$) with the last column of table III one can see the discontinuity of the decay rate at $\alpha=3$ (for $d=5$ and $7$).[]{data-label="table 3"}
-- ----------- --------- ---------- ---------- ---------- ---------- ----------
Theory Numerics Theory Numerics Theory Numerics
Exponent 5.5 5.4993 6 6.002 6 6.0000
Amplitude -0.0731 -0.0696 0.00886 0.00862 0.00886 0.00843
Exponent 7.5 7.4998 8 8.0003 8 7.9999
Amplitude 0.0603 0.0579 -0.00177 -0.00175 -0.00177 -0.00172
Exponent 9.5 9.4999 10 9.9957 10 9.9997
Amplitude -0.1131 -0.1115 0.00152 0.00145 0.00152 0.00149
-- ----------- --------- ---------- ---------- ---------- ---------- ----------
: The exceptional case: comparison of analytic and numerical parameters of the tails for the potential (31b) (the first two columns) and (31a) (the third column) with $\alpha=3$, $\lambda=0.1$, and initial data (\[idn\]). The analytic results are given by the formula (\[tailf1p\]) for the subtype IIa potential, and by the formula (\[gtail2\]) for the type I and IIb potentials. Note that although the dominant tails for the type I and the subtype IIb potentials are theoretically the same, in the case IIb there is an additional first order error due to the subdominant term $\mathcal{O}(\lambda)
t^{-(2l+\alpha+\gamma)}$ which accounts for a slight difference in numerical accuracy between these two cases.[]{data-label="table 3"}
Schwarzschild background
========================
Consider the evolution of the massless scalar field outside the $d+1$ dimensional Schwarzschild black hole $$\label{sch}
ds^2=-\left(1-\frac{1}{r^{d-2}}\right) dt^2 +
\left(1-\frac{1}{r^{d-2}}\right)^{-1}
dr^2 + r^2
d\Omega_{d-1}^2\,,$$ where $d\Omega_{d-1}^2$ is the round metric on the unit sphere $S^{d-1}$ and $d\geq 5$ is odd. Here we use units in which the horizon radius is at $r=1$. Introducing the tortoise coordinate $x$, defined by $dr/dx=1-1/r^{d-2}$, and decomposing the scalar field into multipoles, one obtains the following reduced wave equation for the $j$th multipole [@ik] $$\label{eqsch}
\partial_t^2\psi -\partial_x^2 \psi +U(x) \psi=0, \qquad
U=\left(1-\frac{1}{r^{d-2}}\right) \left(\frac{(2j+d-3)(2j+d-1)}{4r^2}+\frac{(d-1)^2}{4
r^d}\right)\,.$$ Note that (\[eqsch\]) is the $1+1$ dimensional wave equation on the whole axis $-\infty<x<\infty$. For large positive $x$ we have $$\label{exp2}
r =
x+\frac{1}{d-3}\frac{1}{x^{d-3}}-\frac{d-2}{(2d-5)(d-3)}\frac{1}{x^{2d-5}} +\mathcal{O} \left( \frac{1}{x^{3d-7}}\right)\,,$$ which implies that $$\label{expv}
U(x) =
\frac{(2j+d-3)(2j+d-1)}{4x^2}+V(x)\,,\qquad V(x)=
\frac{a}{x^d}+\frac{b}{x^{2d-2}} +\mathcal{O} \left( \frac{1}{x^{3d-4}}\right) \quad \mbox{as}\quad x\rightarrow
\infty\,,$$ with $$\label{wsp} a = - \frac {(d-1) j (j+d-2)} {d-3} \qquad \mbox{and}
\qquad b = - \frac {(2d - 3) ((d-3)(d-2)^2(d-1) - 4 j
(j+d-2)(1+d(d-3)))} {4(2d-5)(d-3)^2}\,.$$ For large negative $x$ (near the horizon) the potential is exponentially small, so one expects that the backscattering off the left edge of the potential can be neglected. If so, the decay rate (but not the amplitude!) should follow from the analysis of section 4. Comparing equation (\[eqsch\]) for large positive $x$ to equation (\[a3\]) with the potential (\[case2\]) and using (\[expv\]) we find that $l=j+(d-3)/2$ and the potential $V$ is of the subtype IIa with $\alpha=d$ and $\gamma=d-2$. Thus, applying (\[tailf1p\]) we get the first-order tail $$\label{tails}
\psi(t,x) \sim t^{-(2j+3d-5)}\,.$$ *Remark 3.* Late-time tails outside higher dimensional Schwarzschild black holes were studied in [@car], however in the even-dimensional case the reasoning presented there is not correct, even though the result agrees with (\[tails\]). The reason is that the analysis of [@car] is based on the application of Ching *et al.* conjecture about the decay of the second-order tail $t^{-(2l+2\alpha-2)}$ which for $l=j+(d-3)/2$ and $\alpha=d$ gives $t^{-(2j+3d-5)}$. Unfortunately, this conjecture does not apply to the problem at hand. For $j=0$ this is evident because the leading term in $V$ (proportional to $x^{-d}$) vanishes (since by (\[wsp\]) $a=0$), while the subleading term (proportional to $x^{-(2d-2)}$) is of generic type. For $j>0$ this follows from the fact that the potential is of the subtype IIa. Thus, for all $j\geq
0$ the dominant (first-order) contribution to the tail comes from the subleading term in the potential. The agreement of the decay rate obtained in [@car] with (\[tails\]) is accidental and due to the fact that the subdominant term in (\[expv\]) (not considered in [@car]) is on a borderline between subtypes IIa and IIb. Admittedly, the handwaving argument leading to (\[tails\]) is far from satisfactory. Unfortunately, we have not been able to carry over the analysis from sections 2-4 in the case of equation (\[eqsch\]). There are two difficulties in this respect. First, in contrast to the spherical case, Huygens’ principle is not valid for the free wave equation in $1+1$ dimensions. Second, there is no natural small parameter in the problem. In the impressive tour de force work [@b] Barack showed how to overcome these difficulties for a restricted class of initial data in four dimensions. It would be interesting to generalize Barack’s approach to higher even-dimensional Schwarzschild spacetimes.
0.3cm **Acknowledgments:** PB thanks Nikodem Szpak for helpful discussions and Leor Barack for clarifying some details of the paper [@b]. AR thanks Prof. Bernd Brügmann for hospitality in his group at FSU Jena, where a part of this work was done. This research was supported in part by the MNII grant 1PO3B01229 and grant 189/6. PR UE/2007/7.
Throughout the appendix we use the notation of [@gkp] in which the square bracket around a logical expression returns a value $1$ if the expression is true and a value $0$ if the expression is false: $$[condition] = \left\{ \begin{array}{ccl} 1 & \mbox{if} & condition =
\mbox{true} \\ 0 & \mbox{if} & condition = \mbox{false} \end{array}
\right.\nonumber$$ In order to derive the asymptotic behavior of the iterates (\[iter1\]) and (\[iter2\]) near timelike infinity (fixed $r$ and $t\rightarrow\infty$) we need to evaluate the following expression $$\mathcal{F} (t,r;\,\beta,\,L) = - \frac {2^{\beta}}{4 r^{l+1}}
\sum_{k=0}^L c_{L,k} \, \int \limits_{-\infty}^{+\infty} d\eta \,
\int \limits_{t-r}^{t+r} d\xi \, \frac {P_l (\mu)}
{(\xi-\eta)^{\beta+L-k}} a^{(k)}(\eta), \label{master}$$ where $$c_{L,k} = \frac {(2 L - k)!} {k! (L-k)!}$$ and $$\label{mu} \mu = \frac {(\xi-t)(t-\eta)+r^2}{r(\xi-\eta)}\, .$$ From (\[f1\]) and (\[iter1\]) we have $$\phi_{1}(t,r) = \mathcal{F} (t,r;\,\alpha,\,l),$$ and from (\[iter2\]) and (\[f1nul\]) we have $$\label{a5}
\phi_{2}(t,r) = \frac {1}{2 (\alpha-1) r^{\alpha +l}}
\sum_{q=0}^{l-\alpha/2+1/2} (l-\alpha/2+1/2)^{\underline{q}} \cdot
\frac {2^q \left(\alpha/2 \right)^{\overline{q}}}
{\alpha^{\overline{q}}} \, \mathcal{F} (t,r;\,2 \alpha - 1 +
q,\,l-1-q).$$ Since $a(\eta)$ has compact support, it is advantageous to begin with integration by parts $$\int \limits_{-\infty}^{+\infty} d\eta \, \frac {P_l (\mu)}
{(\xi-\eta)^{\beta+L-k}} a^{(k)}(\eta)
= \int \limits_{-\infty}^{+\infty} d\eta \, (-1)^k \frac {d^k} {d\eta^k} \left( \frac {P_l (\mu)} {(\xi-\eta)^{\beta+L-k}} \right)
a(\eta)\,.\nonumber$$ For $\mu$ as defined in (\[mu\]) and for any function $g(\mu)$ the following identity holds $$\frac {d^k} {d\eta^k} \left( \frac {g(\mu)} {(\xi-\eta)^{\beta}}
\right) = \sum_{j=0}^k \left( \begin{array}{c} k \\ j
\end{array} \right) (\beta+k-1)^{\underline{k-j}} \left( \frac
{r^2-(t-\xi)^2} {r} \right)^j \frac {g^{(j)}(\mu)}
{(\xi-\eta)^{\beta+k+j}}\,,$$ hence $$\mathcal{F} (t,r;\,\beta,\,L) = - \frac {2^{\beta}}{4 r^{l+1}} \int
\limits_{-\infty}^{+\infty} d\eta \, a(\eta) \, \sum_{0\leq j\leq k
\leq L} (-1)^k \left( \begin{array}{c} k \\ j \end{array} \right)
c_{L,k} (\beta+L-1)^{\underline{k-j}} \frac {1} {r^j} \int
\limits_{t-r}^{t+r} d\xi \, \frac {\left( r^2-(t-\xi)^2 \right)^j}
{(\xi-\eta)^{\beta+L+j}} P^{(j)}_{l} (\mu)\,. \label{master(2)}$$ The sum over $k$ can be evaluated explicitly $$\label{k-sum} \sum_{k=j}^L (-1)^k \left( \begin{array}{c} k
\\ j \end{array} \right) \frac {(2 L - k)!}
{k! (L-k)!} (\beta+L-1)^{\underline{k-j}} = (-1)^L \left( \begin{array}{c} L \\ j \end{array} \right) (\beta-2)^{\underline{L-j}}\,.$$ Let us define $$\label{I}
\mathcal{I} := \frac {1} {r^j} \int \limits_{t-r}^{t+r} d\xi \,
\frac {\left( r^2-(t-\xi)^2 \right)^j} {(\xi-\eta)^{\beta+L+j}}
P^{(j)}_{l} (\mu) \,.$$ Changing the integration variable from $\xi$ to $\mu$ and integrating by parts, we get $$\mathcal{I} = (-1)^j \frac {r^{j+1} (t-\eta)^{\beta-2+L-j}} {\left[
(t-\eta)^2-r^2 \right]^{\beta-1+L}} \int \limits_{-1}^{+1} d\mu \,
P_{l} (\mu) \frac {d^j} {d\mu^j} \left[ (1 - \mu^2)^j \left( 1 -
\frac {r} {t-\eta} \mu \right)^{\beta-2+L-j} \right]. \label{I}$$ Using the identity [@wolfram] $$\label{mu-to-k} \mu^k = \sum_{l=k,k-2,k-4,\dots} \frac {(2l+1) k!}
{2^{(k-l)/2} \left( \frac {k-l} {2} \right)! (k+l+1)!!}\, P_l
(\mu)\,,$$ and expanding $\dfrac {d^j} {d\mu^j} \left[ (1 - \mu^2)^j \left( 1 -
\frac {r} {t-\eta} \mu \right)^{\beta-2+L-j} \right]$ in Taylor series we get $$\begin{aligned}
\mathcal{I} &=& (-1)^j \frac {r^{j+1} (t-\eta)^{\beta-2+L-j}}
{\left[ (t-\eta)^2-r^2 \right]^{\beta-1+L}} \,
\sum_{n=0}^{\beta-2+L} (j+n)^{\underline{j}} \, \int
\limits_{-1}^{+1} d\mu \, P_{l} (\mu) \mu^{n}
\nonumber\\
&\times& \sum_{m=0}^{\lfloor (j+n)/2 \rfloor} \left(
\begin{array}{c} j \\ m \end{array} \right) \left( \begin{array}{c}
\beta-2+L-j \\ j+n-2m \end{array} \right) (-1)^{j+n+m} \left( \frac
{r} {t-\eta} \right)^{j+n-2m}
\nonumber\\
&=& \frac {r^{l+1} (t-\eta)^{\beta-2+L-l}} {\left[ (t-\eta)^2 - r^2
\right]^{\beta-1+L}} \, \sum_{n=0}^{\lfloor (\beta-2+L-l) / 2
\rfloor} (j+l+2n)^{\underline{j}} \, \int \limits_{-1}^{+1} d\mu \,
P_{l} (\mu) \mu^{l+2n}
\nonumber\\
&\times& \sum_{m=0}^{\lfloor (j+l+2n)/2 \rfloor} \left(
\begin{array}{c} j \\ m \end{array} \right) \left(
\begin{array}{c} \beta-2+L-j \\ j+l+2n-2m \end{array} \right)
(-1)^{l+m} \left( \frac {r} {t-\eta} \right)^{2j+2n-2m}
\nonumber\\
&=& \frac {r^{l+1} (t-\eta)^{\beta-2+L-l}} {\left[ (t-\eta)^2 - r^2
\right]^{\beta-1+L}} \, \sum_{n=0}^{\lfloor (\beta-2+L-l) / 2
\rfloor} (j+l+2n)^{\underline{j}} \, \, 2^{l+1} \frac {(l+2n)!
(l+n)!} {n! (2l+2n+1)!}
\nonumber\\
&\times& \sum_{m=0}^{\lfloor (j+l+2n)/2 \rfloor} \left(
\begin{array}{c} j \\ m \end{array} \right) \left(
\begin{array}{c} \beta-2+L-j \\ j+l+2n-2m \end{array} \right)
(-1)^{l+m} \left( \frac {r} {t-\eta} \right)^{2j+2n-2m}
\label{I(2)}\,.\end{aligned}$$ Collecting the results of (\[k-sum\], \[I\], \[I(2)\]) and plugging them into (\[master(2)\]) we get $$\mathcal{F} (t,r;\,\beta,\,L)
= - \frac {2^{\beta+l+1}}{4} \int \limits_{-\infty}^{+\infty}
d\eta \, a(\eta) \, \frac {(t-\eta)^{\beta-2+L-l}} {\left[
(t-\eta)^2 - r^2 \right]^{\beta-1+L}} \sum_{n=0}^{\lfloor
(\beta-2+L-l) / 2 \rfloor} \frac {(l+2n)! (l+n)!} {n! (2l+2n+1)!}
(-1)^{L+l} L! \, S (\beta, L), \label{master(3)}$$ where $$\!\!S (\beta, L) = \sum_{j=0}^{L} \left( \begin{array}{c}
\beta-2 \\ L-j \end{array} \right) \left( \begin{array}{c} j+l+2n \\
j \end{array} \right) \sum_{m=0}^{\lfloor (j+l+2n)/2 \rfloor}\!\!\!
(-1)^{m} \left( \begin{array}{c} j
\\ m \end{array} \right) \left( \begin{array}{c} \beta-2+L-j \\
j+l+2n-2m \end{array} \right) \left( \frac {r} {t-\eta}
\right)^{2j+2n-2m} \,.$$
First-order approximation
-------------------------
To evaluate the first iterate $\phi_1(t,r)$ we apply the formula (\[master(3)\]) with $\beta=\alpha$ and $L=l$. Then $$S(\alpha,l) = \sum_{j=0}^{l} \left( \begin{array}{c} \alpha-2
\\ l-j \end{array} \right) \left( \begin{array}{c} l+2n+j
\\ j \end{array} \right) \sum_{m=(j-l)/2}^{j+n} \!\!\!(-1)^{j+n-m} \left(
\begin{array}{c} j \\ j+n-m \end{array} \right) \left(
\begin{array}{c} \alpha-2+l-j \\ l-j+2m \end{array} \right) \left(
\frac {r} {t-\eta} \right)^{2m}\,, \label{SI}$$ where we shifted the summation index $m\rightarrow j+n-m$. Next, we interchange the order of summation according to $$\begin{aligned}
&& [ 0 \leq j ] [ j \leq l ] [ m-n \leq j ] [ j \leq 2m+l ]
\nonumber\\
&\Leftrightarrow& [ -\frac {l}{2} \leq m < 0 ] [ 0 \leq j \leq l+2m
] \, + \, [ 0 \leq m < n ] [ 0 \leq j \leq l ] \, + \, [ n \leq m
\leq l+n ] [ m-n \leq j \leq l ]\,,\nonumber\end{aligned}$$ and convert the sum over $j$ into a generalized hypergeometric function [@gkp]. Defining $$t_j = (-1)^{j+n-m} \left( \begin{array}{c} \alpha-2 \\ l-j
\end{array} \right) \left( \begin{array}{c} l+2n+j \\ j \end{array}
\right) \left( \begin{array}{c} j \\ j+n-m \end{array} \right)
\left( \begin{array}{c} \alpha-2+l-j \\ l-j+2m \end{array}
\right)\,,$$ we see that $t_0\neq 0$ iff $n=m$, thus the sums for $[ -\frac
{l}{2} \leq m < 0 ]$ and $[ 0 \leq m < n ]$ do not contribute to (\[SI\]) and we are left with $$\begin{aligned}
S (\alpha, l)&=& \sum_{m=n}^{n+l} \left( \frac {r}
{t-\eta} \right)^{2m} \sum_ {j=0}^{l+n-m} (-1)^{j} \left(
\begin{array}{c} \alpha-2 \\ l+n-m-j \end{array} \right) \left(
\begin{array}{c} l+n+m+j \\ j+m-n \end{array} \right)
\nonumber\\
& \times & \left( \begin{array}{c} j+m-n \\ j \end{array} \right)
\left(
\begin{array}{c} \alpha-2+l+n-m-j \\ l+n+m-j \end{array} \right),\end{aligned}$$ where we shifted the summation index $j \rightarrow j+m-n$. Defining $$\tilde{t}_j = (-1)^{j} \left( \begin{array}{c} \alpha-2 \\ l+n-m-j
\end{array} \right) \left( \begin{array}{c} l+n+m+j \\ j+m-n
\end{array} \right) \left( \begin{array}{c} j+m-n \\ j \end{array}
\right) \left( \begin{array}{c} \alpha-2+l+n-m-j \\ l+n+m-j
\end{array} \right)$$, we see that $$\begin{aligned}
\tilde{t}_0 &=& \frac {(\alpha-2)^{\underline{l+n-m}}} {(l+n-m)!}
\cdot \frac {(l+n+m)!} {(m-n)! (l+2n)!} \cdot \frac
{(\alpha-2+l+n-m)^{\underline{l+n+m}}} {(l+n+m)!}\nonumber\end{aligned}$$ and $$\frac {\tilde{t}_{j+1}} {\tilde{t}_j} = \frac {(j - (l+n-m)) (j -
(l+n+m)) (j + (l+n+m+1)) } {(j + ((\alpha-1) - (l+n-m))) (j +
(-(\alpha-2) - (l+n-m)))
(j +1)}\,,$$ hence $$\begin{aligned}
S (\alpha, l) &=& \sum_{m=n}^{n+l} \left( \frac {r} {t-\eta}
\right)^{2m} \frac {(\alpha-2)^{\underline{l+n-m}}} {(l+n-m)!} \cdot
\frac {(\alpha-2+l+n-m)^{\underline{l+n+m}}} {(m-n)! (l+2n)!}
\nonumber\\
&\times& F\left( \left. \begin{array}{c} - (l+n-m),\, - (l+n+m),\,
(l+n+m+1) \\ (\alpha-1) - (l+n-m),\, -(\alpha-2) - (l+n-m)
\end{array} \right| 1 \right)
\nonumber\\
&=& \sum_{m=n}^{n+l} \left( \frac {r} {t-\eta} \right)^{2m} 2^{1 +
2(l+n-m)} \pi \,\frac {(\alpha-2)^{\underline{l+n-m}}} {(l+n-m)!}
\cdot \frac {(\alpha-2+l+n-m)^{\underline{l+n+m}}} {(m-n)! (l+2n)!}
\nonumber\\
&\times& \frac {\Gamma(-(\alpha-2) - (l+n-m)) \Gamma((\alpha-1) -
(l+n-m))} {\Gamma \left( - \frac {\alpha-3} {2} + m \right) \Gamma
\left( - \frac {\alpha-2} {2} - (l+n) \right) \Gamma \left( \frac
{\alpha} {2} + m \right) \Gamma \left( \frac {\alpha-1} {2} - (l+n)
\right)}\,, \label{SI(2)}\end{aligned}$$ where in the last equation we used the identity $$F\left( \left. \begin{array}{c} a+1,\, -a,\, (b+c-1)/2 \\ b,\, c
\end{array} \right| 1 \right) = 2^{2-(b+c)} \pi \frac {\Gamma(b)
\Gamma(c)} {\Gamma \left( \frac {b-a} {2} \right) \Gamma \left(
\frac {c-a} {2} \right)
\Gamma \left( \frac {1+b+a} {2} \right) \Gamma \left( \frac {1+c+a} {2}
\right)}\,.$$ Substituting $$(\alpha-2)^{\underline{l+n-m}} \Gamma((\alpha-1) - (l+n-m)) =
\Gamma(\alpha-1),\nonumber$$ and $$(\alpha-2+l+n-m)^{\underline{l+n+m}} \Gamma(-(\alpha-2) - (l+n-m)) =
(-1)^{l+n+m} \Gamma(-\alpha + 2 + 2m)\nonumber$$ into (\[SI(2)\]) we get $$\begin{aligned}
\!S (\alpha, l) \! &\!=\!&\! \sum_{m=n}^{n+l} \left( \frac {r}
{t-\eta} \right)^{2m}\! \!\frac {(-1)^{l+n+m} 2^{1 + 2(l+n-m)} \pi}
{(l+n-m)! (m-n)! (l+2n)!} \frac {\Gamma(\alpha-1) \Gamma(-\alpha + 2
+ 2m)} {\Gamma \left( \frac {\alpha} {2} + m \right) \Gamma \left(
- \frac {\alpha-3} {2} + m \right) \Gamma \left( - \frac {\alpha-2}
{2} - (l+n) \right) \Gamma \left( \frac {\alpha-1} {2} - (l+n)
\right)}\,. \nonumber \label{SI(3)}\end{aligned}$$ The last equation can be still simplified due to the identity $$\label{GammasId} \frac {\Gamma(\alpha -1) \Gamma(-\alpha + 2)}
{\Gamma \left( \frac {\alpha} {2} \right) \Gamma \left( - \frac
{\alpha-3} {2} \right) \Gamma \left( - \frac {\alpha-2} {2} - l
\right) \Gamma \left( \frac {\alpha-1} {2} - l \right)} = \frac
{(-1)^l} {2 \pi} \left( \frac {\alpha-3} {2} \right)^{\underline{l}}
\left( \frac {\alpha} {2} \right)^{\overline{l}}.$$ We have $$\begin{aligned}
\label{Gamma1} \Gamma(-\alpha + 2 + 2m) &=& (-\alpha +
2)^{\overline{2m}} \Gamma(-\alpha + 2),\nonumber\\
\label{Gamma2} \Gamma \left( - \frac {\alpha-3} {2} + m \right)& =&
\left( - \frac {\alpha-3} {2} \right)^{\overline{m}} \Gamma \left( -
\frac {\alpha-3} {2} \right),\nonumber\\
\label{Gamma3} \Gamma \left( - \frac {\alpha-2} {2} - l - n \right)
&=& \frac {\Gamma \left( - \frac {\alpha-2} {2} - l \right)} {\left(
- \frac {\alpha-2} {2} - l - 1 \right)^{\underline{n}}}\,,\nonumber \\
\label{Gamma4} \Gamma \left( \frac {\alpha-1} {2} - l - n \right)
&=& \frac {\Gamma \left( \frac {\alpha-1} {2} - l \right)} {\left(
\frac {\alpha-1} {2} - l - 1 \right)^{\underline{n}}}\,,\nonumber\\
\label{Gamma5} \Gamma \left( \frac {\alpha} {2} + m \right) &= &
\left( \frac {\alpha} {2} \right)^{\overline{m}} \Gamma \left( \frac
{\alpha} {2} \right),\nonumber\end{aligned}$$ and $$\frac {(-\alpha + 2)^{\overline{2m}}} {\left( - \frac {\alpha-3}
{2} \right)^{\overline{m}}} = 2^{2m} \left( - \frac {\alpha} {2} + 1
\right)^{\overline{m}},\nonumber$$ so finally $$\begin{aligned}
S (\alpha, l) &=& \sum_{m=n}^{n+l} \left( \frac {r} {t-\eta}
\right)^{2m} \frac {(-1)^{n+m} 2^{2(l+n)}} {(l+n-m)! (m-n)! (l+2n)!}
\left( \frac {\alpha-3} {2} \right)^{\underline{l}} \left( \frac
{\alpha} {2} \right)^{\overline{l}}
\nonumber\\
&\times& \frac {\left( - \frac {\alpha} {2} + 1
\right)^{\overline{m}} \left( - \frac {\alpha-2} {2} - l - 1
\right)^{\underline{n}} \left( \frac {\alpha-1} {2} - l - 1
\right)^{\underline{n}}} {\left( \frac {\alpha} {2}
\right)^{\overline{m}}}\,. \label{SI(4)}\end{aligned}$$ Plugging (\[SI(4)\]) into (\[master(3)\]) with $\beta=\alpha$ and $L=l$ we get the expression (\[tailf1\]).
Second-order approximation
--------------------------
The calculation in the second order ($\beta=2\alpha-1+q$ and $L=l-1-q$) is only a slight modification of what we have already done in the first order. Following the same steps which led us from (\[SI\]) to (\[SI(3)\]) we get $$\begin{aligned}
S (\beta, L)
&=& \sum_{m=n}^{n+L} \left( \frac {r} {t-\eta}
\right)^{2m} \frac {(-1)^{l+n+m} 2^{1 + 2(L+n-m)} \pi} {(L+n-m)!
(m-n)! (l+2n)!}
\nonumber\\
&\times& \frac {\Gamma(\beta-1) \Gamma(-\beta + 2 + l-L + 2m)}
{\Gamma \left( \frac {\beta} {2} + \frac{l-L}{2} + m \right) \Gamma
\left( - \frac {\beta-3} {2} + \frac{l-L}{2} + m \right) \Gamma
\left( - \frac {\beta-2} {2} - \left(\frac{l+L}{2}+n\right)
\right)\Gamma \left( \frac {\beta-1} {2} -
\left(\frac{l+L}{2}+n\right) \right)}\,. \label{j-sum(3)II}
\label{SII(3)}\end{aligned}$$ The last equation can be simplified due to the identity $$\begin{aligned}
&& \frac {\Gamma(\beta - 1) \Gamma(-\beta + 2 + l-L)} {\Gamma \left(
\frac {\beta} {2} + \frac{l-L}{2} \right) \Gamma \left( - \frac
{\beta-3} {2} + \frac{l-L}{2} \right) \Gamma \left( - \frac
{\beta-2} {2} - \frac{l+L}{2} \right) \Gamma \left( \frac {\beta-1}
{2} - \frac{l+L}{2} \right)}
\nonumber\\
&=& \frac {(-1)^l} {2 \pi} \left( \frac {\beta-3-(l-L)} {2}
\right)^{\underline{L}} \left( \frac {\beta + l-L} {2}
\right)^{\overline{L}} (\beta-2)^{\underline{l-L}}\,,\end{aligned}$$ which for $L=l$ reduces to (\[GammasId\]). We have $$\begin{aligned}
\label{Gamma1II} \Gamma(-\beta + 2 + l-L + 2m) &=& (-\beta + 2 +
l-L)^{\overline{2m}} \,\Gamma(-\beta + 2 + l-L)\,,\nonumber\\
\label{Gamma2II} \Gamma \left( - \frac {\beta-3} {2} + \frac{l-L}{2}
+ m \right) &=& \left( - \frac {\beta-3} {2} + \frac{l-L}{2}
\right)^{\overline{m}} \Gamma \left( - \frac {\beta-3} {2} +
\frac{l-L}{2} \right)\,,\nonumber\\
\label{Gamma3II} \Gamma \left( - \frac {\beta-2} {2} - \frac{l+L}{2}
- n \right) &=& \frac {\Gamma \left( - \frac {\beta-2} {2} -
\frac{l+L}{2} \right)} {\left( - \frac {\beta-2} {2} - \frac{l+L}{2}
- 1 \right)^{\underline{n}}}\,,\nonumber\\
\label{Gamma4II} \Gamma \left( \frac {\beta-1} {2} - \frac{l+L}{2} -
n \right) &=& \frac {\Gamma \left( \frac {\beta-1} {2} -
\frac{l+L}{2} \right)} {\left( \frac {\beta-1} {2} - \frac{l+L}{2} -
1 \right)^{\underline{n}}}\,,\nonumber\\
\label{Gamma5II} \Gamma \left( \frac {\beta} {2} + \frac{l-L}{2} + m
\right) &=& \left( \frac {\beta} {2} + \frac{l-L}{2}
\right)^{\overline{m}} \Gamma \left( \frac {\beta} {2} +
\frac{l-L}{2} \right)\,,\nonumber\end{aligned}$$ and $$\frac {(-\beta + 2 + l-L)^{\overline{2m}}} {\left( - \frac
{\beta-3} {2} + \frac{l-L}{2} \right)^{\overline{m}}} = 2^{2m}
\left( - \frac {\beta} {2} + \frac{l-L}{2} + 1
\right)^{\overline{m}},\nonumber$$ hence $$\begin{aligned}
S (\beta, L) &\!=\!& \sum_{m=n}^{n+L} \left( \frac {r} {t-\eta}
\right)^{2m} \frac {(-1)^{n+m} 2^{2(L+n)}} {(L+n-m)! (m-n)! (l+2n)!} \label{SII(4)}\\
&\times& \left( \frac {\beta-3-(l-L)} {2} \right)^{\underline{L}}
\left( \frac {\beta + l-L} {2} \right)^{\overline{L}}
(\beta-2)^{\underline{l-L}} \frac {\left( - \frac {\beta} {2} +
\frac{l-L}{2} + 1 \right)^{\overline{m}} \left( - \frac {\beta-2}
{2} - \frac{l+L}{2} - 1 \right)^{\underline{n}} \left( \frac
{\beta-1} {2} - \frac{l+L}{2} - 1 \right)^{\underline{n}}} {\left(
\frac {\beta} {2} + \frac{l-L}{2} \right)^{\overline{m}}}.\nonumber\end{aligned}$$ Plugging (\[SII(4)\]) into (\[master(3)\]) we get $$\begin{aligned}
&& \mathcal{F} (t,r;\,2\alpha-1+q,\,l-1-q) = (-1)^{q} \,\frac
{2^{2\alpha+3l-2-q}}{4} \left( \alpha - \frac {5} {2}
\right)^{\underline{l-1-q}} \left( \alpha - 2 + l
\right)^{\underline{l-1-q}} (2\alpha-3)^{\overline{1+q}}
\nonumber\\
&\times& \int \limits_{-\infty}^{+\infty} d\eta \, a(\eta) \, \frac
{(t-\eta)^{2\alpha-4}} {\left[ (t-\eta)^2 - r^2
\right]^{2\alpha-3+l}}
\\
&\times& \sum_{n=0}^{\alpha-2}\! (-1)^{n} \frac {2^{2n}(l+n)!} {n!
(2l+2n+1)!} \left( -\alpha+1-l \right)^{\underline{n}} \left( \alpha
- \frac {3} {2} - l + q \right)^{\underline{n}} \, \sum_{m=n}^{
n+l-1-q} \!(-1)^{m} \left(
\begin{array}{c} l-1-q \\ m-n \end{array} \right) \frac {\left(
-\alpha + 2 \right)^{\overline{m}}} {\left( \alpha + q
\right)^{\overline{m}}} \left( \frac {r} {t-\eta}
\right)^{2m}.\nonumber\end{aligned}$$ Substituting this into (\[a5\]) and expanding in $1/t$ we have $$\begin{aligned}
\phi_{2}(t,r) &=& \frac {1} {2(\alpha-1)} \cdot \frac
{2^{2\alpha+2l-2}}{4 (2l+1)!!} \cdot \frac {1} {t^{2\alpha+2l-2}}
\left[ A + 2(\alpha+l-1) \frac{B}{t} \, + \mathcal{O} \left(
\frac{1}{t^2} \right) \right]
\nonumber\\
&\times& \left( \sum_{q=0}^{l-(\alpha-1)/2} (-1)^{q}
(l-p)^{\underline{q}} \,\, \frac {2^q \left( \alpha/2
\right)^{\overline{q}}} {\alpha^{\overline{q}}} \left( \alpha -
\frac {5} {2} \right)^{\underline{l-1-q}} \left( \alpha - 2 + l
\right)^{\underline{l-1-q}} (2\alpha-3)^{\overline{1+q}} \right),\end{aligned}$$ with $A$ and $B$ defined in (\[ab\]). Converting the sum over $q$ into the generalized hypergeometric function we get (\[tail2\]).
[10]{}
W. Strauss and K. Tsutaya, Discrete Cont. Dyn. Sys. **3**, 175 (1997).
N. Szpak, arXiv:0708.1185 \[math-ph\]
E. S. C. Ching et al., Phys. Rev. **D52**, 2118 (1995).
R. Price, Phys. Rev. **D5**, 2419 (1972).
E. W. Leaver, Phys. Rev. **D34**, 384 (1986).
C. Gundlach, R. Price and J. Pullin, Phys. Rev. **D49**, 883 (1994).
L. Barack, Phys. Rev. **D59**, 044017 (1999).
M. Dafermos and I. Rodnianski, Invent. Math. **162**, 381 (2005).
V. Cardoso et al., Phys. Rev. **D68**, 061503 (2003).
J. G. Kingston, Quart. Appl. Math. **46**, 775 (1988).
H. Lindblad and C. D. Sogge, Amer. J. Math. **118**, 1047 (1996).
S. Hod, Class. Quantum Grav. **18**, 1311 (2001).
A. Ishibashi and H. Kodama, Prog. Theor. Phys. **110**, 901 (2003).
R. L. Graham, D. E. Knuth and O. Patashnik, *Concrete Mathematics* (Reading, Massachusetts: Addison-Wesley, 1994).
http://mathworld.wolfram.com/
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'Deep neural network-based methods have been proved to achieve outstanding performance on object detection and classification tasks. Despite the significant performance improvement using the deep structures, they still require prohibitive runtime to process images and maintain the highest possible performance for real-time applications. Observing the phenomenon that human visual system (HVS) relies heavily on the temporal dependencies among frames from the visual input to conduct recognition efficiently, we propose a novel framework dubbed as TKD: temporal knowledge distillation. This framework distills the temporal knowledge from a heavy neural network-based model over selected video frames (the perception of the moments) to a light-weight model. To enable the distillation, we put forward two novel procedures: 1) a Long-short Term Memory (LSTM)-based key frame selection method; and 2) a novel teacher-bounded loss design. To validate our approach, we conduct comprehensive empirical evaluations using different object detection methods over multiple datasets including Youtube-Objects and Hollywood scene dataset. Our results show consistent improvement in accuracy-speed trad-offs for object detection over the frames of the dynamic scene, compared to other modern object recognition methods. It can maintain the desired accuracy with the throughput of around $220$ images per second. Implementation: <https://github.com/mfarhadi/TKD-Cloud>.'
author:
- |
Mohammad Farhadi\
Arizona State University\
[[email protected]]{}
- |
Yezhou Yang\
Arizona State University\
[[email protected]]{}
bibliography:
- 'egbib.bib'
title: 'TKD: Temporal Knowledge Distillation for Active Perception'
---
Introduction {#sec:intro}
============
Object detection plays a critical role in a variety of mobile robot tasks such as obstacle avoidance [@carrio2018drone; @yaghoubi2019worst], detection and tracking [@breuers2018detection] and object searching [@ye2018active; @ye2019gaple]. During the last decade, we have witnessed the great success of Convolutional Neural Networks (CNNs)-based methods in the object detection task. This success has led researchers to explore deeper models such as RetinaNet [@lin2018focal] or Faster-RCNN [@ren2015faster], which yield high recognition accuracy. The “secret” sauce behind the success of these deeper and deeper CNNs models is the stacking of repetitive layers and increasing the number of model parameters [@chen2017learning]. This practice becomes possible while the applications are running on infrastructures with high processing capabilities.
![An illustration of our TKD model’s actual performance: F-1 score distribution over example object categories in different environments using TKD.[]{data-label="set-subset"}](./images/jamal.png){width="7.5cm"}
However, the disadvantages of this practice are obvious and the high performance is achieved by the significant growth of the model complexity: stacking up layers and increasing the model parameters which are computationally expensive and also increase the inference time significantly. Hence, these models are not suitable for real-time and embedded visual processing systems, and thus impede their deployment in the era of intelligent robots and autonomous vehicles. The same concerns also lie in the energy conservation and computation limits, since deep models require a large number of matrix multiplications, which are time-consuming and energy-demanding for mobile applications.
The aforementioned concerns trigger various approaches, such as using the alignment of memory and SIMD (Single Instruction, Multiple Data) operations to boost matrix operations [@gong2014compressing]. More recently, studies [@chen2017learning] and [@hinton2015distilling] proposed transferring the knowledge of deep models to shallow models while maintaining the recognition accuracy. Although these approaches do improve the model efficiency, they ignore the temporal dependencies among the frames from dynamic scenes, which is one of the critical capabilities to maintain high recognition accuracy while being energy-aware. The motivation for our TKD model comes from the visual adaptation phenomenon observed in the Human Visual Systems (HVS). Visual adaption involves temporary changes in the human perception system when exposed to intense or new stimulus and by the lingering aftereffects when the stimulus is removed [@webster2015visual]. Other studies from [@webster2015visual] show that the visual system adapts to the changes in the environment and this adjustment can happen in a few milliseconds. More specifically, a study from [@clifford2007visual] reveals that the face recognition process happens at a higher level of cognition, and later at the stage of visual encoding, we observe that the sensory systems adapt itself to the prevailing environment. This shows that HVS relies heavily on the prior estimation of the objects’ appearance distribution to improve the perception capability at the current time-stamp.
Moreover, the adaptation happens both in the “low” and “high” level visual features. The human visual system adapts to the distribution of “low-level” visual features such as color, motion, and texture, as well as the “high-level” visual features such as face classification including identity, gender, expression, or ethnicity [@webster2015visual]. This adaptation can be both short-term and long-term. For instance, our perception system adapts itself to the general visual features of the environment which we are living in for a long time such as faces and colors (like training a model). Also, it can adapt itself dynamically when the environment changes, for example, moving from the indoor environment to the outdoor [@webster2015visual] (like adapting a shallow model). This adaptation capability is essential for our HVS to perform recognition well and efficient, with low energy consumption.
Inspired by the aforementioned findings, we design our TKD framework that utilizes the knowledge distillation techniques. It transfers temporal knowledge from the heavy model to a light model to boost visual processing efficiency while maintaining the heavy model’s (a.k.a., oracle model) performance. Figure \[set-subset\] illustrates the overall goal of this work. In this figure, we show how TKD improves recognition accuracy over different scenes, compared to the oracle model which we assume to be a perfect model. Also, we show the baseline model which is a tiny model with low accuracy compared to oracle recognition due to a much lower number of parameters. TKD achieves higher accuracy by adapting itself to the observed environment. In the case of an indoor scene, the TKD recognition accuracy improves significantly over objects which are more probable to be observed inside a building. In the outdoor case, TKD recognition accuracy improves over the objects such as a car, bus, and truck which are more probable to be observed outside. For a similar amount of model parameters as the baseline tiny model, the TKD will achieve much better performance over the more probable objects by dynamically learning from the oracle model.
To summarize our contributions: 1) we propose an end-to-end trainable framework to transfer the temporal knowledge (a.k.a., the perception of the moment) of the oracle model to the student model; 2) we propose a novel teacher-bonded loss for knowledge distillation which has a simple structure and performs inferences briskly; and 3) we propose an efficient method to select key frames from the dynamic scene, that indicate the right timing to train student model and to improve the detection accuracy. We design and conduct empirical experiments on both the public datasets (the Youtube Object dataset and the Hollywood Scene dataset) as well as on two long videos with multiple scene changes, which validate each of the aforementioned novel design choices, by observing a fast object recognition performance while maintaining high detection accuracy.
Related Work
============
Visual recognition systems, ranging from object recognition [@lin2018focal], action recognition [@lea2016learning], to scene recognition [@zhou2014learning] have gained attention in recent years. Significant improvements in recognition accuracy have resulted in economic and societal benefits in AI applications such as autonomous vehicles [@khayatian2019crossroads+; @khayatian2018rim], and IoT systems [@tonekaboni2018edge; @tonekaboni2018scouts].
**Object Detection:** Object detection methods based on Convolutional Neural Networks (CNNs) have shown promising results over the past years. There are two main types of object recognition systems which are based on CNNs, one-stage, and two-stage. In one-stage methods, we classify and localize objects in one-stage. Images, when forwarded through the network produce a single output which is then used to classify or localize objects. Some examples of one-stage methods are Yolo [@redmon2018yolov3], RetinaNet [@lin2018focal] and DSSD [@fu2017dssd]. These models are faster compared to other methods due to ruining in a single stage. The second types of models are two-stage methods in which classification and localization happen as two different stages, using classification networks and region proposal networks respectively. Two famous two-stage models are FasterRCNN [@ren2015faster], R-FCN [@dai2016r]. These models reach to higher performance with high intersection over union (IOU). However, Redmon et al. [@redmon2018yolov3] showed at lower IOU (IOU=0.5) one-stage models can perform the same accuracy as two-stage models.
**Model Compression:** Another thrust of work has focused on reducing the resources consumption of CNNs (due to expensive computation and memory usage) by compressing the network structures [@han2015learning; @rastegari2016xnor]. Network pruning is one of well-studied approach which removes unnecessary connections from CNN model, to gain inference speedup [@wen2016learning; @iandola2016squeezenet]. Quantizing [@han2015deep; @farhadi2019novel] and binarizing [@rastegari2016xnor; @bank2019polar] are two other methods that have been used to reduce network size and computation load. These methods improve performance at the hardware level by reducing the size of weights at the binary code level. However, the standard GPU implementation remains challenging for these methods to achieve runtime speedup [@han2015learning]. Also, the advantages of these methods over other one-stage methods without the fully connected layers (the network pruning target in [@han2015deep]) is not clear.
**Domain Adaptation:** Object detection in the real world still needs to address challenges such as low image quality, large variance in the backgrounds, illumination variation, etc. These could lead to a significant domain shift between the training, validation and test data. Consequently, the field of domain adaptation has been widely studied in image classification [@tzeng2014deep; @lu2017unsupervised] and object detection [@chen2018domain; @dai2018dark] tasks. These methods improve accuracy on well-known bench-marking datasets. Nevertheless, they typically adopt an offline domain adaptation procedure and do not concern with domain-change during the inference stage.
**Knowledge Distillation:** Knowledge distillation is another approach to boost accuracy in CNNs. Under the knowledge distillation setting, an ensemble of CNN models or a very deep model will serve as the teacher model, which transfers its knowledge to the student model (shallow model). Hinton et al. [@hinton2015distilling] proposed a method to apply teacher prediction as a “soft-label” and distill teacher classifier’s knowledge to the student. Moreover, they proposed a temperature cross entropy instead of $L2$ distance as the loss function. Romero et al. [@romero2014fitnets] proposed a so-called “hint” procedure to guide the training of the student model. There are also other approaches to distill knowledge between different domains such as from RGB to depth images [@gupta2016cross; @su2016adapting]. Knowledge distillation has been also applied to the object detection task. Chen et al. [@chen2017learning] proposed a method which adopts all of the soft labeling (labels generated by the teacher), the hard labeling (the ground truth) and the hint procedure to transfer knowledge from the teacher with deep feature extractor to the student with a shallow feature extractor. They adopt a two-stage method (FasterRCNN [@ren2015faster]) in their system. Mehta et al. [@mehta2018object] applied the same procedure to one stage method (Tiny-Yolo v2).
Mullapudi et al. [@mullapudi2018online] proposed an online model distillation for efficient segmentation. They adopt a light CNN model as a student and a heavy model as a teacher. At the inference time, the student model is trained periodically using the teacher knowledge. However, the naive usage of a fixed period may not be efficient in their approach. Moreover, their shallow model struggles to handle emerging new objects in the scene when these objects are observed in the middle of the fixed period. Here, ours is able to select the period length based on the incoming frames, by which TKD could trigger re-training and thus detecting the emerging new objects, as demonstrated experimentally in Sec. \[sec:exp\].
Temporal Knowledge Distillation {#distillation}
===============================
The conventional use of knowledge distillation has been proposed for training CNNs based classification models. In these models, we have a dataset $(x_{i},y_{i}), i=1,2,...,n$ where $x_{i}$ and $y_{i}$ are input images and the class labels. The student model is trained to optimize the following general loss function (with $\beta$ is a modulation factor): $$\small
\setlength{\abovedisplayskip}{5pt}
\setlength{\belowdisplayskip}{5pt}
\centering
\begin{aligned}
& O_s = Student(x); O_t= Teacher(x), \\
& L(O_s,(y,O_t)) = \beta L_{gt}(O_s,y) + (1-\beta)L_t(O_s,O_t), \label{eq:a}
\end{aligned}$$ where $L_t$ is the loss using teacher output ($O_t$) and $L_{gt}$ is the loss using ground truth $y$ [@mehta2018object; @chen2017learning; @hinton2015distilling].
In addition to the classification task, object detection also could benefit from the knowledge distillation procedure. However, it’s not as straightforward as the classification task. Most notably, the teacher model’s output may yield misleading guidance to the student model [@chen2017learning]. The teacher regression result can be contradictory to the ground truth labels, also the output from the teacher regression module is unbounded. To address these issues, [@chen2017learning] proposed a procedure to only adopt teacher’s output at beneficial times. For a one-stage object detection setting, [@mehta2018object] optimized the student model with a similar loss function to Eq. \[eq:a\].
In this paper, we propose a novel and bio-inspired way of adopting the teacher model’s knowledge. Namely, temporally estimating the expectation of object labels, their sizes, and shapes based on the previously observed frames or $E[y_i| \alpha_1, \alpha_2, ... , \alpha_{i-1}]$ where $y_i$ is our objects label and $\alpha$ our observations. This expectation changes in time by camera or objects movements, and/or the changing of the field of view. Here, we utilize this extracted knowledge to improve object detection performance. Unlike the previous work such as [@mehta2018object; @chen2017learning], we are not aiming to improve the feature extractor and/or the general knowledge of the student model. We optimize the decoder inside the student model to adapt it to the current environment. It is done by increasing the likelihood of objects which are more frequently found from the previous observations. Since the model requires online training during the inference stage, it should be able to address the following challenges:
1. Training is a time consuming procedure, running it at the inference stage hurts model efficiency;
2. Selecting the key frames accurately on which the student model needs to be adapted;
3. Objects with low appearance probability may not be detected by the student model after adaptation;
4. The oracle model still introduces noise at locations where there are no objects. Simply training the student model with noisy oracle output decreases the accuracy.
In the following section, we will introduce our approach to address these challenges respectively.
Our Approach {#sec:TKD_structure}
============
In this work, we adopt Yolo-v3 (as teacher) and Tiny-Yolo v3 (as student) [@redmon2018yolov3] as the base object detection methods. These two models are one-stage object detection models. In both models, object detection is conducted at various layers. The middle layers are used to detect large objects and the last layers to detect small objects. Studies [@redmon2018yolov3], [@mullapudi2018online] and [@lin2018focal] showed that this strategy successfully improves the object detection accuracy with a significant edge.
As mentioned in Section \[sec:intro\], the overall objective of our system is to estimate the expectation of object labels, their sizes, and shapes on the temporal domain and to improve the performance of the student model. Following this intuition, we put forward a mechanism with a combination of an oracle model (which we consider it as the best possible model) and a student model (which is fast but has considerably lower accuracy compared to the oracle). We are transferring the temporal knowledge of the oracle model to the student model at the . By transferring this knowledge, the student model adapts itself to the current environment or scene. Without loss of generality, We select Yolo-v3 object detection model as the oracle model due to its reliable and dominating performance compared with other one-stage methods. We select Tiny-Yolo model [@redmon2018yolov3] as the student model due to its high base frame rate and having a similar model structure with the Yolo-v3.
The TKD Architecture
--------------------
We show our overall framework in Figure \[TKD\_fig\]. In the student model, we include two decoders as the TKD decoder and the general decoder. Then, the pre-trained Yolo-v3 [@redmon2018yolov3] is adopted as the oracle. We run the Oracle model with the input image and the weights of student’s TKD decoders get updates at specific frames from the oracle model’s result. Finally, we design a decision procedure using an LSTM model, to generate the signals that indicate the right timing to use the Oracle knowledge.
Specifically, we train Tiny-Yolo with a general decoder over the COCO dataset [@lin2014microsoft]. The design of Tiny-Yolo has two general decoders to improve the accuracy of different object sizes. We first make a copy of the general decoders bounded together as TKD decoder. The TKD decoder is updated during the inference stage. We only update the last three layers of Tiny-Yolo and treat it as the decoder, since it yields enough performance in practice. We keep the general decoder from Tiny-Yolo together with the TKD decoder to make the final detection. TKD decoder and general decoder are executed in two parallel threads which do not increase the latency. This will preserve the chance of detecting viable objects addressing the challenge (3) in Sec. \[distillation\].
Distillation Loss
-----------------
Before describing our distillation loss, we provide a brief overview of the other distillation loss functions. First, Chen et al. [@chen2017learning] proposed a combination of hint procedure and weighted loss function. They generate boxes and labels using both the student and the teacher model, then calculate two loss values comparing the teacher’s output and the ground truth. In the end, they sum up the weighted loss values. If the student model outperforms the teacher model, they continue training only using ground-truth supervision. More recently, Mehta et al. [@mehta2018object] applied the similar procedure to the one-stage object detection models (Tiny-Yolo v2 with some modification). They generate bounding boxes and labels, and apply Non-Maximum Suppression (NMS) to these boxes and then follow the loss function to optimize the student model. The loss is defined in the following equation: $$\small
\setlength{\abovedisplayskip}{5pt}
\setlength{\belowdisplayskip}{5pt}
\begin{aligned}
L_{final}= L_{bb}^{C}(b_i^{gt}, \hat{b_i}, b_i^{T}, o_i^{T}) + L_{cl}^{C}(p_i^{gt}, \hat{p_i}, p_i^{T}, o_i^{T})\\
+ L_{obj}^{C}(o_i^{gt}, \hat{o_i}, o_i^{T}) ,
\end{aligned}
\label{eq:c}$$ where $L_{bb}^{C},L_{cl}^{C},L_{obj}^{C}$ are objectness loss, classification loss and regression loss which are calculated using both ground truth and the teacher output. Also, $\hat{b_i},\hat{p_i},\hat{o_i}$ are bounding box coordinates, class probability and objectness of the the student model. $b_i^{gt},p_i^{gt},o_i^{gt}$ and $b_i^{T},p_i^{T},o_i^{T}$ are values derived from ground truth and the teacher model output.
In our study of the Yolo-v3 and Tiny-Yolo models, we noticed that the detection layer is the most computationally expensive part. In this layer, several processes are done (sorting, applying softmax to classification cells, removing low confidence boxes, etc.) to produce bounding boxes and then applying NMS to these boxes. These processes are computationally slow due to the multiple steps of processing, and also running over CPU by the implementation. Consequently, directly adopting these loss functions will be also computationally expensive during the inference stage.
With this observation, we adopt the mean square error (MSE) between the tensors generated by the student decoder and the oracle decoder, which should be the fastest method. However, the side effects are also notorious. The oracle model generates noises over some parts of frame which have no object existences; hence directly forcing the student model to retrain will hurt its performance.
[0.6]{}
[0.4]{}
Another approach could be calculating the MSE between the tensor cells which have high confidence of object existence. But, the approach will hurt the student’s recognition accuracy too. By applying this loss function, the student model tends to generate redundant detection boxes which yield a larger number of false positives.
To alleviate the downsides of both loss designs and still to preserve their advantages, we introduce a novel loss by a combination of them in Equation \[eq:b\]: $$\small
\setlength{\abovedisplayskip}{5pt}
\setlength{\belowdisplayskip}{5pt}
\centering
\begin{aligned}
& L_{final}= \sum \lVert T_s^H - T_o^H \rVert_2^2\\
& + \sum \lVert T_s^E - ((\lambda * T_s^E) + ((1-\lambda) * T_o^E)) \rVert_2^2,
\end{aligned}
\label{eq:b}$$ where $T_s^H \& T_o^H$ are the student and oracle cells with a high chance of object existences and $T_s^E \& T_o^E$ are the cells with a low expectation. More specifically, the first part on the left side of Eq. \[eq:b\] calculates the MSE between the parts which have high confidence of objects. The second part calculates a modulated MSE between the cells with a low expectation from both the oracle output tensor and the student output tensor. Here, $\lambda$ is the modulation factor. Figure \[Tensor\_copy\] shows the procedure of creating the target tensor.
By using this loss function, the student model will have a lower chance to generate extra false positives. Also, it would not strictly force the student model to mimic the oracle exactly. We aim to partially address the challenges 1) and 4) in Sec. \[distillation\], with such a fast and effective loss function.
Key Frame Selection
-------------------
Another crucial module to enable TKD working properly is a procedure to demonically select the time instances to train the student model during the inference stage. Specifically, TKD seeks the frames that by training over them the model has a higher expectation of reducing the loss, thus eventually improves the detection accuracy. For the rest of the paper, we denote these frames as the key frames.
Selecting a larger number of frames as the key frames will hurt the performance since re-training is computationally expensive; While selecting too few number of frames will hurt the detection accuracy as the student may not align well with the oracle model in time. Thus, an effective and fast procedure to select the key frames is highly desired to yield a positive effect on the system’s performance.
We propose a key frame selection procedure which is both efficient and also practical. First, we check the training prevention factor $\tau$. If the student model has been trained in any last $\tau$ frames; we will exit the key selection procedure. It is based on the reasonable assumption that if we have an environment change, it typically takes $\tau$ frames that this change to be fully observable. Thus, when we train the student, training for the next $\tau$ frames would not be beneficiary. Second, we start our decision process which we formulate in Equations \[eq:d\]: $$\small
\setlength{\abovedisplayskip}{5pt}
\setlength{\belowdisplayskip}{5pt}
\begin{aligned}
& I \in \{0,1\} \left\{\begin{matrix}
&0 &Do \ not \ distill \ knowledge, \\
&1 &Distill \ knowledge,
\end{matrix}\right.
\\
& I = LSTM(F_s) \ \vee \ I_R, \quad I_R\sim B(2,P_t),
\\
& P_t=\left\{\begin{matrix}
&max((P_{t-1}-0.05),0.05) \ \ \ \ \ \ \ \Delta L <\sigma, & \\
&min(2P_{t-1},1.0) \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \Delta L >\sigma, &
\end{matrix}\right.
\end{aligned}
\label{eq:d}$$ where $I$ is the indicator that denotes our final decision. It takes the disjunction of the LSTM’s output and the random module’s output. We pass the features extracted from the student model $F_s$ (the last layer before the decoder) to the LSTM module (with one LSTM layer & one fully connected layer) which outputs a signal indicating to train the student model or not. Here, it is worth to note that we introduce another binary random module $I_R$ (with binomial distribution $B(2,p_t)$) which decides in a random fashion to train the student model or not. The random procedure is added as a safeguard in case the LSTM model outputs a sequence of erroneous decisions. In the end, If the LSTM makes a correct decision where the observed loss decrease $\Delta L \ < \ \sigma$ wherein our experiments $\sigma=-0.1$, the random factor $P_t$ will be reduced by $0.05$. If the LSTM model makes a wrong decision, we update the LSTM model and double the random factor $P_t$. Figure \[TKD\_loss\] shows an example output of key frames selected by our method. We apply knowledge distillation selectively to a few numbers of frames which partially addresses the aforementioned challenges 1) and 2) in Sec. \[distillation\].
--------------------------------- ---------- ---------- ---------- ---------- ---------- ---------- ---------- ---------- ---------- ---------- ---------- ----------
**AP** **F-1** **AP** **F-1** **AP** **F-1**
Random Selection 0.71 0.75 0.54 0.68 0.48 0.49 0.65 0.65 0.55 0.58 0.35 0.43
Scene Change Detection 0.68 0.58 0.47 0.50 0.23 0.35 0.54 0.58 0.45 0.53 0.35 0.44
Tiny-Yolo [@redmon2018yolov3] 0.45 0.16 0.38 0.14 0.10 0.28 0.37 0.11 0.25 0.10 0.08 0.06
Tiny-Yolo (73%) + Yolo-v3 (27%) 0.60 0.49 0.59 0.49 0.44 0.46 0.58 0.47 0.52 0.46 0.39 0.44
TKD **0.75** **0.76** **0.58** **0.69** **0.49** **0.50** **0.73** **0.67** **0.59** **0.61** **0.40** **0.46**
--------------------------------- ---------- ---------- ---------- ---------- ---------- ---------- ---------- ---------- ---------- ---------- ---------- ----------
Experiments {#sec:exp}
===========
The presented theoretical framework suggests three hypotheses that deserve empirical tests: 1) TKD can perform visual recognition efficiently, without hurting the recognition performance significantly; 2) the novel loss function can improve online training of the decoder; 3) with our TKD frame selector mechanism, the overall system yields the best performance over other key-frame selection mechanisms, by locating the key frames more accurately (frames which training over them can improve TKD accuracy).
To validate these three hypotheses, we evaluate TKD on the Hollywood scene dataset [@marszalek09], YouTube-Objects dataset [@prest2012learning], The Pursuit of Happyness [@Muccino2008] and the office [@Daniels2013]. We have trained all the base models (RetinaNet [@lin2018focal], FasterRCNN [@ren2015faster], Yolo-v3 and Tiny-Yolo [@redmon2018yolov3]) over MS COCO dataset [@lin2014microsoft]. We implemented the TKD as described in Sec. \[sec:TKD\_structure\] with two different configurations. First, we perform the process of inference and distillation sequentially among the same thread; the other way, we perform the distillation in a separate thread and run the student and oracle in parallel, both architecture implemented using the PyTorch environment [@paszke2017automatic]. All experiments are carried out on one single NVIDIA TITAN X Pascal graphics card.
**Hollywood scene dataset [@marszalek09]** has 10 classes of scenes distributed over 1152 video. In this dataset, videos are collected from 69 movies. The length of these video clips are from 5 seconds to 180 seconds. The length and diversity of video clips make this dataset a perfect candidate to evaluate our key selector method and the novel loss function.
**YouTube-Objects dataset [@prest2012learning]** is a weakly annotated dataset from YouTube videos, 10 object classes of the PASCAL VOC Challenge [@everingham2010pascal] has been used in this dataset. It contains 9 and 24 video clips for each object class which length of these videos are between 30 seconds to 3 minutes. We used this dataset to evaluate TKD’s overall performance due to its high-quality objects level annotations.
**The pursuit of happyness [@Muccino2008] & The office [@Daniels2013]** are two famous movie and TV series. These two video clips contain several scenes which have smooth transitions. The Pursuit of happyness serves a great testbed since it has scenes in different locations such as office, street, etc. It is also more close to the real world scenario from a camera of the intelligent agent. Also, the Office is selected as most of the scenes have been recorded in the same location which make it suitable for testing our novel loss function.
Ablation Study {#sec:ablation}
--------------
As shown in table \[TKD\_hollywood\], we compare different strategies to highlight the effectiveness of our proposed novel loss and key frame selector. We consider the output of the oracle model as ground truth and evaluating different methods over it. Here, we compare five methods: 1) TKD with random key frame selection; 2) TKD with Scene Change detection; 3) Tiny-Yolo without any training; 4) Combination of Tiny-Yolo and Yolo-v3 without training; 5) TKD with our proposed key frame selection method.
In the following experiments, we have set the $\lambda$ to be $0.4$ which is obtained heuristically. In \[sec:discussion\], we will go through the findings which we observed in our search for the best $\lambda$.
**Random Selection:** Here, instead of selecting key frames by our proposed method, decision modules selects frames purely randomly for further processing. During the testing phase, the probability is set to be $27\%$ (to make sure it selects more frames than our method ($25\%$ on average)). Random selection achieves $0.75$ $F_1$ score (IOU=$0.5$) in the Hollywood scene dataset and achieves $0.65$ $F_1$ score (IOU=0.5) in the pursuit of happiness. On average, it reaches a frame-rate of $89$ frames per second (FPS).
**Scene Change Detection:**This method uses the content-aware scene detection method [@castellano2018pyscenedetect]. It finds areas where the difference between two subsequent frames exceeds the threshold value and used them as key frames for training the student. We selected the threshold with the highest performance and accuracy to report. This method achieves $0.58$ $F_1$ score and $0.58$ $F_1$ score in the Hollywood scene dataset and The pursuit of happyness respectively. This method selected $24\%$ frames as key frames ultimately. On average, the system yields a $93$ FPS.
**Tiny-Yolo without any training:** We test Tiny-Yolo [@redmon2018yolov3] to show the accuracy of a strong baseline model without temporal knowledge distillation. This model achieves $0.16$ $F_1$ score and $0.11$ $F_1$ score in the Hollywood scene dataset and The pursuit of happyness respectively, which are significantly lower than the other mentioned methods. However, This model has $220$ FPS, the fastest among all.
**Tiny-Yolo $+$ Yolo-v3 without training:** In this configuration, we used Tiny-Yolo and Yolo-v3 v3 [@redmon2018yolov3] together. We designed a random procedure which runs Yolo-v3 with a probability of $27\%$ and Tiny-Yolo for the rest of the times. This model achieves $0.49$ $F_1$ score and $0.47$ $F_1$ score in the Hollywood scene dataset and the pursuit of happyness respectively. Frame-rate approaches 89 FPS.
**TKD with our key frame selection method:** Initially, we set $\tau$ (the training prevention factor) to $2$ (We observe that the transition between two scenes takes at least $2$ frames); along with setting the minimum random selection to $5\%$. In the Hollywood dataset, our method selects around $26\%$ of frames and the $F_1$ score achieves $0.76$ (IOU=0.5). In the pursuit of happyness movie, our method selects around $24\%$ of frames and the $F_1$ score reaches to $0.67$ (IOU=$0.5$). On average, the system achieves a frame-rate of $91$ FPS sequentially and $220$ FPS with running inference and knowledge distillation in parallel.
Table \[TKD\_hollywood\] lists the experimental results we observed with these variants. These experiments show, the TKD, while maintaining a similar frame-rate as other methods, it can achieve higher recognition accuracy. To further validate this claim, we conduct one additional experiment on a single-shot movie [@fish2013], TKD selects $21\%$ and random procedure selects $27\%$ of the total frames for re-training. They reach comparable F1-score (TKD:$0.807$, Random:$0.812$), but our TKD method uses 10400 frames less than the random one.
[ccc]{} &\
& & F-1 score\
& &\
& &\
& &\
& &\
& &\
\
& &\
Overall Performance
-------------------
Table 2 shows mean average precision (mAP) and $F_1$ score for five different object detection models as well as our TKD method over the Youtube object dataset [@prest2012learning]. For the student models without oracle’s supervision, we train them to the best performance we could achieve. Not surprisingly, larger or deeper models with larger numbers of parameters perform better than shallower models, while smaller models run faster than larger ones. However, TKD achieves a high detection accuracy compare to RetinaNet, FasterRCNN, Tiny-Yolo, the combination of Tiny-Yolo and Yolo-v3 (same configuration which is described in Sec. \[sec:ablation\]). TKD’s detection performance also approaches the performance of the oracle model (Yolo-v3). In this experiment, $25\%$ of frames have been selected for training using the proposed key frames selection method.
To illustrate the accuracy-speed trade-off, we further plot them in Figure \[TKD\_plot\], where we can see that the TKD archives higher accuracy compare to other shallow methods while still operating far above the real-time speeds with a $91$ FPS. The oracle model has a better detection accuracy, but it runs much slower than the TKD.
Further Study and Discussions {#sec:discussion}
-----------------------------
In this section, we provide further insight into the loss function design, the general knowledge distillation idea, and suggest an application of the proposed method.
**Loss function:** we studied the $\lambda$ effect over the number of true positives and false positives generated by TKD. All tests are done over an episode from The office [@Daniels2013]. We choose this video since it was recorded in one indoor environment, with a consistent objects distribution. Table \[loss\_tune\] shows the student model’s detection accuracy varies with the different choices of $\lambda$. At $\lambda=0$, we observed a lower number of false positives since a fewer number of frames ($5\%$) selected by the key frame selection module. With a low $\lambda$ (except at $0$), we observe an increase in false positives as the model tries to generate more boxes and loss function doesn’t punish hardly enough onto the student model for generating false positives. With a high $\lambda$, we observe drops in the true positive rates since we are forcing the student to learn noises which are likely introduced by the oracle model. Consequently, $0.4$ is empirically the best choice here, and we set it as the $\lambda$ value for all the experiments.
To validate our loss design, we further compare its performance with the one from Mehta et al. [@mehta2018object], where the proposed loss is based on Non-Maximum Suppression algorithm. It is computationally more expensive in comparison with our approach. Figure \[LossPerf\] depicts that, an increasing number of targets from each frame will result in the increasing of execution time for calculating the loss function in [@mehta2018object]. Our loss design has an almost constant execution time, while the proposed loss function by [@mehta2018object] is linearly growing.
**0** **0.2** **0.4** **0.6** **0.8** **1**
-- --------- --------- --------- ----------- ---------- ------- -------
**AP** 0.47 0.72 0.82 **0.83** 0.79 0.8
**F-1** 0.36 0.649 **0.676** 0.656 0.634 0.643
3353 8570 8371 7806 7274 7438
215 2952 1522 1129 814 841
: Parameter study of $\lambda$ over the TKD.[]{data-label="loss_tune"}
**Temporal knowledge distillation:** Here, we take a closer look at the key selection module. Figure \[TKD\_loss\] shows its performance over two video clips from the Hollywood scene dataset. Red crosses are frames selected by our proposed method as key frames. At peaks, we have a scene change and logically these points would be the best candidate for training. Following this insight, we observe our model has a lag on detecting these points. Here, we argue that training over these frames is not the best one for improving the student model’s accuracy. The scene detection method can identify these points yet table \[TKD\_hollywood\] shows it achieves lower accuracy. Figure \[TKD\_loss\] shows the TKD after detecting a change in loss start stabilizing the model by selecting most of the frames (parts A & C) and for the rest select less number of frames (parts B & D).
The proposed key frame selection method leads to improved performance comparing with [@mullapudi2018online]’s. Figure \[histo\] shows that the number of selected key frames is adjusted based on the domain change. With the fixed camera case in which the domain does not change, the number of selected frames decreases along observing more frames (validated over the UCF Crime dataset [@sultani2018real]). Indeed, for the case of a moving camera, more key frames are selected to adjust the TKD to the specific domain. Here, the method presented in [@mullapudi2018online] relies on a static strategy of selecting frames which are chosen manually at the beginning.
For further evaluation, we applied TKD on one episode of the office TV series. Then, we test the trained student model over another episode without any re-training at the inference time. We observed an increase of precision by $6\%$ comparing to the case in which we use the original student model without applying TKD. The result demonstrates the domain adaption capability of our method. Furthermore, it maintains a high recall over other domains which indicates that unseen objects have a chance to be detected. With the method presented in [@mullapudi2018online], the model loses its generality over unseen objects due to the practice of optimizing the overall model with the new frames.
Conclusion and Future Work
==========================
In this paper, we propose a novel approach to distill temporal knowledge of an accurate but slow object detection model to a tinier model yielding a light and accurate object detection paradigm for robotic applications, called TKD. We conducted experiments on the Hollywood scene dataset, Youtube object dataset, the pursuit of happyness movie and the office TV series, and empirically validate that TKD maintains a high inference efficiency while achieving a high recognition accuracy. The accuracy even approaches the original oracle model for the object detection task.
The promising experimental results we observed suggest several potential lines of future work: 1) the frame selection procedure could be further optimized to be more selective while maintaining the recognition accuracy; 2) we plan to test our TKD model with an oracle model that follows the two-stage object detection manner; 3) TKD performance can future improve by adopting temporal features in video.
**Acknowledgment:** The National Science Foundation under the Robust Intelligence Program (\#1750082), and the IoT Innovation (I-square) fund provided by ASU Fulton Schools of Engineering, GPU and FPGA donations from NVIDIA and Xilinx, are gratefully acknowledged.
| {
"pile_set_name": "ArXiv"
} |
---
abstract: 'We report on experiments with deformed polymer microlasers that have a low refractive index and exhibit unidirectional light emission. We demonstrate that the highly directional emission is due to transport of light rays along the unstable manifold of the chaotic saddle in phase space. Experiments, ray-tracing simulations, and mode calculations show very good agreement.'
author:
- 'M. Schermer'
- 'S. Bittner'
- 'G. Singh'
- 'C. Ulysse'
- 'M. Lebental'
- 'J. Wiersig'
date:
title: 'Unidirectional light emission from low-index polymer microlasers'
---
The physics of optical microcavities is a topical research field for more than one decade [@Vahala03]. Microcavities [@MLSGPL92; @MSJPLHKH07] confine photons for a long time $\tau$ due to total internal reflection at the boundary of the cavity. These so-called whispering-gallery modes (WGMs) have high quality factors $Q = \omega\tau$, where $\omega$ is the resonance frequency. The in-plane light emission from an ideal circular-shaped microlaser is isotropic due to the rotational symmetry. To overcome this disadvantage microcavities with deformed boundaries have been fabricated leading to significantly improved emission patterns [@LSMGPL93; @ND97; @GCNNSFSC98; @SRTCS04; @Cao2015]. Even unidirectional emission is possible, which has been demonstrated for several shapes, e.g., the spiral [@CTSCKJ03; @HK09; @HKB09], cavities with holes [@WH06; @Djellali2009], the [[[limaçon]{}]{}]{} [@WH08; @SCL08; @YWD09; @WYD09; @SHH09; @YKK09; @AHE12], the circle with a point scatterer [@DMS09], and the notched ellipse [@BBS06; @WYY10].
The ray dynamics inside a deformed microcavity is (partially) chaotic, i.e., neighboring ray trajectories deviate from each other exponentially fast. Because of this, deformed microdisks have attracted attention as models for studying ray-wave correspondence in open systems [@ND97]. This is analog to the study of quantum-classical correspondence in the field of quantum chaos [@Stoeckmann00]. In open chaotic systems the long-time behavior of trajectories is governed by the chaotic saddle (or chaotic repeller for noninvertible dynamical systems) and its unstable manifold [@LichLieb92; @LT10]. The chaotic saddle is the set of points in phase space that never visits the leaky region both in forward and backward time evolution. The unstable manifold of a chaotic saddle is the set of points that converges to the saddle in backward time evolution. This unstable manifold therefore describes how trajectories, after a transient time, escape from the open chaotic system. It has been shown both theoretically and experimentally that this kind of unstable manifold in chaotic microcavities determines the far-field pattern of all high-$Q$ modes [@SRTCS04; @LYMLASLK07]. As a consequence, the high-$Q$ modes in such a cavity have very similar far-field patterns. This useful property has been termed universal far-field pattern [@LYMLASLK07]. Based on this concept it was predicted that the light emission from high-$Q$ modes in a [[[limaçon]{}]{}]{} cavity with refractive index $n$ between 2.7 and 3.9 is universal and unidirectional [@WH08]. This was confirmed experimentally by a number of groups [@SCL08; @YWD09; @WYD09; @SHH09; @YKK09; @AHE12].
Microcavities made from materials with low refractive index $n \leq 2$ like polymers are interesting due to their cheap and easy fabrication. For this low-index regime, the “face” cavity has been proposed [@ZouJSTQE13]. Unfortunately, this cavity is not fully chaotic and therefore the universality and directionality of light emission is spoiled by islands of regular motion in phase space. The notched ellipse can work for a rather broad regime of refractive indices by adapting the eccentricity of the ellipse [@WYY10]. In the low-index regime, however, the notch does not function as an efficient scatterer needed for directing the light [@Unterhinninghofen11]. Deformed silica microtoroids can also exhibit directional emission but at the cost of a complicated fabrication process [@Jiang2012].
It is the aim of the present letter to fill this low-index gap by introducing a cavity shape called “shortegg” that is shown in Fig. \[fig:photo+diagramme\](a). For $1.5 \leq n \leq 1.8$, its emission is strongly concentrated in a single direction \[see Fig. \[fig:photo+diagramme\](b)\], the far-field pattern is universal, and the quality factors are reasonably high. Numerical simulations and experimental data show good agreement.
The boundary of the shortegg cavity is given in polar coordinates by $$\label{eq:shortegg}
\rho(\phi) = R [1 + \varepsilon_1 \cos(\phi) + \varepsilon_2 \cos(2\phi) + \varepsilon_3 \cos(3\phi)]$$ with mean radius $R$ and deformation parameters $\varepsilon_1 = 0.16$, $\varepsilon_2 = -0.022$, and $\varepsilon_3 = -0.05$. The ray dynamics inside microcavities is best described in a two-dimensional phase space representation, the so-called Poincaré surface of section. Whenever the ray hits the cavity’s boundary, its position $s$ in terms of the arclength coordinate along the circumference and its tangential momentum $\sin(\chi)$ are recorded. Here, $\chi$ is the angle of incidence measured from the surface normal. An angle $\chi < 0$ ($\chi > 0$) indicates clockwise (counterclockwise) propagation direction. When $|\chi|$ is larger than the critical angle for total internal reflection, $\chi_c = \arcsin{(1/n)}$, the ray is completely reflected. In the leaky region of phase space $|\chi| < \chi_c$ the ray is only partially reflected according to Fresnel’s law.
The ray dynamics inside the shortegg is chaotic except for small islands of regular motion in the leaky region (not shown). Figure \[fig:unstablemanifold\] depicts the unstable manifold of the chaotic saddle for the case of transverse electric (TE) polarization (i.e., electric field parallel to the plane of the cavity) with effective refractive index $n = 1.5$ according to the numerical scheme discussed in Ref. [@WH08]. It looks similar for transverse magnetic (TM) polarization (i.e., magnetic field parallel to the plane). It should be noted that flat dye-based microlasers predominantly exhibit TE polarized modes [@Gozhyk2012]. As in the situation of the [[[limaçon]{}]{}]{} cavity [@WH08], the unstable manifold has only two significant overlaps with the leaky region, one at $s/s_\mathrm{max} \approx 0.35$ and one related by symmetry at $s/s_\mathrm{max} \approx 0.65$. However, in contrast to the [[[limaçon]{}]{}]{} cavity these two overlap regions are rather elongated and it is therefore not clear *a priori* how this can lead to unidirectional emission. In fact, we have optimized the boundary of the cavity such that the arms of the unstable manifold follow precisely the curve of points that are emitted to the far-field angle $0^\circ$ (red curve in Fig. \[fig:unstablemanifold\]). Due to this fact, a highly directed emission can be expected.
The far-field intensity pattern resulting from the ray-tracing simulations is shown as dashed black curve in Fig. \[fig:ffpwave\]. As predicted by the structure of the unstable manifold, the emission is strongly peaked around the far-field angle $\phi = 0^\circ$ with considerably smaller side peaks at $\pm 80^\circ$ and $\pm 150^\circ$. The beam divergence is as small as $\pm 3^\circ$, which is much smaller than the beam divergence from the [[[limaçon]{}]{}]{} cavity [@WH08] and comparable to the beam divergence from the notched ellipse [@WYY10] in the high-index regime. The small oscillations of the far-field intensity pattern in Fig. \[fig:ffpwave\] are due to the nontrivial fine structure of the unstable manifold which is not visible on the scale shown in Fig. \[fig:unstablemanifold\]. Due to the small value of $n$ the emission for TM polarization is similar with additional small side peaks at $\pm 120^\circ$ (not shown). Extensive numerical investigations reveal that the unidirectional emission persists for $1.5 \leq n \leq 1.8$ but not for higher refractive indices. Moreover, the far-field pattern is robust with respect to small variations of the deformation parameters $\varepsilon_1$, $\varepsilon_2$, and $\varepsilon_3$. However, the directionality is considerably spoiled for $\varepsilon_3 = 0$.
Numerical simulations of Maxwell’s equations in the shortegg geometry are performed within the effective refractive index approximation in two dimensions (see, e.g., Ref. [@Lebental2007]). Using the boundary element method [@Wiersig02b] we determine the spatial mode pattern $\psi(x,y)$ corresponding to the $z$-component of the magnetic (TE) and electric (TM) field as well as the complex resonant frequencies $\omega =ck$ where $c$ is the speed of light in vacuum and $k$ is the wave number. The real part is the conventional frequency whereas the imaginary part determines the lifetime $\tau =-1/[2\,{\text{Im}\,\omega}]$ of a given mode. For convenience, the dimensionless complex frequency $\Omega = \omega R/c = kR$ is used.
In the considered frequency regime the computed high-$Q$ modes are WGMs (two examples for $n = 1.5$ are shown as insets in Fig. \[fig:ffpwave\]). In all cases their far-field patterns agree well with the ray-tracing simulations, which clearly demonstrates the universality of the far-field patterns. The fact that WGMs with reasonably high $Q$s are formed in such a strongly deformed cavity with chaotic ray dynamics can be explained by the existence of partial barriers in phase space [@SWC11]. Moreover, we verified that the Husimi functions [@HSS03] (not shown) of the high-$Q$ modes in the leaky region are well localized on the unstable manifold. This presents another indication that the emission directionality in the shortegg is due to the unstable manifold of the chaotic saddle.
Shortegg cavities were experimentally investigated using organic microlasers. The microlasers were fabricated from a $700$ nm thick layer of PMMA [^1] doped with $5$ wt% of the laser dye DCM [^2]. The dye-doped polymer with bulk refractive index $1.54$ was spin-coated on a Si wafer with a $2~\mu$m buffer layer of SiO$_2$ to avoid leakage of the cavity modes into the Si wafer. The cavities were created from the polymer layer by electron-beam lithography which ensured nanometric precision [@Lozenko2012]. They can be considered as two-dimensional systems with effective refractive index $n = 1.50$ (see Ref. [@Lebental2007] for details of the effective refractive index calculation). The cavities were completely supported by the silica layer[@Lozenko2012]. An optical microscope image of a shortegg cavity is shown in Fig. \[fig:photo+diagramme\](a). The results presented here were measured with a cavity with radius $R = 80~\mu$m \[${\text{Re}(\Omega)} \simeq 810$\], and are similar to those obtained with smaller ones. Furthermore, a quadrupolar cavity was investigated for comparison. Its shape is given by Eq. (\[eq:shortegg\]) with $\epsilon_{1, 2}$ as for the shortegg and $\epsilon_3 = 0$. The microlasers were pumped by a pulsed, frequency-doubled Nd:YAG laser ($532$ nm, $0.5$ ns, $10$ Hz). The size of the pump spot was adjusted to cover the whole cavity homogeneously. The lasing emission was collected by a lens and analyzed with a spectrometer. The cavities could be rotated in order to measure the lasing spectra for arbitrary directions in the cavity plane and thus obtain the azimuthal far-field intensity pattern. See Ref. [@Chen2014] for a detailed description of the experimental setup.
The lasing threshold intensity of the shortegg microlaser was $2.4$ MW$\cdot$cm$^{-2}$ and is about three times lower than that of the quadrupolar microlaser. This evidences that the particular shape of the shortegg leads to an enhancement of the quality factors compared to other cavities with similar shapes. The lasing emission was TE polarized.
A spectrum recorded for $\phi = 0^\circ$ just above threshold is shown in Fig. \[fig:spectre\](a) and features a sequence of equidistant lasing modes. Lasing modes with large ${\text{Re}(\Omega)}$ are often related to specific sets of trajectories. If this is the case, then the optical length of these trajectories is inversely proportional to the free spectral range, ${\ell_\mathrm{opt}}= 2 \pi / \Delta k$, and can be determined from the Fourier transform of the spectrum [@Lebental2007]. The Fourier transform of the spectrum in Fig. \[fig:spectre\](a) is presented in Fig. \[fig:spectre\](b). Its first two significant peaks are at ${\ell_\mathrm{opt}}= 755.7~\mu$m and $785.5~\mu$m, and further peaks are found at multiples of these lengths. They correspond to a geometrical length of 455 $\mu$m and 473 $\mu$m, respectively, for a group refractive index of $n_g = 1.66$. These lengths are significantly longer than the length of the diameter, ${\ell_\mathrm{geo}}^\mathrm{(diam)} \simeq 3.912 \, R = 313~\mu$m, and hence evidence that the lasing modes are of the whispering gallery type. On the other hand they are also significantly shorter than the length of the perimeter, ${\ell_\mathrm{geo}}^\mathrm{(per)} \simeq 6.360 \, R = 509~\mu$m, which indicates that the observed modes are WGMs with higher radial excitation. In contrast, the spectrum of the quadrupolar microlaser (not shown) corresponds to the optical length of the diameter orbit. The shortegg cavity exhibits a larger number of resonances for higher pump intensities \[see Fig. \[fig:spectre\](c)\], but the Fourier transform features only peaks at the same optical lengths as in Fig. \[fig:spectre\](b).
The measured far-field intensity pattern in Fig. \[fig:photo+diagramme\](b) shows the maximal intensity of each spectrum as a function of the azimuthal angle. It was recorded for a pump intensity two times higher than the threshold. No significant effect of bleaching was observed during the measurement. The quadrupolar microlaser exhibits a very broad far-field intensity pattern without sharp emission lobes (not shown). On the contrary, the far-field intensity pattern of the shortegg micro-lasers exhibits one dominant emission lobe at $\phi = 0^\circ$ with a divergence of about $\pm 4^\circ$, in very good agreement with the theoretical predictions. Four much smaller lobes at $\pm 80^\circ$ and $\pm 150^\circ$ are also observed. This is consistent with the numerical simulations even though the lasing modes probably have a higher radial excitation than the modes shown in Fig. \[fig:ffpwave\], which once again shows the universality of the emission patterns.
A photograph of the lasing cavity made by a CMOS sensor camera with a high-magnification zoom lens is presented in Fig. \[fig:camera\](a). The observation direction was $\phi = 0^\circ$ and the camera had an inclination angle of $10^{\circ}$ with respect to the plane of the cavity. The photograph shows that the red lasing emission towards $\phi = 0^\circ$ originates from two small regions of the cavity boundary around $\phi = \pm50^\circ$. Figure \[fig:camera\](b) shows a sketch of the cavity where the emitting parts of the cavity are indicated in red. Their positions agree very well with the calculated near-field intensity patterns shown in the insets of Fig. \[fig:ffpwave\].
We have demonstrated both theoretically and experimentally that low-index polymer microlasers with the so-called shortegg shape exhibit lasing modes with highly directional emission. The lasing modes are based on the unstable manifold of the chaotic saddle which leads to a universal (i.e., mode-independent) far-field pattern of the high-Q modes. The theoretical predictions of near- and far-field intensity patterns showed excellent agreement with the experimental findings.
Fruitful discussions with J. Zyss, J.-B. Shim and M. Kraft are acknowledged. M. S. thanks A. Ebersp[ä]{}cher for providing his computer code package. S. B. gratefully acknowledges funding from the European Union Seventh Framework Programme (FP7/2007-2013) under Grant No. 246.556.10. This work was supported by a public grant from the Laboratoire d’Excellence Physics Atom Light Matter (LabEx PALM) overseen by the French National Research Agency (ANR) as part of the Investissements d’Avenir program (reference: ANR-10-LABX-0039).
[39]{}ifxundefined \[1\][ ifx[\#1]{} ]{}ifnum \[1\][ \#1firstoftwo secondoftwo ]{}ifx \[1\][ \#1firstoftwo secondoftwo ]{}““\#1””@noop \[0\][secondoftwo]{}sanitize@url \[0\][‘\
12‘\$12 ‘&12‘\#12‘12‘\_12‘%12]{}@startlink\[1\]@endlink\[0\]@bib@innerbibempty @noop [****, ()]{} [****, ()](\doibase
10.1063/1.106688) [****, ()](\doibase 10.1063/1.2435608) [****, ()](\doibase
10.1063/1.108911) [****, ()](\doibase 10.1038/385045a0) [****, ()](\doibase
10.1126/science.280.5369.1556) [****, ()](\doibase
10.1364/JOSAB.21.000923) [****, ()](\doibase 10.1103/RevModPhys.87.61) [****, ()](\doibase
10.1063/1.1605792) [****, ()](\doibase 10.1364/OL.34.000163) [****, ()](\doibase
10.1364/OE.17.010335) [****, ()](\doibase 10.1103/PhysRevA.73.031802) [****, ()](\doibase 10.1063/1.3205474) [****, ()](\doibase 10.1103/PhysRevLett.100.033901) [****, ()](\doibase 10.1103/PhysRevA.80.041807) [****, ()](\doibase 10.1063/1.3153276) [****, ()](\doibase
10.1088/1367-2630/11/12/125018) [****, ()](\doibase
10.1103/PhysRevA.80.031801) [****, ()](\doibase 10.1063/1.3242014) [****, ()](\doibase 10.1063/1.4733726) [****, ()](\doibase 10.1103/PhysRevA.80.063813) [****, ()](\doibase 10.1109/JSTQE.2005.863002) [****, ()](\doibase 10.1073/pnas.1015386107) @noop [**]{} (, , ) @noop [**]{} (, , ) @noop [**]{} (, , ) [****, ()](\doibase 10.1103/PhysRevA.75.011802) [****, ()](\doibase 10.1109/JSTQE.2012.2220896) **, @noop , () [****, ()](\doibase 10.1002/adma.201201229) [****, ()](\doibase 10.1103/PhysRevA.86.043817) [****, ()](\doibase 10.1103/PhysRevA.76.023830) [****, ()](\doibase 10.1088/1464-4258/5/1/308) [****, ()](\doibase 10.1103/PhysRevE.84.035202) [****, ()](\doibase 10.1209/epl/i2003-00421-1) [****, ()](\doibase 10.1063/1.4720474) [****, ()](\doibase 10.1364/OE.22.012316)
[^1]: Poly(methyl methacrylate) (PMMA A6 resist by Microchem)
[^2]: 4-(Dicyanomethylene)-2-methyl-6-(4-dimethylaminostyryl)-4Hpyran (by Exciton)
| {
"pile_set_name": "ArXiv"
} |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.